Commit
•
12a385e
1
Parent(s):
1d2bded
Training in progress, step 500
Browse files- .gitattributes +1 -0
- .gitignore +1 -0
- config.json +265 -0
- create_model.py +58 -0
- merges.txt +0 -0
- preprocessor_config.json +9 -0
- pytorch_model.bin +3 -0
- run_grid_search.sh +50 -0
- run_librispeech.sh +36 -0
- run_speech_recognition_seq2seq.py +539 -0
- special_tokens_map.json +1 -0
- tokenizer.json +0 -0
- tokenizer_config.json +1 -0
- training_args.bin +3 -0
- vocab.json +0 -0
- wandb/debug-internal.log +1 -0
- wandb/debug.log +1 -0
- wandb/latest-run +1 -0
- wandb/run-20220228_093705-yn2gmwrw/files/config.yaml +0 -0
- wandb/run-20220228_093705-yn2gmwrw/files/output.log +0 -0
- wandb/run-20220228_093705-yn2gmwrw/files/requirements.txt +184 -0
- wandb/run-20220228_093705-yn2gmwrw/files/wandb-metadata.json +60 -0
- wandb/run-20220228_093705-yn2gmwrw/files/wandb-summary.json +0 -0
- wandb/run-20220228_093705-yn2gmwrw/logs/debug-internal.log +0 -0
- wandb/run-20220228_093705-yn2gmwrw/logs/debug.log +27 -0
- wandb/run-20220228_093705-yn2gmwrw/run-yn2gmwrw.wandb +3 -0
.gitattributes
CHANGED
@@ -25,3 +25,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
25 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
26 |
*.zstandard filter=lfs diff=lfs merge=lfs -text
|
27 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
25 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
26 |
*.zstandard filter=lfs diff=lfs merge=lfs -text
|
27 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
28 |
+
wandb/run-20220228_093705-yn2gmwrw/run-yn2gmwrw.wandb filter=lfs diff=lfs merge=lfs -text
|
.gitignore
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
checkpoint-*/
|
config.json
ADDED
@@ -0,0 +1,265 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "./",
|
3 |
+
"architectures": [
|
4 |
+
"SpeechEncoderDecoderModel"
|
5 |
+
],
|
6 |
+
"decoder": {
|
7 |
+
"_name_or_path": "gpt2-medium",
|
8 |
+
"activation_function": "gelu_new",
|
9 |
+
"add_cross_attention": true,
|
10 |
+
"architectures": [
|
11 |
+
"GPT2LMHeadModel"
|
12 |
+
],
|
13 |
+
"attn_pdrop": 0.0,
|
14 |
+
"bad_words_ids": null,
|
15 |
+
"bos_token_id": 50256,
|
16 |
+
"chunk_size_feed_forward": 0,
|
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 |
+
"embd_pdrop": 0.0,
|
23 |
+
"encoder_no_repeat_ngram_size": 0,
|
24 |
+
"eos_token_id": 50256,
|
25 |
+
"finetuning_task": null,
|
26 |
+
"forced_bos_token_id": null,
|
27 |
+
"forced_eos_token_id": null,
|
28 |
+
"id2label": {
|
29 |
+
"0": "LABEL_0",
|
30 |
+
"1": "LABEL_1"
|
31 |
+
},
|
32 |
+
"initializer_range": 0.02,
|
33 |
+
"is_decoder": true,
|
34 |
+
"is_encoder_decoder": false,
|
35 |
+
"label2id": {
|
36 |
+
"LABEL_0": 0,
|
37 |
+
"LABEL_1": 1
|
38 |
+
},
|
39 |
+
"layer_norm_epsilon": 1e-05,
|
40 |
+
"length_penalty": 1.0,
|
41 |
+
"max_length": 20,
|
42 |
+
"min_length": 0,
|
43 |
+
"model_type": "gpt2",
|
44 |
+
"n_ctx": 1024,
|
45 |
+
"n_embd": 1024,
|
46 |
+
"n_head": 16,
|
47 |
+
"n_inner": null,
|
48 |
+
"n_layer": 24,
|
49 |
+
"n_positions": 1024,
|
50 |
+
"n_special": 0,
|
51 |
+
"no_repeat_ngram_size": 0,
|
52 |
+
"num_beam_groups": 1,
|
53 |
+
"num_beams": 1,
|
54 |
+
"num_return_sequences": 1,
|
55 |
+
"output_attentions": false,
|
56 |
+
"output_hidden_states": false,
|
57 |
+
"output_scores": false,
|
58 |
+
"pad_token_id": null,
|
59 |
+
"predict_special_tokens": true,
|
60 |
+
"prefix": null,
|
61 |
+
"problem_type": null,
|
62 |
+
"pruned_heads": {},
|
63 |
+
"remove_invalid_values": false,
|
64 |
+
"reorder_and_upcast_attn": false,
|
65 |
+
"repetition_penalty": 1.0,
|
66 |
+
"resid_pdrop": 0.0,
|
67 |
+
"return_dict": true,
|
68 |
+
"return_dict_in_generate": false,
|
69 |
+
"scale_attn_by_inverse_layer_idx": false,
|
70 |
+
"scale_attn_weights": true,
|
71 |
+
"sep_token_id": null,
|
72 |
+
"summary_activation": null,
|
73 |
+
"summary_first_dropout": 0.0,
|
74 |
+
"summary_proj_to_labels": true,
|
75 |
+
"summary_type": "cls_index",
|
76 |
+
"summary_use_proj": true,
|
77 |
+
"task_specific_params": {
|
78 |
+
"text-generation": {
|
79 |
+
"do_sample": true,
|
80 |
+
