[Examples] Add an official audio classification example (#13722)

* Restore broken merge

* Additional args, DDP, remove CommonLanguage

* Update examples for V100, add training results

* Style

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Remove custom datasets for simplicity, apply suggestions from code review

* Add the attention_mask flag, reorganize README
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

examples/pytorch/audio-classification/README.md

+<!---
+Copyright 2021 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+-->
+
+# Audio classification examples
+
+The following examples showcase how to fine-tune `Wav2Vec2` for audio classification using PyTorch.
+
+Speech recognition models that have been pretrained in unsupervised fashion on audio data alone, 
+*e.g.* [Wav2Vec2](https://huggingface.co/transformers/master/model_doc/wav2vec2.html), 
+[HuBERT](https://huggingface.co/transformers/master/model_doc/hubert.html), 
+[XLSR-Wav2Vec2](https://huggingface.co/transformers/master/model_doc/xlsr_wav2vec2.html), have shown to require only 
+very little annotated data to yield good performance on speech classification datasets.
+
+## Single-GPU 
+
+The following command shows how to fine-tune [wav2vec2-base](https://huggingface.co/facebook/wav2vec2-base) on the 🗣️ [Keyword Spotting subset](https://huggingface.co/datasets/superb#ks) of the SUPERB dataset.
+
+```bash
+python run_audio_classification.py \
+    --model_name_or_path facebook/wav2vec2-base \
+    --dataset_name superb \
+    --dataset_config_name ks \
+    --output_dir wav2vec2-base-keyword-spotting \
+    --overwrite_output_dir \
+    --remove_unused_columns False \
+    --do_train \
+    --do_eval \
+    --fp16 \
+    --learning_rate 3e-5 \
+    --max_length_seconds 1 \
+    --warmup_ratio 0.1 \
+    --num_train_epochs 5 \
+    --per_device_train_batch_size 32 \
+    --gradient_accumulation_steps 4 \
+    --per_device_eval_batch_size 32 \
+    --dataloader_num_workers 4 \
+    --logging_strategy steps \
+    --logging_steps 10 \
+    --evaluation_strategy epoch \
+    --save_strategy epoch \
+    --load_best_model_at_end True \
+    --save_total_limit 3 \
+    --seed 0 \
+    --push_to_hub
+```
+
+On a single V100 GPU (16GB), this script should run in ~10 minutes and yield accuracy of **98.4%**.
+
+👀 See the results here: [anton-l/wav2vec2-base-keyword-spotting](https://huggingface.co/anton-l/wav2vec2-base-keyword-spotting)
+
+## Multi-GPU 
+
+The following command shows how to fine-tune [wav2vec2-base](https://huggingface.co/facebook/wav2vec2-base) for 🌎 **Language Identification** on the [CommonLanguage dataset](https://huggingface.co/datasets/anton-l/common_language).
+
+```bash
+python run_audio_classification.py \
+    --model_name_or_path facebook/wav2vec2-base \
+    --dataset_name anton-l/common_language \
+    --audio_column_name path \
+    --label_column_name language \
+    --output_dir wav2vec2-base-lang-id \
+    --overwrite_output_dir \
+    --remove_unused_columns False \
+    --do_train \
+    --do_eval \
+    --fp16 \
+    --learning_rate 3e-5 \
+    --max_length_seconds 16 \
+    --attention_mask False \
+    --warmup_ratio 0.1 \
+    --num_train_epochs 10 \
+    --per_device_train_batch_size 8 \
+    --gradient_accumulation_steps 4 \
+    --per_device_eval_batch_size 1 \
+    --dataloader_num_workers 8 \
+    --logging_strategy steps \
+    --logging_steps 10 \
+    --evaluation_strategy epoch \
+    --save_strategy epoch \
+    --load_best_model_at_end True \
+    --save_total_limit 3 \
+    --seed 0 \
+    --push_to_hub
+```
+
+On 4 V100 GPUs (16GB), this script should run in ~1 hour and yield accuracy of **79.45%**.
+
+👀 See the results here: [anton-l/wav2vec2-base-lang-id](https://huggingface.co/anton-l/wav2vec2-base-lang-id)
+
+## Sharing your model on 🤗 Hub
+
+0. If you haven't already, [sign up](https://huggingface.co/join) for a 🤗 account
+
+1. Make sure you have `git-lfs` installed and git set up.
+
+```bash
+$ apt install git-lfs
+```
+
+2. Log in with your HuggingFace account credentials using `huggingface-cli`
+
+```bash
+$ huggingface-cli login
+# ...follow the prompts
+```
+
+3. When running the script, pass the following arguments:
+
+```bash
+python run_audio_classification.py \
+    --push_to_hub \
+    --hub_model_id <username/model_id> \
+    ...
+```
\ No newline at end of file

