import functools
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import GPT2Model, GPT2Config, GPT2PreTrainedModel
from transformers.modeling_outputs import CausalLMOutputWithCrossAttentions
from transformers.utils.model_parallel_utils import get_device_map, assert_device_map
from trainer.networks import register_model
from utils.util import opt_get
class ResBlock(nn.Module):
"""
Basic residual convolutional block that uses GroupNorm.
"""
def __init__(self, chan):
super().__init__()
self.net = nn.Sequential(
nn.Conv1d(chan, chan, kernel_size=3, padding=1),
nn.GroupNorm(chan//8, chan),
nn.ReLU(),
nn.Conv1d(chan, chan, kernel_size=3, padding=1),
nn.GroupNorm(chan//8, chan)
)
def forward(self, x):
return F.relu(self.net(x) + x)
class LeanMelEncoder(nn.Module):
"""
Encodes a BxCxS MEL tensor into a latent space suitable for use with a transformer.
"""
def __init__(self, channels, mel_channels=80, resblocks_per_reduction=1):
super().__init__()
self.channels = channels
self.encoder = nn.Sequential(nn.Conv1d(mel_channels, channels//2, kernel_size=5, stride=2, padding=1),
nn.GroupNorm(channels//16, channels//2),
nn.ReLU(),
nn.Sequential(*[ResBlock(channels//2) for _ in range(resblocks_per_reduction)]),
nn.Conv1d(channels//2, channels, kernel_size=3, stride=2, padding=1),
nn.GroupNorm(channels//8, channels),
nn.ReLU(),
nn.Sequential(*[ResBlock(channels) for _ in range(resblocks_per_reduction)]),
nn.Conv1d(channels, channels, kernel_size=3, stride=2, padding=1),
nn.GroupNorm(channels//8, channels),
nn.ReLU(),
nn.Sequential(*[ResBlock(channels) for _ in range(resblocks_per_reduction)]),
)
self.reduction = 8
def forward(self, x):
for e in self.encoder:
x = e(x)
return x
def null_position_embeddings(range, dim):
"""
Helper method which simply returns a range-shaped tensor filled with zeros. Useful for emulating a no-effect
embedding.
"""
return torch.zeros((range.shape[0], range.shape[1], dim), device=range.device)
class GPT2InferenceModel(GPT2PreTrainedModel):
def __init__(self, config, gpt, text_pos_emb, norm, linear):
super().__init__(config)
self.transformer = gpt
self.text_pos_embedding = text_pos_emb
self.lm_head = nn.Sequential(norm, linear)
# Model parallel
self.model_parallel = False
self.device_map = None
self.cached_mel_emb = None
def parallelize(self, device_map=None):
self.device_map = (
get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
if device_map is None
else device_map
)
assert_device_map(self.device_map, len(self.transformer.h))
self.transformer.parallelize(self.device_map)
self.lm_head = self.lm_head.to(self.transformer.first_device)
self.model_parallel = True
def deparallelize(self):
self.transformer.deparallelize()
self.transformer = self.transformer.to("cpu")
self.lm_head = self.lm_head.to("cpu")
self.model_parallel = False
torch.cuda.empty_cache()
def get_output_embeddings(self):
return self.lm_head
def set_output_embeddings(self, new_embeddings):
self.lm_head = new_embeddings
def store_mel_emb(self, mel_emb):
self.cached_mel_emb = mel_emb
def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
token_type_ids = kwargs.get("token_type_ids", None)
# only last token for inputs_ids if past is defined in kwargs
if past:
input_ids = input_ids[:, -1].unsqueeze(-1)
if token_type_ids is not None:
token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
attention_mask = kwargs.get("attention_mask", None)
position_ids = kwargs.get("position_ids", None)
if attention_mask is not None and position_ids is None:
# create position_ids on the fly for batch generation
position_ids = attention_mask.long().cumsum(-1) - 1
position_ids.masked_fill_(attention_mask == 0, 1)
if past:
position_ids = position_ids[:, -1].unsqueeze(-1)
else:
position_ids = None
return {
"input_ids": input_ids,
"past_key_values": past,
"use_cache": kwargs.get("use_cache"),
"position_ids": position_ids,
"attention_mask": attention_mask,
"token_type_ids": token_type_ids,
}
def forward(
self,
input_ids=None,
past_key_values=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
labels=None,
use_cache=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
assert self.cached_mel_emb is not None
assert inputs_embeds is None # Not supported by this inference model.
assert labels is None # Training not supported by this inference model.
