Easy audio classification with torch

This text interprets Daniel Falbel’s ‘Easy Audio Classification’ article from tensorflow/keras to torch/torchaudio. The principle objective is to introduce torchaudio and illustrate its contributions to the torch ecosystem. Right here, we concentrate on a well-liked dataset, the audio loader and the spectrogram transformer. An attention-grabbing aspect product is the parallel between torch and tensorflow, displaying typically the variations, typically the similarities between them.

Downloading and Importing

torchaudio has the speechcommand_dataset in-built. It filters out background_noise by default and lets us select between variations v0.01 and v0.02.

# set an current folder right here to cache the dataset
DATASETS_PATH <- "~/datasets/"

# 1.4GB obtain
df <- speechcommand_dataset(
  root = DATASETS_PATH, 
  url = "speech_commands_v0.01",
  obtain = TRUE
)

# anticipate folder: _background_noise_
df$EXCEPT_FOLDER
# [1] "_background_noise_"

# variety of audio recordsdata
size(df)
# [1] 64721

# a pattern
pattern <- df[1]

pattern$waveform[, 1:10]

torch_tensor
0.0001 *
 0.9155  0.3052  1.8311  1.8311 -0.3052  0.3052  2.4414  0.9155 -0.9155 -0.6104
[ CPUFloatType{1,10} ]

pattern$sample_rate
# 16000
pattern$label
# mattress

plot(pattern$waveform[1], kind = "l", col = "royalblue", important = pattern$label)

Determine 1: A pattern waveform for a ‘mattress’.

Lessons

 [1] "mattress"    "chicken"   "cat"    "canine"    "down"   "eight"  "5"  
 [8] "4"   "go"     "pleased"  "home"  "left"   "marvin" "9"  
[15] "no"     "off"    "on"     "one"    "proper"  "seven"  "sheila"
[22] "six"    "cease"   "three"  "tree"   "two"    "up"     "wow"   
[29] "sure"    "zero"  

Generator Dataloader

torch::dataloader has the identical activity as data_generator outlined within the authentic article. It’s accountable for making ready batches – together with shuffling, padding, one-hot encoding, and so forth. – and for taking good care of parallelism / system I/O orchestration.

In torch we do that by passing the prepare/check subset to torch::dataloader and encapsulating all of the batch setup logic inside a collate_fn() perform.

At this level, dataloader(train_subset) wouldn’t work as a result of the samples usually are not padded. So we have to construct our personal collate_fn() with the padding technique.

I recommend utilizing the next method when implementing the collate_fn():

1. start with collate_fn <- perform(batch) browser().
2. instantiate dataloader with the collate_fn()
3. create an atmosphere by calling enumerate(dataloader) so you possibly can ask to retrieve a batch from dataloader.
4. run atmosphere[[1]][[1]]. Now you ought to be despatched inside collate_fn() with entry to batch enter object.
5. construct the logic.

collate_fn <- perform(batch) {
  browser()
}

ds_train <- dataloader(
  train_subset, 
  batch_size = 32, 
  shuffle = TRUE, 
  collate_fn = collate_fn
)

ds_train_env <- enumerate(ds_train)
ds_train_env[[1]][[1]]

The ultimate collate_fn() pads the waveform to size 16001 after which stacks every little thing up collectively. At this level there are not any spectrograms but. We going to make spectrogram transformation part of mannequin structure.

pad_sequence <- perform(batch) {
    # Make all tensors in a batch the identical size by padding with zeros
    batch <- sapply(batch, perform(x) (x$t()))
    batch <- torch::nn_utils_rnn_pad_sequence(batch, batch_first = TRUE, padding_value = 0.)
    return(batch$permute(c(1, 3, 2)))
  }

# Closing collate_fn
collate_fn <- perform(batch) {
 # Enter construction:
 # checklist of 32 lists: checklist(waveform, sample_rate, label, speaker_id, utterance_number)
 # Transpose it
 batch <- purrr::transpose(batch)
 tensors <- batch$waveform
 targets <- batch$label_index

