sanchit-gandhi HF staff commited on
Commit
1b22ac1
1 Parent(s): 6dc252b

Saving train state of step 80000

Browse files
checkpoint-80000-epoch-5/optimizer.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9d85a6ebcfaea697deb81a9f43cf52b763722cc63c52559be99475194f3b1740
3
+ size 3652769047
checkpoint-80000-epoch-5/pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:64ed77445bae74c0a44de014644335672082db652c167e03307f5aa69f497029
3
+ size 2588465818
checkpoint-80000-epoch-5/random_states_0.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9bee040cadf66c07cc28951d8b63f5309317c467c935239899967dedde236356
3
+ size 16100
checkpoint-80000-epoch-5/random_states_1.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ac2984f6c11e0d6baba6e153c36e02cf7d8849f435a8e1d10db3db9422a5e14f
3
+ size 16100
checkpoint-80000-epoch-5/random_states_2.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0328ba0db514b5427f0cf5728abf82349345bec8a02097c30900ec57f433da11
3
+ size 16100
checkpoint-80000-epoch-5/random_states_3.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:668cc4578435da562a838c3860a81ef62a34fdfed1920ccebb7657eb586c14cb
3
+ size 16100
checkpoint-80000-epoch-5/random_states_4.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c856097a2b0832cc85eba37f9cd1be76e0d395fe4b2c99ef3153199386dae5e5
3
+ size 16100
checkpoint-80000-epoch-5/random_states_5.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:01d11b3bb1f3959f58be8c638fb21de1e89e54c13aa716a178687cb1fc51ac99
3
+ size 16100
checkpoint-80000-epoch-5/random_states_6.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:18447190b31cbfd6ac8565a46dfbdc82123f510bef0eb0a72313d1150001b8f4
3
+ size 16100
checkpoint-80000-epoch-5/random_states_7.pkl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:025b2bf49a9a6fe2a6fd8cb7a98ba98f4eb439e04b5aa68fd3285c8c109bdc4f
3
+ size 16100
checkpoint-80000-epoch-5/scheduler.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fd55300ad4b9da07287969193b88215d1af9b5d03f7bf3a833f1b260aa6434c4
3
+ size 1000
starting_point_0.01_rope.json CHANGED
@@ -10,6 +10,7 @@
10
  "prompt_tokenizer_name":"google/flan-t5-base",
11
 
12
  "report_to": ["wandb"],
 
13
  "overwrite_output_dir": false,
14
  "output_dir": "./",
15
 
 
10
  "prompt_tokenizer_name":"google/flan-t5-base",
11
 
12
  "report_to": ["wandb"],
13
+ "wandb_run_name": "parler-tts-600M-cross-attention-rope",
14
  "overwrite_output_dir": false,
15
  "output_dir": "./",
16
 
training/__pycache__/arguments.cpython-311.pyc CHANGED
Binary files a/training/__pycache__/arguments.cpython-311.pyc and b/training/__pycache__/arguments.cpython-311.pyc differ
 
training/__pycache__/data.cpython-311.pyc CHANGED
Binary files a/training/__pycache__/data.cpython-311.pyc and b/training/__pycache__/data.cpython-311.pyc differ
 
training/__pycache__/eval.cpython-311.pyc CHANGED
Binary files a/training/__pycache__/eval.cpython-311.pyc and b/training/__pycache__/eval.cpython-311.pyc differ
 
training/__pycache__/utils.cpython-311.pyc CHANGED
Binary files a/training/__pycache__/utils.cpython-311.pyc and b/training/__pycache__/utils.cpython-311.pyc differ
 
training/arguments.py CHANGED
@@ -218,7 +218,7 @@ class DataTrainingArguments:
218
  metadata={
219
  "help": (
220
  "If set, filter samples with descriptions that are longer than `max_description_token_length` tokens."
221
- "Also, used to set maximum desription token length if `pad_to_max_length=True`."
222
  )
223
  },
224
  )
@@ -277,6 +277,12 @@ class DataTrainingArguments:
277
  default="parler-speech",
278
  metadata={"help": "The name of the wandb project."},
279
  )
 
 
 
 
 
 
280
  save_to_disk: str = field(
281
  default=None,
282
  metadata={
 
218
  metadata={
219
  "help": (
220
  "If set, filter samples with descriptions that are longer than `max_description_token_length` tokens."
221
+ "Also, used to set maximum description token length if `pad_to_max_length=True`."
222
  )
223
  },
224
  )
 