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# Create embedding
mel_len = self.cached_mel_emb.shape[1]
if input_ids.shape[1] != 1:
text_inputs = input_ids[:, mel_len:]
text_emb = self.transformer.get_input_embeddings()(text_inputs)
text_emb = text_emb + self.text_pos_embedding(torch.arange(text_emb.shape[1], device=text_emb.device))
if self.cached_mel_emb.shape[0] != text_emb.shape[0]:
mel_emb = self.cached_mel_emb.repeat_interleave(text_emb.shape[0]//self.cached_mel_emb.shape[0], 0)
else:
mel_emb = self.cached_mel_emb
emb = torch.cat([mel_emb, text_emb], dim=1)
else:
emb = self.transformer.get_input_embeddings()(input_ids) + \
self.text_pos_embedding(torch.tensor(attention_mask.shape[1]-mel_len, device=attention_mask.device)).unsqueeze(0).unsqueeze(0)
transformer_outputs = self.transformer(
inputs_embeds=emb,
past_key_values=past_key_values,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
hidden_states = transformer_outputs[0]
# Set device for model parallelism
if self.model_parallel:
torch.cuda.set_device(self.transformer.first_device)
hidden_states = hidden_states.to(self.lm_head.weight.device)
lm_logits = self.lm_head(hidden_states)
if not return_dict:
return (lm_logits,) + transformer_outputs[1:]
return CausalLMOutputWithCrossAttentions(
loss=None,
logits=lm_logits,
past_key_values=transformer_outputs.past_key_values,
hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions,
cross_attentions=transformer_outputs.cross_attentions,
)
@staticmethod
def _reorder_cache(past, beam_idx):
"""
This function is used to re-order the :obj:`past_key_values` cache if
:meth:`~transformers.PreTrainedModel.beam_search` or :meth:`~transformers.PreTrainedModel.beam_sample` is
called. This is required to match :obj:`past_key_values` with the correct beam_idx at every generation step.
"""
return tuple(
tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
for layer_past in past
)
class GptAsrHf2(nn.Module):
"""
Core module that encapsulates a set of embeddings, a MEL encoder, a GPT-style transformer and the head needed to
make its output useful.
"""
def __init__(self, layers=8, model_dim=512, heads=8, max_symbols_per_phrase=800, max_mel_frames=3000,
checkpointing=True, number_text_tokens=512, start_token=511, stop_token=0, mel_compression=256):
super().__init__()
self.number_text_tokens = number_text_tokens
self.start_token = start_token
self.stop_token = stop_token
self.max_symbols_per_phrase = max_symbols_per_phrase
self.model_dim = model_dim
self.mel_encoder = LeanMelEncoder(model_dim)
self.max_mel_frames = max_mel_frames // self.mel_encoder.reduction
self.mel_compression = mel_compression
seq_length = 2+self.max_symbols_per_phrase+self.max_mel_frames
self.gpt_config = GPT2Config(vocab_size=self.number_text_tokens,
n_positions=seq_length,
n_ctx=seq_length,
n_embd=model_dim,
n_layer=layers,
n_head=heads,
gradient_checkpointing=checkpointing,
use_cache=not checkpointing)
self.gpt = GPT2Model(self.gpt_config)
# Override the built in positional embeddings
del self.gpt.wpe
self.gpt.wpe = functools.partial(null_position_embeddings, dim=model_dim)
# This model uses its own positional embeddings, which helps discriminate between text and audio MELs.
self.text_pos_embedding = nn.Embedding(self.max_symbols_per_phrase + 1, model_dim)
self.mel_pos_embedding = nn.Embedding(self.max_mel_frames, model_dim)
self.text_solo_embedding = nn.Parameter(torch.randn(1,1,model_dim) * self.gpt.config.initializer_range, requires_grad=True)
# Head layers
self.final_norm = nn.LayerNorm(model_dim)
self.text_head = nn.Linear(model_dim, self.number_text_tokens)
# Initialize the embeddings per the GPT-2 scheme
for module in [self.text_pos_embedding, self.mel_pos_embedding]:
module.weight.data.normal_(mean=0.0, std=self.gpt.config.initializer_range)
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
def build_aligned_inputs_and_targets(self, input, start_token, stop_token):
"""
Helper function for producing inputs and outputs for the GPT model.
"""
inp = F.pad(input, (1,0), value=start_token)
tar = F.pad(input, (0,1), value=stop_token)
return inp, tar
def get_logits(self, mel_inputs, text_emb, get_attns=False):
"""
Helper function for producing text logits.
"""
if mel_inputs is None:
emb = text_emb
mel_len = 0
else:
mel_emb = self.mel_encoder(mel_inputs)
assert mel_emb.shape[-1] <= self.max_mel_frames, f'{mel_emb.shape[-1]} > {self.max_mel_frames}'
mel_emb = mel_emb.permute(0,2,1).contiguous()
mel_emb = mel_emb + self.mel_pos_embedding(torch.arange(mel_emb.shape[1], device=mel_emb.device))
emb = torch.cat([mel_emb, text_emb], dim=1)
mel_len = mel_emb.shape[1]
gpt_out = self.gpt(inputs_embeds=emb, return_dict=True, output_attentions=get_attns)
if get_attns:
return gpt_out.attentions
enc = gpt_out.last_hidden_state
text_logits = self.final_norm(enc[:, mel_len:])
text_logits = self.text_head(text_logits)
text_logits = text_logits.permute(0,2,1)
return text_logits
def forward(self, mel_inputs, wav_lengths, text_inputs, text_lengths, return_attentions=False):
"""
"Normal" forward pass which produces a text loss when given a MEL-encoded audio clip and transcribed text
targets.
