parking cdcc

my long running tiliqua project has cached dilated causal convolutions (CDCC); a large dilated conv net that leans heavily on activation caching to get inference speeds of 192kHz

the initial goal for CDCC was to be an effect, like a delay or reverb, which requires the ability to integrate over a long sequence.

was tinkering with this architecture trying to get some psram stuff working when i suddenly realised something; the cdcc approach is great for long range integration but for a plain "stateless" oscillator you don't need any of that.

an oscillator is way more direct than an effect; it just needs to map from phase and some conditioning variables to a waveshape and that's a classic implicit neural representation (INR)

dataset

the test dataset is the same one i started with for CDCC.

the output waveshape is represented by a 2d embedding. training examples are generated in numpy by sampling two endpoints, then picking a point in between, and generating a constant power cross fade interpolation. we intentionally use waves that have interesting interpolated points. we sample data over a range of octaves.

neural oscillator

my baseline for INR was just 3 inputs; the 1d phase value & the 2d embedding. and it didn't work at all... the network needs more to bite into than just a single phase value!

the next version was 4 inputs; a quadrature sin/cos for the phase & the 2d embedding. this did a lot better ( as i knew from the cdcc work ) and i thought, if sin/cos does well, why not other basis values too? and of course, this is already well studied under random fourier features (RFFs)

the idea is just to sample a random bank of basis frequencies and have the phase sweep through them, calculating the quadrature sin and cos pairs for each. the job of the network then is to represent a mixing of these basis values.

the initial model was...

• 3 values for the input; 1d input phase ( a sawtooth which acts as an index into the bank ) and 2d embeddings
• a mapping from the phase to 64 random fourier basis values; calculating the sin and cos of each to make a 128d feature vector.
• concatenation of the 128d RFF feature vector & the 2d embedding
• finishing with a small MLP

this gives a pretty good result but has some key problems we can fix...

feature imbalance and FiLM to the rescue

the main problem i found is that there is a bit of an imbalance between the inputs.

we have 128D representing the phase and only 2D for the embedding part. it's easy for the model to get a bit stuck just ignoring the embedding.

turns out there's a much better way to use the embeddings to condition the RFFs; rather than concat them, we can use the embeddings to modulate the fourier features.

specifically we learn three projections using dense layers, all to the same output dimension

1. rff -> F
2. embeddings -> gamma
3. embeddings -> beta

and then combine them as output = (1+gamma) . F + beta

this now means the embeddings are being used to rescale and offset the RFFs

( notice that we parameterise as 1+gamma. this allows us to init weights for gamma and beta as 0 turning the entire linear function into a noop at the start of training )

i first saw this in FiLM: Visual Reasoning with a General Conditioning Layer ( Feature-wise Linear Modulation )

the qkeras models looks like ...

__________________________________________________________________________________________________
 Layer (type)                Output Shape                 Param #   Connected to                  
==================================================================================================
 input_1 (InputLayer)        [(None, None, 3)]            0         []

 phase (Lambda)              (None, None, 1)              0         ['input_1[0][0]']             
 rff (QRandomFourierFeature  (None, None, 1024)           512       ['phase_q[0][0]']             
 film0_F (QDense)            (None, None, 32)             32800     ['rff[0][0]']

 embed (Lambda)              (None, None, 2)              0         ['input_1[0][0]']
 film0_gamma (QDense)        (None, None, 32)             96        ['embed_q[0][0]']             
 film0_beta (QDense)         (None, None, 32)             96        ['embed_q[0][0]']             

 film0 (FiLM)                (None, None, 32)             0         ['mlp0[0][0]',                
                                                                     'film0_gamma_q[0][0]',       
                                                                     'film0_beta_q[0][0]']
 qrelu0 (QActivation)        (None, None, 32)             0         ['film0[0][0]']               

 mlp1 (QDense)               (None, None, 32)             1056      ['qrelu0[0][0]']              
 qrelu1 (QActivation)        (None, None, 32)             0         ['mlp1[0][0]']                

 mlp2 (QDense)               (None, None, 32)             1056      ['qrelu1[0][0]']              
 qrelu2 (QActivation)        (None, None, 32)             0         ['mlp2[0][0]']                

 y_pred (QDense)             (None, None, 1)              33        ['qrelu2[0][0]']              

==================================================================================================
Total params: 35651 (139.26 KB)

this gives a much better smoother result.

initially though i thought i wouldn't be able to use this approach since it adds more compute ( the gamma and beta dense layers ) but then i noticed a whooping optimisation!

lookup tables!

the biggest layer by far is the FiLM dense_F; it has to map from 1024 RFFs to a 32d output

but note that phase -> RFF -> FiLM dense_F is only dependant on phase. and since everything here is being quantised there's only a fixed number of values that phase can take. so we can just materialise the whole lot and treat it as a lookup. ( we have more than enough space for it in psram! )

AND note that the size of table depends on only two things

1. the quantisation of the phase ( which dicates the number of entries ) and
2. the size output dim of dense_F ( which dicates the size of each entry ).

which makes the number of RFFs only a concern at training since, for inference, it gets contracted away into lookup table. that's a HUGE win since it was the largest component of everything.

also note that this is something we couldn't do with the concatenation case ( though i guess we could have projected down between the RFFs & the concat and cached that instead... )

for v1 of this i built the table at module startup by sweeping sweep thru the phase values and populating the table in psram. this was fine until the table got big enough that the startup time was getting annoying.... at that point i switched to building it fully offline as a .bin file and adding it to the archive as something to directly load in psram.

activation fn?

did a fair bit of experimenting with what activation function to use for the MLP dense layers.

tried relu vs leakyrelu vs siren ( introduced in Implicit Neural Representations with Periodic Activation Functions )

for this problem at least siren only made a very minor improvement. have already seen it doing much better on ( work in progress ) more complex waveshaping so expect it to be important for the next piece of work.

training

for training i took the same approach as CDCC; it's worked for me a couple of times now..

1. pretrain f32 keras model
2. finetune fixed point qkeras model
3. export weights for amaranth inference

felt like a lot more messing around this time getting bit exactness between qkeras and amaranth, especially with things like the LUTs

siren in particular required another LUT for the trig functions and had to roll my own STE (straight through estimation) stuff for it since qkeras didn't support it.

losses

also spent a bunch of time tinkering with loss functions ( mainly since training for spectral correctness is new to me... )

main loss was a combination of huber & short time fourier transform (STFT) with stft needed a lot of tuning, and i still don't think it's quite right.

( and spectral stuff is so weird, it's bizarre to have cases of low spectral loss, and higher huber loss, and have the wave look worse but sound better (?!?!) )

the key i noticed was to let huber stabilise things then ramp up the STFT loss ( which switches the huber from being the waveshaping to be more about phase locking ). if i started training with stft immediately it was unstable.

also included some aux losses that, to be honest, probably didn't make much difference to things.

• a slope loss; simple first order loss ; instead of the absolute difference, try to matching the rate of change over two steps.
• dc offset loss; output should also be zero centered so added this.

not sure either did anything, was mainly curious.

device inference and util

most of amaranth code was taken from CDCC. did quite a bit of tuning around the lane-parallel and weight banks for the MAC stuff. all fits comfortably on the device.

DP16KD:32%  MULT18X18D:42%  ALU54B:0%  TRELLIS_FF:37%  TRELLIS_COMB:48%
Info: Router1 time: 0h 00m 36s
Info:  '$glbnet$audio_clk': 71.97 MHz (PASS at 50.00 MHz)
Info:        '$glbnet$clk': 89.00 MHz (PASS at 60.00 MHz)

demo

next steps

• v1 of CDCC was keras -> qkeras -> verilog
• v2 of CDCC was keras -> qkeras -> amaranth
• v1 of INR was keras -> qkeras -> amaranth
• maybe FiLM and an MLP even with siren isn't the right approach; maybe i should try RFF -> state space model (SSM) ?
• next version (SSM) will be jax -> amaranth ( since i have to patch qkeras a bit ( in terms of STE for LUT backprop ) have put more and more of
• RFF basis set is just randomly sampled. is there a way to analysis a bunch of the training data to get a better set of starting basis values? maybe a mix of random and some derived ones?
• get back to parametric capture stuff i was working before i got distracted with INR. it's currently tied up with the cdcc work but i think it's worth pulling out to a single project ( to support all these projects )
• am also keen to replicate some of the neural tangent kernel stuff given how well FiLM worked for me!

code

on github

tiliqua builds for r3 and r5

protracker on the amiga was my introduction to making electronic music and its paula chip was the first time i really worked with assembly.

was super excited then to see a c64 sid wrapper done as an example for the tiliqua eurorack fpga module

if tiliqua can run sid, surely it'll be easy enough to run paula too? right? right?!

TL;DR

just want to hear results?

code structure

minimig aga mister is a port of minimig ( a cycle accurate implementation of the amiga ) to the aga_mister board.

at first i was wondering if running this entire thing would be easiest to get paula going but at the same time i wanted to isolate a minimal piece to have more control. i like this code since it has a really clear separation of the paula audio pieces

side note; it wasn't until i had made a decent start on things that i realised it wasn't just paula i needed to deal with but the agnus chip too. though paula is the synth voice it's agnus that manages everything around memory access ( and not just for audio, it handles video and more ). my initial thought was to bring all of it across, but then decided just to implement a minimal fake_agnus to hook up instead.

paula registers

the crux of this project is to

• init a wrapper of the paula_audio verilog
• hook up a sampler to write data that fake_agnus can read from
• read a bunch of midi values to set various registers on paula ( in the mother of all state machines )
• pipe paula audio out via tiliqua

simple. ish.

HUGE SHOUTOUT to the internet archive for hosting the amiga hardware reference manual. i couldn't have done this project without it as a reference and i hugely regret putting my copy in the recycling 20+ years ago when my amiga500 died :( any pNNN references below refer to this

paula has 4 channels; each channel is controlled by a set of registers; of interest for us are ( for x in [0, 1, 2, 3] )

AUDxLEN - sample length. we set an init value of a max sample size ( see sampling below ) and we can vary this down to 1 ( not quite granular in my implementation, i think i have a bug )

AUDxPER - sample period; the number of ticks between each sample fed to the DAC. like the inverse of the pitch. see p140. a lot of the interesting character of the amiga sound comes from this one IMHO.

AUDxVOL - simple 6 bit volume. of note is that i had to add some gain on output based on my sampling.

