Commit
•
0dd32c6
1
Parent(s):
e577347
Training in progress, step 500
Browse files- .gitignore +1 -0
- config.json +252 -0
- create_model.py +47 -0
- preprocessor_config.json +9 -0
- pytorch_model.bin +3 -0
- run_grid_search.sh +49 -0
- run_librispeech.sh +36 -0
- run_speech_recognition_seq2seq.py +539 -0
- runs/Feb23_16-14-28_sanchit--v100/1645632899.7738805/events.out.tfevents.1645632899.sanchit--v100.125825.1 +3 -0
- runs/Feb23_16-14-28_sanchit--v100/events.out.tfevents.1645632899.sanchit--v100.125825.0 +3 -0
- runs/Feb23_16-16-27_sanchit--v100/1645633019.095641/events.out.tfevents.1645633019.sanchit--v100.126289.1 +3 -0
- runs/Feb23_16-16-27_sanchit--v100/events.out.tfevents.1645633019.sanchit--v100.126289.0 +3 -0
- special_tokens_map.json +1 -0
- tokenizer.json +0 -0
- tokenizer_config.json +1 -0
- training_args.bin +3 -0
- vocab.txt +0 -0
.gitignore
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
checkpoint-*/
|
config.json
ADDED
@@ -0,0 +1,252 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "./",
|
3 |
+
"architectures": [
|
4 |
+
"SpeechEncoderDecoderModel"
|
5 |
+
],
|
6 |
+
"decoder": {
|
7 |
+
"_name_or_path": "bert-large-uncased",
|
8 |
+
"add_cross_attention": true,
|
9 |
+
"architectures": [
|
10 |
+
"BertForMaskedLM"
|
11 |
+
],
|
12 |
+
"attention_probs_dropout_prob": 0.0,
|
13 |
+
"bad_words_ids": null,
|
14 |
+
"bos_token_id": null,
|
15 |
+
"chunk_size_feed_forward": 0,
|
16 |
+
"classifier_dropout": null,
|
17 |
+
"cross_attention_hidden_size": null,
|
18 |
+
"decoder_start_token_id": null,
|
19 |
+
"diversity_penalty": 0.0,
|
20 |
+
"do_sample": false,
|
21 |
+
"early_stopping": false,
|
22 |
+
"encoder_no_repeat_ngram_size": 0,
|
23 |
+
"eos_token_id": null,
|
24 |
+
"finetuning_task": null,
|
25 |
+
"forced_bos_token_id": null,
|
26 |
+
"forced_eos_token_id": null,
|
27 |
+
"gradient_checkpointing": false,
|
28 |
+
"hidden_act": "gelu",
|
29 |
+
"hidden_dropout_prob": 0.1,
|
30 |
+
"hidden_size": 1024,
|
31 |
+
"id2label": {
|
32 |
+
"0": "LABEL_0",
|
33 |
+
"1": "LABEL_1"
|
34 |
+
},
|
35 |
+
"initializer_range": 0.02,
|
36 |
+
"intermediate_size": 4096,
|
37 |
+
"is_decoder": true,
|
38 |
+
"is_encoder_decoder": false,
|
39 |
+
"label2id": {
|
40 |
+
"LABEL_0": 0,
|
41 |
+
"LABEL_1": 1
|
42 |
+
},
|
43 |
+
"layer_norm_eps": 1e-12,
|
44 |
+
"length_penalty": 1.0,
|
45 |
+
"max_length": 20,
|
46 |
+
"max_position_embeddings": 512,
|
47 |
+
"min_length": 0,
|
48 |
+
"model_type": "bert",
|
49 |
+
"no_repeat_ngram_size": 0,
|
50 |
+
"num_attention_heads": 16,
|
51 |
+
"num_beam_groups": 1,
|
52 |
+
"num_beams": 1,
|
53 |
+
"num_hidden_layers": 24,
|
54 |
+
"num_return_sequences": 1,
|
55 |
+
"output_attentions": false,
|
56 |
+
"output_hidden_states": false,
|
57 |
+
"output_scores": false,
|
58 |
+
"pad_token_id": 0,
|
59 |
+
"position_embedding_type": "absolute",
|
60 |
+
"prefix": null,
|
61 |
+
"problem_type": null,
|
62 |
+
"pruned_heads": {},
|
63 |
+
"remove_invalid_values": false,
|
64 |
+
"repetition_penalty": 1.0,
|
65 |
+
"return_dict": true,
|
66 |
+
"return_dict_in_generate": false,
|
67 |
+
"sep_token_id": null,
|
68 |
+
"task_specific_params": null,
|
69 |
+
"temperature": 1.0,
|
70 |
+
"tie_encoder_decoder": false,
|
71 |
+
"tie_word_embeddings": true,
|
72 |
+
"tokenizer_class": null,
|
73 |
+
"top_k": 50,
|
74 |
+
"top_p": 1.0,
|
75 |
+
"torch_dtype": null,
|
76 |
+
"torchscript": false,
|
77 |
+
"transformers_version": "4.17.0.dev0",
|
78 |
+
"type_vocab_size": 2,
|
79 |
+
"use_bfloat16": false,
|
80 |
+
"use_cache": false,
|
81 |
+
"vocab_size": 30522
|
82 |
+
},
|
83 |
+
"decoder_start_token_id": 101,
|
84 |
+
"encoder": {
|
85 |
+
"_name_or_path": "facebook/wav2vec2-large-lv60",
|
86 |
+
"activation_dropout": 0.0,
|
87 |
+
"adapter_kernel_size": 3,
|
88 |
+
"adapter_stride": 2,
|
89 |
+
"add_adapter": true,
|
90 |
+
"add_cross_attention": false,
|
91 |
+
"apply_spec_augment": false,
|
92 |
+
"architectures": [
|
93 |
+
"Wav2Vec2ForPreTraining"
|
94 |
+
],
|
95 |
+
"attention_dropout": 0.0,
|
96 |
+
"bad_words_ids": null,
|
97 |
+
"bos_token_id": 1,
|
98 |
+
"chunk_size_feed_forward": 0,
|
99 |
+
"classifier_proj_size": 256,
|
100 |
+
"codevector_dim": 768,
|
101 |
+
"contrastive_logits_temperature": 0.1,
|
102 |
+
"conv_bias": true,
|
103 |
+
"conv_dim": [
|
104 |
+
512,
|
105 |
+
512,
|
106 |
+
512,
|
107 |
+
512,
|
108 |
+
512,
|
109 |
+
512,
|
110 |
+
512
|
111 |
+
],
|
112 |
+
"conv_kernel": [
|
113 |
+
10,
|
114 |
+
3,
|
115 |
+
3,
|
116 |
+
3,
|
117 |
+
3,
|
118 |
+
2,
|
119 |
+