"max_length": 50
|
81 |
+
}
|
82 |
+
},
|
83 |
+
"temperature": 1.0,
|
84 |
+
"tie_encoder_decoder": false,
|
85 |
+
"tie_word_embeddings": true,
|
86 |
+
"tokenizer_class": null,
|
87 |
+
"top_k": 50,
|
88 |
+
"top_p": 1.0,
|
89 |
+
"torch_dtype": null,
|
90 |
+
"torchscript": false,
|
91 |
+
"transformers_version": "4.17.0.dev0",
|
92 |
+
"use_bfloat16": false,
|
93 |
+
"use_cache": false,
|
94 |
+
"vocab_size": 50257
|
95 |
+
},
|
96 |
+
"decoder_start_token_id": 50256,
|
97 |
+
"encoder": {
|
98 |
+
"_name_or_path": "facebook/wav2vec2-large-lv60",
|
99 |
+
"activation_dropout": 0.0,
|
100 |
+
"adapter_kernel_size": 3,
|
101 |
+
"adapter_stride": 2,
|
102 |
+
"add_adapter": true,
|
103 |
+
"add_cross_attention": false,
|
104 |
+
"apply_spec_augment": false,
|
105 |
+
"architectures": [
|
106 |
+
"Wav2Vec2ForPreTraining"
|
107 |
+
],
|
108 |
+
"attention_dropout": 0.0,
|
109 |
+
"bad_words_ids": null,
|
110 |
+
"bos_token_id": 1,
|
111 |
+
"chunk_size_feed_forward": 0,
|
112 |
+
"classifier_proj_size": 256,
|
113 |
+
"codevector_dim": 768,
|
114 |
+
"contrastive_logits_temperature": 0.1,
|
115 |
+
"conv_bias": true,
|
116 |
+
"conv_dim": [
|
117 |
+
512,
|
118 |
+
512,
|
119 |
+
512,
|
120 |
+
512,
|
121 |
+
512,
|
122 |
+
512,
|
123 |
+
512
|
124 |
+
],
|
125 |
+
"conv_kernel": [
|
126 |
+
10,
|
127 |
+
3,
|
128 |
+
3,
|
129 |
+
3,
|
130 |
+
3,
|
131 |
+
2,
|
132 |
+
2
|
133 |
+
],
|
134 |
+
"conv_stride": [
|
135 |
+
5,
|
136 |
+
2,
|
137 |
+
2,
|
138 |
+
2,
|
139 |
+
2,
|
140 |
+
2,
|
141 |
+
2
|
142 |
+
],
|
143 |
+
"cross_attention_hidden_size": null,
|
144 |
+
"ctc_loss_reduction": "sum",
|
145 |
+
"ctc_zero_infinity": false,
|
146 |
+
"decoder_start_token_id": null,
|
147 |
+
"diversity_loss_weight": 0.1,
|
148 |
+
"diversity_penalty": 0.0,
|
149 |
+
"do_sample": false,
|
150 |
+
"do_stable_layer_norm": true,
|
151 |
+
"early_stopping": false,
|
152 |
+
"encoder_no_repeat_ngram_size": 0,
|
153 |
+
"eos_token_id": 2,
|
154 |
+
"feat_extract_activation": "gelu",
|
155 |
+
"feat_extract_dropout": 0.0,
|
156 |
+
"feat_extract_norm": "layer",
|
157 |
+
"feat_proj_dropout": 0.0,
|
158 |
+
"feat_quantizer_dropout": 0.0,
|
159 |
+
"final_dropout": 0.0,
|
160 |
+
"finetuning_task": null,
|
161 |
+
"forced_bos_token_id": null,
|
162 |
+
"forced_eos_token_id": null,
|
163 |
+
"gradient_checkpointing": false,
|
164 |
+
"hidden_act": "gelu",
|
165 |
+
"hidden_dropout": 0.0,
|
166 |
+
"hidden_dropout_prob": 0.0,
|
167 |
+
"hidden_size": 1024,
|
168 |
+
"id2label": {
|
169 |
+
"0": "LABEL_0",
|
170 |
+
"1": "LABEL_1"
|
171 |
+
},
|
172 |
+
"initializer_range": 0.02,
|
173 |
+
"intermediate_size": 4096,
|
174 |
+
"is_decoder": false,
|
175 |
+
"is_encoder_decoder": false,
|
176 |
+
"label2id": {
|
177 |
+
"LABEL_0": 0,
|
178 |
+
"LABEL_1": 1
|
179 |
+
},
|
180 |
+
"layer_norm_eps": 1e-05,
|
181 |
+
"layerdrop": 0.0,
|
182 |
+
"length_penalty": 1.0,
|
183 |
+
"mask_feature_length": 10,
|
184 |
+
"mask_feature_min_masks": 0,
|
185 |
+
"mask_feature_prob": 0.0,
|
186 |
+
"mask_time_length": 10,
|
187 |
+
"mask_time_min_masks": 2,
|
188 |
+
"mask_time_prob": 0.0,
|
189 |
+
"max_length": 20,
|
190 |
+
"min_length": 0,
|
191 |
+
"model_type": "wav2vec2",
|
192 |
+
"no_repeat_ngram_size": 0,
|
193 |
+
"num_adapter_layers": 3,
|
194 |
+
"num_attention_heads": 16,
|
195 |
+
"num_beam_groups": 1,
|
196 |
+
"num_beams": 1,
|
197 |
+
"num_codevector_groups": 2,
|
198 |
+
"num_codevectors_per_group": 320,
|
199 |
+
"num_conv_pos_embedding_groups": 16,
|
200 |
+
"num_conv_pos_embeddings": 128,
|
201 |
+
"num_feat_extract_layers": 7,
|
202 |
+
"num_hidden_layers": 24,
|
203 |
+
"num_negatives": 100,
|
204 |
+
"num_return_sequences": 1,
|
205 |
+
"output_attentions": false,
|
206 |
+
"output_hidden_size": 1024,
|
207 |
+
"output_hidden_states": false,
|
208 |
+
"output_scores": false,
|
209 |
+
"pad_token_id": 0,
|
210 |
+
"prefix": null,
|
211 |
+
"problem_type": null,
|
212 |
+
"proj_codevector_dim": 768,
|
213 |
+
"pruned_heads": {},
|
214 |
+
"remove_invalid_values": false,
|
215 |
+
"repetition_penalty": 1.0,
|
216 |
+
"return_dict": true,
