Delete 1234

1234/app.py

@@ -1,464 +0,0 @@
-#!/usr/bin/env python3
-import sys
-import torch
-import logging
-import gradio as gr
-import speechbrain as sb
-from pathlib import Path
-import os
-import torchaudio
-from hyperpyyaml import load_hyperpyyaml
-from speechbrain.tokenizers.SentencePiece import SentencePiece
-from speechbrain.utils.data_utils import undo_padding
-from speechbrain.utils.distributed import run_on_main
-
-"""Recipe for training a sequence-to-sequence ASR system with CommonVoice.
-The system employs a wav2vec2 encoder and a CTC decoder.
-Decoding is performed with greedy decoding (will be extended to beam search).
-
-To run this recipe, do the following:
-> python train_with_wav2vec2.py hparams/train_with_wav2vec2.yaml
-
-With the default hyperparameters, the system employs a pretrained wav2vec2 encoder.
-The wav2vec2 model is pretrained following the model given in the hprams file.
-It may be dependent on the language.
-
-The neural network is trained with CTC on sub-word units estimated with
-Byte Pairwise Encoding (BPE).
-
-The experiment file is flexible enough to support a large variety of
-different systems. By properly changing the parameter files, you can try
-different encoders, decoders, tokens (e.g, characters instead of BPE),
-training languages (all CommonVoice languages), and many
-other possible variations.
-
-Authors
- * Titouan Parcollet 2021
-"""
-
-logger = logging.getLogger(__name__)
-
-
-# Define training procedure
-class ASR(sb.core.Brain):
-    def compute_forward(self, batch, stage):
-        """Forward computations from the waveform batches to the output probabilities."""
-
-        batch = batch.to(self.device)
-        wavs, wav_lens = batch.sig
-        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
-        if stage == sb.Stage.TRAIN:
-            if hasattr(self.hparams, "augmentation"):
-                wavs = self.hparams.augmentation(wavs, wav_lens)
-
-        # Forward pass
-        feats = self.modules.wav2vec2(wavs, wav_lens)
-        x = self.modules.enc(feats)
-        logits = self.modules.ctc_lin(x)
-        p_ctc = self.hparams.log_softmax(logits)
-
-        return p_ctc, wav_lens
-    
-    def treat_wav(self,sig):
-        feats = self.modules.wav2vec2(sig.to("cpu"), torch.tensor([1]).to("cpu"))
-        feats = self.modules.enc(feats)
-        logits = self.modules.ctc_lin(feats)
-        p_ctc = self.hparams.log_softmax(logits)
-        predicted_words =[]
-        for logs in p_ctc:
-            text = decoder.decode(logs.detach().cpu().numpy())
-            predicted_words.append(text.split(" "))
-        return " ".join(predicted_words[0])
- 
-    def compute_objectives(self, predictions, batch, stage):
-        """Computes the loss (CTC) given predictions and targets."""
-
-        p_ctc, wav_lens = predictions
-
-        ids = batch.id
-        tokens, tokens_lens = batch.tokens
-
-        loss = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
-
-        if stage != sb.Stage.TRAIN:
-            predicted_tokens = sb.decoders.ctc_greedy_decode(
-                p_ctc, wav_lens, blank_id=self.hparams.blank_index
-            )
-            # Decode token terms to words
-            if self.hparams.use_language_modelling:
-                predicted_words = []
-                for logs in p_ctc:
-                    text = decoder.decode(logs.detach().cpu().numpy())
-                    predicted_words.append(text.split(" "))
-            else:
-                predicted_words = [
-                    "".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
-                    for utt_seq in predicted_tokens
-                ]
-            # Convert indices to words
-            target_words = [wrd.split(" ") for wrd in batch.wrd]
-
-            self.wer_metric.append(ids, predicted_words, target_words)
-            self.cer_metric.append(ids, predicted_words, target_words)
-
-        return loss
-
-    def fit_batch(self, batch):
-        """Train the parameters given a single batch in input"""
-        should_step = self.step % self.grad_accumulation_factor == 0
-        # Managing automatic mixed precision
-        # TOFIX: CTC fine-tuning currently is unstable
-        # This is certainly due to CTC being done in fp16 instead of fp32
-        if self.auto_mix_prec:
-            with torch.cuda.amp.autocast():
-                with self.no_sync():
-                    outputs = self.compute_forward(batch, sb.Stage.TRAIN)
-                loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
-            with self.no_sync(not should_step):
-                self.scaler.scale(
-                    loss / self.grad_accumulation_factor
-                ).backward()
-            if should_step:
-
-                if not self.hparams.wav2vec2.freeze:
-                    self.scaler.unscale_(self.wav2vec_optimizer)
-                self.scaler.unscale_(self.model_optimizer)
-                if self.check_gradients(loss):
-                    if not self.hparams.wav2vec2.freeze:
-                        if self.optimizer_step >= self.hparams.warmup_steps:
-                            self.scaler.step(self.wav2vec_optimizer)
-                    self.scaler.step(self.model_optimizer)
-                self.scaler.update()
-                self.zero_grad()
-                self.optimizer_step += 1
-        else:
-            # This is mandatory because HF models have a weird behavior with DDP
-            # on the forward pass
-            with self.no_sync():
-                outputs = self.compute_forward(batch, sb.Stage.TRAIN)
-
-            loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
-
-            with self.no_sync(not should_step):
-                (loss / self.grad_accumulation_factor).backward()
-            if should_step:
-                if self.check_gradients(loss):
-                    if not self.hparams.wav2vec2.freeze:
-                        if self.optimizer_step >= self.hparams.warmup_steps:
-                            self.wav2vec_optimizer.step()
-                    self.model_optimizer.step()
-                self.zero_grad()
-                self.optimizer_step += 1
-
-        self.on_fit_batch_end(batch, outputs, loss, should_step)
-        return loss.detach().cpu()
-
-    def evaluate_batch(self, batch, stage):
-        """Computations needed for validation/test batches"""
-        predictions = self.compute_forward(batch, stage=stage)
-        with torch.no_grad():
-            loss = self.compute_objectives(predictions, batch, stage=stage)
-        return loss.detach()
-
-    def on_stage_start(self, stage, epoch):
-        """Gets called at the beginning of each epoch"""
-        if stage != sb.Stage.TRAIN:
-            self.cer_metric = self.hparams.cer_computer()
-            self.wer_metric = self.hparams.error_rate_computer()
-
-    def on_stage_end(self, stage, stage_loss, epoch):
-        """Gets called at the end of an epoch."""
-        # Compute/store important stats
-        stage_stats = {"loss": stage_loss}
-        if stage == sb.Stage.TRAIN:
-            self.train_stats = stage_stats
-        else:
-            stage_stats["CER"] = self.cer_metric.summarize("error_rate")
-            stage_stats["WER"] = self.wer_metric.summarize("error_rate")
-
-        # Perform end-of-iteration things, like annealing, logging, etc.
-        if stage == sb.Stage.VALID:
-            old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
-                stage_stats["loss"]
-            )
-            old_lr_wav2vec, new_lr_wav2vec = self.hparams.lr_annealing_wav2vec(
-                stage_stats["loss"]
-            )
-            sb.nnet.schedulers.update_learning_rate(
-                self.model_optimizer, new_lr_model
-            )
-            if not self.hparams.wav2vec2.freeze:
-                sb.nnet.schedulers.update_learning_rate(
-                    self.wav2vec_optimizer, new_lr_wav2vec
-                )
-            self.hparams.train_logger.log_stats(
-                stats_meta={
-                    "epoch": epoch,
-                    "lr_model": old_lr_model,
-                    "lr_wav2vec": old_lr_wav2vec,
-                },
-                train_stats=self.train_stats,
-                valid_stats=stage_stats,
-            )
-            self.checkpointer.save_and_keep_only(
-                meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
-            )
-        elif stage == sb.Stage.TEST:
-            self.hparams.train_logger.log_stats(
-                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
-                test_stats=stage_stats,
-            )
-            with open(self.hparams.wer_file, "w") as w:
-                self.wer_metric.write_stats(w)
-
-    def init_optimizers(self):
-        "Initializes the wav2vec2 optimizer and model optimizer"
-
-        # If the wav2vec encoder is unfrozen, we create the optimizer
-        if not self.hparams.wav2vec2.freeze:
-            self.wav2vec_optimizer = self.hparams.wav2vec_opt_class(
-                self.modules.wav2vec2.parameters()
-            )
-            if self.checkpointer is not None:
-                self.checkpointer.add_recoverable(
-                    "wav2vec_opt", self.wav2vec_optimizer
-                )
-