2
|
120 |
+
],
|
121 |
+
"conv_stride": [
|
122 |
+
5,
|
123 |
+
2,
|
124 |
+
2,
|
125 |
+
2,
|
126 |
+
2,
|
127 |
+
2,
|
128 |
+
2
|
129 |
+
],
|
130 |
+
"cross_attention_hidden_size": null,
|
131 |
+
"ctc_loss_reduction": "sum",
|
132 |
+
"ctc_zero_infinity": false,
|
133 |
+
"decoder_start_token_id": null,
|
134 |
+
"diversity_loss_weight": 0.1,
|
135 |
+
"diversity_penalty": 0.0,
|
136 |
+
"do_sample": false,
|
137 |
+
"do_stable_layer_norm": true,
|
138 |
+
"early_stopping": false,
|
139 |
+
"encoder_no_repeat_ngram_size": 0,
|
140 |
+
"eos_token_id": 2,
|
141 |
+
"feat_extract_activation": "gelu",
|
142 |
+
"feat_extract_dropout": 0.0,
|
143 |
+
"feat_extract_norm": "layer",
|
144 |
+
"feat_proj_dropout": 0.0,
|
145 |
+
"feat_quantizer_dropout": 0.0,
|
146 |
+
"final_dropout": 0.0,
|
147 |
+
"finetuning_task": null,
|
148 |
+
"forced_bos_token_id": null,
|
149 |
+
"forced_eos_token_id": null,
|
150 |
+
"gradient_checkpointing": false,
|
151 |
+
"hidden_act": "gelu",
|
152 |
+
"hidden_dropout": 0.0,
|
153 |
+
"hidden_dropout_prob": 0.0,
|
154 |
+
"hidden_size": 1024,
|
155 |
+
"id2label": {
|
156 |
+
"0": "LABEL_0",
|
157 |
+
"1": "LABEL_1"
|
158 |
+
},
|
159 |
+
"initializer_range": 0.02,
|
160 |
+
"intermediate_size": 4096,
|
161 |
+
"is_decoder": false,
|
162 |
+
"is_encoder_decoder": false,
|
163 |
+
"label2id": {
|
164 |
+
"LABEL_0": 0,
|
165 |
+
"LABEL_1": 1
|
166 |
+
},
|
167 |
+
"layer_norm_eps": 1e-05,
|
168 |
+
"layerdrop": 0.0,
|
169 |
+
"length_penalty": 1.0,
|
170 |
+
"mask_feature_length": 10,
|
171 |
+
"mask_feature_min_masks": 0,
|
172 |
+
"mask_feature_prob": 0.0,
|
173 |
+
"mask_time_length": 10,
|
174 |
+
"mask_time_min_masks": 2,
|
175 |
+
"mask_time_prob": 0.0,
|
176 |
+
"max_length": 20,
|
177 |
+
"min_length": 0,
|
178 |
+
"model_type": "wav2vec2",
|
179 |
+
"no_repeat_ngram_size": 0,
|
180 |
+
"num_adapter_layers": 3,
|
181 |
+
"num_attention_heads": 16,
|
182 |
+
"num_beam_groups": 1,
|
183 |
+
"num_beams": 1,
|
184 |
+
"num_codevector_groups": 2,
|
185 |
+
"num_codevectors_per_group": 320,
|
186 |
+
"num_conv_pos_embedding_groups": 16,
|
187 |
+
"num_conv_pos_embeddings": 128,
|
188 |
+
"num_feat_extract_layers": 7,
|
189 |
+
"num_hidden_layers": 24,
|
190 |
+
"num_negatives": 100,
|
191 |
+
"num_return_sequences": 1,
|
192 |
+
"output_attentions": false,
|
193 |
+
"output_hidden_size": 1024,
|
194 |
+
"output_hidden_states": false,
|
195 |
+
"output_scores": false,
|
196 |
+
"pad_token_id": 0,
|
197 |
+
"prefix": null,
|
198 |
+
"problem_type": null,
|
199 |
+
"proj_codevector_dim": 768,
|
200 |
+
"pruned_heads": {},
|
201 |
+
"remove_invalid_values": false,
|
202 |
+
"repetition_penalty": 1.0,
|
203 |
+
"return_dict": true,
|
204 |
+
"return_dict_in_generate": false,
|
205 |
+
"sep_token_id": null,
|
206 |
+
"task_specific_params": null,
|
207 |
+
"tdnn_dilation": [
|
208 |
+
1,
|
209 |
+
2,
|
210 |
+
3,
|
211 |
+
1,
|
212 |
+
1
|
213 |
+
],
|
214 |
+
"tdnn_dim": [
|
215 |
+
512,
|
216 |
+
512,
|
217 |
+
512,
|
218 |
+
512,
|
219 |
+
1500
|
220 |
+
],
|
221 |
+
"tdnn_kernel": [
|
222 |
+
5,
|
223 |
+
3,
|
224 |
+
3,
|
225 |
+
1,
|
226 |
+
1
|
227 |
+
],
|
228 |
+
"temperature": 1.0,
|
229 |
+
"tie_encoder_decoder": false,
|
230 |
+
"tie_word_embeddings": true,
|
231 |
+
"tokenizer_class": null,
|
232 |
+
"top_k": 50,
|
233 |
+
"top_p": 1.0,
|
234 |
+
"torch_dtype": null,
|
235 |
+
"torchscript": false,
|
236 |
+
"transformers_version": "4.17.0.dev0",
|
237 |
+
"use_bfloat16": false,
|
238 |
+
"use_weighted_layer_sum": false,
|
239 |
+
"vocab_size": 32,
|
240 |
+
"xvector_output_dim": 512
|
241 |
+
},
|
242 |
+
"eos_token_id": 102,
|
243 |
+
"is_encoder_decoder": true,
|
244 |
+
"max_length": 50,
|
245 |
+
"model_type": "speech-encoder-decoder",
|
246 |
+
"pad_token_id": 0,
|
247 |
+
"processor_class": "Wav2Vec2Processor",
|
248 |
+
"tie_word_embeddings": false,
|
249 |
+
"torch_dtype": "float32",
|
250 |
+
"transformers_version": null,
|
251 |
+
"use_cache": false
|
252 |
+
}
|
create_model.py
ADDED
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import SpeechEncoderDecoderModel, AutoFeatureExtractor, AutoTokenizer, Wav2Vec2Processor
|
2 |
+
import torch
|
3 |
+
|
4 |
+
# checkpoints to leverage
|
5 |
+
encoder_id = "facebook/wav2vec2-large-lv60"
|
6 |
+
decoder_id = "bert-large-uncased"
|
7 |
+
|
8 |
+
feature_extractor = AutoFeatureExtractor.from_pretrained(encoder_id)
|
9 |
+
feature_extractor.save_pretrained("./")
|
10 |
+
tokenizer = AutoTokenizer.from_pretrained(decoder_id)
|
11 |
+