ADKCON ( & "creative difference" )

there's also a special register that controls cross modulation between channels. ( one of the main reason i wanted to even do this project! )

there are bits for ATPER0, 1, 2 & ATVOL0, 1, 2 that dictate whether chN will modulate chN+1 for either PER ( FM mod ) or VOL ( AM mod )

and i have to say that it was super satisfying after all the messing around with fake_agnus dma crap to implement this piece and have it "just work" #angelsSinging

however... as i was first messing around toggling this bits i thought i had a bug. modulation was all over the place & after reading the reference manual ( p149-150 ) i realised it was working perfectly. it's just that the behaviour is not what i expected.

so how does the modulation work? when you set ch_x to modulate the period, or volume, of ch_y the wiring is just, quite literally, AUDxDAT -> AUDyPER or AUDxDAT -> AUDyVOL.

what interesting is ...

• AUDxVOL is just the values 0 to 64 ( the lower 6 bits )
• AUDxPER is an unsigned value
• AUDxDAT is signed audio

so if you were hand crafting in assembly a channel to be a modulator for, say, VOL, you'd never set bits 7-15. but if you use audio samples as modulators all manner of weird hell breaks loose! :D

this is fun, in it's own right, but i wanted something a bit different. so i changed this behaviour to make things more what i'd expect

• FM modulation ( PER ) is more classic audio FM; modulator=0 => original freq, +ve => faster, -ve => slower.
• AM modulation ( VOL ) is ring modulation.

these changes had to be done in a vendored version of the paula_XXX.v source... sacrilege! i know! i expect to be struck down by the paula gods some time soon...

the change is pretty small to be honest, and if i ever want more realistic playback ( e.g. to playback IFF samples ) i can just put in a toggle.

sampling

i certainly didn't want to write my own wavetable in pseudo assembly so instead i added a simple sampler to fake_agnus.

you can start/stop a recording and then toggle playback per channel. toggling playback makes the whole thing wavetable focussed but would be trivial to switch to one shots instead.

midi

all of this is controlled via midi! setup for my beat step pro but ccs and notes are configurable via a json config ( at compile time )

i use encoders to set range based registers ( AUDxVOL etc ) and pads to toggle various register bits in ADKCON as well as to start/stop recording/playback.

• the blue section represents the range control for AUDxLEN, AUDxPER and AUDxVOL. one set per channel
• the red section represents the pads to start/stop recording or playback for each channel
• the green section toggles the 6 interesting bits of ADKCON; 3 for the PER modulation ( FM ) , 3 for the VOL modulation ( AM )

additionally there are two reset pads ( R ). one to set ADKCON bits to 0 and the other to set all AUDxxxx registers to their defaults.

TODOS

known bugs

• channels can get locked up if you mash the record/play pads :(
• pretty sure i have a shifting bug re: AUDxLEN, the shortest length is still pretty long.

cleanup

midi -> register -> signals

the code started to get really squirrely with a mix of general lists of N channels vs assuming 4 will generalise further if i ever want to run 2 paulas over 8 channels

i did notice when i started this project that the sid code uses a soft cpu for the register setting. perhaps the hellish state machine i've got would be more cleanly represented in rust?

( i might actually do this just for the purpose of learning more around with vexriscv. )

fake agnii

the first version i wrote ( that i subsequently completely rewrote ) used a single fake_agnus.

for the second version i ended up splitting things to have 4 fake angus chips instead ( one for each channel ).

the way i rewrote it just made dev and debugging easier. now that things work i could refactor this back to one fake chip, but am not sure i can be bothered.

there is also a stack of weird and wonderful derived constants in this class trying to accurately represent the various strobe timing and dma handling. 80+% of my time was spent here, i'll sure there is a lot of simplification i could do here...

extensions

clearly there's a truck load of things i could add

wavetable vs one shot everything so far is about building wavetables from inputs. i build short loops ( with clickless zero crossing checks ) but it could instead just do one shot playback by 1) increasing buffer size and 2) instead of toggle just do one shot. trivial change.

v/oct vs AUDxPER currently you tune frequency directly via the AUDxPER register. from an encoder! very playable :/ but it could easily support v/oct. there's a whole set of tables around v/oct -> AUDxPER on p158 onwards i could wire in.

granular aside from the ( potential ) AUDxLEN bug mentioned i think i could "invented" a new register AUDxSTART to do more interesting granular stuff. would slot into fake_agnus easily.

panning standard output from paula is ch0 & 3 -> left & ch 1 & 2 -> right. could expose in paula.v the raw 4 so the routing could just be tiliqua.inN -> paula.chN -> tiliqua.outN

xN paulas why init one paula when you can init two? basically you could have one per eurorack-pmod.

cv mod rather than control AUDx registers with midi they could handled by audio rate cv in. would just have to wire things differently.

modulation ring real paula only allowed chN to modulate chN+1 ( for N in 0,1,2 ) but why not have ch3 modulate ch0 ? works for the lyra8 well enough!

coding agents as over enthusiastic interns

how good are agents?! they definitely helped me in a number of places. i probably could have vibe coded this from scratch but that would have robbed me of the pleasure of finding things out

i used an agent to refactor code to do all sorts of things; separate code into modules, generalise cut and paste code developed for handling ch0 and ch1 into loops to support N=4 channels, etc etc

but the biggest thing was around the DMA between paula and fake_agnus. my original code used the "direct audio output" approach ( p157 ) and i had that working pretty quickly. but when i switched to the DMA approach i definitely hit a wall and had a bunch of small lockup timing bugs. without an agent helping me write loads of debugging code ( e.g. highjacking audio outputs to see various pulse and strobe timing ) i probably would have given up in frustration :/

code

on github

or download a r5 bitstream

video synthesis and manipulation looks super fun but video synthx hardware is very expensive :(

so i was wondering, could we convert video to audio and then use eurorack for glitching things instead?

yes! but with one big caveat; bandwidth! audio bandwidth is measured in kHz whereas video is more commonly Mhz.

so if we want to serialise video to audio and back we have to put up with a pretty low resolution and framerate

data flow

let's consider the following data flow

• capture "video" on a raspberry pi at 320x240 at 5FPS
• resize to 128x75 and send via SPI to tiliqua
• encode frames as audio on tiliqua and send out to eurorack at 192 kHz
• read back audio from eurorack and decode back to frames
• send frames out to HDMI resized to 1024x600 at 25.175 Mhz

demo and code

this all works surprisingly well!

code on github

i've always seen keras as providing two key things; 1) a way to compose layers into models and 2) an opinionated fit loop

for the last few years i've been focussed much more on jax than tensorflow which means i've not really been using keras for any hobby stuff. am super excited that keras3 supports a jax backend because i can go back to using keras for it's clear model definition while having the flexibility to mess around with weird non standard model optimisation.

my day job has some focus around the tflite ecosystem ( recently renamed liteRT ) and this set of tools has had great(†) support for tensorflow conversion, though not-so-great (†) support for jax conversion. † for some definition of great

if you have a "standard" keras model you can do conversion to tflite via the various keras save formats. ( in which case whether the backend is jax or not is not relevant ) however after using keras3/jax for a little while i'm finding myself building composite models that don't fit into this "standard" keras model.

consider my yolz project; it's a model that i train with jax that's actually made up of two keras models, and given the multiple input nature of this model using different nested batching it's not something that fits nicely with the standard keras model.

the crux of it all is; you can compose two keras models into one, but not if one of them has been vmapped ( or has had some other jax function transform on it ) i can define the parts with keras, train them with jax/optax, but i can't put them "back" into a keras model to save for tf lite export.

orbax tflite export

jax to tflite conversion has been through a LOT of churn; i've used libs listed under .experimental. more than not it feels like...

the one i started poking with today is orbax

consider the following keras3 model

from keras.layers import Input, Conv2D, GlobalMaxPooling2D
from keras.models import Model

input = Input((64, 64, 3))
y = Conv2D(filters=8, strides=1, kernel_size=3, activation='relu', padding='same')(input)
y = GlobalMaxPooling2D()(y)
model = Model(input, y)

we can convert this to a jax function for inference with a closure over the non trainable params

params = model.trainable_variables
nt_params = model.non_trainable_variables

def predict(params, x):
  y_pred, _nt_params = model.stateless_call(params, nt_params, x, training=False)
  return y_pred

eg_x = jnp.ones((1, 64, 64, 3))
predict(params, eg_x).shape
# (1, 8)

and then pass this function to the tf lite export with no surprises.

from orbax.export import constants
jax_module = JaxModule(params, k3_predict, input_polymorphic_shape='b, ...')
converter = tf.lite.TFLiteConverter.from_concrete_functions(
   [
       jax_module.methods[constants.DEFAULT_METHOD_KEY].get_concrete_function(
           tf.TensorSpec(shape=(1, 64, 64, 3), dtype=tf.float32, name="input")
       )
   ]
)
tflite_model = converter.convert()
with open('keras.tflite', 'wb') as f:
  f.write(tflite_model)

a vmapped model

more interestingly we can make a vectorised version of the inference function that is called on a set of inputs and then does a reduction ( such as the embedding model does in yolz )

input = Input((64, 64, 3))
y = Conv2D(filters=8, strides=1, kernel_size=3, activation='relu', padding='same')(input)
y = GlobalMaxPooling2D()(y)
model = Model(input, y)

params = model.trainable_variables
nt_params = model.non_trainable_variables

def predict(params, x):
  """ the vanilla version of inference """
  y_pred, _nt_params = model.stateless_call(params, nt_params, x, training=False)
  return y_pred

def v_predict(params, x):
  """ a variant that supports multiple inputs and does a mean over the result """
  y_preds = jax.vmap(predict, [None, 0])(params, x)
  return jnp.mean(y_preds, axis=-2)

eg_x = jnp.ones((1, 4, 64, 64, 3))
v_predict(params, eg_x).shape
# (1, 8)

this time we call with the additional shape axis representing the set of inputs; in this example 4

converter = tf.lite.TFLiteConverter.from_concrete_functions(
   [
       jax_module.methods[constants.DEFAULT_METHOD_KEY].get_concrete_function(
           tf.TensorSpec(shape=(1, 4, 64, 64, 3), dtype=tf.float32, name="input")
       )
   ]
)

and we see the tflite result is correctly compiled as a reshape to collapse the leading axis ( from (B, 4, ...) to (4B, ...) for the convolution and then back to (B, 4, ...) for the reduction )

we also see how the additional plumbing is included for the mean reduction mixed with the pooling

note: all of this is additional the batching we use during training for not just hardware optimisation but also for gradient variance reduction

as i mentioned before in the yolz blog post this is all reshaping, axis plumbing etc we can do manually but it's much easier, especially as we start to nest vmaps, when the compiler does it for us

code

here's the colab

transfer learning

transfer learning is arguably the greatest success story of classifier neural nets. it works based on the idea that an N layer model can be thought of as an N-1 feature extractor followed by a 'simple' logisitic regression. transfer learning is based on the idea that the large N-1 feature extractor is reusable across related domains and you can get a long way just by retraining the classifier to get a second model.

note that this works best when the first model was trained on a set of classes that is, somehow, related to the set of the classes in the second model.