-        self.model_optimizer = self.hparams.model_opt_class(
-            self.hparams.model.parameters()
-        )
-
-        if self.checkpointer is not None:
-            self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
-
-    def zero_grad(self, set_to_none=False):
-        if not self.hparams.wav2vec2.freeze:
-            self.wav2vec_optimizer.zero_grad(set_to_none)
-        self.model_optimizer.zero_grad(set_to_none)
-
-
-# Define custom data procedure
-def dataio_prepare(hparams):
-    """This function prepares the datasets to be used in the brain class.
-    It also defines the data processing pipeline through user-defined functions."""
-
-    # 1. Define datasets
-    data_folder = hparams["data_folder"]
-
-    train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
-        csv_path=hparams["train_csv"], replacements={"data_root": data_folder},
-    )
-
-    if hparams["sorting"] == "ascending":
-        # we sort training data to speed up training and get better results.
-        train_data = train_data.filtered_sorted(
-            sort_key="duration",
-            key_max_value={"duration": hparams["avoid_if_longer_than"]},
-        )
-        # when sorting do not shuffle in dataloader ! otherwise is pointless
-        hparams["dataloader_options"]["shuffle"] = False
-
-    elif hparams["sorting"] == "descending":
-        train_data = train_data.filtered_sorted(
-            sort_key="duration",
-            reverse=True,
-            key_max_value={"duration": hparams["avoid_if_longer_than"]},
-        )
-        # when sorting do not shuffle in dataloader ! otherwise is pointless
-        hparams["dataloader_options"]["shuffle"] = False
-
-    elif hparams["sorting"] == "random":
-        pass
-
-    else:
-        raise NotImplementedError(
-            "sorting must be random, ascending or descending"
-        )
-
-    valid_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
-        csv_path=hparams["valid_csv"], replacements={"data_root": data_folder},
-    )
-    # We also sort the validation data so it is faster to validate
-    valid_data = valid_data.filtered_sorted(sort_key="duration")
-    test_datasets = {}
-    for csv_file in hparams["test_csv"]:
-        name = Path(csv_file).stem
-        test_datasets[name] = sb.dataio.dataset.DynamicItemDataset.from_csv(
-            csv_path=csv_file, replacements={"data_root": data_folder}
-        )
-        test_datasets[name] = test_datasets[name].filtered_sorted(
-            sort_key="duration"
-        )
-
-    datasets = [train_data, valid_data] + [i for k, i in test_datasets.items()]
-
-
-    # 2. Define audio pipeline:
-    @sb.utils.data_pipeline.takes("wav")
-    @sb.utils.data_pipeline.provides("sig")
-    def audio_pipeline(wav):
-        info = torchaudio.info(wav)
-        sig = sb.dataio.dataio.read_audio(wav)
-        resampled = torchaudio.transforms.Resample(
-            info.sample_rate, hparams["sample_rate"],
-        )(sig)
-        return resampled
-
-    sb.dataio.dataset.add_dynamic_item(datasets, audio_pipeline)
-    label_encoder = sb.dataio.encoder.CTCTextEncoder()
-
-    # 3. Define text pipeline:
-    @sb.utils.data_pipeline.takes("wrd")
-    @sb.utils.data_pipeline.provides(
-        "wrd", "char_list", "tokens_list", "tokens"
-    )
-    def text_pipeline(wrd):
-        yield wrd
-        char_list = list(wrd)
-        yield char_list
-        tokens_list = label_encoder.encode_sequence(char_list)
-        yield tokens_list
-        tokens = torch.LongTensor(tokens_list)
-        yield tokens
-
-    sb.dataio.dataset.add_dynamic_item(datasets, text_pipeline)
-    lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
-    special_labels = {
-        "blank_label": hparams["blank_index"],
-        "unk_label": hparams["unk_index"]
-    }
-    label_encoder.load_or_create(
-        path=lab_enc_file,
-        from_didatasets=[train_data],
-        output_key="char_list",
-        special_labels=special_labels,
-        sequence_input=True,
-    )
-
-    # 4. Set output:
-    sb.dataio.dataset.set_output_keys(
-        datasets, ["id", "sig", "wrd", "char_list", "tokens"],
-    )
-    return train_data, valid_data,test_datasets, label_encoder
-
-
-
-# Load hyperparameters file with command-line overrides
-hparams_file, run_opts, overrides = sb.parse_arguments(["train_semi.yaml"])
-with open(hparams_file) as fin:
-    hparams = load_hyperpyyaml(fin, overrides)
-
-# If --distributed_launch then
-# create ddp_group with the right communication protocol
-sb.utils.distributed.ddp_init_group(run_opts)
-
-
-# Create experiment directory
-sb.create_experiment_directory(
-    experiment_directory=hparams["output_folder"],
-    hyperparams_to_save=hparams_file,
-    overrides=overrides,
-)
-
-# Due to DDP, we do the preparation ONLY on the main python process
-# Defining tokenizer and loading it
-# Create the datasets objects as well as tokenization and encoding :-D
-label_encoder = sb.dataio.encoder.CTCTextEncoder()
-
-lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
-special_labels = {
-    "blank_label": hparams["blank_index"],
-    "unk_label": hparams["unk_index"]
-}
-label_encoder.load_or_create(
-    path=lab_enc_file,
-    from_didatasets=[[]],
-    output_key="char_list",
-    special_labels=special_labels,
-    sequence_input=True,
-)
-
-from pyctcdecode import build_ctcdecoder
-ind2lab = label_encoder.ind2lab
-print(ind2lab)
-labels = [ind2lab[x] for x in range(len(ind2lab))]
-labels = [""] + labels[1:-1] + ["1"] 
-# Replace the <blank> token with a blank character, needed for PyCTCdecode
-print(labels)
-decoder = build_ctcdecoder(
-    labels,
-    kenlm_model_path=hparams["ngram_lm_path"],  # .arpa or .bin
-    alpha=0.5,  # Default by KenLM
-    beta=1.0,  # Default by KenLM
-)
-# Trainer initialization
-run_opts["device"] = "cpu"
-asr_brain = ASR(
-    modules=hparams["modules"],
-    hparams=hparams,
-    run_opts=run_opts,
-    checkpointer=hparams["checkpointer"],
-)
-
-# Adding objects to trainer.
-asr_brain.tokenizer = label_encoder
-asr_brain.checkpointer.recover_if_possible(device="cpu")
-asr_brain.modules.eval()
-description = """This is a speechbrain-based Automatic Speech Recognition (ASR) model for Tunisian arabic. It outputs Tunisian Arabic transcriptions written in Arabic characters.
-This model outputs transcriptions in arabic alphabet only and performs poorly with sentences containing foreign words. However if you do  need code-switching in your transcripts, i.e. foreign outputs in latin alphabet,  you would better use the code switched model, available in another space from the same author. (https://huggingface.co/SalahZa/Code_Switched_Tunisian_Speech_Recognition) 
-
-Run is done on CPU to keep it free in this space. This leads to quite long running times on long sequences. If for your project or research, you want to transcribe long sequences, you would better use the model directly from its page, some instructions for inference on a test set have been provided there. (https://huggingface.co/SalahZa/Tunisian_Automatic_Speech_Recognition). If you need help,  feel free to drop an email here : [email protected] 
-
-Authors : 
-* [Salah Zaiem](https://fr.linkedin.com/in/salah-zaiem)
-* [Ahmed Amine Ben Aballah](https://www.linkedin.com/in/aabenz/)
-* [Ata Kaboudi](https://www.linkedin.com/in/ata-kaboudi-63365b1a8)
-* [Amir Kanoun](https://tn.linkedin.com/in/ahmed-amir-kanoun)
-
-More in-depth details and insights are available in a released preprint. Please find the paper [here](https://arxiv.org/abs/2309.11327).
-If you use or refer to this model, please cite : 
-
-```
-@misc{abdallah2023leveraging,
-      title={Leveraging Data Collection and Unsupervised Learning for Code-switched Tunisian Arabic Automatic Speech Recognition}, 
-      author={Ahmed Amine Ben Abdallah and Ata Kabboudi and Amir Kanoun and Salah Zaiem},
-      year={2023},
-      eprint={2309.11327},
-      archivePrefix={arXiv},
-      primaryClass={eess.AS}
-}
-
-
-"""
-title = "Tunisian Speech Recognition"
-
-def treat_wav_file(file_mic,file_upload ,asr=asr_brain, device="cpu") :
-    if (file_mic is not None) and (file_upload is not None):
-        warn_output = "WARNING: You've uploaded an audio file and used the microphone. The recorded file from the microphone will be used and the uploaded audio will be discarded.\n"
-        wav = file_mic
-    elif (file_mic is None) and (file_upload is None):
-        return "ERROR: You have to either use the microphone or upload an audio file"
-    elif file_mic is not None:
-        wav = file_mic
-    else:
-        wav = file_upload
-    info = torchaudio.info(wav)
-    sr = info.sample_rate
-    sig = sb.dataio.dataio.read_audio(wav)
-    if len(sig.shape)>1 :
-        sig = torch.mean(sig, dim=1)
-    sig = torch.unsqueeze(sig, 0)
-    tensor_wav = sig.to(device)
-    resampled = torchaudio.functional.resample( tensor_wav, sr, 16000)
-    sentence = asr.treat_wav(resampled)
-    return sentence
-
-gr.Interface(
-    title = title,
-    description = description,
-    fn=treat_wav_file,
-    inputs=[gr.Audio(source="microphone", type='filepath', label = "record", optional = True),
-            gr.Audio(source="upload", type='filepath', label="filein", optional=True)]
-    ,outputs="text").launch()
-
-