tokenizer.save_pretrained("./")
|
12 |
+
|
13 |
+
model = SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(encoder_id, decoder_id, encoder_add_adapter=True)
|
14 |
+
|
15 |
+
# set all encoder regularisation to zero
|
16 |
+
model.config.encoder.feat_proj_dropout = 0.0
|
17 |
+
model.config.encoder.final_dropout = 0.0
|
18 |
+
model.config.encoder.activation_dropout = 0.0
|
19 |
+
model.config.encoder.apply_spec_augment = False
|
20 |
+
model.config.encoder.attention_dropout = 0.0
|
21 |
+
model.config.encoder.feat_extract_dropout = 0.0
|
22 |
+
model.config.encoder.feat_proj_dropout = 0.0
|
23 |
+
model.config.encoder.hidden_dropout = 0.0
|
24 |
+
model.config.encoder.hidden_dropout_prob = 0.0
|
25 |
+
model.config.encoder.layerdrop = 0.0
|
26 |
+
model.config.encoder.mask_feature_prob = 0.0
|
27 |
+
model.config.encoder.mask_time_prob = 0.0
|
28 |
+
|
29 |
+
# set all decoder regularisation to zero
|
30 |
+
model.config.decoder.attention_probs_dropout_prob = 0.0
|
31 |
+
|
32 |
+
# set special token ids
|
33 |
+
model.config.decoder_start_token_id = tokenizer.cls_token_id
|
34 |
+
model.config.pad_token_id = tokenizer.pad_token_id
|
35 |
+
model.config.eos_token_id = tokenizer.sep_token_id
|
36 |
+
model.config.max_length = 50
|
37 |
+
model.config.num_beams = 1
|
38 |
+
model.config.use_cache = False
|
39 |
+
model.config.decoder.use_cache = False
|
40 |
+
model.config.processor_class = "Wav2Vec2Processor"
|
41 |
+
|
42 |
+
# check if generation works
|
43 |
+
out = model.generate(torch.ones((1, 2000)))
|
44 |
+
|
45 |
+
model.save_pretrained("./")
|
46 |
+
|
47 |
+
|
preprocessor_config.json
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"do_normalize": true,
|
3 |
+
"feature_extractor_type": "Wav2Vec2FeatureExtractor",
|
4 |
+
"feature_size": 1,
|
5 |
+
"padding_side": "right",
|
6 |
+
"padding_value": 0.0,
|
7 |
+
"return_attention_mask": true,
|
8 |
+
"sampling_rate": 16000
|
9 |
+
}
|
pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d95d5c1fe20d610fb07e961cf75b322a7a64f2273f154e8e1e431ffa3b693356
|
3 |
+
size 3081656435
|
run_grid_search.sh
ADDED
@@ -0,0 +1,49 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env bash
|
2 |
+
|
3 |
+
declare -a learning_rates=("1e-5" "3e-5" "1e-4" "3e-4" "1e-3")
|
4 |
+
declare -a batch_sizes=("8" "12" "14" "16")
|
5 |
+
declare -a gradient_accumulation_step_sizes=("2" "4" "8")
|
6 |
+
|
7 |
+
for learning_rate in "${learning_rates[@]}"; do
|
8 |
+
for batch_size in "${batch_sizes[@]}"; do
|
9 |
+
for gradient_accumulation_steps in "${gradient_accumulation_step_sizes[@]}"; do
|
10 |
+
python create_model.py
|
11 |
+
CUDA_VISIBLE_DEVICES=1 python run_speech_recognition_seq2seq.py \
|
12 |
+
--dataset_name="librispeech_asr" \
|
13 |
+
--model_name_or_path="./" \
|
14 |
+
--tokenizer_name="./" \
|
15 |
+
--dataset_config_name="clean" \
|
16 |
+
--train_split_name="train.100" \
|
17 |
+
--eval_split_name="validation" \
|
18 |
+
--output_dir="./" \
|
19 |
+
--preprocessing_num_workers="1" \
|
20 |
+
--length_column_name="input_length" \
|
21 |
+
--overwrite_output_dir \
|
22 |
+
--num_train_epochs="1" \
|
23 |
+
--per_device_train_batch_size=$batch_size \
|
24 |
+
--per_device_eval_batch_size=$batch_size \
|
25 |
+
--gradient_accumulation_steps=$gradient_accumulation_steps \
|
26 |
+
--generation_max_length="40" \
|
27 |
+
--generation_num_beams="1" \
|
28 |
+
--learning_rate=$learning_rate \
|
29 |
+
--warmup_steps="500" \
|
30 |
+
--evaluation_strategy="steps" \
|
31 |
+
--text_column_name="text" \
|
32 |
+
--save_steps="500" \
|
33 |
+
--eval_steps="500" \
|
34 |
+
--logging_steps="1" \
|
35 |
+
--save_total_limit="1" \
|
36 |
+
--freeze_feature_encoder \
|
37 |
+
--gradient_checkpointing \
|
38 |
+
--fp16 \
|
39 |
+
--group_by_length \
|
40 |
+
--predict_with_generate \
|
41 |
+
--do_lower_case \
|
42 |
+
--do_train \
|
43 |
+
--do_eval \
|
44 |
+
--push_to_hub \
|
45 |
+
--use_auth_token
|
46 |
+
done
|
47 |
+
done
|
48 |
+
done
|
49 |
+
|
run_librispeech.sh
ADDED
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env bash
|
2 |
+
CUDA_VISIBLE_DEVICES=1 python run_speech_recognition_seq2seq.py \
|
3 |
+
--dataset_name="librispeech_asr" \
|
4 |
+
--model_name_or_path="./" \
|
5 |
+
--tokenizer_name="./" \
|
6 |
+
--dataset_config_name="clean" \
|
7 |
+
--train_split_name="train.100" \
|
8 |
+
--eval_split_name="validation" \
|
9 |
+
--output_dir="./" \
|
10 |
+
--preprocessing_num_workers="1" \
|
11 |
+
--length_column_name="input_length" \
|
12 |
+
--overwrite_output_dir \
|
13 |
+