( for more details see this talk i did on self supervised learning )

what is few shot learning?

the fact that transfer learning even works is kind of an accident ( in the sense that it was never the objective of training the first model ). however there's no reason you can't make it an objective. techniques such as meta learning focus on this and aim to train a model that can adapted with transfer learning with only a few examples, sometimes as little as a single one per class ( e.g. reptile )

( for more details see this talk i did on learning to learn )

what is zero shot learning?

with transfer learning you have to provide a lot of new examples during retraining. with few shot learning you only need to provide onlt a few examples during retraining. but you can train a model that generalises to new classes without needing to retrain at all?

a zero shot learner aims to do this with a key technique being that the new class to generalise to is somehow provided as part of the model input itself.

( note how LLMs are the ultimate in zero shot learning since the use an input that is one of the richest representations we know, language )

what is zero shot detection

to do zero shot detection we need a model that has takes two inputs

1. examples of the thing we're actually trying to detect ( the 'reference objects' )
2. the image we want to find those object in ( the 'scene' )

and returns as output a binary mask of the location of the object of interest...

since we want the model to be robust to different angles of the object so we'll take as input multiple views for the object reference...

inference model

first let's consider the inference version of this model..

the N object references are run through an embedding network with the "mean" embedding being the result

the scene input is run through a standard u-net like architecture with the object embeddings combined with the scene features in the middle of the u-net ( i.e. after the downsampling, but before the upsampling ).

note we have one embedding to mix into a (10, 10) feature map so we first broadcast the embedding to match the feature map size.

the best performing approach to combining was just simple elementwise addition ( after a non linear projection of both scene features and object embeddings to a joint feature depth ). some simple experiments were run trying to use the object embeddings as a query for self attention against the scene features but it wasn't any better and was more computationally expensive. less is more sometimes i guess, and maybe it'd provide a better result for a more complex version of the problem?

training model

now consider the additional parts required for the training model...

the main change is around the object reference branch.

1. since a scene has many objects of interest we want to train against the masks of all the objects present, not just one. this means the inclusion of an additional C axis.
2. since we want a set of embeddings that will generalise well to novel examples we jointly train the embedding model with a constrastive loss. this requires having to run the embeddings for pairs of examples that will go to the constrastive loss. note: we only pass the anchors down to the scene branch.

the joint training means that the embeddings are forced to both

1. generalise well based on the constrastive loss but also
2. be useful for features to include in the scene branch.

though the main change is around the contrastive loss, we also need to do extra broadcasting in the scene branch. in the inference model we needed to broadcast the embeddings from (E) to (10, 10, E) but now the broadcasting will be (C, E) to (C, 10, 10, E) which requires the scene features to be also broadcast from (10, 10, F) to (C, 10, 10, F)

note here the output across C remains independent predictions, i.e. there is no softmax or anything

contrastive embeddings

as a side note, how do we train the object embeddings? pretty standard...

• sample C classes
• sample N examples of each C as 'anchors' ( with randomly coloured backgrounds )
• calculate their embeddings and take the mean
• sample another set of N examples as 'positives'
• calculate their embeddings and take the mean
• train them so the cosine_sim(mean(anc_i), mean(pos_j)) = 1.0 for i==j and = 0.0 otherwise

example class 1 anchors ( top row ) and positives
example class 2 anchors and positives

( for more details on contrastive learning see this tute i did on keras.io )

this trains an embedding space for set of N examples that generalises to new classes. the random backgrounds were included for robust to the training.

model definition and fitting

this all requires a non trivial fit loop since there's no "standard" sense of a batch anywhere and the contrastive embeddings in this model being a great example of the power of jax.vmap IMHO.

an image model will by default operate with a leading batch dim (B, H, W, 3). since we want to embed N objects this fit nicely as (N, H, W, 3) -> (N, E) after which we take the mean (E) "outside" of the model forward pass.

but now we want to do this for C examples, so our data is (C, N, H, W, 3). the classic approach to this, since the CxN examples are independent, would be to reshape to (CN, H, W, 3), run through to (CN, E) after which we'd need to reshape back to (C, N, E) and finally run the mean over the second axis to get (C, E)

though it totally doable as a manual thing it's a lot simpler in my mind to compose this instead as just a vmap over the operation of (N, H, W, 3) -> (N, E) -> (E) ( especially since that last mean step happend outside the model forward pass of batch=N. )

furthermore there's another bigger reason to get vmap to do this for us....

so far we've looked at a model that has two inputs;

1. object references (C, 2, N, 64, 64, 3) and
2. scene (640, 640, 3)

but to get the best result in terms of gradient variance we actually want to batch the entire composite model and actually give inputs

1. object references (B, C, 2, N, 64, 64, 3) and
2. scene (B, 640, 640, 3)

and this nesting of vmaps inside vmaps completely handles all the required axis remapping for us.

basic results

here we show a basic example of inference on held out data.

• given an input image ( far left )
• top row is reference objects and detection for 3 best performers
• bottom row is reference objects and detection for 3 worst performers

findings

• keras3 and stateless_call with jax is a great combo and i'll be using this more
• focal loss gives a non trival bump to performance of the scene classifier, makes me wonder if an equivalent version can be used to contrastive loss.
• having a stop gradient from the scene loss to the object embeddings gives the result we expect; embeddings model has a lower loss but the overall classifier performs worse.

code

all the code on github

was doing some work with kalman filters recently and, as always, wasn't sure what the best way to tune the filter configuration.

in the past i've taken a simple grid/random search like approach but was curious how i could express this tuning as an optimisation using jax instead.

first though, what is a kalman filter? they are pretty broad concept but the main thing i've used them for is dynamic system prediction.

they operate in a two step fashion...

1. a predict step which predicts something about the system based on some internal state and
2. an update step which integrates new observation information into the filter's state ready for the next predict

in this post we'll use a simple 2D trajectory as the system. e.g. throwing an object

the task of the kalman filter to be the prediction of the objects' position at the next time step

a simple dynamics system

consider then the simple system of throwing an object under a trivial physics model

def simulate_throw(dx, dy):
  x, y = 0, 0
  for _ in range(10):
    yield x, y
    x += dx
    y += dy
    dy -= 1

we can use this to simulate a couple of throws and plot the trajectories...

draw_throw_with_colours(
    [simulate_throw_a(dx=3, dy=5), simulate_throw_a(dx=4, dy=3)],
    ['red', 'green']
)

a numpy kalman filter implementation

can we use a kalman filter to predict the next state of these systems? i do hope so, it's what they were designed for!

implementations of a kalman filter can vary a lot so let's just use this random kalman filter from the internet

note: using this implementation, on any random snippet of code from the internet, for, say, controlling a rocket or something might be correctly considered a generally "bad idea". the correctness, or otherwise, of this filter is irrelevant to the task of jaxifying it :)

# as is from https://machinelearningspace.com/2d-object-tracking-using-kalman-filter/ with some
# minor changes

class KalmanFilter(object):
  def __init__(self, dt, u_x,u_y, std_acc, x_std_meas, y_std_meas):
    """
    :param dt: sampling time (time for 1 cycle)
    :param u_x: acceleration in x-direction
    :param u_y: acceleration in y-direction
    :param std_acc: process noise magnitude
    :param x_std_meas: standard deviation of the measurement in x-direction
    :param y_std_meas: standard deviation of the measurement in y-direction
    """
    # Define sampling time
    self.dt = dt
    # Define the  control input variables
    self.u = np.matrix([[u_x],[u_y]])
    # Intial State
    self.x = np.matrix([[0, 0], [0, 0], [0, 0], [0, 0]])
    # Define the State Transition Matrix A
    self.A = np.matrix([[1, 0, self.dt, 0],
                        [0, 1, 0, self.dt],
                        [0, 0, 1, 0],
                        [0, 0, 0, 1]])
    # Define the Control Input Matrix B
    self.B = np.matrix([[(self.dt**2)/2, 0],
                        [0, (self.dt**2)/2],
                        [self.dt, 0],
                        [0, self.dt]])
    # Define Measurement Mapping Matrix
    self.H = np.matrix([[1, 0, 0, 0],
                        [0, 1, 0, 0]])
    # Initial Process Noise Covariance
    self.Q = np.matrix([[(self.dt**4)/4, 0, (self.dt**3)/2, 0],
                        [0, (self.dt**4)/4, 0, (self.dt**3)/2],
                        [(self.dt**3)/2, 0, self.dt**2, 0],
                        [0, (self.dt**3)/2, 0, self.dt**2]]) * std_acc**2
    # Initial Measurement Noise Covariance
    self.R = np.matrix([[x_std_meas**2,0],
                        [0, y_std_meas**2]])
    # Initial Covariance Matrix
    self.P = np.eye(self.A.shape[1])

  def predict(self):
    # Refer to :Eq.(9) and Eq.(10)
    # Update time state
    #x_k =Ax_(k-1) + Bu_(k-1)     Eq.(9)
    self.x = np.dot(self.A, self.x) + np.dot(self.B, self.u)
    # Calculate error covariance
    # P= A*P*A' + Q               Eq.(10)
    self.P = np.dot(np.dot(self.A, self.P), self.A.T) + self.Q
    return self.x[0]

  def update(self, z):
    # Refer to :Eq.(11), Eq.(12) and Eq.(13)
    # S = H*P*H'+R
    S = np.dot(self.H, np.dot(self.P, self.H.T)) + self.R
    # Calculate the Kalman Gain
    # K = P * H'* inv(H*P*H'+R)
    K = np.dot(np.dot(self.P, self.H.T), np.linalg.inv(S))  #Eq.(11)
    self.x = self.x + np.dot(K, (z - np.dot(self.H, self.x)))   #Eq.(12)
    I = np.eye(self.H.shape[1])
    # Update error covariance matrix
    self.P = (I - (K * self.H)) * self.P   #Eq.(13)

a couple of things to note about this filter

• recall the api is two methods; predict which provides an estimate of x and update which updates the internal state of the filter based on a real observation. the implementations of these are kinda opaque and i wouldn't be surprised if there are a stack of subtle bugs here that a dynamic systems expert could spot. i am happy to take it "as is" for the purpose of poking around in some jax
• since both predict and update change the internal state of the filter it's expected they are called in sequence; predict, update, predict, update etc
• there are a bunch of cryptically named variables; u, B, H etc, some of which are to do with the internal state of the filter ( like P ) with others representing a form of config around how we expect the dynamics of the system to behave ( like A ). these latter matrices are configured based off scalar values such as dt and x_std_meas.

predicting a trajectory with the numpy kalman filter

we can use this filter as is to make predictions about the next time step of a throw, and it's not too bad at it...

