joseangelatm
commited on
Commit
•
c877a65
1
Parent(s):
278690b
First Commit
Browse files- .gitattributes +2 -0
- README.md +1 -0
- data/dataset.txt +3 -0
- data/dev.txt +3 -0
- data/roberta_cached_lm_510_dev.txt +3 -0
- data/roberta_cached_lm_510_train.txt +3 -0
- data/train.txt +3 -0
- models/roberta/config.json +1 -0
- models/roberta/merges.txt +3 -0
- models/roberta/output/checkpoint-10/config.json +55 -0
- models/roberta/output/checkpoint-10/merges.txt +3 -0
- models/roberta/output/checkpoint-10/optimizer.pt +3 -0
- models/roberta/output/checkpoint-10/pytorch_model.bin +3 -0
- models/roberta/output/checkpoint-10/scheduler.pt +3 -0
- models/roberta/output/checkpoint-10/special_tokens_map.json +1 -0
- models/roberta/output/checkpoint-10/tokenizer_config.json +1 -0
- models/roberta/output/checkpoint-10/training_args.bin +3 -0
- models/roberta/output/checkpoint-10/vocab.json +0 -0
- models/roberta/output/checkpoint-20/config.json +55 -0
- models/roberta/output/checkpoint-20/merges.txt +3 -0
- models/roberta/output/checkpoint-20/optimizer.pt +3 -0
- models/roberta/output/checkpoint-20/pytorch_model.bin +3 -0
- models/roberta/output/checkpoint-20/scheduler.pt +3 -0
- models/roberta/output/checkpoint-20/special_tokens_map.json +1 -0
- models/roberta/output/checkpoint-20/tokenizer_config.json +1 -0
- models/roberta/output/checkpoint-20/training_args.bin +3 -0
- models/roberta/output/checkpoint-20/vocab.json +0 -0
- models/roberta/output/config.json +55 -0
- models/roberta/output/eval_results.txt +3 -0
- models/roberta/output/merges.txt +3 -0
- models/roberta/output/pytorch_model.bin +3 -0
- models/roberta/output/special_tokens_map.json +1 -0
- models/roberta/output/tokenizer_config.json +1 -0
- models/roberta/output/training_args.bin +3 -0
- models/roberta/output/vocab.json +0 -0
- models/roberta/tokenizer_config.json +1 -0
- models/roberta/vocab.json +0 -0
- run_language_modeling.py +783 -0
- runs/Dec15_14-18-19_2c94adf95c33/events.out.tfevents.1608041899.2c94adf95c33.522.0 +0 -0
.gitattributes
CHANGED
@@ -6,3 +6,5 @@
|
|
6 |
*.tar.gz filter=lfs diff=lfs merge=lfs -text
|
7 |
*.ot filter=lfs diff=lfs merge=lfs -text
|
8 |
*.onnx filter=lfs diff=lfs merge=lfs -text
|
|
|
|
6 |
*.tar.gz filter=lfs diff=lfs merge=lfs -text
|
7 |
*.ot filter=lfs diff=lfs merge=lfs -text
|
8 |
*.onnx filter=lfs diff=lfs merge=lfs -text
|
9 |
+
*.pt filter=lfs diff=lfs merge=lfs -text
|
10 |
+
*.txt filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
@@ -0,0 +1 @@
|
|
|
1 |
+
Ojalá funcione
|
data/dataset.txt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:8d7fd017b3523613e16ac79dfe8713c6a52ee6ffc13d10504666e121f37298a9
|
3 |
+
size 114354869
|
data/dev.txt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:822cdc6afa8134ea58ef1211a26c008cb19d961f16b3d712d6dca816c2a6cd6e
|
3 |
+
size 616146
|
data/roberta_cached_lm_510_dev.txt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:1af6d44f857cf723a256e3ff6a5f35d3fcba4b6540f0ccb661d469bd461c191e
|
3 |
+
size 381593
|
data/roberta_cached_lm_510_train.txt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:eeb4e7cb3ad81fa73be33d4b6931ff999574d8ea58284cbb80eb746b70413100
|
3 |
+
size 43764725
|
data/train.txt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:25103781d2d1f350cf3c1d9b375f275380a9e1b49381e7c75e02379a016ef0e4
|
3 |
+
size 71099261
|
models/roberta/config.json
ADDED
@@ -0,0 +1 @@
|
|
|
1 |
+