--num_train_epochs="3" \
|
14 |
+
--per_device_train_batch_size="14" \
|
15 |
+
--per_device_eval_batch_size="14" \
|
16 |
+
--gradient_accumulation_steps="8" \
|
17 |
+
--generation_max_length="40" \
|
18 |
+
--generation_num_beams="1" \
|
19 |
+
--learning_rate="1e-3" \
|
20 |
+
--warmup_steps="500" \
|
21 |
+
--evaluation_strategy="steps" \
|
22 |
+
--text_column_name="text" \
|
23 |
+
--save_steps="500" \
|
24 |
+
--eval_steps="500" \
|
25 |
+
--logging_steps="1" \
|
26 |
+
--save_total_limit="1" \
|
27 |
+
--freeze_feature_encoder \
|
28 |
+
--gradient_checkpointing \
|
29 |
+
--fp16 \
|
30 |
+
--group_by_length \
|
31 |
+
--predict_with_generate \
|
32 |
+
--do_lower_case \
|
33 |
+
--do_eval --do_train \
|
34 |
+
--push_to_hub \
|
35 |
+
--use_auth_token
|
36 |
+
|
run_speech_recognition_seq2seq.py
ADDED
@@ -0,0 +1,539 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
# coding=utf-8
|
3 |
+
# Copyright 2021 The HuggingFace Team. All rights reserved.
|
4 |
+
#
|
5 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
6 |
+
# you may not use this file except in compliance with the License.
|
7 |
+
# You may obtain a copy of the License at
|
8 |
+
#
|
9 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
10 |
+
#
|
11 |
+
# Unless required by applicable law or agreed to in writing, software
|
12 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
13 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
14 |
+
# See the License for the specific language governing permissions and
|
15 |
+
# limitations under the License.
|
16 |
+
"""
|
17 |
+
Fine-tuning the library models for sequence to sequence speech recognition.
|
18 |
+
"""
|
19 |
+
# You can also adapt this script on your own sequence to sequence speech
|
20 |
+
# recognition task. Pointers for this are left as comments.
|
21 |
+
|
22 |
+
import logging
|
23 |
+
import os
|
24 |
+
import sys
|
25 |
+
from dataclasses import dataclass, field
|
26 |
+
from typing import Any, Dict, List, Optional, Union
|
27 |
+
|
28 |
+
import datasets
|
29 |
+
import torch
|
30 |
+
from datasets import DatasetDict, load_dataset, load_metric
|
31 |
+
|
32 |
+
import bitsandbytes as bnb
|
33 |
+
import transformers
|
34 |
+
from transformers import (
|
35 |
+
AutoConfig,
|
36 |
+
AutoFeatureExtractor,
|
37 |
+
AutoModelForSpeechSeq2Seq,
|
38 |
+
AutoProcessor,
|
39 |
+
AutoTokenizer,
|
40 |
+
HfArgumentParser,
|
41 |
+
Seq2SeqTrainer,
|
42 |
+
Seq2SeqTrainingArguments,
|
43 |
+
set_seed,
|
44 |
+
)
|
45 |
+
from transformers.trainer_pt_utils import get_parameter_names
|
46 |
+
from transformers.trainer_utils import get_last_checkpoint, is_main_process
|
47 |
+
from transformers.utils import check_min_version
|
48 |
+
from transformers.utils.versions import require_version
|
49 |
+
from transformers.optimization import Adafactor
|
50 |
+
|
51 |
+
|
52 |
+
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
|
53 |
+
check_min_version("4.17.0.dev0")
|
54 |
+
|
55 |
+
require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/summarization/requirements.txt")
|
56 |
+
|
57 |
+
logger = logging.getLogger(__name__)
|
58 |
+
|
59 |
+
|
60 |
+
@dataclass
|
61 |
+
class ModelArguments:
|
62 |
+
"""
|
63 |
+
Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
|
64 |
+
"""
|
65 |
+
|
66 |
+
model_name_or_path: str = field(
|
67 |
+
metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
|
68 |
+
)
|
69 |
+
config_name: Optional[str] = field(
|
70 |
+
default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
|
71 |
+
)
|
72 |
+
tokenizer_name: Optional[str] = field(
|
73 |
+
default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
|
74 |
+
)
|
75 |
+
feature_extractor_name: Optional[str] = field(
|
76 |
+
default=None, metadata={"help": "feature extractor name or path if not the same as model_name"}
|
77 |
+
)
|
78 |
+
cache_dir: Optional[str] = field(
|
79 |
+
default=None,
|
80 |
+
metadata={"help": "Where to store the pretrained models downloaded from huggingface.co"},
|
81 |
+
)
|
82 |
+
use_fast_tokenizer: bool = field(
|
83 |
+
default=True,
|
84 |
+
metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
|
85 |
+
)
|
86 |
+
model_revision: str = field(
|
87 |
+
default="main",
|
88 |
+
metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
|
89 |
+
)
|
90 |
+
use_auth_token: bool = field(
|
91 |
+
default=False,
|
92 |
+
metadata={
|
93 |
+
"help": "Will use the token generated when running `transformers-cli login` (necessary to use this script "
|
94 |
+
"with private models)."