# construct a filter with some config
filter = KalmanFilter(
    dt=1.0, u_x=0, u_y=0,
    std_acc=1.0, x_std_meas=0.1, y_std_meas=0.1)

# simulate a throw
xy_trues = simulate_throw(2.8, 4.8)

# step throw the trajectory
xy_preds = []
for xy_true in xy_trues:
  # make prediction based on filter and record it
  xy_pred = filter.predict()
  xy_preds.append(xy_pred)  
  # update the filter state based on the true value
  filter.update(xy_true)

# plot the pair
# * red denotes true values,
# * green denotes predicted values based on the time step before
xy_preds = np.stack(xy_preds).squeeze()
draw_throw_with_colours(
    [xy_trues, xy_preds],
    ['red',    'green'])

porting to pure functional jax

next let's port this kalman filter to jax. there's a few aspects of this....

params and states

a key concept in the jax port is making predict and update fully functional, and that means taking a state and returning it for the two methods. i.e. something like...

def predict(params, state):
  ...
  return state, xy_pred

def update(params, state, z):
  ...
  return state

we can then have all the P, A, Q, etc go in either params or state.

note: we are going to be explicit about two different types of variables we are dealing with here...

• state represents the internal state of the filter that changes over time based on the sequence of predict, update, predict, ... calls
• params represents the configuration items, based on dt etc, that we eventually want to get gradients for ( with respect to a loss function )

the pairing of predict then update can be expressed as the following that reassigns state each method call

def predict_then_update(params, state, xy_true):
  state, xy_pred = predict(params, state)
  state = update(params, state, xy_true)
  return state, xy_pred

what type are params and state then? we want them to be collections of variables which are well supported in jax under the idea of pytrees and we can get a lot way with just thinking of them as dictionaries...

note for this jax port:

• i switched from jnp.dot to @ which, IMHO, reads easier.
• given no control inputs i've dropped u_x and u_y which were 0.0s anyways. and no u implies no B either...
• the implementation feels a bit clunky with dictionary look ups but oh well...

def predict(params, state):
  state['x'] = params['A'] @ state['x']
  state['P'] = ((params['A'] @ state['P']) @ params['A'].T) + params['Q']
  xy_pred = state['x'][0]
  return state, xy_pred

def update(params, state, z):
  # Define Measurement Mapping Matrix
  H = jnp.array([[1, 0, 0, 0],
                 [0, 1, 0, 0]])
  S = (H @ (state['P'] @ H.T)) + params['R']
  K = (state['P'] @ H.T) @ jnp.linalg.inv(S)
  state['x'] = state['x'] + (K @ (z - (H @ state['x'])))
  I = jnp.eye(4)
  state['P'] = (I - (K @ H)) @ state['P']
  return state

the final missing piece then is how we define the initial values for params and state

def default_params():
  dt = 1.0
  std_acc = 1.0
  x_std_meas, y_std_meas = 0.1, 0.1
  return {
    # Define the State Transition Matrix A
    'A': jnp.array([[1, 0, dt, 0],
                    [0, 1, 0, dt],
                    [0, 0, 1, 0],
                    [0, 0, 0, 1]]),
    # Initial Measurement Noise Covariance
    'R': jnp.array([[x_std_meas**2, 0],
                    [0, y_std_meas**2]]),
    # Initial Process Noise Covariance
    'Q': jnp.array([[(dt**4)/4, 0, (dt**3)/2, 0],
                    [0, (dt**4)/4, 0, (dt**3)/2],
                    [(dt**3)/2, 0, dt**2, 0],
                    [0, (dt**3)/2, 0, dt**2]]) * std_acc**2
  }

def initial_state():
  return {
    # Initial State
    'x': jnp.zeros((4, 2)),
    # Initial Covariance Matrix
    'P': jnp.eye(4),
  }

which all comes together like the numpy one as ...

params = default_params()
state = initial_state()
xy_trues = simulate_throw(2.8, 4.8)
xy_preds = []
for xy_true in xy_trues:
    state, xy_pred = predict_then_update(params, state, xy_true)
    xy_preds.append(xy_pred)
xy_preds = np.stack(xy_preds)

draw_throw_with_colours(
    [xy_trues, xy_preds],
    ['red',    'green'])

a minor diversion regarding rolling out kalman filters and teacher forcing.

before we go any deeper into jax land let's talk about one more aspect of using kalman filters.

the main idea touched on already is that we can use them to make a prediction about the next state of a system before we observe it.

xy_pred_t0 = predict()
# wait for actual xy_true_t0
update(xy_true_t0)
xy_pred_t1 = predict()
# wait for actual xy_true_t1
update(xy_true_t1)

but if we really trust the filter we can use it to alos handle missing observations by passing in the last predicted as the observed one when we don't have it ( for whatever reason )

xy_pred_t0 = predict()
# oh oh! for whatever reason we don't have _t0
update(xy_pred_t0)  # use predicted instead
xy_pred_t1 = predict()
# wait for actual xy_true_t1
update(xy_true_t1)

this can be handy for a signal that noisy and dropping out but it does put more pressure on the filter to be robust to any compounding error it might exhibit.

during training then ( which we still haven't talked about yet ) we can induce this by occasionally randomly dropping out xy_true values and using the prior xy_pred value instead.

( those with a background in RNNs might notice this is the same as teacher forcing )

the code to do this based on a 20% dropout can be....

def predict_then_update(params, state, has_observation, xy_true):
  state, xy_pred = predict(params, state)
  xy_for_update = jnp.where(has_observation, xy_true, xy_pred)
  state = update(params, state, xy_for_update)
  return state, xy_pred

xy_preds = []
has_observations = []
for xy_true in xy_trues:
  has_observation = rng.uniform() > 0.2
  has_observations.append(has_observation)
  state, xy_pred = predict_then_update(params, state, has_observation, xy_true)  
  xy_preds.append(xy_pred)
xy_preds = np.stack(xy_preds)

print("has_observations", has_observations

draw_throw_with_colours(
    [xy_trues, xy_preds],
    ['red', 'green'])

has_observations [True, True, True, True, True, False, False, True, False, False]

notice how the filter shoots way off after that pair of Falses. the default_params values might be good enough to predict one step in the future, but they don't look robust to predicting two or more steps in the future.

time to jax things up!!!

jax.lax.scan

a key first thing to do is to implement the for loop with jax.lax.scan so that jax can more cleanly trace it.

jax.lax.scan provides the classic functional programming idea of iterating over a method with a carried state.

note:

• we move the random allocation of has_observation into the jax function which means having to assign a key to the state that will be carried along between calls to predict_then_update_single
• the scanning works cleanly because each xy_trues sequence is the same length. in the cases where we might have a different roll out length we'd need to use something like zero padding and loss masking or bucketing by sequence length.
• we get the entire set of xy_preds in a single call from xy_trues

def initial_state(seed):
  return {
    # Initial State
    'x': jnp.zeros((4, 2)),
    # Initial Covariance Matrix
    'P': jnp.eye(4),
    # rng key for missing observations
    'key': jax.random.key(seed)
  }

def rolled_out_predict_then_update(params, seed, xy_trues, missing_rate):

  def predict_then_update_single(state, xy_true):
    state, xy_pred = predict(params, state)

    state['key'], subkey = jax.random.split(state['key'])
    has_observation = jax.random.uniform(subkey) > missing_rate
    xy_for_update = jnp.where(has_observation, xy_true, xy_pred)

    state = update(params, state, xy_for_update)
    return state, xy_pred

  _final_state, predictions = jax.lax.scan(predict_then_update_single, initial_state(seed), xy_trues)
  return predictions

xy_trues = np.array(list(simulate_throw(2.8, 4.8)))

seed = 1234414
xy_preds = jax.jit(rolled_out_predict_then_update)(default_params(), seed, xy_trues, missing_rate=0.2)

draw_throw_with_colours(
    [xy_trues, xy_preds],
    ['red',    'green'])

this example has a bunch of missed observations and the filter struggles quite a bit :/

a loss function, jax.grad & a simple training loop

with a function now that takes a list of xy_true values and returns a list of xy_pred values we can start to think about a loss, the simplest being mean square error.

def loss_fn(params, seed, xy_trues, missing_rate):
  predictions = rolled_out_predict_then_update(params, seed, xy_trues, missing_rate)
  squared_difference = (predictions - xy_trues) ** 2
  return jnp.mean(squared_difference)

it's interesting to note how the losses are quite unstable across seeds, especially as we increase the missing_rate.

for missing_rate in np.linspace(0.0, 0.9, 10):
  losses = [loss_fn(default_params(), seed, xy_trues, missing_rate)
            for seed in range(100)]  
  print(missing_rate, np.mean(losses), np.std(losses))

0.0 2.359239  2.3841858e-07
0.1 5.6141863 10.308739
0.2 18.179102 53.249737
0.3 88.48642  323.95425
0.4 173.53473 505.12674
0.5 269.37463 636.1734
0.6 497.91122 774.9014
0.7 712.81384 858.53827
0.8 989.5386  972.14777
0.9 661.3348  749.5342

having a loss function means we can get gradients with respect to the params and use them in a trivial gradient descent update step

@jax.jit
def update_step(params, seed, xy_trues):
  gradients = jax.grad(loss_fn)(params, seed, xy_trues, missing_rate=0.1)

  def apply_gradients(p, g):
    learning_rate = 1e-5
    return p - learning_rate * g

  return jax.tree_util.tree_map(apply_gradients, params, gradients)

this update step allows us to sample a trajectory, run it through the rolled out filter, calculate a loss and gradient and finally update the params...

params = default_params()
seed = 0
for _ in range(1000):
  dx = 2 + np.random.uniform() * 5  # (2, 7)
  dy = 4 + np.random.uniform() * 4  # (4, 8)
  xy_trues = simulate_throw(dx, dy)
  params = update_step(params, seed, xy_trues)
  if i % 100 == 0:
    print("loss", next_seed, loss_fn(params, next_seed, xy_trues, missing_rate=0.1))
  seed += 1

loss 0   2.5613098
loss 100 12.690718
loss 200 4.0097485
loss 300 4.4077697
loss 400 4.8039966
loss 500 3.5365129
loss 600 3.0296855
loss 700 2.361609
loss 800 3.0249815
loss 900 1.7365919

we see the loss has dropped, hooray! but how does the filter behave?

if we plot some examples of the default_params versus these trained params for a range of missing_rate we can see the trained filter is much more robust to missing values

• red represents ground truth
• green represents the filter behaviour with the default params
• yellow represents the filter behaviour with the trained params

xy_trues = simulate_throw_a(3, 5)
seed = 1234414

def throw_img(xy_trues, seed, missing_rate):
  xy_preds_initial_params = rolled_out_predict_then_update(default_params(), seed, xy_trues, missing_rate)
  xy_preds_trained_params = rolled_out_predict_then_update(params, seed, xy_trues, missing_rate)
  return draw_throw_with_colours(
      [xy_trues, xy_preds_initial_params, xy_preds_trained_params],
      ['red',    'green',                 'yellow'])

missing_rate=0.0	missing_rate=0.2	missing_rate=0.4

how do the optimised params differ?

let's look at the differences between the original params and the ones that were learnt....