{"architectures": ["RobertaForMaskedLM"], "attention_probs_dropout_prob": 0.1, "hidden_act": "gelu", "hidden_dropout_prob": 0.1, "hidden_size": 768, "initializer_range": 0.02, "intermediate_size": 3072, "layer_norm_eps": 1e-05, "max_position_embeddings": 514, "model_type": "roberta", "num_attention_heads": 12, "num_hidden_layers": 12, "type_vocab_size": 1, "vocab_size": 50265}
|
models/roberta/merges.txt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:95c856ebae37de8257392fbb140bcda36151bf65eabeb5b9a64beecfac343193
|
3 |
+
size 536544
|
models/roberta/output/checkpoint-10/config.json
ADDED
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_num_labels": 2,
|
3 |
+
"architectures": [
|
4 |
+
"RobertaForMaskedLM"
|
5 |
+
],
|
6 |
+
"attention_probs_dropout_prob": 0.1,
|
7 |
+
"bad_words_ids": null,
|
8 |
+
"bos_token_id": 0,
|
9 |
+
"decoder_start_token_id": null,
|
10 |
+
"do_sample": false,
|
11 |
+
"early_stopping": false,
|
12 |
+
"eos_token_id": 2,
|
13 |
+
"finetuning_task": null,
|
14 |
+
"hidden_act": "gelu",
|
15 |
+
"hidden_dropout_prob": 0.1,
|
16 |
+
"hidden_size": 768,
|
17 |
+
"id2label": {
|
18 |
+
"0": "LABEL_0",
|
19 |
+
"1": "LABEL_1"
|
20 |
+
},
|
21 |
+
"initializer_range": 0.02,
|
22 |
+
"intermediate_size": 3072,
|
23 |
+
"is_decoder": false,
|
24 |
+
"is_encoder_decoder": false,
|
25 |
+
"label2id": {
|
26 |
+
"LABEL_0": 0,
|
27 |
+
"LABEL_1": 1
|
28 |
+
},
|
29 |
+
"layer_norm_eps": 1e-05,
|
30 |
+
"length_penalty": 1.0,
|
31 |
+
"max_length": 20,
|
32 |
+
"max_position_embeddings": 514,
|
33 |
+
"min_length": 0,
|
34 |
+
"model_type": "roberta",
|
35 |
+
"no_repeat_ngram_size": 0,
|
36 |
+
"num_attention_heads": 12,
|
37 |
+
"num_beams": 1,
|
38 |
+
"num_hidden_layers": 12,
|
39 |
+
"num_return_sequences": 1,
|
40 |
+
"output_attentions": false,
|
41 |
+
"output_hidden_states": false,
|
42 |
+
"output_past": true,
|
43 |
+
"pad_token_id": 1,
|
44 |
+
"prefix": null,
|
45 |
+
"pruned_heads": {},
|
46 |
+
"repetition_penalty": 1.0,
|
47 |
+
"task_specific_params": null,
|
48 |
+
"temperature": 1.0,
|
49 |
+
"top_k": 50,
|
50 |
+
"top_p": 1.0,
|
51 |
+
"torchscript": false,
|
52 |
+
"type_vocab_size": 1,
|
53 |
+
"use_bfloat16": false,
|
54 |
+
"vocab_size": 50265
|
55 |
+
}
|
models/roberta/output/checkpoint-10/merges.txt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:babb3dc0405f9d288cc8e6205943f9f3ba1ee91cfadd9d89d1a85972890aa95b
|
3 |
+
size 536506
|
models/roberta/output/checkpoint-10/optimizer.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:c2906549a1c8be7c09d9777124a1fc0afcf0b1b0b1206ef6b201cd978409f894
|
3 |
+
size 997701959
|
models/roberta/output/checkpoint-10/pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:e01a6671f9f0124a8323ea36af6d9c58ee663644b53031e23edbce4dceba701b
|
3 |
+
size 501237143
|
models/roberta/output/checkpoint-10/scheduler.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:2f9cb9a2f68ceffad1228c2a6158aa67d57eca6d08f4ad6996c971cf59374fd5
|
3 |
+
size 623
|
models/roberta/output/checkpoint-10/special_tokens_map.json
ADDED
@@ -0,0 +1 @@
|
|
|
1 |
+
{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "sep_token": "</s>", "pad_token": "<pad>", "cls_token": "<s>", "mask_token": "<mask>"}
|
models/roberta/output/checkpoint-10/tokenizer_config.json
ADDED
@@ -0,0 +1 @@
|
|
|
1 |
+
{"max_len": 512}
|
models/roberta/output/checkpoint-10/training_args.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:64268a85c10e9bcbe053aa9baedfdf45adab1e94a1d25d742604cab763792940
|
3 |
+
size 1519
|
models/roberta/output/checkpoint-10/vocab.json
ADDED
The diff for this file is too large to render.