1234/env.log

@@ -1,520 +0,0 @@
-SpeechBrain system description
-==============================
-Python version:
-3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0]
-==============================
-Installed Python packages:
-absl-py==1.4.0
-aiohttp==3.9.5
-aiosignal==1.3.1
-alabaster==0.7.16
-albumentations==1.3.1
-altair==4.2.2
-annotated-types==0.7.0
-anyio==3.7.1
-argon2-cffi==23.1.0
-argon2-cffi-bindings==21.2.0
-array_record==0.5.1
-arviz==0.15.1
-astropy==5.3.4
-astunparse==1.6.3
-async-timeout==4.0.3
-atpublic==4.1.0
-attrs==23.2.0
-audioread==3.0.1
-autograd==1.6.2
-Babel==2.15.0
-backcall==0.2.0
-beautifulsoup4==4.12.3
-bidict==0.23.1
-bigframes==1.6.0
-bleach==6.1.0
-blinker==1.4
-blis==0.7.11
-blosc2==2.0.0
-bokeh==3.3.4
-bqplot==0.12.43
-branca==0.7.2
-build==1.2.1
-CacheControl==0.14.0
-cachetools==5.3.3
-catalogue==2.0.10
-certifi==2024.2.2
-cffi==1.16.0
-chardet==5.2.0
-charset-normalizer==3.3.2
-chex==0.1.86
-click==8.1.7
-click-plugins==1.1.1
-cligj==0.7.2
-cloudpathlib==0.16.0
-cloudpickle==2.2.1
-cmake==3.27.9
-cmdstanpy==1.2.2
-colorcet==3.1.0
-colorlover==0.3.0
-colour==0.1.5
-community==1.0.0b1
-confection==0.1.4
-cons==0.4.6
-contextlib2==21.6.0
-contourpy==1.2.1
-cryptography==42.0.7
-cuda-python==12.2.1
-cudf-cu12 @ https://pypi.nvidia.com/cudf-cu12/cudf_cu12-24.4.1-cp310-cp310-manylinux_2_28_x86_64.whl#sha256=57366e7ef09dc63e0b389aff20df6c37d91e2790065861ee31a4720149f5b694
-cufflinks==0.17.3
-cupy-cuda12x==12.2.0
-cvxopt==1.3.2
-cvxpy==1.3.4
-cycler==0.12.1
-cymem==2.0.8
-Cython==3.0.10
-dask==2023.8.1
-datascience==0.17.6
-datasets==2.19.1
-db-dtypes==1.2.0
-dbus-python==1.2.18
-debugpy==1.6.6
-decorator==4.4.2
-defusedxml==0.7.1
-dill==0.3.8
-distributed==2023.8.1
-distro==1.7.0
-dlib==19.24.4
-dm-tree==0.1.8
-docstring_parser==0.16
-docutils==0.18.1
-dopamine_rl==4.0.9
-duckdb==0.10.2
-earthengine-api==0.1.403
-easydict==1.13
-ecos==2.0.13
-editdistance==0.6.2
-eerepr==0.0.4
-en-core-web-sm @ https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.7.1/en_core_web_sm-3.7.1-py3-none-any.whl#sha256=86cc141f63942d4b2c5fcee06630fd6f904788d2f0ab005cce45aadb8fb73889
-entrypoints==0.4
-et-xmlfile==1.1.0
-etils==1.7.0
-etuples==0.3.9
-exceptiongroup==1.2.1
-fastai==2.7.15
-fastcore==1.5.38
-fastdownload==0.0.7
-fastjsonschema==2.19.1
-fastprogress==1.0.3
-fastrlock==0.8.2
-filelock==3.14.0
-fiona==1.9.6
-firebase-admin==5.3.0
-Flask==2.2.5
-flatbuffers==24.3.25
-flax==0.8.3
-folium==0.14.0
-fonttools==4.51.0
-frozendict==2.4.4
-frozenlist==1.4.1
-fsspec==2023.6.0
-future==0.18.3
-gast==0.5.4
-gcsfs==2023.6.0
-GDAL==3.6.4
-gdown==5.1.0
-geemap==0.32.1
-gensim==4.3.2
-geocoder==1.38.1
-geographiclib==2.0
-geopandas==0.13.2
-geopy==2.3.0
-gin-config==0.5.0
-glob2==0.7
-google==2.0.3
-google-ai-generativelanguage==0.6.4
-google-api-core==2.11.1
-google-api-python-client==2.84.0
-google-auth==2.27.0
-google-auth-httplib2==0.1.1
-google-auth-oauthlib==1.2.0
-google-cloud-aiplatform==1.52.0
-google-cloud-bigquery==3.21.0
-google-cloud-bigquery-connection==1.12.1
-google-cloud-bigquery-storage==2.25.0
-google-cloud-core==2.3.3
-google-cloud-datastore==2.15.2
-google-cloud-firestore==2.11.1
-google-cloud-functions==1.13.3
-google-cloud-iam==2.15.0
-google-cloud-language==2.13.3
-google-cloud-resource-manager==1.12.3
-google-cloud-storage==2.8.0
-google-cloud-translate==3.11.3
-google-colab @ file:///colabtools/dist/google-colab-1.0.0.tar.gz#sha256=be2293f8ecde760dca63eb67279ba06283ad80f0e13e1c08343f5ef1b33c7116
-google-crc32c==1.5.0
-google-generativeai==0.5.4
-google-pasta==0.2.0
-google-resumable-media==2.7.0
-googleapis-common-protos==1.63.0
-googledrivedownloader==0.4
-graphviz==0.20.3
-greenlet==3.0.3
-grpc-google-iam-v1==0.13.0
-grpcio==1.64.0
-grpcio-status==1.48.2
-gspread==6.0.2
-gspread-dataframe==3.3.1
-gym==0.25.2
-gym-notices==0.0.8
-h5netcdf==1.3.0
-h5py==3.9.0
-holidays==0.49
-holoviews==1.17.1
-html5lib==1.1
-httpimport==1.3.1
-httplib2==0.22.0
-huggingface-hub==0.23.1
-humanize==4.7.0
-hyperopt==0.2.7
-HyperPyYAML==1.2.2
-hypothesis==6.102.6
-ibis-framework==8.0.0
-idna==3.7
-imageio==2.31.6
-imageio-ffmpeg==0.4.9
-imagesize==1.4.1
-imbalanced-learn==0.10.1
-imgaug==0.4.0
-importlib_metadata==7.1.0
-importlib_resources==6.4.0
-imutils==0.5.4
-inflect==7.0.0
-iniconfig==2.0.0
-intel-openmp==2023.2.4
-ipyevents==2.0.2
-ipyfilechooser==0.6.0
-ipykernel==5.5.6
-ipyleaflet==0.18.2
-ipython==7.34.0
-ipython-genutils==0.2.0
-ipython-sql==0.5.0
-ipytree==0.2.2
-ipywidgets==7.7.1
-itsdangerous==2.2.0
-jax==0.4.26
-jaxlib @ https://storage.googleapis.com/jax-releases/cuda12/jaxlib-0.4.26+cuda12.cudnn89-cp310-cp310-manylinux2014_x86_64.whl#sha256=813cf1fe3e7ca4dbf5327d6e7b4fc8521e92d8bba073ee645ae0d5d036a25750
-jeepney==0.7.1
-jellyfish==1.0.3
-jieba==0.42.1
-Jinja2==3.1.4
-jiwer==3.0.4
-joblib==1.4.2
-jsonpickle==3.0.4
-jsonschema==4.19.2
-jsonschema-specifications==2023.12.1
-jupyter-client==6.1.12
-jupyter-console==6.1.0
-jupyter-server==1.24.0
-jupyter_core==5.7.2
-jupyterlab_pygments==0.3.0
-jupyterlab_widgets==3.0.10
-kaggle==1.6.14
-kagglehub==0.2.5
-kenlm==0.2.0
-keras==2.15.0
-keyring==23.5.0
-kiwisolver==1.4.5
-langcodes==3.4.0
-language_data==1.2.0
-launchpadlib==1.10.16
-lazr.restfulclient==0.14.4
-lazr.uri==1.0.6
-lazy_loader==0.4
-libclang==18.1.1
-librosa==0.10.2.post1
-lightgbm==4.1.0
-linkify-it-py==2.0.3
-llvmlite==0.41.1
-locket==1.0.0
-logical-unification==0.4.6
-lxml==4.9.4
-malloy==2023.1067
-marisa-trie==1.1.1
-Markdown==3.6
-markdown-it-py==3.0.0
-MarkupSafe==2.1.5
-matplotlib==3.7.1
-matplotlib-inline==0.1.7
-matplotlib-venn==0.11.10
-mdit-py-plugins==0.4.1
-mdurl==0.1.2
-miniKanren==1.0.3
-missingno==0.5.2
-mistune==0.8.4
-mizani==0.9.3
-mkl==2023.2.0
-ml-dtypes==0.2.0
-mlxtend==0.22.0
-more-itertools==10.1.0
-moviepy==1.0.3
-mpmath==1.3.0
-msgpack==1.0.8
-multidict==6.0.5
-multipledispatch==1.0.0
-multiprocess==0.70.16
-multitasking==0.0.11
-murmurhash==1.0.10