examples/pytorch/audio-classification/requirements.txt

+datasets>=1.12.0
+torchaudio
+torch>=1.6
\ No newline at end of file

examples/pytorch/audio-classification/run_audio_classification.py

+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+from random import randint
+from typing import Optional
+
+import datasets
+import numpy as np
+import torchaudio
+from datasets import DatasetDict, load_dataset
+
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoModelForAudioClassification,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    set_seed,
+)
+from transformers.trainer_utils import get_last_checkpoint
+from transformers.utils import check_min_version
+from transformers.utils.versions import require_version
+
+
+logger = logging.getLogger(__name__)
+
+# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
+check_min_version("4.11.0.dev0")
+
+require_version("datasets>=1.12.1", "To fix: pip install -r examples/pytorch/audio-classification/requirements.txt")
+
+
+def load_audio(path: str, sample_rate: int = 16000):
+    wav, sr = torchaudio.load(path)
+    # convert multi-channel audio to mono
+    wav = wav.mean(0)
+    # standardize sample rate if it varies in the dataset
+    resampler = torchaudio.transforms.Resample(sr, sample_rate)
+    wav = resampler(wav)
+    return wav
+
+
+def random_subsample(wav: np.ndarray, max_length: float, sample_rate: int = 16000):
+    """Randomly sample chunks of `max_length` seconds from the input audio"""
+    sample_length = int(round(sample_rate * max_length))
+    if len(wav) <= sample_length:
+        return wav
+    random_offset = randint(0, len(wav) - sample_length - 1)
+    return wav[random_offset : random_offset + sample_length]
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    Using `HfArgumentParser` we can turn this class
+    into argparse arguments to be able to specify them on
+    the command line.
+    """
+
+    dataset_name: Optional[str] = field(default=None, metadata={"help": "Name of a dataset from the datasets package"})
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    train_file: Optional[str] = field(
+        default=None, metadata={"help": "A file containing the training audio paths and labels."}
+    )
+    eval_file: Optional[str] = field(
+        default=None, metadata={"help": "A file containing the validation audio paths and labels."}
+    )
+    train_split_name: Optional[str] = field(
+        default="train",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    eval_split_name: Optional[str] = field(
+        default="validation",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to "
+            "'validation'"
+        },
+    )
+    audio_column_name: Optional[str] = field(
+        default="file",
+        metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'file'"},
+    )
+    label_column_name: Optional[str] = field(
+        default="label", metadata={"help": "The name of the dataset column containing the labels. Defaults to 'label'"}
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
+            "value if set."
+        },
+    )
+    max_length_seconds: Optional[float] = field(
+        default=20,
+        metadata={"help": "Audio clips will be randomly cut to this length during training if the value is set."},
+    )
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        default="facebook/wav2vec2-base",
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"},
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from the Hub"}
+    )
+    model_revision: str = field(
+        default="main",
+        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
+    )
+    feature_extractor_name: str = field(default=None, metadata={"help": "Name or path of preprocessor config."})
+    freeze_feature_extractor: Optional[bool] = field(
+        default=True, metadata={"help": "Whether to freeze the feature extractor layers of the model."}
+    )
+    attention_mask: Optional[bool] = field(
+        default=True, metadata={"help": "Whether to generate an attention mask in the feature extractor."}
+    )
+    use_auth_token: bool = field(
+        default=False,
+        metadata={
+            "help": "Will use the token generated when running `transformers-cli login` (necessary to use this script "
+            "with private models)."
+        },
+    )
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+
+    log_level = training_args.get_process_log_level()
+    logger.setLevel(log_level)
+    transformers.utils.logging.set_verbosity(log_level)
+    transformers.utils.logging.enable_default_handler()
+    transformers.utils.logging.enable_explicit_format()
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu} "
+        + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # Detecting last checkpoint.
+    last_checkpoint = None
+    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
+        last_checkpoint = get_last_checkpoint(training_args.output_dir)
+        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
+            raise ValueError(
+                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
+                "Use --overwrite_output_dir to train from scratch."
+            )
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
+            logger.info(
+                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
+                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
+            )
+
+    # Initialize our dataset and prepare it for the audio classification task.
+    raw_datasets = DatasetDict()
+    raw_datasets["train"] = load_dataset(
+        data_args.dataset_name, data_args.dataset_config_name, split=data_args.train_split_name
+    )
+    raw_datasets["eval"] = load_dataset(
+        data_args.dataset_name, data_args.dataset_config_name, split=data_args.eval_split_name
+    )
+
+    if data_args.audio_column_name not in raw_datasets["train"].column_names:
+        raise ValueError(