|
217 |
+
"return_dict_in_generate": false,
|
218 |
+
"sep_token_id": null,
|
219 |
+
"task_specific_params": null,
|
220 |
+
"tdnn_dilation": [
|
221 |
+
1,
|
222 |
+
2,
|
223 |
+
3,
|
224 |
+
1,
|
225 |
+
1
|
226 |
+
],
|
227 |
+
"tdnn_dim": [
|
228 |
+
512,
|
229 |
+
512,
|
230 |
+
512,
|
231 |
+
512,
|
232 |
+
1500
|
233 |
+
],
|
234 |
+
"tdnn_kernel": [
|
235 |
+
5,
|
236 |
+
3,
|
237 |
+
3,
|
238 |
+
1,
|
239 |
+
1
|
240 |
+
],
|
241 |
+
"temperature": 1.0,
|
242 |
+
"tie_encoder_decoder": false,
|
243 |
+
"tie_word_embeddings": true,
|
244 |
+
"tokenizer_class": null,
|
245 |
+
"top_k": 50,
|
246 |
+
"top_p": 1.0,
|
247 |
+
"torch_dtype": null,
|
248 |
+
"torchscript": false,
|
249 |
+
"transformers_version": "4.17.0.dev0",
|
250 |
+
"use_bfloat16": false,
|
251 |
+
"use_weighted_layer_sum": false,
|
252 |
+
"vocab_size": 32,
|
253 |
+
"xvector_output_dim": 512
|
254 |
+
},
|
255 |
+
"eos_token_id": 50256,
|
256 |
+
"is_encoder_decoder": true,
|
257 |
+
"max_length": 50,
|
258 |
+
"model_type": "speech-encoder-decoder",
|
259 |
+
"pad_token_id": 50256,
|
260 |
+
"processor_class": "Wav2Vec2Processor",
|
261 |
+
"tie_word_embeddings": false,
|
262 |
+
"torch_dtype": "float32",
|
263 |
+
"transformers_version": null,
|
264 |
+
"use_cache": false
|
265 |
+
}
|
create_model.py
ADDED
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import SpeechEncoderDecoderModel, AutoFeatureExtractor, GPT2Tokenizer
|
2 |
+
import torch
|
3 |
+
|
4 |
+
# checkpoints to leverage
|
5 |
+
encoder_id = "facebook/wav2vec2-large-lv60"
|
6 |
+
decoder_id = "gpt2-medium"
|
7 |
+
|
8 |
+
model = SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(encoder_id, decoder_id, encoder_add_adapter=True)
|
9 |
+
|
10 |
+
# set all encoder regularisation to zero
|
11 |
+
model.config.encoder.feat_proj_dropout = 0.0
|
12 |
+
model.config.encoder.final_dropout = 0.0
|
13 |
+
model.config.encoder.activation_dropout = 0.0
|
14 |
+
model.config.encoder.apply_spec_augment = False
|
15 |
+
model.config.encoder.attention_dropout = 0.0
|
16 |
+
model.config.encoder.feat_extract_dropout = 0.0
|
17 |
+
model.config.encoder.feat_proj_dropout = 0.0
|
18 |
+
model.config.encoder.hidden_dropout = 0.0
|
19 |
+
model.config.encoder.hidden_dropout_prob = 0.0
|
20 |
+
model.config.encoder.layerdrop = 0.0
|
21 |
+
model.config.encoder.mask_feature_prob = 0.0
|
22 |
+
model.config.encoder.mask_time_prob = 0.0
|
23 |
+
|
24 |
+
# set all decoder regularisation to zero
|
25 |
+
model.config.decoder.attn_pdrop = 0.0
|
26 |
+
model.config.decoder.embd_pdrop = 0.0
|
27 |
+
model.config.decoder.resid_pdrop = 0.0
|
28 |
+
model.config.decoder.summary_first_dropout = 0.0
|
29 |
+
|
30 |
+
# force GPT2 to append EOS to begin and end of seq
|
31 |
+
def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
|
32 |
+
outputs = [self.bos_token_id] + token_ids_0 + [self.eos_token_id]
|
33 |
+
return outputs
|
34 |
+
|
35 |
+
GPT2Tokenizer.build_inputs_with_special_tokens = build_inputs_with_special_tokens
|
36 |
+
gpt2_tokenizer = GPT2Tokenizer.from_pretrained(decoder_id)
|
37 |
+
# set pad_token_id to unk_token_id, note: unk_token_id == eos_token_id == bos_token_id
|
38 |
+
gpt2_tokenizer.pad_token = gpt2_tokenizer.unk_token
|
39 |
+
gpt2_tokenizer.save_pretrained("./")
|
40 |
+
|
41 |
+
model.config.pad_token_id = gpt2_tokenizer.pad_token_id
|
42 |
+
model.config.decoder_start_token_id = model.decoder.config.bos_token_id
|
43 |
+
model.config.eos_token_id = model.decoder.config.eos_token_id
|
44 |
+
model.config.max_length = 50
|
45 |
+
model.config.num_beams = 1
|
46 |
+
|
47 |
+
model.config.use_cache = False
|
48 |
+
model.config.decoder.use_cache = False
|
49 |
+
model.config.processor_class = "Wav2Vec2Processor"
|
50 |
+
|
51 |
+
# check if generation works
|
52 |
+
out = model.generate(torch.ones((1, 2000)))
|
53 |
+
|
54 |
+
model.save_pretrained("./")
|
55 |
+
|
56 |
+
feature_extractor = AutoFeatureExtractor.from_pretrained(encoder_id)
|
57 |
+
feature_extractor.save_pretrained("./")
|
58 |
+
|
merges.txt
ADDED
The diff for this file is too large to render.