277
  default="parler-speech",
278
  metadata={"help": "The name of the wandb project."},
279
  )
280
+ wandb_run_name: str = field(
281
+ default=None,
282
+ metadata={
283
+ "help": "If specified, the name of the run. If not specified, wandb will give a random name to this run."
284
+ },
285
+ )
286
  save_to_disk: str = field(
287
  default=None,
288
  metadata={
training/data.py CHANGED
@@ -31,7 +31,12 @@ class DataCollatorEncodecWithPadding:
31
  audios = [feature[self.audio_column_name]["array"] for feature in features]
32
  len_audio = [len(audio) for audio in audios]
33
 
34
- batch = self.feature_extractor(audios, return_tensors="pt", padding=self.padding, max_length=self.max_length)
 
 
 
 
 
35
  batch["len_audio"] = torch.tensor(len_audio).unsqueeze(1)
36
  return batch
37
 
 
31
  audios = [feature[self.audio_column_name]["array"] for feature in features]
32
  len_audio = [len(audio) for audio in audios]
33
 
34
+ # since resampling has already been performed in the 'load_multiple_datasets' function,
35
+ # a fixed sampling_rate(44100hz) is passed to the feature_extractor.
36
+ sampling_rate = self.feature_extractor.sampling_rate
37
+ batch = self.feature_extractor(
38
+ audios, sampling_rate=sampling_rate, return_tensors="pt", padding=self.padding, max_length=self.max_length
39
+ )
40
  batch["len_audio"] = torch.tensor(len_audio).unsqueeze(1)
41
  return batch
42
 
training/eval.py CHANGED
@@ -47,8 +47,7 @@ def wer(asr_model_name_or_path, prompts, audios, device, per_device_eval_batch_s
47
  normalized_references = []
48
 
49
  for pred, ref in zip(transcriptions, prompts):
50
- normalizer = english_normalizer
51
-
52
  norm_ref = normalizer(ref)
53
  if len(norm_ref) > 0:
54
  norm_pred = normalizer(pred["text"])
 
47
  normalized_references = []
48
 
49
  for pred, ref in zip(transcriptions, prompts):
50
+ normalizer = english_normalizer if return_language and pred["chunks"][0]["language"] == "english" else basic_normalizer
 
51
  norm_ref = normalizer(ref)
52
  if len(norm_ref) > 0:
53
  norm_pred = normalizer(pred["text"])
training/run_parler_tts_training.py CHANGED
@@ -98,9 +98,6 @@ def main():
98
 
99
  ####### A. Preparation
100
  kwargs_handlers = [InitProcessGroupKwargs(timeout=timedelta(minutes=60))]
101
- if training_args.torch_compile:
102
- # TODO(YL): add more compile modes?
103
- kwargs_handlers.append(TorchDynamoPlugin(backend="inductor", mode="default")) # reduce-overhead
104
 
105
  accelerator = Accelerator(
106
  gradient_accumulation_steps=training_args.gradient_accumulation_steps,
@@ -129,6 +126,7 @@ def main():
129
  "adam_beta2": training_args.adam_beta2,
130
  "temperature": model_args.temperature,
131
  },
 
132
  )
133
 
134
  # Detecting last checkpoint and eventually continue from last checkpoint
@@ -314,6 +312,7 @@ def main():
314
  token=data_args.token,
315
  trust_remote_code=data_args.trust_remote_code,
316
  )
 
317
 
318
  # enable gradient checkpointing if necessary
319
  if training_args.gradient_checkpointing:
@@ -334,8 +333,8 @@ def main():
334
  feature_extractor_input_name = feature_extractor.model_input_names[0]
335
  audio_encoder_pad_token_id = config.decoder.pad_token_id
336
  audio_encoder_eos_token_id = config.decoder.eos_token_id
337
- audio_encoder_bos_token_id = model.generation_config.decoder_start_token_id
338
- max_length = model.generation_config.max_length
339
  num_codebooks = model.decoder.config.num_codebooks
340
  bandwidth = model_args.bandwidth
341
 
@@ -538,7 +537,7 @@ def main():
538
  logger.info(f"Dataset saved at {data_args.save_to_disk}")
539
 
540
  audio_max_length = None
541
- if training_args.torch_compile:
542
  audio_max_length = max(vectorized_datasets["train"]["target_length"])
543
  with accelerator.main_process_first():
544
  max_sample = vectorized_datasets["train"].filter(
@@ -548,6 +547,18 @@ def main():
548
  )
549
  audio_max_length = torch.tensor(max_sample[0]["labels"]).shape[1]
550
 
 
 
 
 