See raw diff
|
models/roberta/output/checkpoint-20/config.json
ADDED
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_num_labels": 2,
|
3 |
+
"architectures": [
|
4 |
+
"RobertaForMaskedLM"
|
5 |
+
],
|
6 |
+
"attention_probs_dropout_prob": 0.1,
|
7 |
+
"bad_words_ids": null,
|
8 |
+
"bos_token_id": 0,
|
9 |
+
"decoder_start_token_id": null,
|
10 |
+
"do_sample": false,
|
11 |
+
"early_stopping": false,
|
12 |
+
"eos_token_id": 2,
|
13 |
+
"finetuning_task": null,
|
14 |
+
"hidden_act": "gelu",
|
15 |
+
"hidden_dropout_prob": 0.1,
|
16 |
+
"hidden_size": 768,
|
17 |
+
"id2label": {
|
18 |
+
"0": "LABEL_0",
|
19 |
+
"1": "LABEL_1"
|
20 |
+
},
|
21 |
+
"initializer_range": 0.02,
|
22 |
+
"intermediate_size": 3072,
|
23 |
+
"is_decoder": false,
|
24 |
+
"is_encoder_decoder": false,
|
25 |
+
"label2id": {
|
26 |
+
"LABEL_0": 0,
|
27 |
+
"LABEL_1": 1
|
28 |
+
},
|
29 |
+
"layer_norm_eps": 1e-05,
|
30 |
+
"length_penalty": 1.0,
|
31 |
+
"max_length": 20,
|
32 |
+
"max_position_embeddings": 514,
|
33 |
+
"min_length": 0,
|
34 |
+
"model_type": "roberta",
|
35 |
+
"no_repeat_ngram_size": 0,
|
36 |
+
"num_attention_heads": 12,
|
37 |
+
"num_beams": 1,
|
38 |
+
"num_hidden_layers": 12,
|
39 |
+
"num_return_sequences": 1,
|
40 |
+
"output_attentions": false,
|
41 |
+
"output_hidden_states": false,
|
42 |
+
"output_past": true,
|
43 |
+
"pad_token_id": 1,
|
44 |
+
"prefix": null,
|
45 |
+
"pruned_heads": {},
|
46 |
+
"repetition_penalty": 1.0,
|
47 |
+
"task_specific_params": null,
|
48 |
+
"temperature": 1.0,
|
49 |
+
"top_k": 50,
|
50 |
+
"top_p": 1.0,
|
51 |
+
"torchscript": false,
|
52 |
+
"type_vocab_size": 1,
|
53 |
+
"use_bfloat16": false,
|
54 |
+
"vocab_size": 50265
|
55 |
+
}
|
models/roberta/output/checkpoint-20/merges.txt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:babb3dc0405f9d288cc8e6205943f9f3ba1ee91cfadd9d89d1a85972890aa95b
|
3 |
+
size 536506
|
models/roberta/output/checkpoint-20/optimizer.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:007e9598f42cdd29c37a8f1890047f1c9d42f9ce8e93a277fb3a7c9eacc8aaef
|
3 |
+
size 997701959
|
models/roberta/output/checkpoint-20/pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:c9715fa8018cd96e53c7ea8bb4192c131ac64bfebdd93112e2820ecf6a4475d9
|
3 |
+
size 501237143
|
models/roberta/output/checkpoint-20/scheduler.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:ed147c3577ee87e1d3f66769b4fff18b4ad0c6ec05d56bb6878413771b706293
|
3 |
+
size 623
|
models/roberta/output/checkpoint-20/special_tokens_map.json
ADDED
@@ -0,0 +1 @@
|
|
|
1 |
+
{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "sep_token": "</s>", "pad_token": "<pad>", "cls_token": "<s>", "mask_token": "<mask>"}
|
models/roberta/output/checkpoint-20/tokenizer_config.json
ADDED
@@ -0,0 +1 @@
|
|
|
1 |
+
{"max_len": 512}
|
models/roberta/output/checkpoint-20/training_args.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:64268a85c10e9bcbe053aa9baedfdf45adab1e94a1d25d742604cab763792940
|
3 |
+
size 1519
|
models/roberta/output/checkpoint-20/vocab.json
ADDED
The diff for this file is too large to render.
