Training in progress, step 200

.ipynb_checkpoints/eval-checkpoint.py

@@ -0,0 +1,137 @@
+#!/usr/bin/env python3
+import argparse
+import re
+from typing import Dict
+
+import torch
+from datasets import Audio, Dataset, load_dataset, load_metric
+
+from transformers import AutoFeatureExtractor, pipeline
+
+
+def log_results(result: Dataset, args: Dict[str, str]):
+    """DO NOT CHANGE. This function computes and logs the result metrics."""
+
+    log_outputs = args.log_outputs
+    dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
+
+    # load metric
+    wer = load_metric("wer")
+    cer = load_metric("cer")
+
+    # compute metrics
+    wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
+    cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
+
+    # print & log results
+    result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
+    print(result_str)
+
+    with open(f"{dataset_id}_eval_results.txt", "w") as f:
+        f.write(result_str)
+
+    # log all results in text file. Possibly interesting for analysis
+    if log_outputs is not None:
+        pred_file = f"log_{dataset_id}_predictions.txt"
+        target_file = f"log_{dataset_id}_targets.txt"
+
+        with open(pred_file, "w") as p, open(target_file, "w") as t:
+
+            # mapping function to write output
+            def write_to_file(batch, i):
+                p.write(f"{i}" + "\n")
+                p.write(batch["prediction"] + "\n")
+                t.write(f"{i}" + "\n")
+                t.write(batch["target"] + "\n")
+
+            result.map(write_to_file, with_indices=True)
+
+
+def normalize_text(text: str) -> str:
+    """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
+
+    chars_to_ignore_regex = '[,?.!\-\;\:"“%‘”�—’…–]'  # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
+
+    text = re.sub(chars_to_ignore_regex, "", text.lower())
+
+    # In addition, we can normalize the target text, e.g. removing new lines characters etc...
+    # note that order is important here!
+    token_sequences_to_ignore = ["\n\n", "\n", "   ", "  "]
+
+    for t in token_sequences_to_ignore:
+        text = " ".join(text.split(t))
+
+    return text
+
+
+def main(args):
+    # load dataset
+    dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
+
+    # for testing: only process the first two examples as a test
+    # dataset = dataset.select(range(10))
+
+    # load processor
+    feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
+    sampling_rate = feature_extractor.sampling_rate
+
+    # resample audio
+    dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
+
+    # load eval pipeline
+    if args.device is None:
+        args.device = 0 if torch.cuda.is_available() else -1
+    asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
+
+    # map function to decode audio
+    def map_to_pred(batch):
+        prediction = asr(
+            batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
+        )
+
+        batch["prediction"] = prediction["text"].replace("<s>","")
+        batch["target"] = normalize_text(batch["sentence"])
+        return batch
+
+    # run inference on all examples
+    result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
+
+    # compute and log_results
+    # do not change function below
+    log_results(result, args)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
+    )
+    parser.add_argument(
+        "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'`  for Common Voice"
+    )
+    parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
+    parser.add_argument(
+        "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
+    )
+    parser.add_argument(
+        "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
+    )
+    parser.add_argument(
+        "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
+    )
+    parser.add_argument(
+        "--device",
+        type=int,
+        default=None,
+        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
+    )
+    args = parser.parse_args()
+
+    main(args)

.ipynb_checkpoints/run-checkpoint.sh

@@ -4,13 +4,13 @@ python run_speech_recognition_ctc.py  \
 --dataset_config_name="hi" \
 --output_dir="./" \
 --overwrite_output_dir \
---max_steps="5000" \
+--max_steps="8000" \
 --per_device_train_batch_size="16" \
 --learning_rate="3e-4" \
- --warmup_steps="200" \
+ --warmup_steps="500" \
  --save_steps="200" \
  --eval_steps="400" \
---save_total_limit="1" \
+--save_total_limit="3" \
 --evaluation_strategy="steps" \
 --text_column_name="sentence" \
 --length_column_name="input_length"  \

