PyTorch implementation of "Conformer: Convolution-augmented Transformer for Speech Recognition" (INTERSPEECH 2020)

@sooftware Can you kindly explain to me why the output lengths and targets are so different? :/ (also in outputs I get negative floats). Example shown below

The outputs are of shape [32,490,16121] (where 16121 is the len of my vocab) What is the 490 dimensions Also the outputs are probabilities right?

(outputs)
tensor([[[-9.7001, -9.6490, -9.6463,  ..., -9.6936, -9.6430, -9.7431],
         [-9.6997, -9.6487, -9.6470,  ..., -9.6903, -9.6450, -9.7416],
         [-9.6999, -9.6477, -9.6479,  ..., -9.6898, -9.6453, -9.7417],
         ...,
         [-9.7006, -9.6449, -9.6513,  ..., -9.6889, -9.6477, -9.7405],
         [-9.7003, -9.6448, -9.6512,  ..., -9.6893, -9.6477, -9.7410],
         [-9.7007, -9.6453, -9.6513,  ..., -9.6892, -9.6466, -9.7403]],

        [[-9.6844, -9.6316, -9.6387,  ..., -9.6880, -9.6269, -9.7657],
         [-9.6834, -9.6299, -9.6404,  ..., -9.6872, -9.6283, -9.7642],
         [-9.6834, -9.6334, -9.6387,  ..., -9.6864, -9.6290, -9.7616],
         ...,
         [-9.6840, -9.6299, -9.6431,  ..., -9.6830, -9.6304, -9.7608],
         [-9.6838, -9.6297, -9.6428,  ..., -9.6834, -9.6303, -9.7609],
         [-9.6842, -9.6300, -9.6428,  ..., -9.6837, -9.6292, -9.7599]],

        [[-9.6966, -9.6386, -9.6458,  ..., -9.6896, -9.6375, -9.7521],
         [-9.6974, -9.6374, -9.6462,  ..., -9.6890, -9.6369, -9.7516],
         [-9.6974, -9.6405, -9.6456,  ..., -9.6876, -9.6378, -9.7491],
         ...,
         [-9.6978, -9.6336, -9.6493,  ..., -9.6851, -9.6419, -9.7490],
         [-9.6971, -9.6334, -9.6487,  ..., -9.6863, -9.6411, -9.7501],
         [-9.6972, -9.6338, -9.6489,  ..., -9.6867, -9.6396, -9.7497]],

        ...,

        [[-9.7005, -9.6249, -9.6588,  ..., -9.6762, -9.6557, -9.7555],
         [-9.7028, -9.6266, -9.6597,  ..., -9.6765, -9.6574, -9.7542],
         [-9.7016, -9.6240, -9.6605,  ..., -9.6761, -9.6576, -9.7553],
         ...,
         [-9.7036, -9.6237, -9.6624,  ..., -9.6728, -9.6590, -9.7524],
         [-9.7034, -9.6235, -9.6620,  ..., -9.6735, -9.6589, -9.7530],
         [-9.7038, -9.6240, -9.6622,  ..., -9.6738, -9.6582, -9.7524]],

        [[-9.7058, -9.6305, -9.6566,  ..., -9.6739, -9.6557, -9.7466],
         [-9.7061, -9.6273, -9.6569,  ..., -9.6774, -9.6564, -9.7499],
         [-9.7046, -9.6280, -9.6576,  ..., -9.6772, -9.6575, -9.7498],
         ...,
         [-9.7060, -9.6263, -9.6609,  ..., -9.6714, -9.6561, -9.7461],
         [-9.7055, -9.6262, -9.6605,  ..., -9.6723, -9.6558, -9.7469],
         [-9.7058, -9.6270, -9.6606,  ..., -9.6725, -9.6552, -9.7460]],

        [[-9.7101, -9.6312, -9.6570,  ..., -9.6736, -9.6551, -9.7420],
         [-9.7102, -9.6307, -9.6579,  ..., -9.6733, -9.6576, -9.7418],
         [-9.7078, -9.6281, -9.6598,  ..., -9.6704, -9.6596, -9.7418],
         ...,
         [-9.7084, -9.6288, -9.6605,  ..., -9.6706, -9.6588, -9.7399],
         [-9.7081, -9.6286, -9.6600,  ..., -9.6714, -9.6584, -9.7406],
         [-9.7085, -9.6291, -9.6601,  ..., -9.6717, -9.6577, -9.7398]]],
       device='cuda:0', grad_fn=<LogSoftmaxBackward0>)