+            f"--audio_column_name {data_args.audio_column_name} not found in dataset '{data_args.dataset_name}'. "
+            "Make sure to set `--audio_column_name` to the correct audio column - one of "
+            f"{', '.join(raw_datasets['train'].column_names)}."
+        )
+
+    if data_args.label_column_name not in raw_datasets["train"].column_names:
+        raise ValueError(
+            f"--label_column_name {data_args.label_column_name} not found in dataset '{data_args.dataset_name}'. "
+            "Make sure to set `--label_column_name` to the correct text column - one of "
+            f"{', '.join(raw_datasets['train'].column_names)}."
+        )
+
+    # Setting `return_attention_mask=True` is the way to get a correctly masked mean-pooling over
+    # transformer outputs in the classifier, but it doesn't always lead to better accuracy
+    feature_extractor = AutoFeatureExtractor.from_pretrained(
+        model_args.feature_extractor_name or model_args.model_name_or_path,
+        return_attention_mask=model_args.attention_mask,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+
+    def train_transforms(batch):
+        """Apply train_transforms across a batch."""
+        output_batch = {"input_values": []}
+        for f in batch[data_args.audio_column_name]:
+            wav = load_audio(f, sample_rate=feature_extractor.sampling_rate)
+            wav = random_subsample(
+                wav, max_length=data_args.max_length_seconds, sample_rate=feature_extractor.sampling_rate
+            )
+            output_batch["input_values"].append(wav)
+        output_batch["labels"] = [label for label in batch[data_args.label_column_name]]
+
+        return output_batch
+
+    def val_transforms(batch):
+        """Apply val_transforms across a batch."""
+        output_batch = {"input_values": []}
+        for f in batch[data_args.audio_column_name]:
+            wav = load_audio(f, sample_rate=feature_extractor.sampling_rate)
+            output_batch["input_values"].append(wav)
+        output_batch["labels"] = [label for label in batch[data_args.label_column_name]]
+
+        return output_batch
+
+    # Prepare label mappings.
+    # We'll include these in the model's config to get human readable labels in the Inference API.
+    labels = raw_datasets["train"].features[data_args.label_column_name].names
+    label2id, id2label = dict(), dict()
+    for i, label in enumerate(labels):
+        label2id[label] = str(i)
+        id2label[str(i)] = label
+
+    # Load the accuracy metric from the datasets package
+    metric = datasets.load_metric("accuracy")
+
+    # Define our compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with
+    # `predictions` and `label_ids` fields) and has to return a dictionary string to float.
+    def compute_metrics(eval_pred):
+        """Computes accuracy on a batch of predictions"""
+        predictions = np.argmax(eval_pred.predictions, axis=1)
+        return metric.compute(predictions=predictions, references=eval_pred.label_ids)
+
+    config = AutoConfig.from_pretrained(
+        model_args.config_name or model_args.model_name_or_path,
+        num_labels=len(labels),
+        label2id=label2id,
+        id2label=id2label,
+        finetuning_task="audio-classification",
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+    model = AutoModelForAudioClassification.from_pretrained(
+        model_args.model_name_or_path,
+        from_tf=bool(".ckpt" in model_args.model_name_or_path),
+        config=config,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+
+    # freeze the convolutional waveform encoder
+    if model_args.freeze_feature_extractor:
+        model.freeze_feature_extractor()
+
+    if training_args.do_train:
+        if "train" not in raw_datasets:
+            raise ValueError("--do_train requires a train dataset")
+        if data_args.max_train_samples is not None:
+            raw_datasets["train"] = (
+                raw_datasets["train"].shuffle(seed=training_args.seed).select(range(data_args.max_train_samples))
+            )
+        # Set the training transforms
+        raw_datasets["train"].set_transform(train_transforms, output_all_columns=False)
+
+    if training_args.do_eval:
+        if "eval" not in raw_datasets:
+            raise ValueError("--do_eval requires a validation dataset")
+        if data_args.max_eval_samples is not None:
+            raw_datasets["eval"] = (
+                raw_datasets["eval"].shuffle(seed=training_args.seed).select(range(data_args.max_eval_samples))
+            )
+        # Set the validation transforms
+        raw_datasets["eval"].set_transform(val_transforms, output_all_columns=False)
+
+    # Initialize our trainer
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        train_dataset=raw_datasets["train"] if training_args.do_train else None,
+        eval_dataset=raw_datasets["eval"] if training_args.do_eval else None,
+        compute_metrics=compute_metrics,
+        tokenizer=feature_extractor,
+    )
+
+    # Training
+    if training_args.do_train:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
+            checkpoint = last_checkpoint
+        train_result = trainer.train(resume_from_checkpoint=checkpoint)
+        trainer.save_model()
+        trainer.log_metrics("train", train_result.metrics)
+        trainer.save_metrics("train", train_result.metrics)
+        trainer.save_state()
+
+    # Evaluation
+    if training_args.do_eval:
+        metrics = trainer.evaluate()
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # Write model card and (optionally) push to hub
+    kwargs = {
+        "finetuned_from": model_args.model_name_or_path,
+        "tasks": "audio-classification",
+        "dataset": data_args.dataset_name,
+        "tags": ["audio-classification"],
+    }
+    if training_args.push_to_hub:
+        trainer.push_to_hub(**kwargs)
+    else:
+        trainer.create_model_card(**kwargs)
+
+
+if __name__ == "__main__":
+    main()