.ipynb_checkpoints/run_wav2vec2_lm-checkpoint.py

@@ -0,0 +1,68 @@
+#!/usr/bin/env python3
+import sys
+import torch
+import re
+from datasets import load_dataset, load_metric
+from transformers import Wav2Vec2Processor, AutoModelForCTC, Wav2Vec2ProcessorWithLM
+# from transformers.models.wav2vec2.processing_wav2vec2_with_lm import Wav2Vec2ProcessorWithLM
+import torchaudio.functional as F
+import torch
+
+# decide if lm should be used for decoding or not via command line
+do_lm = bool(int(sys.argv[1]))
+eval_size = int(sys.argv[2])
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+model_path = "./"
+
+wer = load_metric("wer")
+cer = load_metric("cer")
+
+# load model and processor
+processor = Wav2Vec2ProcessorWithLM.from_pretrained(model_path) if do_lm else Wav2Vec2Processor.from_pretrained(model_path)
+model = AutoModelForCTC.from_pretrained(model_path).to(device)
+
+ds = load_dataset("common_voice", "es", split="test", streaming=True)
+ds_iter = iter(ds)
+
+
+references = []
+predictions = []
+
+
+CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", "；", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
+                   "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
+                   "{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。",
+                   "、", "﹂", "﹁", "‧", "～", "﹏", "，", "｛", "｝", "（", "）", "［", "］", "【", "】", "‥", "〽",
+                   "『", "』", "〝", "〟", "⟨", "⟩", "〜", "：", "！", "？", "♪", "؛", "/", "\\", "º", "−", "^", "ʻ", "ˆ"]
+chars_to_ignore_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"
+
+
+for _ in range(eval_size):
+    sample = next(ds_iter)
+    resampled_audio = F.resample(torch.tensor(sample["audio"]["array"]), 48_000, 16_000).numpy()
+
+    input_values = processor(resampled_audio, return_tensors="pt", sampling_rate=16_000).input_values
+    with torch.no_grad():
+        logits = model(input_values.to(device)).logits.cpu()
+
+    if do_lm:
+        output_str = processor.batch_decode(logits)[0].lower()
+    else:
+        pred_ids = torch.argmax(logits, dim=-1)
+        output_str = processor.batch_decode(pred_ids)[0].lower()
+
+    ref_str = re.sub(chars_to_ignore_regex, "", sample["sentence"]).lower()
+
+    # replace long empty strings by a single string
+    ref_str = " ".join(ref_str.split())
+
+    print(f"Pred: {output_str} | Target: {ref_str}")
+    print(50 * "=")
+
+    references.append(ref_str)
+    predictions.append(output_str)
+
+print(f"WER: {wer.compute(predictions=predictions, references=references) * 100}")
+print(f"CER: {cer.compute(predictions=predictions, references=references) * 100}")

eval.py

@@ -0,0 +1,137 @@
+#!/usr/bin/env python3
+import argparse
+import re
+from typing import Dict
+
+import torch
+from datasets import Audio, Dataset, load_dataset, load_metric
+
+from transformers import AutoFeatureExtractor, pipeline
+
+
+def log_results(result: Dataset, args: Dict[str, str]):
+    """DO NOT CHANGE. This function computes and logs the result metrics."""
+
+    log_outputs = args.log_outputs
+    dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
+
+    # load metric
+    wer = load_metric("wer")
+    cer = load_metric("cer")
+
+    # compute metrics
+    wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
+    cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
+
+    # print & log results
+    result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
+    print(result_str)
+
+    with open(f"{dataset_id}_eval_results.txt", "w") as f:
+        f.write(result_str)
+
+    # log all results in text file. Possibly interesting for analysis
+    if log_outputs is not None:
+        pred_file = f"log_{dataset_id}_predictions.txt"
+        target_file = f"log_{dataset_id}_targets.txt"
+
+        with open(pred_file, "w") as p, open(target_file, "w") as t:
+
+            # mapping function to write output
+            def write_to_file(batch, i):
+                p.write(f"{i}" + "\n")
+                p.write(batch["prediction"] + "\n")
+                t.write(f"{i}" + "\n")
+                t.write(batch["target"] + "\n")
+
+            result.map(write_to_file, with_indices=True)
+
+
+def normalize_text(text: str) -> str:
+    """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
+
+    chars_to_ignore_regex = '[,?.!\-\;\:"“%‘”�—’…–]'  # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
+
+    text = re.sub(chars_to_ignore_regex, "", text.lower())
+
+    # In addition, we can normalize the target text, e.g. removing new lines characters etc...
+    # note that order is important here!
+    token_sequences_to_ignore = ["\n\n", "\n", "   ", "  "]
+
+    for t in token_sequences_to_ignore:
+        text = " ".join(text.split(t))
+
+    return text
+
+
+def main(args):
+    # load dataset
+    dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
+
+    # for testing: only process the first two examples as a test
+    # dataset = dataset.select(range(10))
+
+    # load processor
+    feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
+    sampling_rate = feature_extractor.sampling_rate
+
+    # resample audio
+    dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
+
+    # load eval pipeline
+    if args.device is None:
+        args.device = 0 if torch.cuda.is_available() else -1
+    asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
+
+    # map function to decode audio
+    def map_to_pred(batch):
+        prediction = asr(
+            batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
+        )
+
+        batch["prediction"] = prediction["text"].replace("<s>","")
+        batch["target"] = normalize_text(batch["sentence"])
+        return batch
+
+    # run inference on all examples
+    result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
+
+    # compute and log_results
+    # do not change function below
+    log_results(result, args)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
+    )
+    parser.add_argument(
+        "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'`  for Common Voice"
+    )
+    parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
+    parser.add_argument(
+        "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
+    )
+    parser.add_argument(
+        "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
+    )
+    parser.add_argument(
+        "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
+    )
+    parser.add_argument(
+        "--device",
+        type=int,
+        default=None,
+        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
+    )
+    args = parser.parse_args()
+
+    main(args)

