initial model files

Dockerfile

@@ -0,0 +1,27 @@
+# Adapted from https://github.com/baaastijn/Dockerimages/tree/main/Hugginface_challenge_speech to use custom
+# Transformers branch.
+
+# Base image. Here we take one from OVHcloud with Jupyter inside and pytorch
+FROM ovhcom/ai-training-pytorch:latest
+
+# Install git, audio loader ang git lfs
+RUN apt-get update && \
+    apt-get install -y git && \
+    apt-get install -y libsndfile1-dev sox && \
+    apt-get install -y git-lfs && \
+    git lfs install
+
+# Install required python libraries. We install transformers from source to get latest version
+RUN pip install --upgrade pip && \
+    pip install git+https://github.com/jsnfly/transformers.git@speech-challenge-experiments && \
+    pip install git+https://github.com/huggingface/datasets && \
+    pip install torchaudio librosa jiwer && \
+    pip install pandas numpy nano gradio
+
+# Create a HOME dedicated to the OVHcloud user (42420:42420)
+RUN chown -R 42420:42420 /workspace
+ENV HOME /workspace
+WORKDIR /workspace
+
+# Copy a folder of example notebooks into another folder in remote workspace
+# COPY notebooks /workspace/

README.md

@@ -0,0 +1,42 @@
+---
+language:
+- de
+license: apache-2.0
+tags:
+- automatic-speech-recognition
+- mozilla-foundation/common_voice_7_0
+- de
+datasets:
+- mozilla-foundation/common_voice_7_0
+model-index:
+- name: Wav2Vec2-Large-XLSR-53-German-GPT2
+  results:
+  - task:
+      name: Automatic Speech Recognition
+      type: automatic-speech-recognition
+    dataset:
+      name: Common Voice 7
+      type: mozilla-foundation/common_voice_7_0
+      args: de
+    metrics:
+       - name: Test WER
+         type: wer
+         value: 11.49
+       - name: Test CER
+         type: cer
+         value: 5.6
+---
+
+# Wav2Vec2-Large-XLSR-53-German-GPT2
+
+This is an encoder-decoder model for automatic speech recognition trained on on the
+MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - DE dataset. The encoder was initialized from
+[jonatasgrosman/wav2vec2-large-xlsr-53-german](https://huggingface.co/jonatasgrosman/wav2vec2-large-xlsr-53-german) and
+the decoder from [dbmdz/german-gpt2](https://huggingface.co/dbmdz/german-gpt2).
+
+It was trained using a two step process:
+* fine-tuning only the cross-attention weights and the decoder using the pre-computed outputs of the Wav2Vec-Modell
+* fine-tuning the model end-to-end
+
+There is also one trick, which seemed to improve performance significantly: adding position embeddings to the
+encoder outputs and initializing them with the pre-trained position embeddings of the GPT2 model (See `eval.py`).