(output_lengths)
tensor([312, 260, 315, 320, 317, 275, 308, 291, 272, 300, 262, 227, 303, 252,
        298, 256, 303, 251, 284, 259, 263, 286, 209, 262, 166, 194, 149, 212,
        121, 114, 110,  57], device='cuda:0', dtype=torch.int32)

(target_lengths)
tensor([57, 55, 54, 50, 49, 49, 49, 48, 48, 47, 43, 42, 41, 40, 40, 39, 37, 37,
        36, 36, 36, 35, 34, 33, 29, 27, 26, 24, 20, 19, 17,  9])

I am using the following code for training and evaluation

import torch
import time
import sys
from google.colab import output
import torch.nn as nn
from conformer import Conformer
import torchmetrics
import random

cuda = torch.cuda.is_available()  
device = torch.device('cuda' if cuda else 'cpu')
print('Device:', device)

################################################################################

def train_model(model, optimizer, criterion, loader, metric):
  running_loss = 0.0
  for i, (audio,audio_len, translations, translation_len) in enumerate(loader):
    # with output.use_tags('some_outputs'):
    #   sys.stdout.write('Batch: '+ str(i+1)+'/290')
    #   sys.stdout.flush();

    #sorting inputs and targets to have targets in descending order based on len
    sorted_list,sorted_indices=torch.sort(translation_len,descending=True)

    sorted_audio=torch.zeros((32,201,1963),dtype=torch.float)
    sorted_audio_len=torch.zeros(32,dtype=torch.int)
    sorted_translations=torch.zeros((32,78),dtype=torch.int)
    sorted_translation_len=sorted_list

    for index, contentof in enumerate(translation_len):
      sorted_audio[index]=audio[sorted_indices[index]]
      sorted_audio_len[index]=audio_len[sorted_indices[index]]
      sorted_translations[index]=translations[sorted_indices[index]]

    #transpose inputs from (batch, dim, seq_len) to (batch, seq_len, dim)
    inputs=sorted_audio.to(device)
    inputs=torch.transpose(inputs, 1, 2)
    input_lengths=sorted_audio_len
    targets=sorted_translations.to(device)
    target_lengths=sorted_translation_len

    optimizer.zero_grad()
  
    # Forward propagate
    outputs, output_lengths = model(inputs, input_lengths)
    # print(outputs)

    # Calculate CTC Loss
    loss = criterion(outputs.transpose(0, 1), targets, output_lengths, target_lengths)

    loss.backward()
    optimizer.step()

    # print statistics
    running_loss += loss.item()

    output.clear(output_tags='some_outputs')

  loss_per_epoch=running_loss/(i+1)
  # print(f'Loss: {loss_per_epoch:.3f}')

  return loss_per_epoch

################################################################################

def eval_model(model, optimizer, criterion, loader, metric):
  running_loss = 0.0
  wer_calc=0.0
  random_index_per_epoch= random.randint(0, 178)

  for i, (audio,audio_len, translations, translation_len) in enumerate(loader):
    # with output.use_tags('some_outputs'):
    #   sys.stdout.write('Batch: '+ str(i+1)+'/72')
    #   sys.stdout.flush();

    #sorting inputs and targets to have targets in descending order based on len
    sorted_list,sorted_indices=torch.sort(translation_len,descending=True)

    sorted_audio=torch.zeros((32,201,1963),dtype=torch.float)
    sorted_audio_len=torch.zeros(32,dtype=torch.int)
    sorted_translations=torch.zeros((32,78),dtype=torch.int)
    sorted_translation_len=sorted_list

    for index, contentof in enumerate(translation_len):
      sorted_audio[index]=audio[sorted_indices[index]]
      sorted_audio_len[index]=audio_len[sorted_indices[index]]
      sorted_translations[index]=translations[sorted_indices[index]]

    #transpose inputs from (batch, dim, seq_len) to (batch, seq_len, dim)
    inputs=sorted_audio.to(device)
    inputs=torch.transpose(inputs, 1, 2)
    input_lengths=sorted_audio_len
    targets=sorted_translations.to(device)
    target_lengths=sorted_translation_len