 
 
 
 
 
 
 
 
551
  # for large datasets it is advised to run the preprocessing on a
552
  # single machine first with ``args.preprocessing_only`` since there will mostly likely
553
  # be a timeout when running the script in distributed mode.
@@ -670,6 +681,8 @@ def main():
670
  checkpoint = last_checkpoint
671
 
672
  if accelerator.is_main_process:
 
 
673
  if training_args.push_to_hub:
674
  api = HfApi(token=training_args.hub_token)
675
 
@@ -682,8 +695,6 @@ def main():
682
  with open(os.path.join(training_args.output_dir, ".gitignore"), "w+") as gitignore:
683
  if "wandb" not in gitignore:
684
  gitignore.write("wandb\n")
685
- elif training_args.output_dir is not None:
686
- os.makedirs(training_args.output_dir, exist_ok=True)
687
  accelerator.wait_for_everyone()
688
 
689
  # Now save everything to be able to create a single processor later
@@ -740,7 +751,13 @@ def main():
740
  "do_sample": model_args.do_sample,
741
  "temperature": model_args.temperature,
742
  "max_length": model_args.max_length,
 
 
 
 
743
  }
 
 
744
 
745
  # Define gradient update step fn
746
  def train_step(
@@ -869,9 +886,11 @@ def main():
869
  # safe_serialization=False to avoid shared tensors saving issue (TODO(YL): it's a temporary fix)
870
  # https://github.com/huggingface/transformers/issues/27293#issuecomment-1872560074
871
  accelerator.save_state(output_dir=intermediate_dir, safe_serialization=False)
 
 
872
  accelerator.wait_for_everyone()
873
  if accelerator.is_main_process:
874
- rotate_checkpoints(
875
  training_args.save_total_limit, output_dir=training_args.output_dir, logger=logger
876
  )
877
 
@@ -886,6 +905,7 @@ def main():
886
  folder_path=training_args.output_dir,
887
  commit_message=f"Saving train state of step {cur_step}",
888
  run_as_future=True,
 
889
  )
890
 
891
  if training_args.do_eval and (cur_step % eval_steps == 0 or cur_step == total_train_steps):
 
98
 
99
  ####### A. Preparation
100
  kwargs_handlers = [InitProcessGroupKwargs(timeout=timedelta(minutes=60))]
 
 
 
101
 
102
  accelerator = Accelerator(
103
  gradient_accumulation_steps=training_args.gradient_accumulation_steps,
 
126
  "adam_beta2": training_args.adam_beta2,
127
  "temperature": model_args.temperature,
128
  },
129
+ init_kwargs={"wandb": {"name": data_args.wandb_run_name}} if data_args.wandb_run_name else {},
130
  )
131
 
132
  # Detecting last checkpoint and eventually continue from last checkpoint
 
312
  token=data_args.token,
313
  trust_remote_code=data_args.trust_remote_code,
314
  )
315
+ generation_config = model.generation_config
316
 
317
  # enable gradient checkpointing if necessary
318
  if training_args.gradient_checkpointing:
 
333
  feature_extractor_input_name = feature_extractor.model_input_names[0]
334
  audio_encoder_pad_token_id = config.decoder.pad_token_id
335
  audio_encoder_eos_token_id = config.decoder.eos_token_id
336
+ audio_encoder_bos_token_id = generation_config.decoder_start_token_id
337
+ max_length = generation_config.max_length
338
  num_codebooks = model.decoder.config.num_codebooks
339
  bandwidth = model_args.bandwidth
340
 
 
537
  logger.info(f"Dataset saved at {data_args.save_to_disk}")
538
 
539
  audio_max_length = None
540
+ if padding == "max_length":
541
  audio_max_length = max(vectorized_datasets["train"]["target_length"])
542
  with accelerator.main_process_first():
543
  max_sample = vectorized_datasets["train"].filter(
 
547
  )
548
  audio_max_length = torch.tensor(max_sample[0]["labels"]).shape[1]
549
 
550
+ if training_args.group_by_length:
551
+ # apply a simple heuristic to take into account audio and text lengths
552
+ def add_target_lengths(target_length, prompt, description):
553
+ return {"target_length": target_length + len(prompt) + len(description)}
554
+
555
+ with accelerator.main_process_first():
556
+ vectorized_datasets = vectorized_datasets.map(
557
+ add_target_lengths,
558
+ num_proc=num_workers,
559
+ input_columns=["target_length", "prompt_input_ids", "input_ids"],
560
+ )
561
+
562
  # for large datasets it is advised to run the preprocessing on a
563
  # single machine first with ``args.preprocessing_only`` since there will mostly likely
564
  # be a timeout when running the script in distributed mode.
 