added_tokens.json

@@ -0,0 +1 @@
+{"<|endoftext|>": 50265}

config.json

@@ -0,0 +1,266 @@
+{
+  "architectures": [
+    "SpeechEncoderDecoderModel"
+  ],
+  "decoder": {
+    "_name_or_path": "dbmdz/german-gpt2",
+    "activation_function": "gelu_new",
+    "add_cross_attention": true,
+    "architectures": [
+      "GPT2LMHeadModel"
+    ],
+    "attn_pdrop": 0.0,
+    "bad_words_ids": null,
+    "bos_token_id": 50256,
+    "chunk_size_feed_forward": 0,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "early_stopping": false,
+    "embd_pdrop": 0.0,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 50256,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "gradient_checkpointing": false,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_range": 0.02,
+    "is_decoder": true,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "layer_norm_epsilon": 1e-05,
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "min_length": 0,
+    "model_type": "gpt2",
+    "n_ctx": 1024,
+    "n_embd": 768,
+    "n_head": 12,
+    "n_inner": null,
+    "n_layer": 12,
+    "n_positions": 1024,
+    "no_repeat_ngram_size": 0,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": null,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "reorder_and_upcast_attn": false,
+    "repetition_penalty": 1.0,
+    "resid_pdrop": 0.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "scale_attn_by_inverse_layer_idx": false,
+    "scale_attn_weights": true,
+    "sep_token_id": null,
+    "summary_activation": null,
+    "summary_first_dropout": 0.1,
+    "summary_proj_to_labels": true,
+    "summary_type": "cls_index",
+    "summary_use_proj": true,
+    "task_specific_params": {
+      "text-generation": {
+        "do_sample": true,
+        "max_length": 50
+      }
+    },
+    "temperature": 1.0,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": "float32",
+    "torchscript": false,
+    "transformers_version": "4.17.0.dev0",
+    "use_bfloat16": false,
+    "use_cache": true,
+    "vocab_size": 50265
+  },
+  "decoder_start_token_id": 98,
+  "encoder": {
+    "_name_or_path": "jonatasgrosman/wav2vec2-large-xlsr-53-german",
+    "activation_dropout": 0.05,
+    "adapter_kernel_size": 3,
+    "adapter_stride": 2,
+    "add_adapter": false,
+    "add_cross_attention": false,
+    "apply_spec_augment": true,
+    "architectures": [
+      "Wav2Vec2ForCTC"
+    ],
+    "attention_dropout": 0.1,
+    "bad_words_ids": null,
+    "bos_token_id": 1,
+    "chunk_size_feed_forward": 0,
+    "classifier_proj_size": 256,
+    "codevector_dim": 768,
+    "contrastive_logits_temperature": 0.1,
+    "conv_bias": true,
+    "conv_dim": [
+      512,
+      512,
+      512,
+      512,
+      512,
+      512,
+      512
+    ],
+    "conv_kernel": [
+      10,
+      3,
+      3,
+      3,
+      3,
+      2,
+      2
+    ],
+    "conv_stride": [
+      5,
+      2,
+      2,
+      2,
+      2,
+      2,
+      2
+    ],
+    "cross_attention_hidden_size": null,
+    "ctc_loss_reduction": "mean",
+    "ctc_zero_infinity": true,
+    "decoder_start_token_id": null,
+    "diversity_loss_weight": 0.1,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "do_stable_layer_norm": true,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 2,
+    "feat_extract_activation": "gelu",
+    "feat_extract_dropout": 0.0,
+    "feat_extract_norm": "layer",
+    "feat_proj_dropout": 0.05,
+    "feat_quantizer_dropout": 0.0,
+    "final_dropout": 0.0,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "hidden_act": "gelu",
+    "hidden_dropout": 0.05,
+    "hidden_size": 1024,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_range": 0.02,
+    "intermediate_size": 4096,
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "layer_norm_eps": 1e-05,
+    "layerdrop": 0.05,
+    "length_penalty": 1.0,
+    "mask_channel_length": 10,
+    "mask_channel_min_space": 1,
+    "mask_channel_other": 0.0,
+    "mask_channel_prob": 0.0,
+    "mask_channel_selection": "static",
+    "mask_feature_length": 10,
+    "mask_feature_min_masks": 0,
+    "mask_feature_prob": 0.0,
+    "mask_time_length": 10,
+    "mask_time_min_masks": 2,
+    "mask_time_min_space": 1,
+    "mask_time_other": 0.0,
+    "mask_time_prob": 0.05,
+    "mask_time_selection": "static",
+    "max_length": 20,
+    "min_length": 0,
+    "model_type": "wav2vec2",
+    "no_repeat_ngram_size": 0,
+    "num_adapter_layers": 3,
+    "num_attention_heads": 16,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_codevector_groups": 2,
+    "num_codevectors_per_group": 320,
+    "num_conv_pos_embedding_groups": 16,
+    "num_conv_pos_embeddings": 128,
+    "num_feat_extract_layers": 7,
+    "num_hidden_layers": 24,
+    "num_negatives": 100,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_size": 1024,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": 0,
+    "prefix": null,
+    "problem_type": null,
+    "proj_codevector_dim": 768,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "sep_token_id": null,
+    "task_specific_params": null,
+    "tdnn_dilation": [
+      1,
+      2,
+      3,
+      1,
+      1
+    ],
+    "tdnn_dim": [
+      512,
+      512,
+      512,
+      512,
+      1500
+    ],
+    "tdnn_kernel": [
+      5,
+      3,
+      3,
+      1,
+      1
+    ],
+    "temperature": 1.0,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": null,
+    "torchscript": false,
+    "transformers_version": "4.17.0.dev0",
+    "use_bfloat16": false,
+    "use_weighted_layer_sum": false,
+    "vocab_size": 38,
+    "xvector_output_dim": 512
+  },
+  "is_encoder_decoder": true,
+  "max_length": 35,
+  "model_type": "speech-encoder-decoder",
+  "pad_token_id": 67,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float32",
+  "transformers_version": null
+}