See raw diff
|
|
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 sha256:80dae0dc0d3975dc2f652e3e057ef9899f203485ebd52fa73614213714c5dcde
|
3 |
+
size 3210531882
|
run_grid_search.sh
ADDED
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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 "${gradient_accumulation_step_sizes[@]}"; do
|
10 |
+
python create_model.py
|
11 |
+
CUDA_VISIBLE_DEVICES=0 python run_speech_recognition_seq2seq.py \
|
12 |
+
--dataset_name="librispeech_asr" \
|
13 |
+
--model_name_or_path="./" \
|
14 |
+
--tokenizer_name="./" \
|
15 |
+
--dataset_config_name="clean" \
|
16 |
+
--train_split_name="train.100" \
|
17 |
+
--eval_split_name="validation" \
|
18 |
+
--output_dir="./" \
|
19 |
+
--preprocessing_num_workers="1" \
|
20 |
+
--length_column_name="input_length" \
|
21 |
+
--overwrite_output_dir \
|
22 |
+
--num_train_epochs="1" \
|
23 |
+
--per_device_train_batch_size=$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 |
+
--report_to="wandb" \
|
45 |
+
--push_to_hub \
|
46 |
+
--use_auth_token
|
47 |
+
done
|
48 |
+
done
|
49 |
+
done
|
50 |
+
|
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()
|
special_tokens_map.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"bos_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "eos_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "unk_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "pad_token": "<|endoftext|>"}
|
tokenizer.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
tokenizer_config.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"unk_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "bos_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "eos_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "add_prefix_space": false, "errors": "replace", "model_max_length": 1024, "special_tokens_map_file": null, "name_or_path": "./", "tokenizer_class": "GPT2Tokenizer"}
|
training_args.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:bdbb29101684863ff261d24c35a1939e62b99d8c9255d77d06750453f301290b
|
3 |
+
size 3119
|
vocab.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
wandb/debug-internal.log
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
run-20220228_093705-yn2gmwrw/logs/debug-internal.log
|
wandb/debug.log
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
run-20220228_093705-yn2gmwrw/logs/debug.log
|
wandb/latest-run
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
run-20220228_093705-yn2gmwrw
|
wandb/run-20220228_093705-yn2gmwrw/files/config.yaml
ADDED
The diff for this file is too large to render.
See raw diff
|
|
wandb/run-20220228_093705-yn2gmwrw/files/output.log
ADDED
The diff for this file is too large to render.
See raw diff
|
|
wandb/run-20220228_093705-yn2gmwrw/files/requirements.txt
ADDED
@@ -0,0 +1,184 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
absl-py==1.0.0
|
2 |
+
aiohttp==3.8.1
|
3 |
+
aiosignal==1.2.0
|
4 |
+
anyio==3.5.0
|
5 |
+
appdirs==1.4.4
|
6 |
+
argon2-cffi-bindings==21.2.0
|
7 |
+
argon2-cffi==21.3.0
|
8 |
+
asttokens==2.0.5
|
9 |
+
async-timeout==4.0.2
|
10 |
+
attrs==21.4.0
|
11 |
+
audioread==2.1.9
|
12 |
+
babel==2.9.1
|
13 |
+
backcall==0.2.0
|
14 |
+
bitsandbytes-cuda113==0.26.0
|
15 |
+
black==22.1.0
|
16 |
+
bleach==4.1.0
|
17 |
+
cachetools==5.0.0
|
18 |
+
certifi==2021.10.8
|
19 |
+
cffi==1.15.0
|
20 |
+
charset-normalizer==2.0.11
|
21 |
+
chex==0.1.0
|
22 |
+
click==8.0.3
|
23 |
+
clldutils==3.10.1
|
24 |
+
colorlog==6.6.0
|
25 |
+
csvw==1.11.0
|
26 |
+
cycler==0.11.0
|
27 |
+
datasets==1.18.3
|
28 |
+
debugpy==1.5.1
|
29 |
+
decorator==5.1.1
|
30 |
+
defusedxml==0.7.1
|
31 |
+
dill==0.3.4
|
32 |
+
dlinfo==1.2.1
|
33 |
+
dm-tree==0.1.6
|
34 |
+
docker-pycreds==0.4.0
|
35 |
+
entrypoints==0.4
|
36 |
+
executing==0.8.2
|
37 |
+
filelock==3.4.2
|
38 |
+
flatbuffers==2.0
|
39 |
+
flax==0.4.0
|
40 |
+
fonttools==4.29.1
|
41 |
+
frozenlist==1.3.0
|
42 |
+
fsspec==2022.1.0
|
43 |
+
gitdb==4.0.9
|
44 |
+
gitpython==3.1.27
|
45 |
+
google-auth-oauthlib==0.4.6
|
46 |
+
google-auth==2.6.0
|
47 |
+
grpcio==1.43.0
|
48 |
+
huggingface-hub==0.4.0
|
49 |
+
hypothesis==6.36.1
|
50 |
+
idna==3.3
|
51 |
+
importlib-metadata==4.10.1
|
52 |
+
ipykernel==6.8.0
|
53 |
+
ipython-genutils==0.2.0
|
54 |
+
ipython==8.0.1
|
55 |
+
ipywidgets==7.6.5
|
56 |
+
isodate==0.6.1
|
57 |
+
jax==0.2.28
|
58 |
+
jaxlib==0.1.76+cuda11.cudnn82
|
59 |
+
jedi==0.18.1
|
60 |
+
jinja2==3.0.3
|
61 |
+
jiwer==2.3.0
|
62 |
+
joblib==1.1.0
|
63 |
+
json5==0.9.6
|
64 |
+
jsonschema==4.4.0
|
65 |
+
jupyter-client==7.1.2
|
66 |
+
jupyter-console==6.4.0
|
67 |
+
jupyter-core==4.9.1
|
68 |
+
jupyter-server==1.13.5
|
69 |
+
jupyter==1.0.0
|
70 |
+
jupyterlab-pygments==0.1.2
|
71 |
+
jupyterlab-server==2.10.3
|
72 |
+
jupyterlab-widgets==1.0.2
|
73 |
+
jupyterlab==3.2.9