681
  checkpoint = last_checkpoint
682
 
683
  if accelerator.is_main_process:
684
+ if training_args.output_dir is not None:
685
+ os.makedirs(training_args.output_dir, exist_ok=True)
686
  if training_args.push_to_hub:
687
  api = HfApi(token=training_args.hub_token)
688
 
 
695
  with open(os.path.join(training_args.output_dir, ".gitignore"), "w+") as gitignore:
696
  if "wandb" not in gitignore:
697
  gitignore.write("wandb\n")
 
 
698
  accelerator.wait_for_everyone()
699
 
700
  # Now save everything to be able to create a single processor later
 
751
  "do_sample": model_args.do_sample,
752
  "temperature": model_args.temperature,
753
  "max_length": model_args.max_length,
754
+ # Because of the delayed pattern mask, generation might stop earlier because of unexpected behaviour
755
+ # on the first tokens of the codebooks that are delayed.
756
+ # This fix the issue.
757
+ "min_new_tokens": num_codebooks + 1,
758
  }
759
+ for key in gen_kwargs:
760
+ generation_config.key = gen_kwargs[key]
761
 
762
  # Define gradient update step fn
763
  def train_step(
 
886
  # safe_serialization=False to avoid shared tensors saving issue (TODO(YL): it's a temporary fix)
887
  # https://github.com/huggingface/transformers/issues/27293#issuecomment-1872560074
888
  accelerator.save_state(output_dir=intermediate_dir, safe_serialization=False)
889
+ config.save_pretrained(intermediate_dir)
890
+ generation_config.save_pretrained(intermediate_dir)
891
  accelerator.wait_for_everyone()
892
  if accelerator.is_main_process:
893
+ checkpoints_to_be_deleted = rotate_checkpoints(
894
  training_args.save_total_limit, output_dir=training_args.output_dir, logger=logger
895
  )
896
 
 
905
  folder_path=training_args.output_dir,
906
  commit_message=f"Saving train state of step {cur_step}",
907
  run_as_future=True,
908
+ delete_patterns=checkpoints_to_be_deleted,
909
  )
910
 
911
  if training_args.do_eval and (cur_step % eval_steps == 0 or cur_step == total_train_steps):
training/utils.py CHANGED
@@ -3,7 +3,7 @@ import re
3
  import shutil
4
  from pathlib import Path
5
  from dataclasses import field
6
- from typing import Dict, List
7
 
8
  import torch
9
  from wandb import Audio
@@ -44,7 +44,7 @@ def sorted_checkpoints(output_dir=None, checkpoint_prefix="checkpoint") -> List[
44
  return checkpoints_sorted
45
 
46
 
47
- def rotate_checkpoints(save_total_limit=None, output_dir=None, checkpoint_prefix="checkpoint", logger=None) -> None:
48
  """Helper function to delete old checkpoints."""
49
  if save_total_limit is None or save_total_limit <= 0:
50
  return
@@ -58,6 +58,8 @@ def rotate_checkpoints(save_total_limit=None, output_dir=None, checkpoint_prefix
58
  for checkpoint in checkpoints_to_be_deleted:
59
  logger.info(f"Deleting older checkpoint [{checkpoint}] due to args.save_total_limit")
60
  shutil.rmtree(checkpoint, ignore_errors=True)
 
 
61
 
62
 
63
  def log_metric(
 
3
  import shutil
4
  from pathlib import Path
5
  from dataclasses import field
6
+ from typing import Dict, List, Union
7
 
8
  import torch
9
  from wandb import Audio
 
44
  return checkpoints_sorted
45
 
46
 
47
+ def rotate_checkpoints(save_total_limit=None, output_dir=None, checkpoint_prefix="checkpoint", logger=None) -> Union[List, None]:
48
  """Helper function to delete old checkpoints."""
49
  if save_total_limit is None or save_total_limit <= 0:
50
  return
 
58
  for checkpoint in checkpoints_to_be_deleted:
59
  logger.info(f"Deleting older checkpoint [{checkpoint}] due to args.save_total_limit")
60
  shutil.rmtree(checkpoint, ignore_errors=True)
61
+ checkpoints_to_be_deleted = [f"*{Path(checkpoint).absolute().name}*" for checkpoint in checkpoints_to_be_deleted]
62
+ return checkpoints_to_be_deleted
63
 
64
 
65
  def log_metric(