 # Group the checklist of tensors right into a batched tensor
 tensors <- pad_sequence(tensors)
 
 # goal encoding
 targets <- torch::torch_stack(targets)

 checklist(tensors = tensors, targets = targets) # (64, 1, 16001)
}

Batch construction is:

• batch[[1]]: waveforms – tensor with dimension (32, 1, 16001)
• batch[[2]]: targets – tensor with dimension (32, 1)

Additionally, torchaudio comes with 3 loaders, av_loader, tuner_loader, and audiofile_loader– extra to come back. set_audio_backend() is used to set one in every of them because the audio loader. Their performances differ based mostly on audio format (mp3 or wav). There isn’t a good world but: tuner_loader is finest for mp3, audiofile_loader is finest for wav, however neither of them has the choice of partially loading a pattern from an audio file with out bringing all the info into reminiscence first.

For a given audio backend we’d like go it to every employee by worker_init_fn() argument.

ds_train <- dataloader(
  train_subset, 
  batch_size = 128, 
  shuffle = TRUE, 
  collate_fn = collate_fn,
  num_workers = 16,
  worker_init_fn = perform(.) {torchaudio::set_audio_backend("audiofile_loader")},
  worker_globals = c("pad_sequence") # pad_sequence is required for collect_fn
)

ds_test <- dataloader(
  test_subset, 
  batch_size = 64, 
  shuffle = FALSE, 
  collate_fn = collate_fn,
  num_workers = 8,
  worker_globals = c("pad_sequence") # pad_sequence is required for collect_fn
)

Mannequin definition

As a substitute of keras::keras_model_sequential(), we’re going to outline a torch::nn_module(). As referenced by the unique article, the mannequin is predicated on this structure for MNIST from this tutorial, and I’ll name it ‘DanielNN’.

dan_nn <- torch::nn_module(
  "DanielNN",
  
  initialize = perform(
    window_size_ms = 30, 
    window_stride_ms = 10
  ) {
    
    # spectrogram spec
    window_size <- as.integer(16000*window_size_ms/1000)
    stride <- as.integer(16000*window_stride_ms/1000)
    fft_size <- as.integer(2^trunc(log(window_size, 2) + 1))
    n_chunks <- size(seq(0, 16000, stride))
    
    self$spectrogram <- torchaudio::transform_spectrogram(
      n_fft = fft_size, 
      win_length = window_size, 
      hop_length = stride, 
      normalized = TRUE, 
      energy = 2
    )
    
    # convs 2D
    self$conv1 <- torch::nn_conv2d(in_channels = 1, out_channels = 32, kernel_size = c(3,3))
    self$conv2 <- torch::nn_conv2d(in_channels = 32, out_channels = 64, kernel_size = c(3,3))
    self$conv3 <- torch::nn_conv2d(in_channels = 64, out_channels = 128, kernel_size = c(3,3))
    self$conv4 <- torch::nn_conv2d(in_channels = 128, out_channels = 256, kernel_size = c(3,3))
    
    # denses
    self$dense1 <- torch::nn_linear(in_features = 14336, out_features = 128)
    self$dense2 <- torch::nn_linear(in_features = 128, out_features = 30)
  },
  
  ahead = perform(x) {
    x %>% # (64, 1, 16001)
      self$spectrogram() %>% # (64, 1, 257, 101)
      torch::torch_add(0.01) %>%
      torch::torch_log() %>%
      self$conv1() %>%
      torch::nnf_relu() %>%
      torch::nnf_max_pool2d(kernel_size = c(2,2)) %>%
      
      self$conv2() %>%
      torch::nnf_relu() %>%
      torch::nnf_max_pool2d(kernel_size = c(2,2)) %>%
      
      self$conv3() %>%
      torch::nnf_relu() %>%
      torch::nnf_max_pool2d(kernel_size = c(2,2)) %>%
      