eval.py

@@ -0,0 +1,277 @@
+import argparse
+import re
+from typing import Dict
+
+import torch
+from datasets import Audio, Dataset, load_dataset, load_metric
+
+from transformers import AutoFeatureExtractor, AutoTokenizer, SpeechEncoderDecoderModel, pipeline
+
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from transformers.models.encoder_decoder.modeling_encoder_decoder import shift_tokens_right
+from transformers.modeling_outputs import Seq2SeqLMOutput
+
+
+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."""
+
+    # From https://huggingface.co/jonatasgrosman/wav2vec2-large-xlsr-53-german.
+    CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", "；", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
+                       "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
+                       "{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。",
+                       "、", "﹂", "﹁", "‧", "～", "﹏", "，", "｛", "｝", "（", "）", "［", "］", "【", "】", "‥", "〽",
+                       "『", "』", "〝", "〟", "⟨", "⟩", "〜", "：", "！", "？", "♪", "؛", "/", "\\", "º", "−", "^", "ʻ", "ˆ"]
+    chars_to_ignore_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"
+    text = re.sub(chars_to_ignore_regex, "", text.lower())
+
+    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 tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(args.model_id)
+
+    # load model
+    model = Wav2VecGPT2Model.from_pretrained(args.model_id)
+
+    # load eval pipeline
+    if args.device is None:
+        args.device = 0 if torch.cuda.is_available() else -1
+    asr = pipeline("automatic-speech-recognition", model=model, device=args.device,
+                   feature_extractor=feature_extractor, tokenizer=tokenizer)
+
+    # 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"] = normalize_text(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)
+
+
+class Wav2VecGPT2Model(SpeechEncoderDecoderModel):
+    """
+    Basically the same as `SpeechEncoderDecoderModel` but position embeddings (initialized with GPT2's position
+    embeddings) are added to encoder output
+    """
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.encoder_outputs_pos_emb = nn.Embedding(1024, self.decoder.config.hidden_size)
+        with torch.no_grad():
+            self.encoder_outputs_pos_emb.weight.copy_(self.decoder.transformer.wpe.weight)
+        self.enc_to_dec_proj_ln = nn.LayerNorm(self.decoder.config.hidden_size,
+                                               eps=self.decoder.config.layer_norm_epsilon)
+
+    def __getattribute__(self, name):
+        # Fake class so it is recognized as seq2seq model.
+        if name == '__class__':
+            return SpeechEncoderDecoderModel
+        return SpeechEncoderDecoderModel.__getattribute__(self, name)
+
+    def forward(
+        self,
+        inputs=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        decoder_inputs_embeds=None,
+        labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        input_values=None,
+        input_features=None,
+        return_dict=None,
+        **kwargs,
+    ):
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        kwargs_encoder = {argument: value for argument, value in kwargs.items() if not argument.startswith("decoder_")}
+
+        kwargs_decoder = {
+            argument[len("decoder_") :]: value for argument, value in kwargs.items() if argument.startswith("decoder_")
+        }
+
+        if encoder_outputs is None and inputs is None:
+            if input_values is not None and input_features is not None:
+                raise ValueError("You cannot specify both input_values and input_features at the same time")
+            elif input_values is not None:
+                inputs = input_values
+            elif input_features is not None:
+                inputs = input_features
+            else:
+                raise ValueError("You have to specify either input_values or input_features")
+
+            encoder_outputs = self.encoder(
+                inputs,
+                attention_mask=attention_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                **kwargs_encoder,
+            )
+
+        encoder_hidden_states = encoder_outputs[0]
+
+        # optionally project encoder_hidden_states
+        if (
+            self.encoder_output_dim != self.decoder.config.hidden_size
+            and self.decoder.config.cross_attention_hidden_size is None
+        ):
+            encoder_hidden_states = self.enc_to_dec_proj(encoder_hidden_states)
+            encoder_hidden_states += self.encoder_outputs_pos_emb(
+                torch.arange(0, encoder_hidden_states.shape[1], device=encoder_hidden_states.device)
+            )
+            encoder_hidden_states = self.enc_to_dec_proj_ln(encoder_hidden_states)
+
+        # compute correct encoder attention mask
+        if attention_mask is not None:
+            encoder_attention_mask = self.encoder._get_feature_vector_attention_mask(
+                encoder_hidden_states.shape[1], attention_mask
+            )
+        else:
+            encoder_attention_mask = None
+
+        if (labels is not None) and (decoder_input_ids is None and decoder_inputs_embeds is None):
+            decoder_input_ids = shift_tokens_right(
+                labels, self.config.pad_token_id, self.config.decoder_start_token_id
+            )
+
+        # Decode
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            inputs_embeds=decoder_inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            use_cache=use_cache,
+            past_key_values=past_key_values,
+            return_dict=return_dict,
+            **kwargs_decoder,
+        )
+
+        # Compute loss independent from decoder (as some shift the logits inside them)
+        loss = None
+        if labels is not None:
+            logits = decoder_outputs.logits if return_dict else decoder_outputs[0]
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.reshape(-1, self.decoder.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            if loss is not None:
+                return (loss,) + decoder_outputs + encoder_outputs
+            else:
+                return decoder_outputs + encoder_outputs
+
+        return Seq2SeqLMOutput(
+            loss=loss,
+            logits=decoder_outputs.logits,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+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)

preprocessor_config.json

@@ -0,0 +1,9 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "return_attention_mask": true,
+  "sampling_rate": 16000
+}

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "~", "eos_token": "<|endoftext|>", "unk_token": "<|endoftext|>", "pad_token": "_"}

tokenizer_config.json

@@ -0,0 +1 @@
+{"unk_token": "<|endoftext|>", "bos_token": "<|endoftext|>", "eos_token": "<|endoftext|>", "add_prefix_space": false, "special_tokens_map_file": null, "name_or_path": "dbmdz/german-gpt2", "tokenizer_class": "GPT2Tokenizer"}