File size: 4,143 Bytes
eab7921 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 |
{
"run_name": "Wav2Vec-fine-tuning-NURC-SP",
"run_description": "Fine tuning NURC-SP",
"seed": 42,
// AUDIO PARAMS
"sampling_rate": 16000,
// VOCABULARY PARAMETERS
"vocab":{
"vocab_path": "example/vocab_example.json", // generic vocab for Portuguese
"blank": "<pad>", // blank token for padding
"silence": "|", // token between words
"unk": "<unk>" // unk token
},
// TRAINING
"batch_size": 8, // Batch size for training.
"mixed_precision": true, // level of optimization with NVIDIA's apex feature for automatic mixed FP16/FP32 precision (AMP), NOTE: currently only O1 is supported, and use "O1" to activate.
"early_stop_epochs": 10, // If 0 disabled else Number of epochs for stop training with validation loss dont decrease
"preprocess_dataset": false, // if true, the dataset will be pre-processed and saved in disk, otherwise the audio files will be loaded in each step. Preprocessing makes training faster, but requires much more disk space.
// OPTIMIZER
"epochs": 50, // total number of epochs to train.
"lr": 0.00003, // Initial learning rate.
"gradient_accumulation_steps": 24,
// LOGGING
"logging_steps": 100, // Number of steps to plot.
"load_best_model_at_end": true,
"save_total_limit": 3,
"warmup_ratio": 0.05, // 0 disable Ratio of total training steps used for a linear warmup from 0 to learning_rate
"warmup_steps": 0, // 0 disable Number of steps used for a linear warmup from 0 to learning_rate
// DATA LOADING
"num_loader_workers": 8, // number of training data loader processes. Don't set it too big. 4-8 are goo
// MODEL
"freeze_feature_encoder": true, // Whether to freeze the feature encoder layers of the model.
"attention_dropout": 0.1, // The dropout ratio for the attention probabilities.
"activation_dropout": 0.1, // The dropout ratio for activations inside the fully connected layer.
"hidden_dropout": 0.1, // The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
"feat_proj_dropout": 0.1, // The dropout probabilitiy for all 1D convolutional layers in feature encoder.
"mask_time_prob": 0.05, // Propability of each feature vector along the time axis to be chosen as the start of the vector span to be masked.
"layerdrop": 0.0, // The LayerDrop probability.
"gradient_checkpointing": true, // If True, use gradient checkpointing to save memory at the expense of slower backward pass.
// PATHS
"output_path": "../checkpoints/Wav2Vec/NURC-SP/final-version/train/",
// CACHE
"dataset_cache": "../datasets/",
// DATASETS
"datasets":{
"files_path": "../datasets/NURC-SP/audios/", // relative path for audios It's will be join with the audio path CSV
"train":
[
// this dicts is pass directly for the load dataset see the documentation: https://huggingface.co/docs/datasets/package_reference/loading_methods.html#datasets.load_dataset
{
"name": "csv",
"path": "csv",
"data_files": ["../datasets/NURC-SP/corpus_2_train.csv"], // csv files
"text_column": "text",
"path_column": "file_path"
}
]
,
"devel":
[
{
"name": "csv",
"path": "csv",
"data_files": ["../datasets/NURC-SP/corpus_2_dev.csv"], // csv files
"text_column": "text",
"path_column": "file_path"
}
]
}//,
// used only for test
// "KenLM":{
// "kenlm_model_path": "../../kenLM/binaries/subtitle/4-gram/lm.binary", // Path for KenLM model
// "lexicon_path": "example/lexicon.lst", // file with all words for limit the decoder search
// "beam": 2048,
// "nbest": 1,
// "beam_threshold": 25,
// "lm_weight": 1,
// "word_score": -1,
// "sil_weight": 0
// }
}
|