-music21==9.1.0
-natsort==8.4.0
-nbclassic==1.0.0
-nbclient==0.10.0
-nbconvert==6.5.4
-nbformat==5.10.4
-nest-asyncio==1.6.0
-networkx==3.3
-nibabel==4.0.2
-nltk==3.8.1
-notebook==6.5.5
-notebook_shim==0.2.4
-numba==0.58.1
-numexpr==2.10.0
-numpy==1.25.2
-nvidia-cublas-cu12==12.1.3.1
-nvidia-cuda-cupti-cu12==12.1.105
-nvidia-cuda-nvrtc-cu12==12.1.105
-nvidia-cuda-runtime-cu12==12.1.105
-nvidia-cudnn-cu12==8.9.2.26
-nvidia-cufft-cu12==11.0.2.54
-nvidia-curand-cu12==10.3.2.106
-nvidia-cusolver-cu12==11.4.5.107
-nvidia-cusparse-cu12==12.1.0.106
-nvidia-nccl-cu12==2.20.5
-nvidia-nvjitlink-cu12==12.5.40
-nvidia-nvtx-cu12==12.1.105
-nvtx==0.2.10
-oauth2client==4.1.3
-oauthlib==3.2.2
-opencv-contrib-python==4.8.0.76
-opencv-python==4.8.0.76
-opencv-python-headless==4.9.0.80
-openpyxl==3.1.2
-opt-einsum==3.3.0
-optax==0.2.2
-orbax-checkpoint==0.4.4
-osqp==0.6.2.post8
-packaging==24.0
-pandas==2.0.3
-pandas-datareader==0.10.0
-pandas-gbq==0.19.2
-pandas-stubs==2.0.3.230814
-pandocfilters==1.5.1
-panel==1.3.8
-param==2.1.0
-parso==0.8.4
-parsy==2.1
-partd==1.4.2
-pathlib==1.0.1
-patsy==0.5.6
-peewee==3.17.5
-pexpect==4.9.0
-pickleshare==0.7.5
-Pillow==9.4.0
-pip-tools==6.13.0
-platformdirs==4.2.2
-plotly==5.15.0
-plotnine==0.12.4
-pluggy==1.5.0
-polars==0.20.2
-pooch==1.8.1
-portpicker==1.5.2
-prefetch-generator==1.0.3
-preshed==3.0.9
-prettytable==3.10.0
-proglog==0.1.10
-progressbar2==4.2.0
-prometheus_client==0.20.0
-promise==2.3
-prompt-toolkit==3.0.43
-prophet==1.1.5
-proto-plus==1.23.0
-protobuf==3.20.3
-psutil==5.9.5
-psycopg2==2.9.9
-ptyprocess==0.7.0
-py-cpuinfo==9.0.0
-py4j==0.10.9.7
-pyarrow==14.0.2
-pyarrow-hotfix==0.6
-pyasn1==0.6.0
-pyasn1_modules==0.4.0
-pycocotools==2.0.7
-pycparser==2.22
-pyctcdecode==0.5.0
-pydantic==2.7.1
-pydantic_core==2.18.2
-pydata-google-auth==1.8.2
-pydot==1.4.2
-pydot-ng==2.0.0
-pydotplus==2.0.2
-PyDrive==1.3.1
-PyDrive2==1.6.3
-pyerfa==2.0.1.4
-pygame==2.5.2
-Pygments==2.16.1
-PyGObject==3.42.1
-pygtrie==2.5.0
-PyJWT==2.3.0
-pymc==5.10.4
-pymystem3==0.2.0
-pynvjitlink-cu12==0.2.3
-PyOpenGL==3.1.7
-pyOpenSSL==24.1.0
-pyparsing==3.1.2
-pyperclip==1.8.2
-pyproj==3.6.1
-pyproject_hooks==1.1.0
-pyshp==2.3.1
-PySocks==1.7.1
-pytensor==2.18.6
-pytest==7.4.4
-python-apt @ file:///backend-container/containers/python_apt-0.0.0-cp310-cp310-linux_x86_64.whl#sha256=b209c7165d6061963abe611492f8c91c3bcef4b7a6600f966bab58900c63fefa
-python-box==7.1.1
-python-dateutil==2.8.2
-python-louvain==0.16
-python-slugify==8.0.4
-python-utils==3.8.2
-pytz==2023.4
-pyviz_comms==3.0.2
-PyWavelets==1.6.0
-PyYAML==6.0.1
-pyzmq==24.0.1
-qdldl==0.1.7.post2
-qudida==0.0.4
-rapidfuzz==3.9.1
-ratelim==0.1.6
-referencing==0.35.1
-regex==2023.12.25
-requests==2.31.0
-requests-oauthlib==1.3.1
-requirements-parser==0.9.0
-rich==13.7.1
-rmm-cu12==24.4.0
-rpds-py==0.18.1
-rpy2==3.4.2
-rsa==4.9
-ruamel.yaml==0.18.6
-ruamel.yaml.clib==0.2.8
-safetensors==0.4.3
-scikit-image==0.19.3
-scikit-learn==1.2.2
-scipy==1.11.4
-scooby==0.10.0
-scs==3.2.4.post1
-seaborn==0.13.1
-SecretStorage==3.3.1
-Send2Trash==1.8.3
-sentencepiece==0.1.99
-shapely==2.0.4
-six==1.16.0
-sklearn-pandas==2.2.0
-smart-open==6.4.0
-sniffio==1.3.1
-snowballstemmer==2.2.0
-sortedcontainers==2.4.0
-soundfile==0.12.1
-soupsieve==2.5
-soxr==0.3.7
-spacy==3.7.4
-spacy-legacy==3.0.12
-spacy-loggers==1.0.5
-speechbrain==0.5.16
-Sphinx==5.0.2
-sphinxcontrib-applehelp==1.0.8
-sphinxcontrib-devhelp==1.0.6
-sphinxcontrib-htmlhelp==2.0.5
-sphinxcontrib-jsmath==1.0.1
-sphinxcontrib-qthelp==1.0.7
-sphinxcontrib-serializinghtml==1.1.10
-SQLAlchemy==2.0.30
-sqlglot==20.11.0
-sqlparse==0.5.0
-srsly==2.4.8
-stanio==0.5.0
-statsmodels==0.14.2
-StrEnum==0.4.15
-sympy==1.12
-tables==3.8.0
-tabulate==0.9.0
-tbb==2021.12.0
-tblib==3.0.0
-tenacity==8.3.0
-tensorboard==2.15.2
-tensorboard-data-server==0.7.2
-tensorflow @ https://storage.googleapis.com/colab-tf-builds-public-09h6ksrfwbb9g9xv/tensorflow-2.15.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl#sha256=a2ec79931350b378c1ef300ca836b52a55751acb71a433582508a07f0de57c42
-tensorflow-datasets==4.9.4
-tensorflow-estimator==2.15.0
-tensorflow-gcs-config==2.15.0
-tensorflow-hub==0.16.1
-tensorflow-io-gcs-filesystem==0.37.0
-tensorflow-metadata==1.15.0
-tensorflow-probability==0.23.0
-tensorstore==0.1.45
-termcolor==2.4.0
-terminado==0.18.1
-text-unidecode==1.3
-textblob==0.17.1
-tf-slim==1.1.0
-tf_keras==2.15.1
-thinc==8.2.3
-threadpoolctl==3.5.0
-tifffile==2024.5.10
-tinycss2==1.3.0
-tokenizers==0.19.1
-toml==0.10.2
-tomli==2.0.1
-toolz==0.12.1
-torch @ https://download.pytorch.org/whl/cu121/torch-2.3.0%2Bcu121-cp310-cp310-linux_x86_64.whl#sha256=0a12aa9aa6bc442dff8823ac8b48d991fd0771562eaa38593f9c8196d65f7007
-torchaudio @ https://download.pytorch.org/whl/cu121/torchaudio-2.3.0%2Bcu121-cp310-cp310-linux_x86_64.whl#sha256=38b49393f8c322dcaa29d19e5acbf5a0b1978cf1b719445ab670f1fb486e3aa6
-torchsummary==1.5.1
-torchtext==0.18.0
-torchvision @ https://download.pytorch.org/whl/cu121/torchvision-0.18.0%2Bcu121-cp310-cp310-linux_x86_64.whl#sha256=13e1b48dc5ce41ccb8100ab3dd26fdf31d8f1e904ecf2865ac524493013d0df5
-tornado==6.3.3
-tqdm==4.66.4
-traitlets==5.7.1
-traittypes==0.2.1
-transformers==4.41.0
-triton==2.3.0
-tweepy==4.14.0
-typer==0.9.4
-types-pytz==2024.1.0.20240417
-types-setuptools==70.0.0.20240523
-typing_extensions==4.11.0
-tzdata==2024.1
-tzlocal==5.2
-uc-micro-py==1.0.3
-uritemplate==4.1.1
-urllib3==2.0.7
-vega-datasets==0.9.0
-wadllib==1.3.6
-wasabi==1.1.2
-wcwidth==0.2.13
-weasel==0.3.4
-webcolors==1.13
-webencodings==0.5.1
-websocket-client==1.8.0
-Werkzeug==3.0.3
-widgetsnbextension==3.6.6
-wordcloud==1.9.3
-wrapt==1.14.1
-xarray==2023.7.0
-xarray-einstats==0.7.0
-xgboost==2.0.3
-xlrd==2.0.1
-xxhash==3.4.1
-xyzservices==2024.4.0
-yarl==1.9.4
-yellowbrick==1.5
-yfinance==0.2.40
-zict==3.0.0
-zipp==3.18.2
-==============================
-Could not get git revision==============================
-CUDA version:
-12.1