|
95 |
+
},
|
96 |
+
)
|
97 |
+
freeze_feature_encoder: bool = field(
|
98 |
+
default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
|
99 |
+
)
|
100 |
+
|
101 |
+
|
102 |
+
@dataclass
|
103 |
+
class DataTrainingArguments:
|
104 |
+
"""
|
105 |
+
Arguments pertaining to what data we are going to input our model for training and eval.
|
106 |
+
"""
|
107 |
+
|
108 |
+
dataset_name: str = field(
|
109 |
+
default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
|
110 |
+
)
|
111 |
+
dataset_config_name: Optional[str] = field(
|
112 |
+
default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
|
113 |
+
)
|
114 |
+
text_column: Optional[str] = field(
|
115 |
+
default=None,
|
116 |
+
metadata={"help": "The name of the column in the datasets containing the full texts (for summarization)."},
|
117 |
+
)
|
118 |
+
overwrite_cache: bool = field(
|
119 |
+
default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
|
120 |
+
)
|
121 |
+
preprocessing_num_workers: Optional[int] = field(
|
122 |
+
default=None,
|
123 |
+
metadata={"help": "The number of processes to use for the preprocessing."},
|
124 |
+
)
|
125 |
+
max_train_samples: Optional[int] = field(
|
126 |
+
default=None,
|
127 |
+
metadata={
|
128 |
+
"help": "For debugging purposes or quicker training, truncate the number of training examples to this "
|
129 |
+
"value if set."
|
130 |
+
},
|
131 |
+
)
|
132 |
+
max_eval_samples: Optional[int] = field(
|
133 |
+
default=None,
|
134 |
+
metadata={
|
135 |
+
"help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
|
136 |
+
"value if set."
|
137 |
+
},
|
138 |
+
)
|
139 |
+
audio_column_name: str = field(
|
140 |
+
default="audio",
|
141 |
+
metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
|
142 |
+
)
|
143 |
+
text_column_name: str = field(
|
144 |
+
default="text",
|
145 |
+
metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
|
146 |
+
)
|
147 |
+
max_duration_in_seconds: float = field(
|
148 |
+
default=20.0,
|
149 |
+
metadata={
|
150 |
+
"help": "Truncate audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
|
151 |
+
},
|
152 |
+
)
|
153 |
+
min_duration_in_seconds: float = field(
|
154 |
+
default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
|
155 |
+
)
|
156 |
+
preprocessing_only: bool = field(
|
157 |
+
default=False,
|
158 |
+
metadata={
|
159 |
+
"help": "Whether to only do data preprocessing and skip training. "
|
160 |
+
"This is especially useful when data preprocessing errors out in distributed training due to timeout. "
|
161 |
+
"In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
|
162 |
+
"so that the cached datasets can consequently be loaded in distributed training"
|
163 |
+
},
|
164 |
+
)
|
165 |
+
train_split_name: str = field(
|
166 |
+
default="train",
|
167 |
+
metadata={
|
168 |
+
"help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
|
169 |
+
},
|
170 |
+
)
|
171 |
+
eval_split_name: str = field(
|
172 |
+
default="test",
|
173 |
+
metadata={
|
174 |
+
"help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
|
175 |
+
},
|
176 |
+
)
|
177 |
+
do_lower_case: bool = field(
|
178 |
+
default=True,
|
179 |
+
metadata={"help": "Whether the target text should be lower cased."},
|
180 |
+
)
|
181 |
+
|
182 |
+
|
183 |
+
@dataclass
|
184 |
+
class DataCollatorSpeechSeq2SeqWithPadding:
|
185 |
+
"""
|
186 |
+
Data collator that will dynamically pad the inputs received.
|
187 |
+
Args:
|
188 |
+
processor ([`Wav2Vec2Processor`])
|
189 |
+
The processor used for proccessing the data.
|
190 |
+
decoder_start_token_id (`int`)
|
191 |
+
The begin-of-sentence of the decoder.