[[1. 0. 1. 0.] 
 [0. 1. 0. 1.] 
 [0. 0. 1. 0.] 
 [0. 0. 0. 1.]]

[[ 0.958  0.014  0.897  0.038] 
 [ 0.024  1.016  0.021  1.020] 
 [ 0.004 -0.019  0.967 -0.003] 
 [-0.026 -0.020  0.024  1.025]]

[[0.25 0.   0.5  0.  ] 
 [0.   0.25 0.   0.5 ] 
 [0.5  0.   1.   0.  ] 
 [0.   0.5  0.   1.  ]]

[[ 0.278  0.064  0.470 -0.035] 
 [-0.001  0.279  0.013  0.513] 
 [ 0.504 -0.027  1.031  0.042] 
 [-0.002  0.465  0.003  1.005]]

[[0.01 0.  ] 
 [0.   0.01]]

[[0.069 0.029] 
 [0.011 0.137]]

we can see that A and Q had minor changes but R was changed much more, particularly the value related to y_std_meas

but what are we actually optimising?

seeing this result made me realise something. by providing the full A matrix to be optimised we end up with non-zero and non-one values where as i really only wanted to tune for dt

it is interesting that the model has tuned the transistion matrix fully but maybe in some cases it's better to constrain things and only allow it to change dt

to do this we just need to be more explict about what the actual parameter set is..

def default_params():
  return {
    'dt': 1.0,
    'std_acc': 1.0,
    'x_std_meas': 0.1,
    'y_std_meas': 0.1
  }

and then materialise A, Q and R from the params as required in predict and update

def predict(params, state):
  dt = params['dt']
  # State Transition Matrix
  A = jnp.array([[1, 0, dt, 0],
                 [0, 1, 0, dt],
                 [0, 0, 1, 0],
                 [0, 0, 0, 1]])
  # Process Noise Covariance
  Q = jnp.array([[(dt**4)/4, 0, (dt**3)/2, 0],
                 [0, (dt**4)/4, 0, (dt**3)/2],
                 [(dt**3)/2, 0, dt**2, 0],
                 [0, (dt**3)/2, 0, dt**2]]) * params['std_acc']**2

  state['x'] = A @ state['x']
  state['P'] = ((A @ state['P']) @ A.T) + Q
  xy_pred = state['x'][0]
  return state, xy_pred

def update(params, state, z):
  # Define Measurement Mapping Matrix
  H = jnp.array([[1, 0, 0, 0],
                 [0, 1, 0, 0]])
  # Measurement Noise Covariance
  R = jnp.array([[params['x_std_meas'] **2, 0],
                 [0, params['y_std_meas'] **2]])

  S = (H @ (state['P'] @ H.T)) + R
  K = (state['P'] @ H.T) @ jnp.linalg.inv(S)
  state['x'] = state['x'] + (K @ (z - (H @ state['x'])))
  I = jnp.eye(4)
  state['P'] = (I - (K @ H)) @ state['P']
  return state

doing this makes for a minor difference in the loss, and it's not really even visible in the trace visualisation.

to be honest though given the instability of the loss there are bigger things at play here in terms of ways to improve...

some extensions

jax.vmap

running an update step per example is often troublesome. not only is it slow (since we're not making use of as much vectorisation as possible) it can also suffer a lot from gradient variance problems.

the general approach is to batch things and calculate gradients with respect to multiple examples before an update step. jax.vmap is perfect for this...

we can express the loss with respect to multiple examples by using jax to make a version of the rollout function that rolls out multiple trajectories at once...

note: in_axes denotes we want to vmap transform to vectorise over the second and third arg of the loss function ( seed & xy_trues ) while broadcasting the first and last arg ( params & missing_rate )

def loss_fn(params, seeds, xy_truess, missing_rate):
  v_rolled_out_predict_then_update = jax.vmap(rolled_out_predict_then_update, in_axes=[None, 0, 0, None])
  predictions = v_rolled_out_predict_then_update(params, seeds, xy_truess, missing_rate)
  squared_difference = (predictions - xy_truess) ** 2
  return jnp.mean(squared_difference)

this loss is called as before but instead with a batch of seeds and xy_trues

batch_size = 8
xy_truess = []
seeds = []
for _ in range(batch_size):
  dx = 2 + np.random.uniform() * 5  # (2, 7)
  dy = 4 + np.random.uniform() * 4  # (4, 8)
  xy_truess.append(simulate_throw_a(dx, dy))
  seeds.append(next_seed)
  next_seed += 1
xy_truess = np.stack(xy_truess)
seeds = np.array(seeds)
...
params = update_step(params, seeds, xy_truess)

interestingly i found this didn't actually help! usually for a neural network this is a big win, for speed as well as gradient variance, but in this case it was behaving worse for me :/

why you no optax?

rolling your own optimisation step is generally (*) a bad idea, better to use an update step using an optax optimiser.

params = default_params()
opt = optax.adam(1e-6)
opt_state = opt.init(params)
...
@jax.jit
def update_step(params, opt_state, seeds, xy_truess):
  gradients = jax.grad(loss_fn)(params, seeds, xy_truess, missing_rate=0.1)  
  updates, opt_state = opt.update(gradients, opt_state)
  params = optax.apply_updates(params, updates)
  return params, opt_state
...
for i in range(1000):
  ...
  params, opt_state = update_step(params, opt_state, seeds, xy_truess)
  ...

having said that trying a couple of different optimisers gave no better result than the simple hand rolled update step from above!!! i guess the dynamics are weird enough that they aren't fitting the expected default space of adam, rmsprop, etc (?) am betting on the noisy behaviour of the rollout to be at fault...

cross product of possible data

sometimes the best way to avoid noisy samples is go nuts on a fixed large dataset size e.g. if we just materialise a large range of cross product values for dx, dy ...

N = 40
SEEDS_PER_DX_DY = 50
xy_truess = []
seeds = []
for dx_l in np.linspace(0, 1, N):  
  for dy_l in np.linspace(0, 1, N):    
    xy_trues = simulate_throw_a(dx=2+(dx_l*5), dy=4+(dy_l*4))
    for _ in range(SEEDS_PER_DX_DY):
      xy_truess.append(xy_trues)
xy_truess = np.stack(xy_truess)
seeds = np.array(range(len(xy_truess)))
print("xy_truess", xy_truess.shape, "seeds", seeds.shape)

xy_truess (80000, 10, 2) seeds (80000,)

we can then jit this fixed data into the update step

@jax.jit
def update_step(params, opt_state):
  gradients = jax.grad(loss_fn)(params, seeds, xy_truess, missing_rate=0.1)
  updates, opt_state = opt.update(gradients, opt_state)
  params = optax.apply_updates(params, updates)
  return params, opt_state

for i in range(1000):
  params, opt_state = update_step(params, opt_state)
  if i % 100 == 0:
    print("loss", i, loss_fn(params, seeds, xy_truess, missing_rate=0.1))

and we get a fast update and the most stable loss, though it flattens out pretty quick

loss 0 9.521738
loss 100 5.65931
loss 200 5.651896
loss 300 5.6428213
loss 400 5.636997
...

curiously in this formulation we've ended up with no gradients with respect to y_std_meas (?)

so i've either got a bug or it's the subtle way it's being used.. TODO!

anyways, that's my time up. check out the prototype colab i made for this post

in my last post, wavenet on an mcu, we talked about running a cached optimised version of wavenet on an MCU at 48kHz. this time we're going to get things running on an FPGA instead!

note: this post assumes you've read that last one; it describes the model architecture, some caching tricks, and describes a waveshaping task.

neural nets on an FPGA?

when doing some research on how to make things faster i stumbled on the eurorack-pmod project which is a piece of hardware that connects an FPGA with a eurorack setup. it includes 192kHz analog-digital-analog conversion and supports 4 channels in and out. just add an FPGA! perfect!

i'd never done anything beyond blinking LEDs with an FPGA but it can't be that hard.... right? right?!?!

well. turns out it wasn't that simple... i couldn't find any open source examples of compiling a neural net to an FPGA. the closest was HLS4ML but, if i understand it correctly, it works only with (expensive) propertiary FPGA toolchains :(

HLS4ML it did at least though put me onto qkeras which is a key component we'll talk about in a second.

( not interested in details? jump straight to the bottom for the demos )

multiple versions for training and optimised inference

recall from the last post that an important aspect of the MCU project was having multiple versions for training vs inference

• a keras version; the main version for training the weights. code
• a cmsisdsp_py version; a python prototype used to develop the caching inference & understand the cmsis api. code
• an inference firmware; the final c++ firmware for running on the MCU using cmsis-dsp for all the matrix math code

the FPGA version ended with a similar collection of variants;

• a qkeras version; main version for training quantised weights. code
• a fxpmath version; a python prototype of hand rolled fixed point inference ( reusing the caching inference from the MCU version ). code
• a verilog version; an FPGA version running first in simulation and then on eurorack pmod hardware. code

for both the MCU and FPGA versions we use a single firmware running on the daisy patch for training data collection.

let's expand on each part of the FPGA version...

qkeras

neural networks end up doing a lot of matrix multiplication. a lot.

for training it's common to work with a floating point representation since it gives the best continuous view of the loss space.

but for inference we can often reduce the precision of the weights and use integer arithmetic. we're motivated to do this since, for a lot of systems, integer math is much faster to run than full floating point math.