1234/hyperparams.yaml

@@ -1,179 +0,0 @@
-# Generated 2024-05-27 from:
-# /content/drive/MyDrive/TASR/finetuning.yaml
-# yamllint disable
-# ################################
-# Model: wav2vec2 + DNN + CTC
-# Augmentation: SpecAugment
-# Authors: Titouan Parcollet 2021
-# ################################
-
-# Seed needs to be set at top of yaml, before objects with parameters are made
-seed: 1234
-__set_seed: !!python/object/apply:torch.manual_seed [1234]
-output_folder: semi_wavlm_large_tunisian_ctc/1234
-wer_file: semi_wavlm_large_tunisian_ctc/1234/wer.txt
-save_folder: semi_wavlm_large_tunisian_ctc/1234/save
-train_log: semi_wavlm_large_tunisian_ctc/1234/train_log.txt
-
-# URL for the biggest LeBenchmark wav2vec french.
-wav2vec2_folder: semi_wavlm_large_tunisian_ctc/1234/save/wav2vec2_checkpoint
-
-# Data files
-data_folder: /path/to/data  # e.g, /localscratch/cv-corpus-5.1-2020-06-22/fr
-train_tsv_file: /path/to/data/train.tsv       # Standard CommonVoice .tsv files
-dev_tsv_file: /path/to/data/dev.tsv       # Standard CommonVoice .tsv files
-test_tsv_file: /path/to/data/test.tsv       # Standard CommonVoice .tsv files
-accented_letters: true
-language: fr # use 'it' for Italian, 'rw' for Kinyarwanda, 'en' for english
-train_csv: Data/train_wavs/train.csv
-valid_csv: Data/dev_wavs/dev.csv
-test_csv:
-- Data/test_wavs/test.csv
-
-skip_prep: true # Skip data preparation
-
-use_language_modelling: true
-ngram_lm_path: outdomain.arpa
-
-# We remove utterance slonger than 10s in the train/dev/test sets as
-# longer sentences certainly correspond to "open microphones".
-avoid_if_longer_than: 10.0
-avoid_if_shorter_than: 1.2
-
-
-# Training parameters
-number_of_epochs: 20
-lr: 1.0
-lr_wav2vec: 0.0001
-sorting: ascending
-auto_mix_prec: false
-sample_rate: 16000
-ckpt_interval_minutes: 30 # save checkpoint every N min
-
-# With data_parallel batch_size is split into N jobs
-# With DDP batch_size is multiplied by N jobs
-# Must be 6 per GPU to fit 16GB of VRAM
-batch_size: 6
-test_batch_size: 4
-
-dataloader_options:
-  batch_size: 6
-  num_workers: 6
-test_dataloader_options:
-  batch_size: 4
-  num_workers: 6
-
-# BPE parameters
-token_type: char  # ["unigram", "bpe", "char"]
-character_coverage: 1.0
-
-# Model parameters
-# activation: !name:torch.nn.LeakyReLU
-wav2vec_output_dim: 1024
-dnn_neurons: 1024
-freeze_wav2vec: false
-freeze_feature_extractor: true
-dropout: 0.15
-warmup_steps: 500 # The wav2vec 2 model isn't updated for this amount of steps
-
-# Outputs
-output_neurons: 40  # BPE size, index(blank/eos/bos) = 0
-
-# Decoding parameters
-# Be sure that the bos and eos index match with the BPEs ones
-blank_index: 0
-unk_index: 1
-
-#
-# Functions and classes
-#
-epoch_counter: &id007 !new:speechbrain.utils.epoch_loop.EpochCounter
-
-  limit: 20
-
-augmentation: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
-  sample_rate: 16000
-  speeds: [95, 100, 105]
-
-enc: &id002 !new:speechbrain.nnet.containers.Sequential
-  input_shape: [null, null, 1024]
-  linear1: !name:speechbrain.nnet.linear.Linear
-    n_neurons: 1024
-    bias: true
-  bn1: !name:speechbrain.nnet.normalization.BatchNorm1d
-  activation: !new:torch.nn.LeakyReLU
-  drop: !new:torch.nn.Dropout
-    p: 0.15
-  linear2: !name:speechbrain.nnet.linear.Linear
-    n_neurons: 1024
-    bias: true
-  bn2: !name:speechbrain.nnet.normalization.BatchNorm1d
-  activation2: !new:torch.nn.LeakyReLU
-  drop2: !new:torch.nn.Dropout
-    p: 0.15
-  linear3: !name:speechbrain.nnet.linear.Linear
-    n_neurons: 1024
-    bias: true
-  bn3: !name:speechbrain.nnet.normalization.BatchNorm1d
-  activation3: !new:torch.nn.LeakyReLU
-
-wav2vec2: &id001 !new:speechbrain.lobes.models.huggingface_wav2vec.HuggingFaceWav2Vec2
-  source: wavlm-large/
-  output_norm: false
-  freeze: false
-  freeze_feature_extractor: true
-  save_path: semi_wavlm_large_tunisian_ctc/1234/save/wav2vec2_checkpoint
-
-
-ctc_lin: &id003 !new:speechbrain.nnet.linear.Linear
-
-  input_size: 1024
-  n_neurons: 40
-
-log_softmax: !new:speechbrain.nnet.activations.Softmax
-  apply_log: true
-
-ctc_cost: !name:speechbrain.nnet.losses.ctc_loss
-  blank_index: 0
-
-modules:
-  wav2vec2: *id001
-  enc: *id002
-  ctc_lin: *id003
-model: &id004 !new:torch.nn.ModuleList
-- [*id002, *id003]
-model_opt_class: !name:torch.optim.Adadelta
-  lr: 1.0
-  rho: 0.95
-  eps: 1.e-8
-
-wav2vec_opt_class: !name:torch.optim.Adam
-  lr: 0.0001
-
-lr_annealing_model: &id005 !new:speechbrain.nnet.schedulers.NewBobScheduler
-  initial_value: 1.0
-  improvement_threshold: 0.0025
-  annealing_factor: 0.8
-  patient: 0
-
-lr_annealing_wav2vec: &id006 !new:speechbrain.nnet.schedulers.NewBobScheduler
-  initial_value: 0.0001
-  improvement_threshold: 0.0025
-  annealing_factor: 0.9
-  patient: 0
-
-checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
-  checkpoints_dir: semi_wavlm_large_tunisian_ctc/1234/save
-  recoverables:
-    wav2vec2: *id001
-    model: *id004
-    scheduler_model: *id005
-    scheduler_wav2vec: *id006
-    counter: *id007
-train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
-  save_file: semi_wavlm_large_tunisian_ctc/1234/train_log.txt
-
-error_rate_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
-
-cer_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
-  split_tokens: true

1234/save/CKPT+2024-05-27+00-52-30+00/CKPT.yaml

@@ -1,4 +0,0 @@
-# yamllint disable
-WER: 22.820037105751393
-end-of-epoch: true
-unixtime: 1716771150.6156993

1234/save/CKPT+2024-05-27+00-52-30+00/brain.ckpt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:17a40ee07b922d0dc2b48662b8e4fd2f6032c946ffdd21cee17047c2a4a6feca
-size 51

1234/save/CKPT+2024-05-27+00-52-30+00/counter.ckpt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:f5ca38f748a1d6eaf726b8a42fb575c3c71f1864a8143301782de13da2d9202b
-size 2

1234/save/CKPT+2024-05-27+00-52-30+00/dataloader-TRAIN.ckpt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:459535faa370a3b5f8b87203b089623c7aeb9325abf241ec8a685b9c325047a3
-size 3

1234/save/CKPT+2024-05-27+00-52-30+00/model.ckpt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:766f5e1023404b4364d9f5d4a270b0699af3a15cb47684f33dd4b8098bd000b6
-size 12814875

1234/save/CKPT+2024-05-27+00-52-30+00/modelopt.ckpt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:59e5f239bff4d0686120f4e4a9c5d54524a622089908b7e3abc8ebbb1b68f113
-size 25576098

1234/save/CKPT+2024-05-27+00-52-30+00/scheduler_model.ckpt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:94041bc093d9222795932e865c54f3c3a2d48cf44677968ce689d2867af20a92
-size 1152

1234/save/CKPT+2024-05-27+00-52-30+00/scheduler_wav2vec.ckpt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:c8c35e64e340922dc41304565bc0c1429bdf06ca931ed9b0bbecd4d7a3d1547c
-size 1160

1234/save/CKPT+2024-05-27+00-52-30+00/wav2vec2.ckpt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:ebad96920b7643616c39cf455ad89689b844603bf08c821a55004e4553bf37fa
-size 1262006670

1234/save/CKPT+2024-05-27+00-52-30+00/wav2vec_opt.ckpt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:af8c6950a6c04fd90519da5f40d9e0d1a3ace6aaa1212be88009c04a6fa46a39
-size 2490362300

1234/save/label_encoder.txt

@@ -1,44 +0,0 @@
-'ب' => 38
-'ا' => 1
-'ه' => 2
-'ي' => 3
-'و' => 4
-'ن' => 5
-'أ' => 6
-' ' => 7
-'م' => 8
-'ش' => 9
-'ل' => 10
-'س' => 11
-'ت' => 12
-'د' => 13
-'ر' => 14
-'ى' => 15
-'ح' => 16
-'ط' => 17
-'ع' => 18
-'ك' => 19
-'ف' => 20
-'ق' => 21
-'آ' => 22
-'ة' => 23
-'ج' => 24
-'ض' => 25
-'ز' => 26
-'ص' => 27
-'إ' => 28
-'ث' => 29
-'خ' => 30
-'ڨ' => 31
-'ذ' => 32
-'ظ' => 33
-'ء' => 34
-'غ' => 35
-'ئ' => 36
-'ؤ' => 37
-'<blank>' => 0
-1 => 39
-================
-'starting_index' => 0
-'unk_label' => 1
-'blank_label' => '<blank>'