    # Forward propagate
    outputs, output_lengths = model(inputs, input_lengths)
    # print(outputs)

    # Calculate CTC Loss
    loss = criterion(outputs.transpose(0, 1), targets, output_lengths, target_lengths)

    print(output_lengths)
    print(target_lengths)
    # outputs_in_words=words_vocab.convert_pred_to_words(outputs.transpose(0, 1))
    # targets_in_words=words_vocab.convert_pred_to_words(targets)
    # wer=metrics_calculation(metric, outputs_in_words,targets_in_words)
    
    break

    if (i==random_index_per_epoch):
        print(outputs_in_words,targets_in_words)

    running_loss += loss.item()
    # wer_calc += wer

    output.clear(output_tags='some_outputs')

  loss_per_epoch=running_loss/(i+1)
  wer_per_epoch=wer_calc/(i+1)

  return loss_per_epoch, wer_per_epoch

################################################################################

def train_eval_model(epochs):
  #conformer model init
  model = nn.DataParallel(Conformer(num_classes=16121, input_dim=201, encoder_dim=32, num_encoder_layers=1)).to(device)

  # Optimizers specified in the torch.optim package
  optimizer = torch.optim.Adam(model.parameters(), lr=0.0001, betas=(0.9, 0.98), eps=1e-9)

  #loss function
  criterion = nn.CTCLoss().to(device)

  #metrics init
  metric=torchmetrics.WordErrorRate()

  for epoch in range(epochs):
    print("Epoch", epoch+1)

    ############################################################################
    #TRAINING      
    model.train()
    print("Training")

    # epoch_loss=train_model(model=model,optimizer=optimizer, criterion=criterion, loader=train_loader, metric=metric)

    # print(f'Loss: {epoch_loss:.3f}')
    # print(f'WER: {epoch_wer:.3f}')

    ############################################################################
    #EVALUATION
    model.train(False)
    print("Validation")

    epoch_val_loss, epoch_val_wer=eval_model(model=model,optimizer=optimizer, criterion=criterion, loader=test_loader, metric=metric)
    
    print(f'Loss: {epoch_val_loss:.3f}')     
    print(f'WER: {epoch_val_wer:.3f}')   

################################################################################

def metrics_calculation(metric, predictions, targets):
    print(predictions)
    print(targets)
    wer=metric(predictions, targets)

    return wer



train_eval_model(1)

Hi @sooftware, thank you for coding this repo. I have a question about the relative shift function: https://github.com/sooftware/conformer/blob/c76ff16d01b149ae518f3fe66a3dd89c9ecff2fc/conformer/attention.py#L105 I don't quite understand how this function works. Could you elaborate on this?

An example input and output of size 4 is shown below, which does not really make sense to me.

Input:

tensor([[[[-0.9623, -0.3168, -1.1478, -1.3076],
          [ 0.5907, -0.0391, -0.1849, -0.6368],
          [-0.3956,  0.2142, -0.6415,  0.2196],
          [-0.8194, -0.2601,  1.1337, -0.3478]]]])

output:

tensor([[[[-1.3076,  0.0000,  0.5907, -0.0391],
          [-0.1849, -0.6368,  0.0000, -0.3956],
          [ 0.2142, -0.6415,  0.2196,  0.0000],
          [-0.8194, -0.2601,  1.1337, -0.3478]]]])

Thank you!

Hi, thanks for the great repo.

the README Usage example gives outputs as a torch tensor of ints. How would you suggest decoding these to strings (the actual speech)?

Thanks!

These are the shapes of my input, input_len, target, target_len where batch size=27

This is the setup I am running (only using first batch to check that is working before training with all the batches)

This is the error I am getting

I need some assistance here please:)