See raw diff
|
models/roberta/output/config.json
ADDED
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_num_labels": 2,
|
3 |
+
"architectures": [
|
4 |
+
"RobertaForMaskedLM"
|
5 |
+
],
|
6 |
+
"attention_probs_dropout_prob": 0.1,
|
7 |
+
"bad_words_ids": null,
|
8 |
+
"bos_token_id": 0,
|
9 |
+
"decoder_start_token_id": null,
|
10 |
+
"do_sample": false,
|
11 |
+
"early_stopping": false,
|
12 |
+
"eos_token_id": 2,
|
13 |
+
"finetuning_task": null,
|
14 |
+
"hidden_act": "gelu",
|
15 |
+
"hidden_dropout_prob": 0.1,
|
16 |
+
"hidden_size": 768,
|
17 |
+
"id2label": {
|
18 |
+
"0": "LABEL_0",
|
19 |
+
"1": "LABEL_1"
|
20 |
+
},
|
21 |
+
"initializer_range": 0.02,
|
22 |
+
"intermediate_size": 3072,
|
23 |
+
"is_decoder": false,
|
24 |
+
"is_encoder_decoder": false,
|
25 |
+
"label2id": {
|
26 |
+
"LABEL_0": 0,
|
27 |
+
"LABEL_1": 1
|
28 |
+
},
|
29 |
+
"layer_norm_eps": 1e-05,
|
30 |
+
"length_penalty": 1.0,
|
31 |
+
"max_length": 20,
|
32 |
+
"max_position_embeddings": 514,
|
33 |
+
"min_length": 0,
|
34 |
+
"model_type": "roberta",
|
35 |
+
"no_repeat_ngram_size": 0,
|
36 |
+
"num_attention_heads": 12,
|
37 |
+
"num_beams": 1,
|
38 |
+
"num_hidden_layers": 12,
|
39 |
+
"num_return_sequences": 1,
|
40 |
+
"output_attentions": false,
|
41 |
+
"output_hidden_states": false,
|
42 |
+
"output_past": true,
|
43 |
+
"pad_token_id": 1,
|
44 |
+
"prefix": null,
|
45 |
+
"pruned_heads": {},
|
46 |
+
"repetition_penalty": 1.0,
|
47 |
+
"task_specific_params": null,
|
48 |
+
"temperature": 1.0,
|
49 |
+
"top_k": 50,
|
50 |
+
"top_p": 1.0,
|
51 |
+
"torchscript": false,
|
52 |
+
"type_vocab_size": 1,
|
53 |
+
"use_bfloat16": false,
|
54 |
+
"vocab_size": 50265
|
55 |
+
}
|
models/roberta/output/eval_results.txt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:7206c5471e7c697fbb01e10dacd4517ce2ea62aa78f1e1800e67e7c5865a741f
|
3 |
+
size 31
|
models/roberta/output/merges.txt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:babb3dc0405f9d288cc8e6205943f9f3ba1ee91cfadd9d89d1a85972890aa95b
|
3 |
+
size 536506
|
models/roberta/output/pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:40827c152b9a025377ceb13b97b0bd1e3ad6c210864e93f271deaf9584d95eb0
|
3 |
+
size 501237143
|
models/roberta/output/special_tokens_map.json
ADDED
@@ -0,0 +1 @@
|
|
|
1 |
+
{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "sep_token": "</s>", "pad_token": "<pad>", "cls_token": "<s>", "mask_token": "<mask>"}
|
models/roberta/output/tokenizer_config.json
ADDED
@@ -0,0 +1 @@
|
|
|
1 |
+
{"max_len": 512}
|
models/roberta/output/training_args.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:64268a85c10e9bcbe053aa9baedfdf45adab1e94a1d25d742604cab763792940
|
3 |
+
size 1519
|
models/roberta/output/vocab.json
ADDED
The diff for this file is too large to render.