1234/train_log.txt

@@ -1,14 +0,0 @@
-Epoch loaded: 12 - test loss: 4.76e-01, test CER: 9.62, test WER: 26.35
-epoch: 13, lr_model: 3.28e-01, lr_wav2vec: 5.90e-05 - train loss: 3.66e-01 - valid loss: 3.36e-01, valid CER: 7.90, valid WER: 20.96
-epoch: 14, lr_model: 3.28e-01, lr_wav2vec: 5.90e-05 - train loss: 3.07e-01 - valid loss: 3.48e-01, valid CER: 8.34, valid WER: 23.38
-epoch: 15, lr_model: 2.62e-01, lr_wav2vec: 5.31e-05 - train loss: 2.66e-01 - valid loss: 3.58e-01, valid CER: 7.79, valid WER: 21.89
-epoch: 16, lr_model: 2.10e-01, lr_wav2vec: 4.78e-05 - train loss: 2.41e-01 - valid loss: 3.61e-01, valid CER: 8.09, valid WER: 22.82
-epoch: 17, lr_model: 1.68e-01, lr_wav2vec: 4.30e-05 - train loss: 2.21e-01 - valid loss: 3.75e-01, valid CER: 7.83, valid WER: 22.45
-epoch: 18, lr_model: 1.34e-01, lr_wav2vec: 3.87e-05 - train loss: 2.08e-01 - valid loss: 3.98e-01, valid CER: 8.12, valid WER: 22.82
-epoch: 19, lr_model: 1.07e-01, lr_wav2vec: 3.49e-05 - train loss: 1.94e-01 - valid loss: 4.06e-01, valid CER: 8.05, valid WER: 22.63
-epoch: 20, lr_model: 8.59e-02, lr_wav2vec: 3.14e-05 - train loss: 1.87e-01 - valid loss: 4.21e-01, valid CER: 7.72, valid WER: 22.82
-Epoch loaded: 20 - test loss: 4.38e-01, test CER: 9.18, test WER: 24.78
-Epoch loaded: 20 - test loss: 4.38e-01, test CER: 9.18, test WER: 24.78
-Epoch loaded: 20 - test loss: 4.38e-01, test CER: 9.18, test WER: 24.78
-Epoch loaded: 20 - test loss: 4.38e-01, test CER: 9.18, test WER: 24.78
-Epoch loaded: 20 - test loss: 4.38e-01, test CER: 9.18, test WER: 24.78