      self$conv4() %>%
      torch::nnf_relu() %>%
      torch::nnf_max_pool2d(kernel_size = c(2,2)) %>%
      
      torch::nnf_dropout(p = 0.25) %>%
      torch::torch_flatten(start_dim = 2) %>%
      
      self$dense1() %>%
      torch::nnf_relu() %>%
      torch::nnf_dropout(p = 0.5) %>%
      self$dense2() 
  }
)

mannequin <- dan_nn()


system <- torch::torch_device(if(torch::cuda_is_available()) "cuda" else "cpu")
mannequin$to(system = system)

print(mannequin)

An `nn_module` containing 2,226,846 parameters.

── Modules ──────────────────────────────────────────────────────
● spectrogram: <Spectrogram> #0 parameters
● conv1: <nn_conv2d> #320 parameters
● conv2: <nn_conv2d> #18,496 parameters
● conv3: <nn_conv2d> #73,856 parameters
● conv4: <nn_conv2d> #295,168 parameters
● dense1: <nn_linear> #1,835,136 parameters
● dense2: <nn_linear> #3,870 parameters

Mannequin becoming

In contrast to in tensorflow, there is no such thing as a mannequin %>% compile(...) step in torch, so we’re going to set loss criterion, optimizer technique and analysis metrics explicitly within the coaching loop.

loss_criterion <- torch::nn_cross_entropy_loss()
optimizer <- torch::optim_adadelta(mannequin$parameters, rho = 0.95, eps = 1e-7)
metrics <- checklist(acc = yardstick::accuracy_vec)

Coaching loop

library(glue)
library(progress)

pred_to_r <- perform(x) {
  courses <- issue(df$courses)
  courses[as.numeric(x$to(device = "cpu"))]
}

set_progress_bar <- perform(complete) {
  progress_bar$new(
    complete = complete, clear = FALSE, width = 70,
    format = ":present/:complete [:bar] - :elapsed - loss: :loss - acc: :acc"
  )
}

epochs <- 20
losses <- c()
accs <- c()

for(epoch in seq_len(epochs)) {
  pb <- set_progress_bar(size(ds_train))
  pb$message(glue("Epoch {epoch}/{epochs}"))
  coro::loop(for(batch in ds_train) {
    optimizer$zero_grad()
    predictions <- mannequin(batch[[1]]$to(system = system))
    targets <- batch[[2]]$to(system = system)
    loss <- loss_criterion(predictions, targets)
    loss$backward()
    optimizer$step()
    
    # eval experiences
    prediction_r <- pred_to_r(predictions$argmax(dim = 2))
    targets_r <- pred_to_r(targets)
    acc <- metrics$acc(targets_r, prediction_r)
    accs <- c(accs, acc)
    loss_r <- as.numeric(loss$merchandise())
    losses <- c(losses, loss_r)
    
    pb$tick(tokens = checklist(loss = spherical(imply(losses), 4), acc = spherical(imply(accs), 4)))
  })
}



# check
predictions_r <- c()
targets_r <- c()
coro::loop(for(batch_test in ds_test) {
  predictions <- mannequin(batch_test[[1]]$to(system = system))
  targets <- batch_test[[2]]$to(system = system)
  predictions_r <- c(predictions_r, pred_to_r(predictions$argmax(dim = 2)))
  targets_r <- c(targets_r, pred_to_r(targets))
})
val_acc <- metrics$acc(issue(targets_r, ranges = 1:30), issue(predictions_r, ranges = 1:30))
cat(glue("val_acc: {val_acc}nn"))