See raw diff
|
models/roberta/tokenizer_config.json
ADDED
@@ -0,0 +1 @@
|
|
|
1 |
+
{"max_len": 512}
|
models/roberta/vocab.json
ADDED
The diff for this file is too large to render.
See raw diff
|
run_language_modeling.py
ADDED
@@ -0,0 +1,783 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
|
3 |
+
# Copyright (c) 2018, NVIDIA CORPORATION. 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 language modeling on a text file (GPT, GPT-2, BERT, RoBERTa).
|
18 |
+
GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT and RoBERTa are fine-tuned
|
19 |
+
using a masked language modeling (MLM) loss.
|
20 |
+
"""
|
21 |
+
|
22 |
+
|
23 |
+
import argparse
|
24 |
+
import glob
|
25 |
+
import logging
|
26 |
+
import os
|
27 |
+
import pickle
|
28 |
+
import random
|
29 |
+
import re
|
30 |
+
import shutil
|
31 |
+
from typing import Dict, List, Tuple
|
32 |
+
|
33 |
+
import numpy as np
|
34 |
+
import torch
|
35 |
+
from torch.nn.utils.rnn import pad_sequence
|
36 |
+
from torch.utils.data import DataLoader, Dataset, RandomSampler, SequentialSampler
|
37 |
+
from torch.utils.data.distributed import DistributedSampler
|
38 |
+
from tqdm import tqdm, trange
|
39 |
+
|
40 |
+
from transformers import (
|
41 |
+
MODEL_WITH_LM_HEAD_MAPPING,
|
42 |
+
WEIGHTS_NAME,
|
43 |
+
AdamW,
|
44 |
+
AutoConfig,
|
45 |
+
AutoModelWithLMHead,
|
46 |
+
AutoTokenizer,
|
47 |
+
PreTrainedModel,
|
48 |
+
PreTrainedTokenizer,
|
49 |
+
get_linear_schedule_with_warmup,
|
50 |
+
)
|
51 |
+
|
52 |
+
|
53 |
+
try:
|
54 |
+
from torch.utils.tensorboard import SummaryWriter
|
55 |
+
except ImportError:
|
56 |
+
from tensorboardX import SummaryWriter
|
57 |
+
|
58 |
+
|
59 |
+
logger = logging.getLogger(__name__)
|
60 |
+
|
61 |
+
|
62 |
+
MODEL_CONFIG_CLASSES = list(MODEL_WITH_LM_HEAD_MAPPING.keys())
|
63 |
+
MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
|
64 |
+
|
65 |
+
|
66 |
+
class TextDataset(Dataset):
|
67 |
+
def __init__(self, tokenizer: PreTrainedTokenizer, args, file_path: str, block_size=512):
|
68 |
+
assert os.path.isfile(file_path)
|
69 |
+
|
70 |
+
block_size = block_size - (tokenizer.max_len - tokenizer.max_len_single_sentence)
|
71 |
+
|
72 |
+
directory, filename = os.path.split(file_path)
|
73 |
+
cached_features_file = os.path.join(
|
74 |
+
directory, args.model_type + "_cached_lm_" + str(block_size) + "_" + filename
|
75 |
+
)
|
76 |
+
|
77 |
+
if os.path.exists(cached_features_file) and not args.overwrite_cache:
|
78 |
+
logger.info("Loading features from cached file %s", cached_features_file)
|
79 |
+
with open(cached_features_file, "rb") as handle:
|
80 |
+
self.examples = pickle.load(handle)
|
81 |
+
else:
|
82 |
+
logger.info("Creating features from dataset file at %s", directory)
|
83 |
+
|
84 |
+
self.examples = []
|
85 |
+
with open(file_path, encoding="utf-8") as f:
|
86 |
+
text = f.read()
|
87 |
+
|
88 |
+
tokenized_text = tokenizer.convert_tokens_to_ids(tokenizer.tokenize(text))
|
89 |
+
|
90 |
+
for i in range(0, len(tokenized_text) - block_size + 1, block_size): # Truncate in block of block_size
|
91 |
+
self.examples.append(tokenizer.build_inputs_with_special_tokens(tokenized_text[i : i + block_size]))