1234/train_with_wavlm.py

@@ -1,399 +0,0 @@
-#!/usr/bin/env python3
-import sys
-import torch
-import logging
-import speechbrain as sb
-from pathlib import Path
-import os
-import torchaudio
-from hyperpyyaml import load_hyperpyyaml
-from speechbrain.tokenizers.SentencePiece import SentencePiece
-from speechbrain.utils.data_utils import undo_padding
-from speechbrain.utils.distributed import run_on_main
-
-"""Recipe for training a sequence-to-sequence ASR system with CommonVoice.
-The system employs a wav2vec2 encoder and a CTC decoder.
-Decoding is performed with greedy decoding (will be extended to beam search).
-
-To run this recipe, do the following:
-> python train_with_wav2vec2.py hparams/train_with_wav2vec2.yaml
-
-With the default hyperparameters, the system employs a pretrained wav2vec2 encoder.
-The wav2vec2 model is pretrained following the model given in the hprams file.
-It may be dependent on the language.
-
-The neural network is trained with CTC on sub-word units estimated with
-Byte Pairwise Encoding (BPE).
-
-The experiment file is flexible enough to support a large variety of
-different systems. By properly changing the parameter files, you can try
-different encoders, decoders, tokens (e.g, characters instead of BPE),
-training languages (all CommonVoice languages), and many
-other possible variations.
-
-Authors
- * Titouan Parcollet 2021
-"""
-
-logger = logging.getLogger(__name__)
-
-
-# Define training procedure
-class ASR(sb.core.Brain):
-    def compute_forward(self, batch, stage):
-        """Forward computations from the waveform batches to the output probabilities."""
-
-        batch = batch.to(self.device)
-        wavs, wav_lens = batch.sig
-        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
-        if stage == sb.Stage.TRAIN:
-            if hasattr(self.hparams, "augmentation"):
-                wavs = self.hparams.augmentation(wavs, wav_lens)
-
-        # Forward pass
-        feats = self.modules.wav2vec2(wavs, wav_lens)
-        x = self.modules.enc(feats)
-        logits = self.modules.ctc_lin(x)
-        p_ctc = self.hparams.log_softmax(logits)
-
-        return p_ctc, wav_lens
-
-    def compute_objectives(self, predictions, batch, stage):
-        """Computes the loss (CTC) given predictions and targets."""
-
-        p_ctc, wav_lens = predictions
-
-        ids = batch.id
-        tokens, tokens_lens = batch.tokens
-
-        loss = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
-
-        if stage != sb.Stage.TRAIN:
-            predicted_tokens = sb.decoders.ctc_greedy_decode(
-                p_ctc, wav_lens, blank_id=self.hparams.blank_index
-            )
-            # Decode token terms to words
-            if self.hparams.use_language_modelling:
-                predicted_words = []
-                for logs in p_ctc:
-                    text = decoder.decode(logs.detach().cpu().numpy())
-                    predicted_words.append(text.split(" "))
-            else:
-                predicted_words = [
-                    "".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
-                    for utt_seq in predicted_tokens
-                ]
-            # Convert indices to words
-            target_words = [wrd.split(" ") for wrd in batch.wrd]
-
-            self.wer_metric.append(ids, predicted_words, target_words)
-            self.cer_metric.append(ids, predicted_words, target_words)
-
-        return loss
-
-    def fit_batch(self, batch):
-        """Train the parameters given a single batch in input"""
-        should_step = self.step % self.grad_accumulation_factor == 0
-        # Managing automatic mixed precision
-        # TOFIX: CTC fine-tuning currently is unstable
-        # This is certainly due to CTC being done in fp16 instead of fp32
-        if self.auto_mix_prec:
-            with torch.cuda.amp.autocast():
-                with self.no_sync():
-                    outputs = self.compute_forward(batch, sb.Stage.TRAIN)
-                loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
-            with self.no_sync(not should_step):
-                self.scaler.scale(
-                    loss / self.grad_accumulation_factor
-                ).backward()
-            if should_step:
-
-                if not self.hparams.wav2vec2.freeze:
-                    self.scaler.unscale_(self.wav2vec_optimizer)
-                self.scaler.unscale_(self.model_optimizer)
-                if self.check_gradients(loss):
-                    if not self.hparams.wav2vec2.freeze:
-                        if self.optimizer_step >= self.hparams.warmup_steps:
-                            self.scaler.step(self.wav2vec_optimizer)
-                    self.scaler.step(self.model_optimizer)
-                self.scaler.update()
-                self.zero_grad()
-                self.optimizer_step += 1
-        else:
-            # This is mandatory because HF models have a weird behavior with DDP
-            # on the forward pass
-            with self.no_sync():
-                outputs = self.compute_forward(batch, sb.Stage.TRAIN)
-
-            loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
-
-            with self.no_sync(not should_step):
-                (loss / self.grad_accumulation_factor).backward()
-            if should_step:
-                if self.check_gradients(loss):
-                    if not self.hparams.wav2vec2.freeze:
-                        if self.optimizer_step >= self.hparams.warmup_steps:
-                            self.wav2vec_optimizer.step()
-                    self.model_optimizer.step()
-                self.zero_grad()
-                self.optimizer_step += 1
-
-        self.on_fit_batch_end(batch, outputs, loss, should_step)
-        return loss.detach().cpu()
-
-    def evaluate_batch(self, batch, stage):
-        """Computations needed for validation/test batches"""
-        predictions = self.compute_forward(batch, stage=stage)
-        with torch.no_grad():
-            loss = self.compute_objectives(predictions, batch, stage=stage)
-        return loss.detach()
-
-    def on_stage_start(self, stage, epoch):
-        """Gets called at the beginning of each epoch"""
-        if stage != sb.Stage.TRAIN:
-            self.cer_metric = self.hparams.cer_computer()
-            self.wer_metric = self.hparams.error_rate_computer()
-
-    def on_stage_end(self, stage, stage_loss, epoch):
-        """Gets called at the end of an epoch."""
-        # Compute/store important stats
-        stage_stats = {"loss": stage_loss}
-        if stage == sb.Stage.TRAIN:
-            self.train_stats = stage_stats
-        else:
-            stage_stats["CER"] = self.cer_metric.summarize("error_rate")
-            stage_stats["WER"] = self.wer_metric.summarize("error_rate")
-
-        # Perform end-of-iteration things, like annealing, logging, etc.
-        if stage == sb.Stage.VALID:
-            old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
-                stage_stats["loss"]
-            )
-            old_lr_wav2vec, new_lr_wav2vec = self.hparams.lr_annealing_wav2vec(
-                stage_stats["loss"]
-            )
-            sb.nnet.schedulers.update_learning_rate(
-                self.model_optimizer, new_lr_model
-            )
-            if not self.hparams.wav2vec2.freeze:
-                sb.nnet.schedulers.update_learning_rate(
-                    self.wav2vec_optimizer, new_lr_wav2vec
-                )
-            self.hparams.train_logger.log_stats(
-                stats_meta={
-                    "epoch": epoch,
-                    "lr_model": old_lr_model,
-                    "lr_wav2vec": old_lr_wav2vec,
-                },