|
192 |
+
"""
|
193 |
+
|
194 |
+
processor: Any
|
195 |
+
decoder_start_token_id: int
|
196 |
+
|
197 |
+
def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
|
198 |
+
# split inputs and labels since they have to be of different lenghts and need
|
199 |
+
# different padding methods
|
200 |
+
input_features = [{"input_values": feature["input_values"]} for feature in features]
|
201 |
+
label_features = [{"input_ids": feature["labels"]} for feature in features]
|
202 |
+
|
203 |
+
batch = self.processor.feature_extractor.pad(input_features, return_tensors="pt")
|
204 |
+
|
205 |
+
labels_batch = self.processor.tokenizer.pad(label_features, return_tensors="pt")
|
206 |
+
|
207 |
+
# replace padding with -100 to ignore loss correctly
|
208 |
+
labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
|
209 |
+
|
210 |
+
# if bos token is appended in previous tokenization step,
|
211 |
+
# cut bos token here as it's append later anyways
|
212 |
+
if (labels[:, 0] == self.decoder_start_token_id).all().cpu().item():
|
213 |
+
labels = labels[:, 1:]
|
214 |
+
|
215 |
+
batch["labels"] = labels
|
216 |
+
|
217 |
+
return batch
|
218 |
+
|
219 |
+
|
220 |
+
def main():
|
221 |
+
# 1. Parse input arguments
|
222 |
+
# See all possible arguments in src/transformers/training_args.py
|
223 |
+
# or by passing the --help flag to this script.
|
224 |
+
# We now keep distinct sets of args, for a cleaner separation of concerns.
|
225 |
+
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
|
226 |
+
|
227 |
+
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
|
228 |
+
# If we pass only one argument to the script and it's the path to a json file,
|
229 |
+
# let's parse it to get our arguments.
|
230 |
+
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
|
231 |
+
else:
|
232 |
+
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
|
233 |
+
|
234 |
+
# 2. Setup logging
|
235 |
+
logging.basicConfig(
|
236 |
+
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
|
237 |
+
datefmt="%m/%d/%Y %H:%M:%S",
|
238 |
+
handlers=[logging.StreamHandler(sys.stdout)],
|
239 |
+
)
|
240 |
+
log_level = training_args.get_process_log_level()
|
241 |
+
logger.setLevel(log_level)
|
242 |
+
datasets.utils.logging.set_verbosity(log_level)
|
243 |
+
transformers.utils.logging.set_verbosity(log_level)
|
244 |
+
transformers.utils.logging.enable_default_handler()
|
245 |
+
transformers.utils.logging.enable_explicit_format()
|
246 |
+
|
247 |
+
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
|
248 |
+
|
249 |
+
# Log on each process the small summary:
|
250 |
+
logger.warning(
|
251 |
+
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
|
252 |
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
|
253 |
+
)
|
254 |
+
logger.info(f"Training/evaluation parameters {training_args}")
|
255 |
+
|
256 |
+
# Set the verbosity to info of the Transformers logger (on main process only):
|
257 |
+
if is_main_process(training_args.local_rank):
|
258 |
+
transformers.utils.logging.set_verbosity_info()
|
259 |
+
logger.info("Training/evaluation parameters %s", training_args)
|
260 |
+
|
261 |
+
# 3. Detecting last checkpoint and eventualy continue from last checkpoint
|
262 |
+
last_checkpoint = None
|
263 |
+
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
|
264 |
+
last_checkpoint = get_last_checkpoint(training_args.output_dir)
|
265 |
+
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
|
266 |
+
raise ValueError(
|
267 |
+
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
|
268 |
+
"Use --overwrite_output_dir to overcome."
|
269 |
+
)
|
270 |
+
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
|
271 |
+
logger.info(
|
272 |
+
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
|
273 |
+
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
|
274 |
+
)
|
275 |
+
|
276 |
+
# Set seed before initializing model.
|
277 |
+
set_seed(training_args.seed)
|
278 |
+
|
279 |
+
# 4. Load dataset
|
280 |
+
raw_datasets = DatasetDict()
|
281 |
+
|
282 |
+
if training_args.do_train:
|
283 |
+
raw_datasets["train"] = load_dataset(
|
284 |
+
data_args.dataset_name, data_args.dataset_config_name, split=data_args.train_split_name
|
285 |
+
)
|
286 |
+
|
287 |
+
if training_args.do_eval:
|
288 |
+
raw_datasets["eval"] = load_dataset(
|
289 |
+
data_args.dataset_name, data_args.dataset_config_name, split=data_args.eval_split_name
|
290 |
+
)
|
291 |
+
|
292 |
+
if data_args.audio_column_name not in next(iter(raw_datasets.values())).column_names:
|
293 |
+
raise ValueError(
|
294 |
+
f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
|
295 |
+
"Make sure to set `--audio_column_name` to the correct audio column - one of "
|
296 |
+
f"{', '.join(next(iter(raw_datasets.values())).column_names)}."
|
297 |
+
)
|
298 |
+
|
299 |
+
if data_args.text_column_name not in next(iter(raw_datasets.values())).column_names:
|
300 |
+
raise ValueError(
|
301 |
+
f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
|
302 |
+
"Make sure to set `--text_column_name` to the correct text column - one of "
|
303 |
+
f"{', '.join(next(iter(raw_datasets.values())).column_names)}."
|
304 |
+
)
|
305 |
+
|
306 |
+
# 5. Load pretrained model, tokenizer, and feature extractor
|
307 |
+
#
|
308 |
+
# Distributed training:
|
309 |
+
# The .from_pretrained methods guarantee that only one local process can concurrently
|
310 |
+
config = AutoConfig.from_pretrained(
|
311 |
+
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
|
312 |
+
cache_dir=model_args.cache_dir,
|
313 |
+
revision=model_args.model_revision,
|
314 |
+
use_auth_token=True if model_args.use_auth_token else None,
|
315 |
+
)
|
316 |
+
|
317 |
+
feature_extractor = AutoFeatureExtractor.from_pretrained(
|
318 |
+
model_args.feature_extractor_name if model_args.feature_extractor_name else model_args.model_name_or_path,
|
319 |
+
cache_dir=model_args.cache_dir,
|
320 |
+
revision=model_args.model_revision,
|
321 |
+
use_auth_token=True if model_args.use_auth_token else None,
|
322 |
+
)
|
323 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
324 |
+
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
|
325 |
+
cache_dir=model_args.cache_dir,
|
326 |
+
use_fast=model_args.use_fast_tokenizer,
|
327 |
+
revision=model_args.model_revision,
|
328 |
+
use_auth_token=True if model_args.use_auth_token else None,
|
329 |
+
)
|
330 |
+
model = AutoModelForSpeechSeq2Seq.from_pretrained(
|
331 |
+
model_args.model_name_or_path,
|
332 |
+
config=config,
|
333 |
+
cache_dir=model_args.cache_dir,
|
334 |
+
revision=model_args.model_revision,
|
335 |
+
use_auth_token=True if model_args.use_auth_token else None,
|
336 |
+
)
|
337 |
+
|
338 |
+
if model.config.decoder_start_token_id is None:
|
339 |
+
raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
|
340 |
+
|
341 |
+
if model_args.freeze_feature_encoder:
|
342 |
+
model.freeze_feature_encoder()
|
343 |
+
|
344 |
+
# 6. Resample speech dataset if necassary
|
345 |
+
dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
|
346 |
+
if dataset_sampling_rate != feature_extractor.sampling_rate:
|
347 |
+
raw_datasets = raw_datasets.cast_column(
|
348 |
+
data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
|
349 |
+
)