|
74 |
+
kiwisolver==1.3.2
|
75 |
+
librosa==0.8.1
|
76 |
+
llvmlite==0.38.0
|
77 |
+
markdown==3.3.6
|
78 |
+
markupsafe==2.0.1
|
79 |
+
matplotlib-inline==0.1.3
|
80 |
+
matplotlib==3.5.1
|
81 |
+
mistune==0.8.4
|
82 |
+
msgpack==1.0.3
|
83 |
+
multidict==6.0.2
|
84 |
+
multiprocess==0.70.12.2
|
85 |
+
mypy-extensions==0.4.3
|
86 |
+
nbclassic==0.3.5
|
87 |
+
nbclient==0.5.10
|
88 |
+
nbconvert==6.4.1
|
89 |
+
nbformat==5.1.3
|
90 |
+
nest-asyncio==1.5.4
|
91 |
+
notebook==6.4.8
|
92 |
+
numba==0.55.1
|
93 |
+
numpy==1.21.5
|
94 |
+
oauthlib==3.2.0
|
95 |
+
opt-einsum==3.3.0
|
96 |
+
optax==0.1.0
|
97 |
+
packaging==21.3
|
98 |
+
pandas==1.4.0
|
99 |
+
pandocfilters==1.5.0
|
100 |
+
parso==0.8.3
|
101 |
+
pathspec==0.9.0
|
102 |
+
pathtools==0.1.2
|
103 |
+
pexpect==4.8.0
|
104 |
+
phonemizer==3.0.1
|
105 |
+
pickleshare==0.7.5
|
106 |
+
pillow==9.0.0
|
107 |
+
pip==22.0.2
|
108 |
+
pkg-resources==0.0.0
|
109 |
+
platformdirs==2.4.1
|
110 |
+
pooch==1.6.0
|
111 |
+
prometheus-client==0.13.1
|
112 |
+
promise==2.3
|
113 |
+
prompt-toolkit==3.0.26
|
114 |
+
protobuf==3.19.4
|
115 |
+
psutil==5.9.0
|
116 |
+
ptyprocess==0.7.0
|
117 |
+
pure-eval==0.2.2
|
118 |
+
pyarrow==6.0.1
|
119 |
+
pyasn1-modules==0.2.8
|
120 |
+
pyasn1==0.4.8
|
121 |
+
pycparser==2.21
|
122 |
+
pyctcdecode==0.3.0
|
123 |
+
pygments==2.11.2
|
124 |
+
pygtrie==2.4.2
|
125 |
+
pyparsing==3.0.7
|
126 |
+
pyrsistent==0.18.1
|
127 |
+
python-dateutil==2.8.2
|
128 |
+
python-levenshtein==0.12.2
|
129 |
+
pytz==2021.3
|
130 |
+
pyyaml==6.0
|
131 |
+
pyzmq==22.3.0
|
132 |
+
qtconsole==5.2.2
|
133 |
+
qtpy==2.0.1
|
134 |
+
regex==2022.1.18
|
135 |
+
requests-oauthlib==1.3.1
|
136 |
+
requests==2.27.1
|
137 |
+
resampy==0.2.2
|
138 |
+
rfc3986==2.0.0
|
139 |
+
rsa==4.8
|
140 |
+
sacremoses==0.0.47
|
141 |
+
scikit-learn==1.0.2
|
142 |
+
scipy==1.7.3
|
143 |
+
segments==2.2.0
|
144 |
+
send2trash==1.8.0
|
145 |
+
sentry-sdk==1.5.6
|
146 |
+
setuptools==44.1.1
|
147 |
+
shortuuid==1.0.8
|
148 |
+
six==1.16.0
|
149 |
+
smmap==5.0.0
|
150 |
+
sniffio==1.2.0
|
151 |
+
sortedcontainers==2.4.0
|
152 |
+
soundfile==0.10.3.post1
|
153 |
+
stack-data==0.1.4
|
154 |
+
tabulate==0.8.9
|
155 |
+
tensorboard-data-server==0.6.1
|
156 |
+
tensorboard-plugin-wit==1.8.1
|
157 |
+
tensorboard==2.8.0
|
158 |
+
termcolor==1.1.0
|
159 |
+
terminado==0.13.1
|
160 |
+
testpath==0.5.0
|
161 |
+
threadpoolctl==3.1.0
|
162 |
+
tokenizers==0.11.4
|
163 |
+
tomli==2.0.0
|
164 |
+
toolz==0.11.2
|
165 |
+
torch==1.10.2+cu113
|
166 |
+
torchaudio==0.10.2+cu113
|
167 |
+
tornado==6.1
|
168 |
+
tqdm==4.62.3
|
169 |
+
traitlets==5.1.1
|
170 |
+
transformers==4.17.0.dev0
|
171 |
+
typing-extensions==3.10.0.2
|
172 |
+
uritemplate==4.1.1
|
173 |
+
urllib3==1.26.8
|
174 |
+
wandb==0.12.10
|
175 |
+
wcwidth==0.2.5
|
176 |
+
webencodings==0.5.1
|
177 |
+
websocket-client==1.2.3
|
178 |
+
werkzeug==2.0.2
|
179 |
+
wheel==0.37.1
|
180 |
+
widgetsnbextension==3.5.2
|
181 |
+
xxhash==2.0.2
|
182 |
+
yarl==1.7.2
|
183 |
+
yaspin==2.1.0
|
184 |
+
zipp==3.7.0
|
wandb/run-20220228_093705-yn2gmwrw/files/wandb-metadata.json
ADDED
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"os": "Linux-5.11.0-1028-gcp-x86_64-with-glibc2.33",
|
3 |
+
"python": "3.9.5",
|
4 |
+
"heartbeatAt": "2022-02-28T09:37:06.811386",
|
5 |
+
"startedAt": "2022-02-28T09:37:05.566873",
|
6 |
+
"docker": null,
|
7 |
+
"gpu": "Tesla V100-SXM2-16GB",
|
8 |
+
"gpu_count": 2,
|
9 |
+
"cpu_count": 16,
|
10 |
+
"cuda": null,
|
11 |
+
"args": [
|
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=8",
|
24 |
+
"--per_device_eval_batch_size=8",
|
25 |
+
"--gradient_accumulation_steps=2",
|
26 |
+
"--generation_max_length=40",
|
27 |
+
"--generation_num_beams=1",
|
28 |
+
"--learning_rate=1e-5",
|
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 |
+
"--report_to=wandb",
|
45 |
+
"--push_to_hub",
|
46 |
+
"--use_auth_token"
|
47 |
+
],
|
48 |
+
"state": "running",
|
49 |
+
"program": "/home/sanchit_huggingface_co/wav2vec2-gpt2-wandb-grid-search/run_speech_recognition_seq2seq.py",
|
50 |
+
"codePath": "run_speech_recognition_seq2seq.py",
|
51 |
+
"git": {
|
52 |
+
"remote": "https://huggingface.co/sanchit-gandhi/wav2vec2-gpt2-wandb-grid-search",
|
53 |
+
"commit": "1d2bded5b2a9a272e2b53e924e0812d0636afc84"
|
54 |
+
},
|
55 |
+
"email": "sanchit@huggingface.co",
|
56 |
+
"root": "/home/sanchit_huggingface_co/wav2vec2-gpt2-wandb-grid-search",
|
57 |
+
"host": "sanchit--v100",
|
58 |
+
"username": "sanchit_huggingface_co",
|
59 |
+
"executable": "/home/sanchit_huggingface_co/gcp/bin/python"
|
60 |
+
}
|
wandb/run-20220228_093705-yn2gmwrw/files/wandb-summary.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
wandb/run-20220228_093705-yn2gmwrw/logs/debug-internal.log
ADDED
The diff for this file is too large to render.