-                train_stats=self.train_stats,
-                valid_stats=stage_stats,
-            )
-            self.checkpointer.save_and_keep_only(
-                meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
-            )
-        elif stage == sb.Stage.TEST:
-            self.hparams.train_logger.log_stats(
-                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
-                test_stats=stage_stats,
-            )
-            with open(self.hparams.wer_file, "w") as w:
-                self.wer_metric.write_stats(w)
-
-    def init_optimizers(self):
-        "Initializes the wav2vec2 optimizer and model optimizer"
-
-        # If the wav2vec encoder is unfrozen, we create the optimizer
-        if not self.hparams.wav2vec2.freeze:
-            self.wav2vec_optimizer = self.hparams.wav2vec_opt_class(
-                self.modules.wav2vec2.parameters()
-            )
-            if self.checkpointer is not None:
-                self.checkpointer.add_recoverable(
-                    "wav2vec_opt", self.wav2vec_optimizer
-                )
-
-        self.model_optimizer = self.hparams.model_opt_class(
-            self.hparams.model.parameters()
-        )
-
-        if self.checkpointer is not None:
-            self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
-
-    def zero_grad(self, set_to_none=False):
-        if not self.hparams.wav2vec2.freeze:
-            self.wav2vec_optimizer.zero_grad(set_to_none)
-        self.model_optimizer.zero_grad(set_to_none)
-
-
-# Define custom data procedure
-def dataio_prepare(hparams):
-    """This function prepares the datasets to be used in the brain class.
-    It also defines the data processing pipeline through user-defined functions."""
-
-    # 1. Define datasets
-    data_folder = hparams["data_folder"]
-
-    train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
-        csv_path=hparams["train_csv"], replacements={"data_root": data_folder},
-    )
-
-    if hparams["sorting"] == "ascending":
-        # we sort training data to speed up training and get better results.
-        train_data = train_data.filtered_sorted(
-            sort_key="duration",
-            key_max_value={"duration": hparams["avoid_if_longer_than"]},
-        )
-        # when sorting do not shuffle in dataloader ! otherwise is pointless
-        hparams["dataloader_options"]["shuffle"] = False
-
-    elif hparams["sorting"] == "descending":
-        train_data = train_data.filtered_sorted(
-            sort_key="duration",
-            reverse=True,
-            key_max_value={"duration": hparams["avoid_if_longer_than"]},
-        )
-        # when sorting do not shuffle in dataloader ! otherwise is pointless
-        hparams["dataloader_options"]["shuffle"] = False
-
-    elif hparams["sorting"] == "random":
-        pass
-
-    else:
-        raise NotImplementedError(
-            "sorting must be random, ascending or descending"
-        )
-
-    valid_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
-        csv_path=hparams["valid_csv"], replacements={"data_root": data_folder},
-    )
-    # We also sort the validation data so it is faster to validate
-    valid_data = valid_data.filtered_sorted(sort_key="duration")
-    test_datasets = {}
-    for csv_file in hparams["test_csv"]:
-        name = Path(csv_file).stem
-        test_datasets[name] = sb.dataio.dataset.DynamicItemDataset.from_csv(
-            csv_path=csv_file, replacements={"data_root": data_folder}
-        )
-        test_datasets[name] = test_datasets[name].filtered_sorted(
-            sort_key="duration"
-        )
-
-    datasets = [train_data, valid_data] + [i for k, i in test_datasets.items()]
-
-
-    # 2. Define audio pipeline:
-    @sb.utils.data_pipeline.takes("wav")
-    @sb.utils.data_pipeline.provides("sig")
-    def audio_pipeline(wav):
-        info = torchaudio.info(wav)
-        sig = sb.dataio.dataio.read_audio(wav)
-        resampled = torchaudio.transforms.Resample(
-            info.sample_rate, hparams["sample_rate"],
-        )(sig)
-        return resampled
-
-    sb.dataio.dataset.add_dynamic_item(datasets, audio_pipeline)
-    label_encoder = sb.dataio.encoder.CTCTextEncoder()
-
-    # 3. Define text pipeline:
-    @sb.utils.data_pipeline.takes("wrd")
-    @sb.utils.data_pipeline.provides(
-        "wrd", "char_list", "tokens_list", "tokens"
-    )
-    def text_pipeline(wrd):
-        yield wrd
-        char_list = list(wrd)
-        yield char_list
-        tokens_list = label_encoder.encode_sequence(char_list)
-        yield tokens_list
-        tokens = torch.LongTensor(tokens_list)
-        yield tokens
-
-    sb.dataio.dataset.add_dynamic_item(datasets, text_pipeline)
-    lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
-    special_labels = {
-        "blank_label": hparams["blank_index"],
-        "unk_label": hparams["unk_index"]
-    }
-    label_encoder.load_or_create(
-        path=lab_enc_file,
-        from_didatasets=[train_data],
-        output_key="char_list",
-        special_labels=special_labels,
-        sequence_input=True,
-    )
-
-    # 4. Set output:
-    sb.dataio.dataset.set_output_keys(
-        datasets, ["id", "sig", "wrd", "char_list", "tokens"],
-    )
-    return train_data, valid_data,test_datasets, label_encoder
-
-
-if __name__ == "__main__":
-
-    # Load hyperparameters file with command-line overrides
-    hparams_file, run_opts, overrides = sb.parse_arguments(sys.argv[1:])
-    with open(hparams_file) as fin:
-        hparams = load_hyperpyyaml(fin, overrides)
-
-    # If --distributed_launch then
-    # create ddp_group with the right communication protocol
-    sb.utils.distributed.ddp_init_group(run_opts)
-
-
-    # Create experiment directory
-    sb.create_experiment_directory(
-        experiment_directory=hparams["output_folder"],
-        hyperparams_to_save=hparams_file,
-        overrides=overrides,
-    )
-
-    # Due to DDP, we do the preparation ONLY on the main python process
-    # Defining tokenizer and loading it
-    # Create the datasets objects as well as tokenization and encoding :-D
-    train_data, valid_data, test_datasets, label_encoder = dataio_prepare(hparams)
-    if hparams["use_language_modelling"]:
-        print("using langauge_modeeling")
-        from pyctcdecode import build_ctcdecoder
-        ind2lab = label_encoder.ind2lab
-        print(ind2lab)
-        labels = [ind2lab[x] for x in range(len(ind2lab))]
-        labels = [""] + labels[1:-1] + ["1"] 
-        # Replace the <blank> token with a blank character, needed for PyCTCdecode
-        print(labels)
-        decoder = build_ctcdecoder(
-            labels,
-            kenlm_model_path=hparams["ngram_lm_path"],  # .arpa or .bin
-            alpha=0.5,  # Default by KenLM
-            beta=1.0,  # Default by KenLM
-        )
-    # Trainer initialization
-    asr_brain = ASR(
-        modules=hparams["modules"],
-        hparams=hparams,
-        run_opts=run_opts,
-        checkpointer=hparams["checkpointer"],
-    )
-
-    # Adding objects to trainer.
-    asr_brain.tokenizer = label_encoder
-
-    # Training
-    asr_brain.fit(
-        asr_brain.hparams.epoch_counter,
-        train_data,
-        valid_data,
-        train_loader_kwargs=hparams["dataloader_options"],
-        valid_loader_kwargs=hparams["test_dataloader_options"],
-    )
-
-    # Test
-    for k in test_datasets.keys():  # keys are test_clean, test_other etc
-        asr_brain.hparams.wer_file = os.path.join(
-            hparams["output_folder"], "wer_{}.txt".format(k)
-        )
-        asr_brain.evaluate(
-            test_datasets[k], test_loader_kwargs=hparams["test_dataloader_options"]
-        )
-