|
350 |
+
|
351 |
+
# 7. Preprocessing the datasets.
|
352 |
+
# We need to read the audio files as arrays and tokenize the targets.
|
353 |
+
max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
|
354 |
+
min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
|
355 |
+
audio_column_name = data_args.audio_column_name
|
356 |
+
num_workers = data_args.preprocessing_num_workers
|
357 |
+
text_column_name = data_args.text_column_name
|
358 |
+
model_input_name = feature_extractor.model_input_names[0]
|
359 |
+
do_lower_case = data_args.do_lower_case
|
360 |
+
|
361 |
+
if data_args.max_train_samples is not None:
|
362 |
+
raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
|
363 |
+
|
364 |
+
if data_args.max_eval_samples is not None:
|
365 |
+
raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
|
366 |
+
|
367 |
+
def prepare_dataset(batch):
|
368 |
+
# process audio
|
369 |
+
sample = batch[audio_column_name]
|
370 |
+
inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
|
371 |
+
# process audio length
|
372 |
+
batch[model_input_name] = inputs.input_values[0]
|
373 |
+
batch["input_length"] = len(batch["input_values"])
|
374 |
+
|
375 |
+
# process targets
|
376 |
+
input_str = batch[text_column_name].lower() if do_lower_case else batch[text_column_name]
|
377 |
+
batch["labels"] = tokenizer(input_str).input_ids
|
378 |
+
return batch
|
379 |
+
|
380 |
+
with training_args.main_process_first(desc="dataset map pre-processing"):
|
381 |
+
vectorized_datasets = raw_datasets.map(
|
382 |
+
prepare_dataset,
|
383 |
+
remove_columns=next(iter(raw_datasets.values())).column_names,
|
384 |
+
num_proc=data_args.preprocessing_num_workers,
|
385 |
+
desc="preprocess train dataset",
|
386 |
+
)
|
387 |
+
|
388 |
+
# filter data that is shorter than min_input_length or longer than
|
389 |
+
# max_input_length
|
390 |
+
def is_audio_in_length_range(length):
|
391 |
+
return length > min_input_length and length < max_input_length
|
392 |
+
|
393 |
+
vectorized_datasets = vectorized_datasets.filter(
|
394 |
+
is_audio_in_length_range,
|
395 |
+
num_proc=num_workers,
|
396 |
+
input_columns=["input_length"],
|
397 |
+
)
|
398 |
+
|
399 |
+
# for large datasets it is advised to run the preprocessing on a
|
400 |
+
# single machine first with `args.preprocessing_only` since there will mostly likely
|
401 |
+
# be a timeout when running the script in distributed mode.
|
402 |
+
# In a second step `args.preprocessing_only` can then be set to `False` to load the
|
403 |
+
# cached dataset
|
404 |
+
if data_args.preprocessing_only:
|
405 |
+
cache = {k: v.cache_files for k, v in vectorized_datasets.items()}
|
406 |
+
logger.info(f"Data preprocessing finished. Files cached at {cache}.")
|
407 |
+
return
|
408 |
+
|
409 |
+
# 8. Load Metric
|
410 |
+
metric = load_metric("wer")
|
411 |
+
|
412 |
+
def compute_metrics(pred):
|
413 |
+
pred_ids = pred.predictions
|
414 |
+
|
415 |
+
pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
|
416 |
+
|
417 |
+
pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)
|
418 |
+
# we do not want to group tokens when computing the metrics
|
419 |
+
label_str = tokenizer.batch_decode(pred.label_ids, skip_special_tokens=True)
|
420 |
+
|
421 |
+
wer = metric.compute(predictions=pred_str, references=label_str)
|
422 |
+
|
423 |
+
return {"wer": wer}
|
424 |
+
|
425 |
+
# 9. Create a single speech processor
|
426 |
+
if is_main_process(training_args.local_rank):
|
427 |
+
# save feature extractor, tokenizer and config
|
428 |
+
feature_extractor.save_pretrained(training_args.output_dir)
|
429 |
+
tokenizer.save_pretrained(training_args.output_dir)
|
430 |
+
config.save_pretrained(training_args.output_dir)
|
431 |
+
|
432 |
+
processor = AutoProcessor.from_pretrained(training_args.output_dir)
|
433 |
+
|
434 |
+
# 10. Define data collator
|
435 |
+
data_collator = DataCollatorSpeechSeq2SeqWithPadding(
|
436 |
+
processor=processor, decoder_start_token_id=model.config.decoder_start_token_id
|
437 |
+
)
|
438 |
+
|
439 |
+
decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
|
440 |
+
decay_parameters = [name for name in decay_parameters if "bias" not in name]
|
441 |
+
optimizer_grouped_parameters = [
|
442 |
+
{
|
443 |
+
"params": [p for n, p in model.named_parameters() if n in decay_parameters],
|
444 |
+
"weight_decay": training_args.weight_decay,
|
445 |
+
},
|
446 |
+
{
|
447 |
+
"params": [p for n, p in model.named_parameters() if n not in decay_parameters],
|
448 |
+
"weight_decay": 0.0,
|
449 |
+
},
|
450 |
+
]
|
451 |
+
|
452 |
+
optimizer = bnb.optim.Adam8bit(
|
453 |
+
params=optimizer_grouped_parameters,
|
454 |
+
lr=training_args.learning_rate,
|
455 |
+
betas=(training_args.adam_beta1, training_args.adam_beta2),
|