See raw diff
|
|
wandb/run-20220228_093705-yn2gmwrw/logs/debug.log
ADDED
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
2022-02-28 09:37:05,568 INFO MainThread:219058 [wandb_setup.py:_flush():75] Loading settings from /home/sanchit_huggingface_co/.config/wandb/settings
|
2 |
+
2022-02-28 09:37:05,569 INFO MainThread:219058 [wandb_setup.py:_flush():75] Loading settings from /home/sanchit_huggingface_co/wav2vec2-gpt2-wandb-grid-search/wandb/settings
|
3 |
+
2022-02-28 09:37:05,569 INFO MainThread:219058 [wandb_setup.py:_flush():75] Loading settings from environment variables: {}
|
4 |
+
2022-02-28 09:37:05,569 INFO MainThread:219058 [wandb_setup.py:_flush():75] Inferring run settings from compute environment: {'program_relpath': 'run_speech_recognition_seq2seq.py', 'program': '/home/sanchit_huggingface_co/wav2vec2-gpt2-wandb-grid-search/run_speech_recognition_seq2seq.py'}
|
5 |
+
2022-02-28 09:37:05,569 INFO MainThread:219058 [wandb_setup.py:_flush():75] Applying login settings: {'api_key': '***REDACTED***'}
|
6 |
+
2022-02-28 09:37:05,569 INFO MainThread:219058 [wandb_init.py:_log_setup():386] Logging user logs to /home/sanchit_huggingface_co/wav2vec2-gpt2-wandb-grid-search/wandb/run-20220228_093705-yn2gmwrw/logs/debug.log
|
7 |
+
2022-02-28 09:37:05,569 INFO MainThread:219058 [wandb_init.py:_log_setup():387] Logging internal logs to /home/sanchit_huggingface_co/wav2vec2-gpt2-wandb-grid-search/wandb/run-20220228_093705-yn2gmwrw/logs/debug-internal.log
|
8 |
+
2022-02-28 09:37:05,569 INFO MainThread:219058 [wandb_init.py:init():420] calling init triggers
|
9 |
+
2022-02-28 09:37:05,569 INFO MainThread:219058 [wandb_init.py:init():425] wandb.init called with sweep_config: {}
|
10 |
+
config: {}
|
11 |
+
2022-02-28 09:37:05,569 INFO MainThread:219058 [wandb_init.py:init():471] starting backend
|
12 |
+
2022-02-28 09:37:05,569 INFO MainThread:219058 [backend.py:_multiprocessing_setup():99] multiprocessing start_methods=fork,spawn,forkserver, using: spawn
|
13 |
+
2022-02-28 09:37:05,647 INFO MainThread:219058 [backend.py:ensure_launched():219] starting backend process...
|
14 |
+
2022-02-28 09:37:05,722 INFO MainThread:219058 [backend.py:ensure_launched():224] started backend process with pid: 219340
|
15 |
+
2022-02-28 09:37:05,724 INFO MainThread:219058 [wandb_init.py:init():480] backend started and connected
|
16 |
+
2022-02-28 09:37:05,736 INFO MainThread:219058 [wandb_init.py:init():550] updated telemetry
|
17 |
+
2022-02-28 09:37:05,917 INFO MainThread:219058 [wandb_init.py:init():581] communicating current version
|
18 |
+
2022-02-28 09:37:06,640 INFO MainThread:219058 [wandb_init.py:init():586] got version response
|
19 |
+
2022-02-28 09:37:06,640 INFO MainThread:219058 [wandb_init.py:init():596] communicating run to backend with 30 second timeout
|
20 |
+
2022-02-28 09:37:06,797 INFO MainThread:219058 [wandb_init.py:init():624] starting run threads in backend
|
21 |
+
2022-02-28 09:37:06,922 INFO MainThread:219058 [wandb_run.py:_console_start():1827] atexit reg
|
22 |
+
2022-02-28 09:37:06,923 INFO MainThread:219058 [wandb_run.py:_redirect():1701] redirect: SettingsConsole.REDIRECT
|
23 |
+
2022-02-28 09:37:06,923 INFO MainThread:219058 [wandb_run.py:_redirect():1706] Redirecting console.
|
24 |
+
2022-02-28 09:37:06,925 INFO MainThread:219058 [wandb_run.py:_redirect():1762] Redirects installed.