examples/pytorch/test_examples.py

@@ -40,12 +40,14 @@ SRC_DIRS = [
         "translation",
         "image-classification",
         "speech-recognition",
+        "audio-classification",
     ]
 ]
 sys.path.extend(SRC_DIRS)
 
 
 if SRC_DIRS is not None:
+    import run_audio_classification
     import run_clm
     import run_generation
     import run_glue
@@ -410,3 +412,38 @@ class ExamplesTests(TestCasePlus):
             run_speech_recognition_ctc.main()
             result = get_results(tmp_dir)
             self.assertLess(result["eval_loss"], result["train_loss"])
+
+    def test_run_audio_classification(self):
+        stream_handler = logging.StreamHandler(sys.stdout)
+        logger.addHandler(stream_handler)
+
+        tmp_dir = self.get_auto_remove_tmp_dir()
+        testargs = f"""
+            run_audio_classification.py
+            --output_dir {tmp_dir}
+            --model_name_or_path hf-internal-testing/tiny-random-wav2vec2
+            --dataset_name anton-l/superb_demo
+            --dataset_config_name ks
+            --train_split_name test
+            --eval_split_name test
+            --audio_column_name file
+            --label_column_name label
+            --do_train
+            --do_eval
+            --learning_rate 1e-4
+            --per_device_train_batch_size 2
+            --per_device_eval_batch_size 1
+            --remove_unused_columns False
+            --overwrite_output_dir True
+            --num_train_epochs 10
+            --max_steps 50
+            --seed 42
+        """.split()
+
+        if is_cuda_and_apex_available():
+            testargs.append("--fp16")
+
+        with patch.object(sys, "argv", testargs):
+            run_audio_classification.main()
+            result = get_results(tmp_dir)
+            self.assertLess(result["eval_loss"], result["train_loss"])