456 |
+
eps=training_args.adam_epsilon,
|
457 |
+
)
|
458 |
+
|
459 |
+
"""optimizer = Adafactor(
|
460 |
+
params=optimizer_grouped_parameters,
|
461 |
+
lr=training_args.learning_rate,
|
462 |
+
beta1=training_args.adam_beta1,
|
463 |
+
eps=training_args.adam_epsilon,
|
464 |
+
relative_step=False,
|
465 |
+
)"""
|
466 |
+
|
467 |
+
|
468 |
+
optimizers = (optimizer, None)
|
469 |
+
|
470 |
+
|
471 |
+
#11. Initialize Trainer
|
472 |
+
|
473 |
+
trainer = Seq2SeqTrainer(
|
474 |
+
model=model,
|
475 |
+
args=training_args,
|
476 |
+
train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
|
477 |
+
eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
|
478 |
+
tokenizer=feature_extractor,
|
479 |
+
data_collator=data_collator,
|
480 |
+
compute_metrics=compute_metrics if training_args.predict_with_generate else None,
|
481 |
+
optimizers=optimizers,
|
482 |
+
)
|
483 |
+
|
484 |
+
# 12. Training
|
485 |
+
if training_args.do_train:
|
486 |
+
checkpoint = None
|
487 |
+
if training_args.resume_from_checkpoint is not None:
|
488 |
+
checkpoint = training_args.resume_from_checkpoint
|
489 |
+
elif last_checkpoint is not None:
|
490 |
+
checkpoint = last_checkpoint
|
491 |
+
train_result = trainer.train(resume_from_checkpoint=checkpoint)
|
492 |
+
trainer.save_model() # Saves the feature extractor too for easy upload
|
493 |
+
|
494 |
+
metrics = train_result.metrics
|
495 |
+
max_train_samples = (
|
496 |
+
data_args.max_train_samples
|
497 |
+
if data_args.max_train_samples is not None
|
498 |
+
else len(vectorized_datasets["train"])
|
499 |
+
)
|
500 |
+
metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
|
501 |
+
trainer.log_metrics("train", metrics)
|
502 |
+
trainer.save_metrics("train", metrics)
|
503 |
+
trainer.save_state()
|
504 |
+
|
505 |
+
# 13. Evaluation
|
506 |
+
results = {}
|
507 |
+
if training_args.do_eval:
|
508 |
+
logger.info("*** Evaluate ***")
|
509 |
+
metrics = trainer.evaluate(
|
510 |
+
metric_key_prefix="eval", max_length=model.config.max_length, num_beams=model.config.num_beams
|
511 |
+
)
|
512 |
+
max_eval_samples = (
|
513 |
+
data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
|
514 |
+
)
|
515 |
+
metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
|
516 |
+
|
517 |
+
trainer.log_metrics("eval", metrics)
|
518 |
+
trainer.save_metrics("eval", metrics)
|
519 |
+
|
520 |
+
# 14. Write Training Stats
|
521 |
+
kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "speech recognition"}
|
522 |
+
if data_args.dataset_name is not None:
|
523 |
+
kwargs["dataset_tags"] = data_args.dataset_name
|
524 |
+
if data_args.dataset_config_name is not None:
|
525 |
+
kwargs["dataset_args"] = data_args.dataset_config_name
|
526 |
+
kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
|
527 |
+
else:
|
528 |
+
kwargs["dataset"] = data_args.dataset_name
|
529 |
+
|
530 |
+
if training_args.push_to_hub:
|
531 |
+
trainer.push_to_hub(**kwargs)
|
532 |
+
else:
|
533 |
+
trainer.create_model_card(**kwargs)
|
534 |
+
|
535 |
+
return results
|
536 |
+
|
537 |
+
|
538 |
+
if __name__ == "__main__":
|
539 |
+
main()
|
runs/Feb23_16-14-28_sanchit--v100/1645632899.7738805/events.out.tfevents.1645632899.sanchit--v100.125825.1
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:884b404003673c14b75e877f8d5dcfcb2e1f40260dc2287325d6830919a2b862
|
3 |
+
size 4964
|
runs/Feb23_16-14-28_sanchit--v100/events.out.tfevents.1645632899.sanchit--v100.125825.0
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:85e96986a472c5528b3cea6f4ca20d22c7101ab142f2ff6b2f914a9e54884862
|
3 |
+
size 10277
|
runs/Feb23_16-16-27_sanchit--v100/1645633019.095641/events.out.tfevents.1645633019.sanchit--v100.126289.1
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:35a833766a4ed973cfe30234ffc2a806b3f3681fe2018b3bc688d907289c8b24
|
3 |
+
size 4964
|
runs/Feb23_16-16-27_sanchit--v100/events.out.tfevents.1645633019.sanchit--v100.126289.0
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:ade303cca69b2ddee63da896a024ceca749734225d6fce24386547afb6ecf0ae
|
3 |
+
size 87325
|
special_tokens_map.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]"}
|
tokenizer.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
tokenizer_config.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"do_lower_case": true, "unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]", "tokenize_chinese_chars": true, "strip_accents": null, "model_max_length": 512, "special_tokens_map_file": null, "name_or_path": "./", "tokenizer_class": "BertTokenizer"}
|
training_args.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:2663f1bf9edea3ff69acdf258bbcaf21d944625d57894f89a589487fbc45ebb4
|
3 |
+
size 3119
|
vocab.txt
ADDED
The diff for this file is too large to render.
See raw diff
|
|