|
25 |
+
2022-02-28 09:37:06,925 INFO MainThread:219058 [wandb_init.py:init():651] run started, returning control to user process
|
26 |
+
2022-02-28 09:37:06,927 INFO MainThread:219058 [wandb_run.py:_config_callback():966] config_cb None None {'return_dict': True, 'output_hidden_states': False, 'output_attentions': False, 'torchscript': False, 'torch_dtype': 'torch.float32', 'use_bfloat16': False, 'pruned_heads': {}, 'tie_word_embeddings': False, 'is_encoder_decoder': True, 'is_decoder': False, 'cross_attention_hidden_size': None, 'add_cross_attention': False, 'tie_encoder_decoder': False, 'max_length': 50, 'min_length': 0, 'do_sample': False, 'early_stopping': False, 'num_beams': 1, 'num_beam_groups': 1, 'diversity_penalty': 0.0, 'temperature': 1.0, 'top_k': 50, 'top_p': 1.0, 'repetition_penalty': 1.0, 'length_penalty': 1.0, 'no_repeat_ngram_size': 0, 'encoder_no_repeat_ngram_size': 0, 'bad_words_ids': None, 'num_return_sequences': 1, 'chunk_size_feed_forward': 0, 'output_scores': False, 'return_dict_in_generate': False, 'forced_bos_token_id': None, 'forced_eos_token_id': None, 'remove_invalid_values': False, 'architectures': ['SpeechEncoderDecoderModel'], 'finetuning_task': None, 'id2label': {0: 'LABEL_0', 1: 'LABEL_1'}, 'label2id': {'LABEL_0': 0, 'LABEL_1': 1}, 'tokenizer_class': None, 'prefix': None, 'bos_token_id': None, 'pad_token_id': 50256, 'eos_token_id': 50256, 'sep_token_id': None, 'decoder_start_token_id': 50256, 'task_specific_params': None, 'problem_type': None, '_name_or_path': './', 'transformers_version': None, 'decoder': {'vocab_size': 50257, 'n_positions': 1024, 'n_embd': 1024, 'n_layer': 24, 'n_head': 16, 'n_inner': None, 'activation_function': 'gelu_new', 'resid_pdrop': 0.0, 'embd_pdrop': 0.0, 'attn_pdrop': 0.0, 'layer_norm_epsilon': 1e-05, 'initializer_range': 0.02, 'summary_type': 'cls_index', 'summary_use_proj': True, 'summary_activation': None, 'summary_first_dropout': 0.0, 'summary_proj_to_labels': True, 'scale_attn_weights': True, 'use_cache': False, 'scale_attn_by_inverse_layer_idx': False, 'reorder_and_upcast_attn': False, 'bos_token_id': 50256, 'eos_token_id': 50256, 'return_dict': True, 'output_hidden_states': False, 'output_attentions': False, 'torchscript': False, 'torch_dtype': None, 'use_bfloat16': False, 'pruned_heads': {}, 'tie_word_embeddings': True, 'is_encoder_decoder': False, 'is_decoder': True, 'cross_attention_hidden_size': None, 'add_cross_attention': True, 'tie_encoder_decoder': False, 'max_length': 20, 'min_length': 0, 'do_sample': False, 'early_stopping': False, 'num_beams': 1, 'num_beam_groups': 1, 'diversity_penalty': 0.0, 'temperature': 1.0, 'top_k': 50, 'top_p': 1.0, 'repetition_penalty': 1.0, 'length_penalty': 1.0, 'no_repeat_ngram_size': 0, 'encoder_no_repeat_ngram_size': 0, 'bad_words_ids': None, 'num_return_sequences': 1, 'chunk_size_feed_forward': 0, 'output_scores': False, 'return_dict_in_generate': False, 'forced_bos_token_id': None, 'forced_eos_token_id': None, 'remove_invalid_values': False, 'architectures': ['GPT2LMHeadModel'], 'finetuning_task': None, 'id2label': {0: 'LABEL_0', 1: 'LABEL_1'}, 'label2id': {'LABEL_0': 0, 'LABEL_1': 1}, 'tokenizer_class': None, 'prefix': None, 'pad_token_id': None, 'sep_token_id': None, 'decoder_start_token_id': None, 'task_specific_params': {'text-generation': {'do_sample': True, 'max_length': 50}}, 'problem_type': None, '_name_or_path': 'gpt2-medium', 'transformers_version': '4.17.0.dev0', 'n_ctx': 1024, 'n_special': 0, 'predict_special_tokens': True, 'model_type': 'gpt2'}, 'encoder': {'return_dict': True, 'output_hidden_states': False, 'output_attentions': False, 'torchscript': False, 'torch_dtype': None, 'use_bfloat16': False, 'pruned_heads': {}, 'tie_word_embeddings': True, 'is_encoder_decoder': False, 'is_decoder': False, 'cross_attention_hidden_size': None, 'add_cross_attention': False, 'tie_encoder_decoder': False, 'max_length': 20, 'min_length': 0, 'do_sample': False, 'early_stopping': False, 'num_beams': 1, 'num_beam_groups': 1, 'diversity_penalty': 0.0, 'temperature': 1.0, 'top_k': 50, 'top_p': 1.0, 'repetition_penalty': 1.0, 'length_penalty': 1.0, 'no_repeat_ngram_size': 0, 'encoder_no_repeat_ngram_size': 0, 'bad_words_ids': None, 'num_return_sequences': 1, 'chunk_size_feed_forward': 0, 'output_scores': False, 'return_dict_in_generate': False, 'forced_bos_token_id': None, 'forced_eos_token_id': None, 'remove_invalid_values': False, 'architectures': ['Wav2Vec2ForPreTraining'], 'finetuning_task': None, 'id2label': {0: 'LABEL_0', 1: 'LABEL_1'}, 'label2id': {'LABEL_0': 0, 'LABEL_1': 1}, 'tokenizer_class': None, 'prefix': None, 'bos_token_id': 