( and, actually, in a lot of systems we just simply might not have the ability to do floating point math at all! )

the systems we use for converting from floats during training to something simpler for inference is called quantisation and we'll look at two flavours for this project.

side note 1; fixed point arithmetic

for this project i mainly used fixed point numbers.

fixed point numbers are a simpler representation of floating point numbers with some constraints around range and precision but they allow the multiplication to be done as if it were integer multiplication. ( this is the first project i've done with fixed point math and i'm hooked, it's perfect for neural nets! )

the high level idea is you specify a total number of bits and then how many of those bits you want to use for the integer part of the number, and how many you want to use from representing the fractional part.

in this project all inputs, outputs, weights and biases are 16 bits in total with 4 bits for the integer and 12 bits for the fractional part. ( i.e. FP4.12 )

with "only" 4 bits for the integer part this means the range of values is +/- 2^4 = 8. though this might seem limiting, it's actually ok for a network, where generally activations etc are centered on zero ( especially if we add a bit of L2 regularisation along the way )

with 12 bits allocated to the fractional part we are able to describe numbers with a precision of 2^-12 = 0.00024.

here are some examples; we show the 16 bit binary number with a decimal point after the 4th bit to denote the change from the integer part to the fractional part.

the purpose of the qkeras model then is to train in full float32 but provide the ability to export the weights and biases in this configured fixed point configuration.

notes...

• using twos compliment for negative number means it's not quite +/-8 but it's close :)
• though the weights, biases and activations are in FP4.12 the multiplications and accumulations are calculated with FP8.24 ( since multiplying two FP4.12 numbers produces an FP8.24 result ). so in all the following we accumulate as FP8.24 and then "slice" the middle bits out to make it a FP4.12 number again.

side note 2; power of 2 quantisation

the options for quantisation can get pretty crazy too!

qkeras provides another scheme called power-of-two quantisation where all values are quantised to be only powers of two. i.e. depending on the fixed point config, a weight/bias can only only be one of [ +/- 1, 0, +/- 1/2, +/- 1/4, +/- 1/8, ...]

though this seems overly restrictive it has one HUGE important benefit... when a weight is a power of two then the "multiplication" of a feature by that weight can be simply done with a bit shift operation. and bit shifting is VERY fast.

and there are ways to "recover" representational power too, the best one i found being based around matrix factorisation.

if we have, say, a restricted weight matrix W of shape (8, 8) it can only contain those fixed values. but if we instead represent W as a a product of matrices, say, (8, 32). (32, 8) then we can see that, even though all the individual weights are restricted, the product of the matrices, an effective (8, 8) matrix, has many more possible values.

the pro is that the weights can take many more values, all the combos of w*w. the con though is we have to do two mat muls. depending on how much space we have for allocating the shift operations compared to the number of multiply unit we have, this tradeoff might be ok!

i messed around with this a lot and though it was interesting, and generally worked, it turned on that for the FPGA sizing ( i'm using at least ) the best result was to just use fixed point multiplication instead. :/ am guessing this is fault on my part in terms of poor verilog design and i still have some ideas to try at least...

fxpmath version

anyways, back to the models. the next model after the qkeras one is a fxpmath one.

the fxpmath version connects qkeras model fixed point export with the caching approach the inference logic that will go into the verilog design

the activation caching has two two elements that are basically the same as the MCU version; the left shift buffer for handling the first input and an activation cache for handling the activations between each convolutional layer.

the big difference comes in with the implementation of the convolution which i had to roll from scratch :/ but at least it allows for some very optimised parallelisation.

convolution 1d

consider a 1D convolution with kernel size K=4 & input / output feature depth of 16.

since we don't intend to stride this convolution at all ( that's handled implicitly by the activation caching ) we can treat this convolution as the following steps...

• 4 instance of an input by weight matrix multiplication; i.e. 4x a (1, 16) . (16, 16) matrix multiply.
• an accumulation of these K=4 results
• the adding of a bias
• applying an activation ( just relu for this project )

each of the x4 matrix multiplications are actually just a row by matrix multiplication, so can be decomposed into k=16 independent dot products ( we'll call this a row_by_matrix_multiply from now on )

and each of those dot products can be decomposed into k=16 independent multiplications followed by an accumulation.

the reason for being so explicit about what is independent, and what isn't, comes into play with the verilog version.

verilog version

verilog is a language used to "program" FPGAs and 's not at all like other languages, it's been really interesting to learn.

in the MPU version the two big concerns were

1. is there enough RAM to hold the samples and network? ( was never a problem ) and
2. does the code run fast enough to process a sample before the next comes in?

but the FPGA version is a little different. instead we more flexibility to design things based on what we want to run in parallel, vs what we want to run sequentially. for neural networks this gives lots of options for design!

side note 3; mats brutally short intro to verilog for matrix math

at a 30,000" view of verilog we have two main concerns; 1) executing code in parallel 2) executing blocks of code sequentially.

introducing verilog with a dot product

e.g. consider a dot product A.B with |A|=|B|=4

normally we'd think of this as simply something like a[0]*b[0] + a[1]*b[1] + a[2]*b[2] + a[3]*b[3] provided by a single np.dot call.

but if we're writing verilog we have to think of things in terms of hardware, and that means considering parallel vs sequential.

the simplest sequential way would be like the following psuedo code...

// WARNING PSEUDO CODE!
case(state):
  multiply_0:
    // recall; the following three statements are run _in parallel_ 
    accum <= 0                    // set accumulator to 0
    product <= a[0] * b[0]        // set intermediate product variable to a0.b0
    state <= multiply_1           // set the next state
  multiply_1:
    accum <= accum + product      // update accumulator with the product value from the last state
    product <= a[1] * b[1]        // set product variable to a0.b0
    state <= multiply_2           // set the next state
  multiply_2:
    accum <= accum + product
    product <= a[2] * b[2]
    state <= multiply_1
  multiply_3:
    accum <= accum + product
    product <= a[3] * b[3]
    state <= final_accumulate
  final_accumulate:
    accum <= accum + product
    state <= result_ready
  done:
    // final result available in `accum`
    state <= done

doing things this way does the dot product in 5 cycles...

an important aside...

an important aspect of how verilog works is to note that the statements in one of those case clauses all run in parallel. i.e. all right hand sides are evaluated and then assigned to the left hand side

e.g. the following would swap a and b

a <= b;
b <= a;

and if functionally equivalent to ...

b <= a;
a <= b;

so thinking in terms of sequential vs parallel we have the option to do more than one multiplication at any given time. this requires the hardware to be able to support 2 multiplications per clock cycle, but saves some clock cycles. not much in this case, but it add ups as |a| and |b| increase...

// WARNING PSEUDO CODE!
case(state):
  multiply_01:
    accum_0 <= 0
    accum_1 <= 0
    product_0 <= a[0] * b[0]
    product_1 <= a[1] * b[1]    
    state <= multiply_23
  multiply_23:
    accum_0 <= accum_0 + product_0
    accum_1 <= accum_1 + product_1  
    product_0 <= a[2] * b[2]
    product_1 <= a[3] * b[3]    
    state <= accumulate_0
  accumulate_0:
    accum_0 <= accum_0 + product_0
    accum_1 <= accum_1 + product_1    
    state <= accumulate_1
  accumulate_1:
    accum_0 <= accum_0 + accum_1
    state <= done
  done:
    // final result available in `accum_0`
    state <= done

or, if we want to go nuts, and, again, can support it, we can do all the elements multiplications at the same time, and then hierarchically accumulate the result into one.

// WARNING PSEUDO CODE!
case(state):
  multiply_all:
    // calculate all 4 elements of dot product in parallel; ( note: requires 4 available multiplication units )
    product_0 <= a[0] * b[0]
    product_1 <= a[1] * b[1]
    product_2 <= a[2] * b[2]
    product_3 <= a[3] * b[3]
    state <= accumulate_0
  accumulate_0:
    // add p0 and p1 at the same time as p2 and p3
    accum_0 <= product_0 + product_1
    accum_1 <= product_2 + product_3
  accumulate_1:
    // final add of the two
    accum_0 <= accum_0 + accum_1
    state <= done
  done:
    // final result available in `accum0`
    state <= done

this general idea of how much we do in a single clock cycles versus making values available in the next clock gives a lot of flexibility for a design

conv1d design in verilog

specifically for this neural net i've represented the K=4 conv 1d as

• a dot_product module code where the elements of dot products are calculated sequentially; i.e. x[0]*w[0] in the first step, x[1]*w[1] in the second, etc. so a dot product of N values takes N*M clock cycles ( where M is the number of cycles the multiple unit takes ) + some parallel accumulation.
• a row_by_matrix_multiply module code which runs the j=16 dot products required for each row_by_matrix_multiply in parallel
• and a conv1d module code that runs the K=4 row_by_matrix_multiplys are also in parallel, as well as handling the state machine for accumulating results with a bias and applying relu.

so we end up having all j=16 * K=4 = 64 dot products run in parallel, all together.

having said this, there are a number of ways to restructure this; e.g. if there were too many dot products to run in parallel for the 4 row_by_matrix_multiply we could run 2 of them in parallel, and then when they were finished, run the other 2. there are loads of trade offs between the number of multiple units available vs the time required to run them.

a slightly modified data set

in the MCU version i was only feeding in samples based on the embedding corner points, one of 4 types sampled randomly from...

for the first FPGA version i did this but the model was large enough that it was quickly overfitting this and basically outputing noise for the intermediate points. as such for training of this model i changed things a bit to include interpolated data.

basically we emit corner points, say (e0=0, e1=1, sine wave) as well as interpolated points, where we pick two waves, say sine and ramp, and a random point between them and train for that point as an interpolated wave between the two ( using constant power interpolation )

though i couldn't get the MCU model to converge well with this kind of data, the larger FPGA variant has no problems.

i also messed around with a 3d input embeddings; to translate between any pairing, but it didn't add anything really so i stuck with 2d.

final model

where as the final MCU model was ...

---------------------------------------------------------------------------
 Layer (type)         Output Shape      Par#  Conv1D params
---------------------------------------------------------------------------
 input (InputLayer)   [(None, 256, 3)]  0
 c0a (Conv1D)         (None, 64, 4)     52    F=4, K=4, D=1, P=causal
 c0b (Conv1D)         (None, 64, 4)     20    F=4, K=1
 c1a (Conv1D)         (None, 16, 4)     68    F=4, K=4, D=4, P=causal
 c1b (Conv1D)         (None, 16, 4)     20    F=4, K=1
 c2a (Conv1D)         (None, 4, 4)      68    F=4, K=4, D=16, P=causal
 c2b (Conv1D)         (None, 4, 4)      20    F=4, K=1
 c3a (Conv1D)         (None, 1, 8)      136   F=8, K=4, D=64, P=causal
 c3b (Conv1D)         (None, 1, 8)      72    F=8, K=1
 y_pred (Conv1D)      (None, 1, 1)      13    F=1, K=1
---------------------------------------------------------------------------
Trainable params: 465
---------------------------------------------------------------------------

... the current FPGA version i'm running is ...