|
92 |
+
# Note that we are loosing the last truncated example here for the sake of simplicity (no padding)
|
93 |
+
# If your dataset is small, first you should loook for a bigger one :-) and second you
|
94 |
+
# can change this behavior by adding (model specific) padding.
|
95 |
+
|
96 |
+
logger.info("Saving features into cached file %s", cached_features_file)
|
97 |
+
with open(cached_features_file, "wb") as handle:
|
98 |
+
pickle.dump(self.examples, handle, protocol=pickle.HIGHEST_PROTOCOL)
|
99 |
+
|
100 |
+
def __len__(self):
|
101 |
+
return len(self.examples)
|
102 |
+
|
103 |
+
def __getitem__(self, item):
|
104 |
+
return torch.tensor(self.examples[item], dtype=torch.long)
|
105 |
+
|
106 |
+
|
107 |
+
class LineByLineTextDataset(Dataset):
|
108 |
+
def __init__(self, tokenizer: PreTrainedTokenizer, args, file_path: str, block_size=512):
|
109 |
+
assert os.path.isfile(file_path)
|
110 |
+
# Here, we do not cache the features, operating under the assumption
|
111 |
+
# that we will soon use fast multithreaded tokenizers from the
|
112 |
+
# `tokenizers` repo everywhere =)
|
113 |
+
logger.info("Creating features from dataset file at %s", file_path)
|
114 |
+
|
115 |
+
with open(file_path, encoding="utf-8") as f:
|
116 |
+
lines = [line for line in f.read().splitlines() if (len(line) > 0 and not line.isspace())]
|
117 |
+
|
118 |
+
self.examples = tokenizer.batch_encode_plus(lines, add_special_tokens=True, max_length=block_size)["input_ids"]
|
119 |
+
|
120 |
+
def __len__(self):
|
121 |
+
return len(self.examples)
|
122 |
+
|
123 |
+
def __getitem__(self, i):
|
124 |
+
return torch.tensor(self.examples[i], dtype=torch.long)
|
125 |
+
|
126 |
+
|
127 |
+
def load_and_cache_examples(args, tokenizer, evaluate=False):
|
128 |
+
file_path = args.eval_data_file if evaluate else args.train_data_file
|
129 |
+
if args.line_by_line:
|
130 |
+
return LineByLineTextDataset(tokenizer, args, file_path=file_path, block_size=args.block_size)
|
131 |
+
else:
|
132 |
+
return TextDataset(tokenizer, args, file_path=file_path, block_size=args.block_size)
|
133 |
+
|
134 |
+
|
135 |
+
def set_seed(args):
|
136 |
+
random.seed(args.seed)
|
137 |
+
np.random.seed(args.seed)
|
138 |
+
torch.manual_seed(args.seed)
|
139 |
+
if args.n_gpu > 0:
|
140 |
+
torch.cuda.manual_seed_all(args.seed)
|
141 |
+
|
142 |
+
|
143 |
+
def _sorted_checkpoints(args, checkpoint_prefix="checkpoint", use_mtime=False) -> List[str]:
|
144 |
+
ordering_and_checkpoint_path = []
|
145 |
+
|
146 |
+
glob_checkpoints = glob.glob(os.path.join(args.output_dir, "{}-*".format(checkpoint_prefix)))
|
147 |
+
|
148 |
+
for path in glob_checkpoints:
|
149 |
+
if use_mtime:
|
150 |
+
ordering_and_checkpoint_path.append((os.path.getmtime(path), path))
|
151 |
+
else:
|
152 |
+
regex_match = re.match(".*{}-([0-9]+)".format(checkpoint_prefix), path)
|
153 |
+
if regex_match and regex_match.groups():
|
154 |
+
ordering_and_checkpoint_path.append((int(regex_match.groups()[0]), path))
|
155 |
+
|
156 |
+
checkpoints_sorted = sorted(ordering_and_checkpoint_path)
|
157 |
+
checkpoints_sorted = [checkpoint[1] for checkpoint in checkpoints_sorted]
|
158 |
+
return checkpoints_sorted
|
159 |
+
|
160 |
+
|
161 |
+
def _rotate_checkpoints(args, checkpoint_prefix="checkpoint", use_mtime=False) -> None:
|
162 |
+
if not args.save_total_limit:
|
163 |
+
return
|
164 |
+
if args.save_total_limit <= 0:
|
165 |
+
return
|
166 |
+
|
167 |
+
# Check if we should delete older checkpoint(s)
|
168 |
+
checkpoints_sorted = _sorted_checkpoints(args, checkpoint_prefix, use_mtime)
|
169 |
+
if len(checkpoints_sorted) <= args.save_total_limit:
|
170 |
+
return
|
171 |
+
|
172 |
+
number_of_checkpoints_to_delete = max(0, len(checkpoints_sorted) - args.save_total_limit)
|
173 |
+
checkpoints_to_be_deleted = checkpoints_sorted[:number_of_checkpoints_to_delete]
|
174 |
+
for checkpoint in checkpoints_to_be_deleted:
|
175 |
+
logger.info("Deleting older checkpoint [{}] due to args.save_total_limit".format(checkpoint))
|
176 |
+
shutil.rmtree(checkpoint)
|
177 |
+
|
178 |
+
|
179 |
+
def mask_tokens(inputs: torch.Tensor, tokenizer: PreTrainedTokenizer, args) -> Tuple[torch.Tensor, torch.Tensor]:
|
180 |
+
""" Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original. """
|
181 |
+
|
182 |
+
if tokenizer.mask_token is None:
|
183 |
+
raise ValueError(
|
184 |
+
"This tokenizer does not have a mask token which is necessary for masked language modeling. Remove the --mlm flag if you want to use this tokenizer."
|
185 |
+
)
|
186 |
+
|
187 |
+
labels = inputs.clone()
|
188 |
+
# We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
|
189 |
+
probability_matrix = torch.full(labels.shape, args.mlm_probability)
|
190 |
+
special_tokens_mask = [
|
191 |
+
tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()
|
192 |
+
]
|
193 |
+
probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)
|
194 |
+
if tokenizer._pad_token is not None:
|
195 |
+
padding_mask = labels.eq(tokenizer.pad_token_id)
|
196 |
+
probability_matrix.masked_fill_(padding_mask, value=0.0)
|
197 |
+
masked_indices = torch.bernoulli(probability_matrix).bool()
|
198 |
+
labels[~masked_indices] = -100 # We only compute loss on masked tokens
|
199 |
+
|
200 |
+
# 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
|
201 |
+
indices_replaced = torch.bernoulli(torch.full(labels.shape, 0.8)).bool() & masked_indices
|
202 |
+
inputs[indices_replaced] = tokenizer.convert_tokens_to_ids(tokenizer.mask_token)
|
203 |
+
|
204 |
+
# 10% of the time, we replace masked input tokens with random word
|
205 |
+
indices_random = torch.bernoulli(torch.full(labels.shape, 0.5)).bool() & masked_indices & ~indices_replaced
|
206 |
+
random_words = torch.randint(len(tokenizer), labels.shape, dtype=torch.long)
|
207 |
+
inputs[indices_random] = random_words[indices_random]
|
208 |
+
|
209 |
+
# The rest of the time (10% of the time) we keep the masked input tokens unchanged
|
210 |
+
return inputs, labels
|
211 |
+
|
212 |
+
|
213 |
+
def train(args, train_dataset, model: PreTrainedModel, tokenizer: PreTrainedTokenizer) -> Tuple[int, float]:
|
214 |
+
""" Train the model """
|
215 |
+
if args.local_rank in [-1, 0]:
|
216 |
+
tb_writer = SummaryWriter()
|
217 |
+
|
218 |
+
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
|
219 |
+
|
220 |
+
def collate(examples: List[torch.Tensor]):
|
221 |
+
if tokenizer._pad_token is None:
|
222 |
+
return pad_sequence(examples, batch_first=True)
|
223 |
+
return pad_sequence(examples, batch_first=True, padding_value=tokenizer.pad_token_id)
|
224 |
+
|