1, 'pad_token_id': 0, 'eos_token_id': 2, 'sep_token_id': None, 'decoder_start_token_id': None, 'task_specific_params': None, 'problem_type': None, '_name_or_path': 'facebook/wav2vec2-large-lv60', 'transformers_version': '4.17.0.dev0', 'feat_extract_dropout': 0.0, 'gradient_checkpointing': False, 'hidden_dropout_prob': 0.0, 'num_feat_extract_layers': 7, 'hidden_size': 1024, 'feat_extract_norm': 'layer', 'feat_extract_activation': 'gelu', 'conv_dim': [512, 512, 512, 512, 512, 512, 512], 'conv_stride': [5, 2, 2, 2, 2, 2, 2], 'conv_kernel': [10, 3, 3, 3, 3, 2, 2], 'conv_bias': True, 'num_conv_pos_embeddings': 128, 'num_conv_pos_embedding_groups': 16, 'num_hidden_layers': 24, 'intermediate_size': 4096, 'hidden_act': 'gelu', 'num_attention_heads': 16, 'hidden_dropout': 0.0, 'attention_dropout': 0.0, 'activation_dropout': 0.0, 'feat_proj_dropout': 0.0, 'final_dropout': 0.0, 'layerdrop': 0.0, 'layer_norm_eps': 1e-05, 'initializer_range': 0.02, 'vocab_size': 32, 'do_stable_layer_norm': True, 'use_weighted_layer_sum': False, 'apply_spec_augment': False, 'mask_time_prob': 0.0, 'mask_time_length': 10, 'mask_time_min_masks': 2, 'mask_feature_prob': 0.0, 'mask_feature_length': 10, 'mask_feature_min_masks': 0, 'num_codevectors_per_group': 320, 'num_codevector_groups': 2, 'contrastive_logits_temperature': 0.1, 'feat_quantizer_dropout': 0.0, 'num_negatives': 100, 'codevector_dim': 768, 'proj_codevector_dim': 768, 'diversity_loss_weight': 0.1, 'ctc_loss_reduction': 'sum', 'ctc_zero_infinity': False, 'add_adapter': True, 'adapter_kernel_size': 3, 'adapter_stride': 2, 'num_adapter_layers': 3, 'output_hidden_size': 1024, 'classifier_proj_size': 256, 'tdnn_dim': [512, 512, 512, 512, 1500], 'tdnn_kernel': [5, 3, 3, 1, 1], 'tdnn_dilation': [1, 2, 3, 1, 1], 'xvector_output_dim': 512, 'model_type': 'wav2vec2'}, 'model_type': 'speech-encoder-decoder', 'processor_class': 'Wav2Vec2Processor', 'use_cache': False, 'output_dir': './', 'overwrite_output_dir': True, 'do_train': True, 'do_eval': True, 'do_predict': False, 'evaluation_strategy': 'steps', 'prediction_loss_only': False, 'per_device_train_batch_size': 8, 'per_device_eval_batch_size': 8, 'per_gpu_train_batch_size': 'None', 'per_gpu_eval_batch_size': 'None', 'gradient_accumulation_steps': 2, 'eval_accumulation_steps': 'None', 'learning_rate': 1e-05, 'weight_decay': 0.0, 'adam_beta1': 0.9, 'adam_beta2': 0.999, 'adam_epsilon': 1e-08, 'max_grad_norm': 1.0, 'num_train_epochs': 1.0, 'max_steps': -1, 'lr_scheduler_type': 'linear', 'warmup_ratio': 0.0, 'warmup_steps': 500, 'log_level': -1, 'log_level_replica': -1, 'log_on_each_node': True, 'logging_dir': './runs/Feb28_09-34-39_sanchit--v100', 'logging_strategy': 'steps', 'logging_first_step': False, 'logging_steps': 1, 'logging_nan_inf_filter': True, 'save_strategy': 'steps', 'save_steps': 500, 'save_total_limit': 1, 'save_on_each_node': False, 'no_cuda': False, 'seed': 42, 'bf16': False, 'fp16': True, 'fp16_opt_level': 'O1', 'half_precision_backend': 'amp', 'bf16_full_eval': False, 'fp16_full_eval': False, 'tf32': 'None', 'local_rank': -1, 'xpu_backend': 'None', 'tpu_num_cores': 'None', 'tpu_metrics_debug': False, 'debug': '[]', 'dataloader_drop_last': False, 'eval_steps': 500, 'dataloader_num_workers': 0, 'past_index': -1, 'run_name': './', 'disable_tqdm': False, 'remove_unused_columns': True, 'label_names': 'None', 'load_best_model_at_end': False, 'metric_for_best_model': 'None', 'greater_is_better': 'None', 'ignore_data_skip': False, 'sharded_ddp': '[]', 'deepspeed': 'None', 'label_smoothing_factor': 0.0, 'optim': 'adamw_hf', 'adafactor': False, 'group_by_length': True, 'length_column_name': 'input_length', 'report_to': "['wandb']", 'ddp_find_unused_parameters': 'None', 'ddp_bucket_cap_mb': 'None', 'dataloader_pin_memory': True, 'skip_memory_metrics': True, 'use_legacy_prediction_loop': False, 'push_to_hub': True, 'resume_from_checkpoint': 'None', 'hub_model_id': 'None', 'hub_strategy': 'every_save', 'hub_token': '<HUB_TOKEN>', 'gradient_checkpointing': True, 'fp16_backend': 'auto', 'push_to_hub_model_id': 'None', 'push_to_hub_organization': 'None', 'push_to_hub_token': '<PUSH_TO_HUB_TOKEN>', '_n_gpu': 1, 'mp_parameters': '', 'sortish_sampler': False, 'predict_with_generate': True, 'generation_max_length': 40, 'generation_num_beams': 1, 'train_batch_size': 8, 'eval_batch_size': 8}
|
27 |
+
2022-02-28 09:37:06,930 INFO MainThread:219058 [wandb_watch.py:watch():43] Watching
|
wandb/run-20220228_093705-yn2gmwrw/run-yn2gmwrw.wandb
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:4cf2f32b0b0693b4f1f8320df3f48129079f1c57c38c2bfad1d0bdb669828532
|
3 |
+
size 16060545
|