-----------------------------------------------------------------
 Layer (type)                Output Shape             Param #   
-----------------------------------------------------------------
 input_1 (InputLayer)        [(None, 64, 4)]          0                                                                         
 qconv_0 (QConv1D)           (None, 16, 16)           272                                                                        
 qrelu_0 (QActivation)       (None, 16, 16)           0                                                                          
 qconv_1 (QConv1D)           (None, 4, 16)            1040                                                                       
 qrelu_1 (QActivation)       (None, 4, 16)            0                                                                          
 qconv_2 (QConv1D)           (None, 1, 4)             260                                                                        
-----------------------------------------------------------------
Trainable params: 1,572
-----------------------------------------------------------------

it has the following differences

• have lowered the depth for the model from 4 to 3; a receptive field of 64 is enough to handle this waveshaping
• have dropped the secondary 1x1 conv and am just running 1 conv1d per layer; primarily because the code is just simpler this way
• all the internal filter sizes have been increase to 16d.

so compared to the MCU version

• it has x3 the params
• it's running at 192kHz instead of 32kHz ( so x4 faster )
• and is running at a utilisation of 30% instead of 88%

to be honest utilisation is a bit harder to compare; the trade off between compute and space is quite different with an FPGA design..

for each sample coming in at 192kHz the FPGA is running a simple state machine of 1) accept next sample 2) run the sequence of qconvs and activation caches, then 3) output the result and sit in a while-true loop until the next sample. when i say above the FPGA is running at 30% what i really should say is that it's spending 70% of the time in the post sample processing while-true loop waiting for the next sample.

looking at the device utilisation we have the following..

Info: Device utilisation:
Info:             TRELLIS_IO:    11/  365     3%
Info:                   DCCA:     5/   56     8%
Info:                 DP16KD:    24/  208    11%
Info:             MULT18X18D:   134/  156    85%
Info:                EHXPLLL:     1/    4    25%
Info:             TRELLIS_FF: 21081/83640    25%
Info:           TRELLIS_COMB: 44647/83640    53%
Info:           TRELLIS_RAMW:   192/10455     1%

the pieces of interest are...

• DP16KD: which is the amount of ( one type ) of RAM being used; this looks to be dominated by the activation cache, so with only 11% being used there is a lot of room for having more layers.
• MULT18X18D: is the big one, it's the max number of multiplication DSP units being used at any one time. in this model that qconv1 with in_dim = out_dim = 16. since it's already 85% if we wanted to increase the filter size much more we might be forced to not run all 16 dot products of the 16x16 row_by_matrix_multiply at once but instead, say, do 8 in parallel, then the other 8. this would incur a latency hit, but that's totally fine given we still have a lot of clock time available to do work between samples. the trouble is the code as written would end up being tricky to refactor.

currently things are setup so that the entire network has to run before the next sample comes in. this is just because it was the simplest thing to do while i'm learning verilog and it seems like the FPGA is fast enough for it. but with a neural net it doesn't have to be like that; we really just need to finish the first layer before next sample comes, not the whole network. as long as we don't mind a little bit of output latency we can run a layer per sample clock tick. doing this would actually delay the output by number-of-layer sample clock ticks, but at 192kHz that'd be fine :)

another way to run a bigger network is to continue using the same naive MULT18X18D dsp allocation but just use an intermediate layer twice; e.g. if you have a network input -> conv0 -> output you can get extra depth but running input -> conv0 -> conv0 -> output instead. you lose a bit of representation power, since the same layer needs to model two layers, but sometimes it's worth it. in this model we'd get extra depth without having to worry about more allocation, and we've got plenty of headroom to do more compute.

the po2 work in progress model

the work in progress model i've been tinkering with for the power of two quantisation is the following...

_________________________________________________________________
 Layer (type)                Output Shape              Param #  
_________________________________________________________________
 input_1 (InputLayer)        [(None, 640, 4)]          0                                                                          
 qconv_0_qb (QConv1D)        (None, 640, 8)            136       
 qrelu_0 (QActivation)       (None, 640, 8)            0         
 qconv_1_qb (QConv1D)        (None, 640, 8)            264       
 qrelu_1 (QActivation)       (None, 640, 8)            0         
 qconv_1_1a_po2 (QConv1D)    (None, 640, 16)           144       
 qconv_1_1b_po2 (QConv1D)    (None, 640, 8)            136       
 qrelu_1_1 (QActivation)     (None, 640, 8)            0         
 qconv_1_2a_po2 (QConv1D)    (None, 640, 16)           144       
 qconv_1_2b_po2 (QConv1D)    (None, 640, 8)            136       
 qrelu_1_2 (QActivation)     (None, 640, 8)            0         
 qconv_2_qb (QConv1D)        (None, 640, 4)            132       
_________________________________________________________________
Trainable params: 1,092
_________________________________________________________________

• the layers postfixed with _qb are the normal quantised bits layers which are fixed point weights and use MULT18X18D units for inference.
• the layers postfixed with _po2 have power-of-two weights and use just shift operators for inference.

the output quality is the same and it uses less MULT18X18D units but doesn't quite fit :)

Info: Device utilisation:
Info:             TRELLIS_IO:    11/  365     3%
Info:                   DCCA:     5/   56     8%
Info:                 DP16KD:    12/  208     5%
Info:             MULT18X18D:    70/  156    44%
Info:                EHXPLLL:     1/    4    25%
Info:             TRELLIS_FF: 37356/83640    44%
Info:           TRELLIS_COMB: 85053/83640   101%   close!!
Info:           TRELLIS_RAMW:    96/10455     0%

i spend a bunch of timing trying various combos of reuse of the modules but never had a design that would meet timing for the FPGA :/

i still feel i'm doing something wrong here, and might come back to it.

waveshaping wave examples

waveforms generated by the model across the embedding space

the above images show the range of waveforms generated by the model across the two dimensional embedding space. of note...

• the corners represent the original logged waveshapes of ( from top left clockwise ) sine, zigzag, square and ramp. 50% of the training examples are of this type.
• the outer edges represent the examples given to the model during training, interpolated ( by constant power cross fade ) from pairs of corners. these represent the other 50% of training examples.
• the inner examples are all interpolations created by the model. note that the closer you get to the middle the further you are from a training example and the noisier the waves get...

waveshaping audio examples

lets look at some examples!

• green trace; core triangle wave being wave shaped
• blue trace; embedding x value, range -1 to 1
• red trace; embedding y value; range -1 to 1
• yellow; neural net output ( and the audio we hear )

( make sure subtitles are enabled! )

corners of the embedding space

an example of a triangle core wave as input and a manual transistion of the embedding values between the corners of the 2d space.

modulating an embedding point at audio rates

modulating the embedding x value at audio rates makes for some great timbres! the FPGA and the eurorack pmod have no problems handling this.

discontinuity glitches

since the model was trained only on a triangle wave if you give it something discontinuous, like a ramp or square, it glitches! :)

neural net techno

and if you sequence things, add an envelope to the embedding point & stick in some 909 kick & hats.... what do you have? neural net techno! doff. doff. some delay on the hats and clap, but no effects on the oscillator. ( mixed the hats too loud as well, story of my life )

TODOs

there is a lot of room to optimise for a larger network. see the issues on github

code

the code for training and verilog simulation is in this github repo

whereas the code representing this as core on the eurorack pmod is in this github repo on a branch

the electro smith daisy patch is a eurorack module made for prototyping; it provides a powerful audio focussed microcontroller setup (the daisy 'seed' with arm cortex-m7 running at 480MHz) along with all the connectivity required to be a eurorack module.

pretty much the first thing i thought when i saw one was; "could it run a neural net?" microcontrollers are fast, but so too is audio rate.

note: see also the updated version of this project running on an FPGA though it's worth reading this first to understand the model and waveshaping task

after a bit of research i found there are a couple of people who have done some real time audio effects processing on the daisy seed; e.g. guitarML's neural seed

all the examples i found though were quite small recurrent models and i was keen to give wavenet a go instead ( i'm a big fan of causal dilated convolutions )

wavenet 101

the wavenet architecture is a 1D convolutional network designed to operate on timeseries data. it is composed of two key structures;

• it uses causal padding to ensure that each convolution never depends on anything from the future and
• it increases dilation exponentially each layer to give the following structure

usually it's just dilation convolutions stacked but for my variant i also include one extra 1x1 conv between each dilation; a 1x1 doesn't break any of the dilation structure wavenet needs and including another non linearity is almost always a good thing.

training

the first step was to write a firmware that just recorded some audio and/or control voltages into buffers and then streamed them out over a serial connection. it's kinda clumsy, and i'm sure there's a cleaner way, but it works without having to mess around too much. the code is datalogger_firmware/

from there a wavenet could be trained ( see keras_model.py ) and it was trivial to quantise and export to a c++ lib for the microcontroller using edge impulse's bring-your-own-model

from there i got a bit stuck though integrating tensorflow with the daisy optimised arm_math.h stuff ( Make files and such aren't my expertise ) so instead i thought i'd use this chance to write my own inference. convolutions don't need much, it's all just a bunch of matrix math after all.

while poking around doing this it suddenly occured to me that you could do heavy heavy caching of the convolution activations! i was super excited, thinking i was doing some novel, and it wasn't until i was actually finished that i found out someone else had thought of the same idea basically, what they call fast-wavenet :/

oh well, it was fun to discover independently i suppose, and i ended up implementing things a bit differently.

caching activations for inference

so let's walk through the caching optimisation...

consider the following wavenet like network; it has 16 inputs being integrated thru 3 layers to a single prediction

if we consider a sliding window input at time step 4 we see that the output of node 0 ( the node with the red 0 ) will be the processed values from [0, 1, 2, 3]

a bit later on an interesting things happens at time step 8; node 1 now gets the same inputs that node 0 had 4 steps ago.

so we don't need to calculate it! when processing a timeseries we can see that the node 1 output just lags node 0 by 4 steps, and nodes 2 and 3 lag another 4 each. as long as we've got the memory for caching we can store all these in a circular buffer.

and this whole thing is stackable! in fact we only ever need to run the right hand side convolutions, as long as we have the memory to cache.

one big win for this on the daisy is that it can all run fast and small enough that we can use full float32 math everywhere. hoorah!

in terms of coding the whole thing becomes an exercise in

1. preparing the data in the right way to pass to some crazy optimised linear algebra math and
2. doing so with as little memory copying as possible :/

representative architectures

the following model has an input of 3 values => a receptive field of 256 steps, and a single output. it runs on the daisy using the above caching technique at 48KHz @88% CPU ( ~20 micro seconds per inference ). we'll describe what the 3 input values are in a bit. it's the model we'll use the examples at the end.