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5bb4a4004eb1787be92384131690112c8cddd0e2583120210818b583b8f91543
+oid sha256:602976fa45f06c0d1a97cb892978c576eaaeb5dd4f45a332752ddaecdc256eb2
 size 1262321393

run.sh

@@ -4,13 +4,13 @@ python run_speech_recognition_ctc.py  \
 --dataset_config_name="hi" \
 --output_dir="./" \
 --overwrite_output_dir \
---max_steps="5000" \
+--max_steps="8000" \
 --per_device_train_batch_size="16" \
 --learning_rate="3e-4" \
- --warmup_steps="200" \
+ --warmup_steps="500" \
  --save_steps="200" \
  --eval_steps="400" \
---save_total_limit="1" \
+--save_total_limit="3" \
 --evaluation_strategy="steps" \
 --text_column_name="sentence" \
 --length_column_name="input_length"  \

run_wav2vec2_lm.py

@@ -0,0 +1,68 @@
+#!/usr/bin/env python3
+import sys
+import torch
+import re
+from datasets import load_dataset, load_metric
+from transformers import Wav2Vec2Processor, AutoModelForCTC, Wav2Vec2ProcessorWithLM
+# from transformers.models.wav2vec2.processing_wav2vec2_with_lm import Wav2Vec2ProcessorWithLM
+import torchaudio.functional as F
+import torch
+
+# decide if lm should be used for decoding or not via command line
+do_lm = bool(int(sys.argv[1]))
+eval_size = int(sys.argv[2])
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+model_path = "./"
+
+wer = load_metric("wer")
+cer = load_metric("cer")
+
+# load model and processor
+processor = Wav2Vec2ProcessorWithLM.from_pretrained(model_path) if do_lm else Wav2Vec2Processor.from_pretrained(model_path)
+model = AutoModelForCTC.from_pretrained(model_path).to(device)
+
+ds = load_dataset("common_voice", "es", split="test", streaming=True)
+ds_iter = iter(ds)
+
+
+references = []
+predictions = []
+
+
+CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", "；", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
+                   "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
+                   "{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。",
+                   "、", "﹂", "﹁", "‧", "～", "﹏", "，", "｛", "｝", "（", "）", "［", "］", "【", "】", "‥", "〽",
+                   "『", "』", "〝", "〟", "⟨", "⟩", "〜", "：", "！", "？", "♪", "؛", "/", "\\", "º", "−", "^", "ʻ", "ˆ"]
+chars_to_ignore_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"
+
+
+for _ in range(eval_size):
+    sample = next(ds_iter)
+    resampled_audio = F.resample(torch.tensor(sample["audio"]["array"]), 48_000, 16_000).numpy()
+
+    input_values = processor(resampled_audio, return_tensors="pt", sampling_rate=16_000).input_values
+    with torch.no_grad():
+        logits = model(input_values.to(device)).logits.cpu()
+
+    if do_lm:
+        output_str = processor.batch_decode(logits)[0].lower()
+    else:
+        pred_ids = torch.argmax(logits, dim=-1)
+        output_str = processor.batch_decode(pred_ids)[0].lower()
+
+    ref_str = re.sub(chars_to_ignore_regex, "", sample["sentence"]).lower()
+
+    # replace long empty strings by a single string
+    ref_str = " ".join(ref_str.split())
+
+    print(f"Pred: {output_str} | Target: {ref_str}")
+    print(50 * "=")
+
+    references.append(ref_str)
+    predictions.append(output_str)
+
+print(f"WER: {wer.compute(predictions=predictions, references=references) * 100}")
+print(f"CER: {cer.compute(predictions=predictions, references=references) * 100}")

special_tokens_map.json

@@ -1 +1 @@
-{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "[UNK]", "pad_token": "[PAD]", "additional_special_tokens": [{"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}]}
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "[UNK]", "pad_token": "[PAD]", "additional_special_tokens": [{"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}]}

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:14da007bc161c5e872cf751d1aff466e7935b1a747b2fd05d274af66cf01f6f9
+oid sha256:016a31cdd0a756dd0bed3fa48205873370275d7ddb0e90527bd97c46b6284c3c
 size 2991