"""
assert text_inputs.shape[1] <= self.max_symbols_per_phrase, str(text_inputs.shape[1])
assert text_inputs.max() <= self.number_text_tokens, str(text_inputs.max())
# Trim off excessive inputs to speed training. This might seem odd, but consider that this model is fed microbatches
# which are padded at the macro-batch level.
max_text_len = text_lengths.max()
text_inputs = text_inputs[:, :max_text_len]
max_mel_len = wav_lengths.max() // self.mel_compression
mel_inputs = mel_inputs[:, :, :max_mel_len]
text_inputs, text_targets = self.build_aligned_inputs_and_targets(text_inputs, self.start_token, self.stop_token)
text_emb = self.gpt.get_input_embeddings()(text_inputs) + \
self.text_pos_embedding(torch.arange(text_inputs.shape[1], device=text_inputs.device))
text_logits = self.get_logits(mel_inputs, text_emb, get_attns=return_attentions)
if return_attentions:
return text_logits # These weren't really the logits.
loss_text = F.cross_entropy(text_logits, text_targets.long())
return loss_text.mean(), text_logits
def text_only(self, text_inputs, text_lengths):
"""
Used to train on only text inputs.
"""
assert text_inputs.shape[1] <= self.max_symbols_per_phrase, str(text_inputs.shape[1])
assert text_inputs.max() <= self.number_text_tokens, str(text_inputs.max())
# Trim off excessive inputs to speed training. This might seem odd, but consider that this model is fed microbatches
# which are padded at the macro-batch level.
max_text_len = text_lengths.max()
text_inputs = text_inputs[:, :max_text_len]
text_inputs, text_targets = self.build_aligned_inputs_and_targets(text_inputs, self.start_token, self.stop_token)
text_emb = self.gpt.get_input_embeddings()(text_inputs) + \
self.text_pos_embedding(torch.arange(text_inputs.shape[1], device=text_inputs.device)) + \
self.text_solo_embedding
text_logits = self.get_logits(None, text_emb)
loss_text = F.cross_entropy(text_logits, text_targets.long())
return loss_text.mean(), text_logits
def inference(self, mel_inputs, wav_lengths, do_sample=False, temperature=1.0, num_beams=8):
"""
Performs inference by transcribing mel_inputs into text. Returns the text tokens.
"""
if not hasattr(self, 'inference_model'):
self.inference_model = GPT2InferenceModel(self.gpt_config, self.gpt, self.text_pos_embedding, self.final_norm, self.text_head)
# TODO: get rid of this..
max_mel_len = wav_lengths.max() // self.mel_compression
mel_inputs = mel_inputs[:, :, :max_mel_len]
mel_emb = self.mel_encoder(mel_inputs)
assert mel_emb.shape[-1] <= self.max_mel_frames
mel_emb = mel_emb.permute(0,2,1).contiguous()
mel_emb = mel_emb + self.mel_pos_embedding(torch.arange(mel_emb.shape[1], device=mel_emb.device))
self.inference_model.store_mel_emb(mel_emb)
# "fake_inputs" are stand-ins for the MEL frames, which will be injected with the prep_inputs function above.
fake_inputs = torch.full((mel_emb.shape[0],mel_emb.shape[1]+1,), fill_value=1, dtype=torch.long, device=mel_inputs.device)
fake_inputs[:,-1] = self.start_token
gen = self.inference_model.generate(fake_inputs, do_sample=do_sample, bos_token_id=self.start_token, pad_token_id=0, eos_token_id=0,
max_length=self.max_symbols_per_phrase+mel_emb.shape[1], temperature=temperature, num_beams=num_beams, use_cache=True)
return gen[:, mel_emb.shape[1]+1:]
@register_model
def register_gpt_asr_hf2(opt_net, opt):
return GptAsrHf2(**opt_get(opt_net, ['kwargs'], {}))
# Quick script that loads a model and halves the number of layers, then saves that model.
def distill():
gpt = GptAsrHf2(max_symbols_per_phrase=250, max_mel_frames=1400, layers=12, model_dim=512, heads=8)
gpt.load_state_dict(torch.load('X:\\dlas\\experiments\\train_gpt_asr_mass_hf\\models\\48000_gpt_ema.pth'))
rc = 0
i = 0
while i < len(gpt.gpt.h):
if rc % 2 != 0:
del gpt.gpt.h[i]
else:
i += 1
rc += 1
torch.save(gpt.state_dict(), 'X:\\dlas\\experiments\\train_gpt_asr_mass_hf\\models\\48000_gpt_distilled.pth')
if __name__ == '__main__':
#distill()
gpt = GptAsrHf2(max_symbols_per_phrase=250, max_mel_frames=1400, layers=16, model_dim=512, heads=8)
l = gpt(torch.randn(2,80,640), torch.tensor([100*256,20*256]), torch.randint(high=100, size=(2,80)), torch.tensor([15,60]))
gpt.text_only(torch.randint(high=100, size=(2,120)), torch.tensor([30,33]))
#start = time()
#gpt.inference(torch.randn(1,80,350), num_beams=1)
#print(f"Elapsed: {time()-start}")