Model save

.ipynb_checkpoints/README-checkpoint.md

@@ -0,0 +1,100 @@
+---
+language:
+- de
+license: apache-2.0
+tags:
+- automatic-speech-recognition
+- mozilla-foundation/common_voice_8_0
+- generated_from_trainer
+datasets:
+- common_voice
+model-index:
+- name: ''
+  results: []
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+# 
+
+This model is a fine-tuned version of [facebook/wav2vec2-xls-r-1b](https://huggingface.co/facebook/wav2vec2-xls-r-1b) on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - DE dataset.
+It achieves the following results on the evaluation set:
+- Loss: 0.1355
+- Wer: 0.1532
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 7.5e-05
+- train_batch_size: 8
+- eval_batch_size: 8
+- seed: 42
+- gradient_accumulation_steps: 4
+- total_train_batch_size: 32
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_steps: 2000
+- num_epochs: 2.5
+- mixed_precision_training: Native AMP
+
+### Training results
+
+| Training Loss | Epoch | Step  | Validation Loss | Wer    |
+|:-------------:|:-----:|:-----:|:---------------:|:------:|
+| 1.0826        | 0.07  | 1000  | 0.4637          | 0.4654 |
+| 1.118         | 0.15  | 2000  | 0.2595          | 0.2687 |
+| 1.1268        | 0.22  | 3000  | 0.2635          | 0.2661 |
+| 1.0919        | 0.29  | 4000  | 0.2417          | 0.2566 |
+| 1.1013        | 0.37  | 5000  | 0.2414          | 0.2567 |
+| 1.0898        | 0.44  | 6000  | 0.2546          | 0.2731 |
+| 1.0808        | 0.51  | 7000  | 0.2399          | 0.2535 |
+| 1.0719        | 0.59  | 8000  | 0.2353          | 0.2528 |
+| 1.0446        | 0.66  | 9000  | 0.2427          | 0.2545 |
+| 1.0347        | 0.73  | 10000 | 0.2266          | 0.2402 |
+| 1.0457        | 0.81  | 11000 | 0.2290          | 0.2448 |
+| 1.0124        | 0.88  | 12000 | 0.2295          | 0.2448 |
+| 1.025         | 0.95  | 13000 | 0.2138          | 0.2345 |
+| 1.0107        | 1.03  | 14000 | 0.2108          | 0.2294 |
+| 0.9758        | 1.1   | 15000 | 0.2019          | 0.2204 |
+| 0.9547        | 1.17  | 16000 | 0.2000          | 0.2178 |
+| 0.986         | 1.25  | 17000 | 0.2018          | 0.2200 |
+| 0.9588        | 1.32  | 18000 | 0.1992          | 0.2138 |
+| 0.9413        | 1.39  | 19000 | 0.1898          | 0.2049 |
+| 0.9339        | 1.47  | 20000 | 0.1874          | 0.2056 |
+| 0.9268        | 1.54  | 21000 | 0.1797          | 0.1976 |
+| 0.9194        | 1.61  | 22000 | 0.1743          | 0.1905 |
+| 0.8987        | 1.69  | 23000 | 0.1738          | 0.1932 |
+| 0.8884        | 1.76  | 24000 | 0.1703          | 0.1873 |
+| 0.8939        | 1.83  | 25000 | 0.1633          | 0.1831 |
+| 0.8629        | 1.91  | 26000 | 0.1549          | 0.1750 |
+| 0.8607        | 1.98  | 27000 | 0.1550          | 0.1738 |
+| 0.8316        | 2.05  | 28000 | 0.1512          | 0.1709 |
+| 0.8321        | 2.13  | 29000 | 0.1481          | 0.1657 |
+| 0.825         | 2.2   | 30000 | 0.1446          | 0.1627 |
+| 0.8115        | 2.27  | 31000 | 0.1396          | 0.1583 |
+| 0.7959        | 2.35  | 32000 | 0.1389          | 0.1569 |
+| 0.7835        | 2.42  | 33000 | 0.1362          | 0.1545 |
+| 0.7959        | 2.49  | 34000 | 0.1355          | 0.1531 |
+
+
+### Framework versions
+
+- Transformers 4.17.0.dev0
+- Pytorch 1.10.2+cu102
+- Datasets 1.18.2.dev0
+- Tokenizers 0.11.0

.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 = '[,?.!\-\;\:"“%‘”�—’…–$&()*+.\/=@\[\]_`¡§«°´µ·»×àáâãåæçèéêëìíîïðñòóôõøùúûýþāăąćčďđēėęěğġħīıłńņňōŏőœřśşšťūůźżžơǐǔșțəʻʾʿ̥̆̇авеикморсфчшѹאבנעש་ནḫṟṣṭạảắằếễệọồộụứ‑‚„‟′″‹›→−≡⟨⟩カ东临乡关合城孙尣幺支比毛泽無生臣辶道镇黃]'  # 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"]
+        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/eval_results-checkpoint.json

@@ -0,0 +1,9 @@
+{
+    "epoch": 2.5,
+    "eval_loss": 0.13549980521202087,
+    "eval_runtime": 1012.9633,
+    "eval_samples": 15995,
+    "eval_samples_per_second": 15.79,
+    "eval_steps_per_second": 1.974,
+    "eval_wer": 0.15316115109878853
+}

.ipynb_checkpoints/run_training-checkpoint.sh

@@ -3,11 +3,10 @@ python run_speech_recognition_ctc.py \
 	--model_name_or_path="facebook/wav2vec2-xls-r-1b" \
 	--dataset_config_name="de" \
 	--output_dir="./" \
-    --overwrite_output_dir \
-	--num_train_epochs="2.5" \
-	--per_device_train_batch_size="8" \
+	--num_train_epochs="3.0" \
+	--per_device_train_batch_size="64" \
 	--per_device_eval_batch_size="8" \
-	--gradient_accumulation_steps="4" \
+	--gradient_accumulation_steps="1" \
 	--learning_rate="7.5e-5" \
 	--warmup_steps="2000" \
 	--length_column_name="input_length" \

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:adab4f8eb5a7f62adcf02be9f38b9ff41939347692ddf7d2d0fc29a4f749467b
+oid sha256:ba797e4f54e852c7386ff48804957732ea4acb2243c819d84a1e4f15828ea35b
 size 3850681649

run_training.sh

@@ -3,11 +3,10 @@ python run_speech_recognition_ctc.py \
 	--model_name_or_path="facebook/wav2vec2-xls-r-1b" \
 	--dataset_config_name="de" \
 	--output_dir="./" \
-    --overwrite_output_dir \
-	--num_train_epochs="2.5" \
-	--per_device_train_batch_size="8" \
+	--num_train_epochs="3.0" \
+	--per_device_train_batch_size="64" \
 	--per_device_eval_batch_size="8" \
-	--gradient_accumulation_steps="4" \
+	--gradient_accumulation_steps="1" \
 	--learning_rate="7.5e-5" \
 	--warmup_steps="2000" \
 	--length_column_name="input_length" \

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}]}
+{"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}]}

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3d671fb0f181e146452d1d68a46c3b54df59aa573465bc6cf0a59cb0e02b849a
+oid sha256:b3a67d83004654b3fcd534d3d282d2ef1f937289dc06f807157fc0a415fad82f
 size 2991