cNa - denotes the dilated convolutions
cNb - denotes the 1x1s convolutions that follow cNa
F - number of filters
K - kernel size; either 4 for cNa or 1 for cNb
D - dilation; K^layer# for cNa, or 1 for cNb
P - padding; 'causal' or the layer dft 'valid'

____________________________________________________________________________________________
 Layer (type)         Output Shape      Par#  Conv1D params   
============================================================================================
 input (InputLayer)   [(None, 256, 3)]  0                                   
 c0a (Conv1D)         (None, 64, 4)     52    F=4, K=4, D=1, P=causal
 c0b (Conv1D)         (None, 64, 4)     20    F=4, K=1
 c1a (Conv1D)         (None, 16, 4)     68    F=4, K=4, D=4, P=causal
 c1b (Conv1D)         (None, 16, 4)     20    F=4, K=1
 c2a (Conv1D)         (None, 4, 4)      68    F=4, K=4, D=16, P=causal
 c2b (Conv1D)         (None, 4, 4)      20    F=4, K=1
 c3a (Conv1D)         (None, 1, 8)      136   F=8, K=4, D=64, P=causal        
 c3b (Conv1D)         (None, 1, 8)      72    F=8, K=1
 y_pred (Conv1D)      (None, 1, 1)      13    F=1, K=1
============================================================================================
Total params: 465
____________________________________________________________________________________________

as a side note, the daisy can actually run at 96kHz but at this rate i could only run a smaller [c0a, c0b, c1a, c2b] model with just 2 filters each. it runs, but i couldn't get it to train on the examples i show below, so i didn't use it. shame because naming the blog post "at (almost) 100,000 inferences a second" would have had a nicer ring to it :D

it can also run slower, e.g. 32kHz, which allows either more filters per step, or even more depth => large receptive field.

but these combos demonstrate an interesting set of trade offs we have between

• depth ( which dictates the receptive field / input size )
• number of filters we can manage and
• what audio rate we can run at.

for training i just use keras.layer.Conv1D everywhere but then exported to the device in two passes; a python prototype and then the c++ code for the device.

cmsis-dsp

the final code on the daisy uses the cmsis-dsp library for all the matrix math. only three pieces end up being used though

• arm_mat_init_f32 for making the matrix structures,
• arm_mat_mult_f32 for the actual matrix multiplications and
• arm_add_f32 for the kernel accumulation and adding biases here and there.

to be honest none of these were benchmarked and i assume they are faster than if the multiplications were just rolled out (???)

python cmsis-dsp prototype

the first pass was to get the inference working as a prototype in python. the cmsisdsp python lib was used ( a pure python api equivalent ) and the code is cmsisdsp_py_version/.

having cmsisdsp was very handy to prototype, especially as i've not done any cmsis stuff before.

cmsis on the daisy

the code for the daisy is under inference_firmware/ and is a port of the python version. it's where i probably spent the most amount of time, especially block.h

it has 4 main parts

• left_shift_buffer.h to handle the shifted InputLayer
• block.h to handle the sequential running of each cNa and cNb pair
• rolling_cache.h which represents the circular buffer used for activation lag and
• regression.h which is just a simple y=mx+b using the weights from the final 1x1 Conv1D regression used during training

it all ends up being pretty straightforward but i did iterate a bit; any code involving memcpy and pointer arithmetic needs careful attention.

the "best" bit of the code is where the initial kera model is trained in a python notebook and then exported to be used in the c++ code by having python code construct a model_defn.h with a bunch of print statements :/ it's actually written to /tmp/model_defn.h no less! what a hack, lol. may god have mercy on my soul.

an example waveshaper

for a demo project i wanted to make a little waveshaper; something that takes one waveform and outputs another.

so 5 waveforms were collected from another eurorack oscillator; a triangle, sine, saw, square and a weird zigzag thing. during data collection random voltages were sampled to change the oscillator's frequency.

to convert to actual training dataset the triangle wave is used as input with one of the other waves acting as output. which output is decided by including two additional selector variables in the inputs. these variables are only ever {0, 1} during training but can act (very loosely) as an embedding since any float value can be passed in the range (0, 1) when running on the device.

the model was trained for < 1 min ( the model is pretty small, and there's not much variety in the data ).

it does well on held out test data...

but the much more interesting thing is what happens when we input values for x2 and x3 that weren't see during training data

e.g. what if we choose a point that is 20% between sine (0, 0) and square (1, 0) ? we end up with some weird non defined part of the input space.

these out of training distribution things are always the most interesting to me; values for x2 and x3 inbetween (0, 1) give a weird sort-of-interpolation. generally these results are never a smooth transistion unless there's some aspect of the loss that directs it to be so (and in this case there isn't).

we get the classic model hallucination stuff that i've always loved. ( see my older post on hallucinating softmaxs for more info )

we could encourage the model to make full use of the space by including a GAN like discriminator loss. this would be trained on random samples of values for (x2, x3). i've seen this kind of training force the model to be much more consistent for interpolated inputs.

this video of the network actually running give a better idea of the "interpolation". the green wave shows the input core triangle wave, the blue wave shows the output waveshaped result. the daisy module display shows the values for x2 and x3; with these values controlled by hand from the module on the right.

the video transistions across values of (x2, x3), moving between the corner (0, 1) values. at each corner the frequency of the core triangle input wave was adjusted over a range as well. something i hadn't considered is that when we change the frequency to be outside the range of what was seen during training the model does some weird extrapolation.

the interpolation extrapolation stuff makes some weird things, but it all sounds cool :)

appendix

further ideas could be...

• use mulaw encoding? the originl deepmind wavenet paper talks about the model output being a distribution across mulaw encoded values. that's interesting; i've used mulaw for compression in loopers and delays, but never as a categorical thing to classify!
• there's other things we can add in the block; e.g. is this model too small to use a skip connection around the 1x1 block?
• further study on relationship of audio rate vs dilation/receptive field
• GAN discriminator across randomly sampled (x2, x3)
• this is all running in float32 math but there's a number of quantised operators available too. how much bigger/faster could the network be if we swap toi quantised calculations?

note: see also the updated version of this project running on an FPGA

code on github

here's a recording; check it out!

and here's a pdf of the slides

eurosat/all is a dataset of 27,000 64x64 satellite images taken with 13 spectral bands. each image is labelled one of ten classes.

for the purpose of classification these 13 aren't equally useful, and the information in them varies across resolutions. if we were designing a sensor we might choose to use different channels in different resolutions.

how can we explore the trade off between mixed resolutions and whether to use a channel at all?

a simple baseline model

let's start with a simple baseline to see what performance we get. we won't spend too much time on this model, we just want something that we can iterate on quickly.

the simple model shown below trained on a 'training' split with adam hits 0.942 top 1 accuracy on a 2nd 'validation' split in 5 epochs. that'll do for a start.

what is the benefit of each channel?

let's check the effect of including different combos of input channels. we'll do so by introducing a channel mask.

a mask of all ones denotes using all channels and gives our baseline performance

a mask of all zeros denotes using no channels and acts as a sanity check; it gives the performance of random chance which is in line with what we expect give the balanced training set. (note: we standardise the input data so that it has zero mean per channel (with the mean, standard deviation parameters fit against training data only) so we can get this effect)

but what about if we drop just one channel? i.e. a mask of all ones except for a single zero.

from this we can see that the performance hit we get from losing a single channel is not always the same. in particular consider channel 11; if we drop that channel we get a huge hit! does that mean that if we keep only 11 that should give reasonable performance?

bbbzzzttt (or other appropriate annoying buzzer noise). channel 11 is contributing to the classification but it's not being used independently. in general this is exactly the behaviour we want from a neural network but what should we do to explore the effect of not having this dependence?

dropping out channels

consider using a dropout idea, just with input channels instead of intermediate nodes.

what behaviour do we get if we drop channels out during training? i.e. with 50% probability we replace an entire input channel with 0s?

things take longer to train and we get a slight hit in accuracy...

...but now when we mask out one channel at a time we don't get a big hit for losing any particular one.

evolving the channel selection

now that we have a model that is robust to any combo of channels what do we see if we use a simple genetic algorithm (GA) to evolve the channel mask to use with this pre trained network? a mask that represents using all channels will be the best right? right?

we'll evolve the GA using the network trained above but based on it's performance on a 3rd "ga_train" split using the inverse loss as a fitness function.

amusingly the GA finds that a mask of [1,1,0,1,0,0,0,1,1,0,1,0,1] does better marginally better than all channels, but only uses 1/2 of them!

important note: we can imagine the best performance overall would be to have the GA evolve not the channels to use from this model, but the channels to use when training from scratch. this would though require a lot more model training, basically a full training cycle per fitness evaluation :( in the approach we describe here we only have to train a single model and then have the GA just run inference.

what about different resolutions?

taking the idea of channel selection a step further, what if we got the GA to not only decide whether to use a channel or not, but also what resolution it should be in?

consider some example images across resolutions....

example images (just RGB channels shown)
orig x64	x32	x16	x8

we could then weight the use of a channel based on resolution; the higher the resolution the more the channel "costs" to use, with not using the channel at all being "free".

to support this we can change the GA to represent members not as a string of {0, 1}s but instead a sequence of {0, x8, x16, x32, x64} values per channel where these represent...

the change in the encoding of our GA is trivial, just 5 values per channel instead of 2, but before we look at that; how do we change our network?

we can do it without having to add too many extra parameters by using the magic of fully convolutional networks :)

notice how the main trunk of our first network was a series of 2d convolutions with a global spatial mean. this network will simply take as input all the resolutions we need! we can simply reuse it multiple times!

so we can have our network...

1. take the original x64 input
2. downsample it multiple times to x32, x16 and x8
3. mask out the channels so that each channel is only represented in one of the resolutions (or not represented at all if we want to ignore that channel)
4. run the main trunk network with shared parameters on each of the masked resolutions
5. combine the outputs with a simple channel concatenation
6. do one more non linear mixing (because, why not..)
7. finish with the logits

note: try as i might i can't get steps 2 to 4 to run parallelised in a pmap. asked on github about it and looks to be something you can't do at the moment.

the channel cost vs loss pareto front

when we consider channel cost vs loss there is no single best solution, it's a classic example of a pareto front where we see a tradeoff between the channel_cost and loss.

consider this sampling of 1,000 random channel masks...

rerunning the GA

the GA needs to operate with a fitness that's a single scalar; for now we just use a simple combo of (1.0 / loss) - channel_cost

running with this fitness function we evolve the solution [x16, x64, x64, x16, x32, ignore, x8, x64, x8, ignore, x8, ignore, x32]

it's on the pareto front, as we'd hope, and it's interesting that it includes a mix of resolutions including ignoring 3 channels completely :)

different mixings of loss and channel_cost would result in different GA solutions along the front

code

all the code is on github