Running the code results in an error:

import torch
print(torch.__version__)
import torch.nn as nn
from conformer import Conformer

batch_size, sequence_length, dim = 3, 12345, 80

cuda = torch.cuda.is_available()  
device = torch.device('cuda' if cuda else 'cpu')

inputs = torch.rand(batch_size, sequence_length, dim).to(device)
input_lengths = torch.IntTensor([12345, 12300, 12000])
targets = torch.LongTensor([[1, 3, 3, 3, 3, 3, 4, 5, 6, 2],
                            [1, 3, 3, 3, 3, 3, 4, 5, 2, 0],
                            [1, 3, 3, 3, 3, 3, 4, 2, 0, 0]]).to(device)
target_lengths = torch.LongTensor([9, 8, 7])

model = nn.DataParallel(Conformer(num_classes=10, input_dim=dim, 
                                  encoder_dim=32, num_encoder_layers=3, 
                                  decoder_dim=32, device=device)).to(device)

# Forward propagate
outputs = model(inputs, input_lengths, targets, target_lengths)

# Recognize input speech
outputs = model.module.recognize(inputs, input_lengths)

1.9.0+cu111
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-12-eea3aeffaf58> in <module>
     21 
     22 # Forward propagate
---> 23 outputs = model(inputs, input_lengths, targets, target_lengths)
     24 
     25 # Recognize input speech

/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
   1049         if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1050                 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1051             return forward_call(*input, **kwargs)
   1052         # Do not call functions when jit is used
   1053         full_backward_hooks, non_full_backward_hooks = [], []

/opt/conda/lib/python3.8/site-packages/torch/nn/parallel/data_parallel.py in forward(self, *inputs, **kwargs)
    167             replicas = self.replicate(self.module, self.device_ids[:len(inputs)])
    168             outputs = self.parallel_apply(replicas, inputs, kwargs)
--> 169             return self.gather(outputs, self.output_device)
    170 
    171     def replicate(self, module, device_ids):

/opt/conda/lib/python3.8/site-packages/torch/nn/parallel/data_parallel.py in gather(self, outputs, output_device)
    179 
    180     def gather(self, outputs, output_device):
--> 181         return gather(outputs, output_device, dim=self.dim)
    182 
    183 

/opt/conda/lib/python3.8/site-packages/torch/nn/parallel/scatter_gather.py in gather(outputs, target_device, dim)
     76     # Setting the function to None clears the refcycle.
     77     try:
---> 78         res = gather_map(outputs)
     79     finally:
     80         gather_map = None

/opt/conda/lib/python3.8/site-packages/torch/nn/parallel/scatter_gather.py in gather_map(outputs)
     61         out = outputs[0]
     62         if isinstance(out, torch.Tensor):
---> 63             return Gather.apply(target_device, dim, *outputs)
     64         if out is None:
     65             return None

/opt/conda/lib/python3.8/site-packages/torch/nn/parallel/_functions.py in forward(ctx, target_device, dim, *inputs)
     73             ctx.unsqueezed_scalar = False
     74         ctx.input_sizes = tuple(i.size(ctx.dim) for i in inputs)
---> 75         return comm.gather(inputs, ctx.dim, ctx.target_device)
     76 
     77     @staticmethod

/opt/conda/lib/python3.8/site-packages/torch/nn/parallel/comm.py in gather(tensors, dim, destination, out)
    233                 'device object or string instead, e.g., "cpu".')
    234         destination = _get_device_index(destination, allow_cpu=True, optional=True)
--> 235         return torch._C._gather(tensors, dim, destination)
    236     else:
    237         if destination is not None:

RuntimeError: Input tensor at index 1 has invalid shape [1, 3085, 8, 10], but expected [1, 3085, 9, 10]

I am using version Python 3.8.8. Which version should it work with?

The inputs.to(self.device) in ConformerConvmodule and FeedForwardModule will cause the network graph in tensorboard to fork and appear kind of messy. Is there any special reason to write like that? Since in most cases we should have send both the model and tensor to the device before we input the tensor to the model, probably no more sending action is needed?

Hello sooftware, thank you very much for your wonderful work!