Epoch 1/20                                                            
[W SpectralOps.cpp:590] Warning: The perform torch.rfft is deprecated and will likely be eliminated in a future PyTorch launch. Use the brand new torch.fft module features, as a substitute, by importing torch.fft and calling torch.fft.fft or torch.fft.rfft. (perform operator())
354/354 [=========================] -  1m - loss: 2.6102 - acc: 0.2333
Epoch 2/20                                                            
354/354 [=========================] -  1m - loss: 1.9779 - acc: 0.4138
Epoch 3/20                                                            
354/354 [============================] -  1m - loss: 1.62 - acc: 0.519
Epoch 4/20                                                            
354/354 [=========================] -  1m - loss: 1.3926 - acc: 0.5859
Epoch 5/20                                                            
354/354 [==========================] -  1m - loss: 1.2334 - acc: 0.633
Epoch 6/20                                                            
354/354 [=========================] -  1m - loss: 1.1135 - acc: 0.6685
Epoch 7/20                                                            
354/354 [=========================] -  1m - loss: 1.0199 - acc: 0.6961
Epoch 8/20                                                            
354/354 [=========================] -  1m - loss: 0.9444 - acc: 0.7181
Epoch 9/20                                                            
354/354 [=========================] -  1m - loss: 0.8816 - acc: 0.7365
Epoch 10/20                                                           
354/354 [=========================] -  1m - loss: 0.8278 - acc: 0.7524
Epoch 11/20                                                           
354/354 [=========================] -  1m - loss: 0.7818 - acc: 0.7659
Epoch 12/20                                                           
354/354 [=========================] -  1m - loss: 0.7413 - acc: 0.7778
Epoch 13/20                                                           
354/354 [=========================] -  1m - loss: 0.7064 - acc: 0.7881
Epoch 14/20                                                           
354/354 [=========================] -  1m - loss: 0.6751 - acc: 0.7974
Epoch 15/20                                                           
354/354 [=========================] -  1m - loss: 0.6469 - acc: 0.8058
Epoch 16/20                                                           
354/354 [=========================] -  1m - loss: 0.6216 - acc: 0.8133
Epoch 17/20                                                           
354/354 [=========================] -  1m - loss: 0.5985 - acc: 0.8202
Epoch 18/20                                                           
354/354 [=========================] -  1m - loss: 0.5774 - acc: 0.8263
Epoch 19/20                                                           
354/354 [==========================] -  1m - loss: 0.5582 - acc: 0.832
Epoch 20/20                                                           
354/354 [=========================] -  1m - loss: 0.5403 - acc: 0.8374
val_acc: 0.876705979296493

Making predictions

We have already got all predictions calculated for test_subset, let’s recreate the alluvial plot from the unique article.

library(dplyr)
library(alluvial)
df_validation <- knowledge.body(
  pred_class = df$courses[predictions_r],
  class = df$courses[targets_r]
)
x <-  df_validation %>%
  mutate(right = pred_class == class) %>%
  rely(pred_class, class, right)

alluvial(
  x %>% choose(class, pred_class),
  freq = x$n,
  col = ifelse(x$right, "lightblue", "pink"),
  border = ifelse(x$right, "lightblue", "pink"),
  alpha = 0.6,
  conceal = x$n < 20
)

Determine 2: Mannequin efficiency: true labels <–> predicted labels.

Mannequin accuracy is 87,7%, considerably worse than tensorflow model from the unique publish. However, all conclusions from authentic publish nonetheless maintain.

Reuse

Textual content and figures are licensed beneath Artistic Commons Attribution CC BY 4.0. The figures which have been reused from different sources do not fall beneath this license and may be acknowledged by a observe of their caption: “Determine from …”.

Quotation

For attribution, please cite this work as

Damiani (2021, Feb. 4). Posit AI Weblog: Easy audio classification with torch. Retrieved from https://blogs.rstudio.com/tensorflow/posts/2021-02-04-simple-audio-classification-with-torch/

BibTeX quotation

@misc{athossimpleaudioclassification,
  writer = {Damiani, Athos},
  title = {Posit AI Weblog: Easy audio classification with torch},
  url = {https://blogs.rstudio.com/tensorflow/posts/2021-02-04-simple-audio-classification-with-torch/},
  12 months = {2021}
}