When I run the sample code in Usage of README:

import torch
import torch.nn as nn
from conformer import Conformer

batch_size, sequence_length, dim = 3, 12345, 80

cuda = torch.cuda.is_available()  
device = torch.device('cuda' if cuda else 'cpu')

inputs = torch.rand(batch_size, sequence_length, dim).to(device)
input_lengths = torch.IntTensor([12345, 12300, 12000])
targets = torch.LongTensor([[1, 3, 3, 3, 3, 3, 4, 5, 6, 2],
                            [1, 3, 3, 3, 3, 3, 4, 5, 2, 0],
                            [1, 3, 3, 3, 3, 3, 4, 2, 0, 0]]).to(device)
target_lengths = torch.LongTensor([9, 8, 7])

model = nn.DataParallel(Conformer(num_classes=10, input_dim=dim, 
                                  encoder_dim=32, num_encoder_layers=3, 
                                  decoder_dim=32, device=device)).to(device)

# Forward propagate
outputs = model(inputs, input_lengths, targets, target_lengths)

# Recognize input speech
outputs = model.module.recognize(inputs, input_lengths)

I got this error:

Traceback (most recent call last):
  File "/home/xuchutian/ASR/sooftware-conformer/try.py", line 36, in <module>
    outputs = model(inputs, input_lengths, targets, target_lengths)
  File "/home/yangyi/anaconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 550, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/yangyi/anaconda3/lib/python3.8/site-packages/torch/nn/parallel/data_parallel.py", line 156, in forward
    return self.gather(outputs, self.output_device)
  File "/home/yangyi/anaconda3/lib/python3.8/site-packages/torch/nn/parallel/data_parallel.py", line 168, in gather
    return gather(outputs, output_device, dim=self.dim)
  File "/home/yangyi/anaconda3/lib/python3.8/site-packages/torch/nn/parallel/scatter_gather.py", line 68, in gather
    res = gather_map(outputs)
  File "/home/yangyi/anaconda3/lib/python3.8/site-packages/torch/nn/parallel/scatter_gather.py", line 55, in gather_map
    return Gather.apply(target_device, dim, *outputs)
  File "/home/yangyi/anaconda3/lib/python3.8/site-packages/torch/nn/parallel/_functions.py", line 68, in forward
    return comm.gather(inputs, ctx.dim, ctx.target_device)
  File "/home/yangyi/anaconda3/lib/python3.8/site-packages/torch/cuda/comm.py", line 165, in gather
    return torch._C._gather(tensors, dim, destination)
RuntimeError: Gather got an input of invalid size: got [1, 3085, 8, 10], but expected [1, 3085, 9, 10]

May I ask how to solve this error?

Thank you very much.

The import path is now absolute, which requires users to install or configuring the python path before using. However, this can be improved with relative import, so users can use the package without installing it first.

This PR solve #33 by removing device from the argument list, which will require the user to manually put input tensors to device as done in the example code in README.

The property solution mentioned in #33 is not adopted as it does work with nn.DataParallel.

When the devices of input tensor and module parameters match, the following to device on the input tensor is not required, which are removed in this PR:

https://github.com/sooftware/conformer/blob/348e8af6c156dae19e311697cbb22b9581880a12/conformer/encoder.py#L117

Besides, as positional encoding is created from a buffer whose device is changed with the module, we don't have to call to device here, which is also removed in this PR.

https://github.com/sooftware/conformer/blob/610a77667aafe533a85001298c522e7079503da4/conformer/attention.py#L147

Hi. I notice the model requires passing the device as an argument, which may have not been decided yet at the point of the module initialization. Once the device is decided, it seems we cannot easily change it. Do you consider making the device switchable? One solution may be instead of passing the device, add an attribute:

@property
def device(self):
    return next(self.parameters()).device

Hi when I tried to import conformer, I got this issue >>> from conformer import Conformer Traceback (most recent call last): File "<stdin>", line 1, in <module> File "/workspace/bert/conformer/conformer.py", line 3, in <module> from conformer import Conformer ImportError: cannot import name 'Conformer' from partially initialized module 'conformer' (most likely due to a circular import) (/workspace/bert/conformer/conformer.py) I did as the installation instruction. Would you please see where I might be wrong? Thanks.

I referred to fairseq's conformer layer multi-head attention. [code] I also confirmed that it is training.

1. math.sqrt(dim) -> math.sqrt(d_head)
2. Add relative positional encoding module
3. Fix _relative_shift method - input : B X n_head X T X 2T-1 - output : B X n_head X T X T

Is there any possibility to use pre-trained conformer model for feature extraction on another speech dataset. Have you uploaded your pre-trained model and is there any tutorial how to extract embeddings ? Thank you

Hi, I am a little confused, if I want to export the onnx, should I use the forward or the recognize function? The difference seems to be that in the recognize function, the decoder loop num is adaptive according to the encoder outputs