225 |
+
train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
|
226 |
+
train_dataloader = DataLoader(
|
227 |
+
train_dataset, sampler=train_sampler, batch_size=args.train_batch_size, collate_fn=collate
|
228 |
+
)
|
229 |
+
|
230 |
+
if args.max_steps > 0:
|
231 |
+
t_total = args.max_steps
|
232 |
+
args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
|
233 |
+
else:
|
234 |
+
t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
|
235 |
+
|
236 |
+
model = model.module if hasattr(model, "module") else model # Take care of distributed/parallel training
|
237 |
+
model.resize_token_embeddings(len(tokenizer))
|
238 |
+
|
239 |
+
# Prepare optimizer and schedule (linear warmup and decay)
|
240 |
+
no_decay = ["bias", "LayerNorm.weight"]
|
241 |
+
optimizer_grouped_parameters = [
|
242 |
+
{
|
243 |
+
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
|
244 |
+
"weight_decay": args.weight_decay,
|
245 |
+
},
|
246 |
+
{"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
|
247 |
+
]
|
248 |
+
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
|
249 |
+
scheduler = get_linear_schedule_with_warmup(
|
250 |
+
optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
|
251 |
+
)
|
252 |
+
|
253 |
+
# Check if saved optimizer or scheduler states exist
|
254 |
+
if (
|
255 |
+
args.model_name_or_path
|
256 |
+
and os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt"))
|
257 |
+
and os.path.isfile(os.path.join(args.model_name_or_path, "scheduler.pt"))
|
258 |
+
):
|
259 |
+
# Load in optimizer and scheduler states
|
260 |
+
optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
|
261 |
+
scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
|
262 |
+
|
263 |
+
if args.fp16:
|
264 |
+
try:
|
265 |
+
from apex import amp
|
266 |
+
except ImportError:
|
267 |
+
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
|
268 |
+
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
|
269 |
+
|
270 |
+
# multi-gpu training (should be after apex fp16 initialization)
|
271 |
+
if args.n_gpu > 1:
|
272 |
+
model = torch.nn.DataParallel(model)
|
273 |
+
|
274 |
+
# Distributed training (should be after apex fp16 initialization)
|
275 |
+
if args.local_rank != -1:
|
276 |
+
model = torch.nn.parallel.DistributedDataParallel(
|
277 |
+
model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True
|
278 |
+
)
|
279 |
+
|
280 |
+
# Train!
|
281 |
+
logger.info("***** Running training *****")
|
282 |
+
logger.info(" Num examples = %d", len(train_dataset))
|
283 |
+
logger.info(" Num Epochs = %d", args.num_train_epochs)
|
284 |
+
logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
|
285 |
+
logger.info(
|
286 |
+
" Total train batch size (w. parallel, distributed & accumulation) = %d",
|
287 |
+
args.train_batch_size
|
288 |
+
* args.gradient_accumulation_steps
|
289 |
+
* (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
|
290 |
+
)
|
291 |
+
logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
|
292 |
+
logger.info(" Total optimization steps = %d", t_total)
|
293 |
+
|
294 |
+
global_step = 0
|
295 |
+
epochs_trained = 0
|
296 |
+
steps_trained_in_current_epoch = 0
|
297 |
+
# Check if continuing training from a checkpoint
|
298 |
+
if args.model_name_or_path and os.path.exists(args.model_name_or_path):
|
299 |
+
try:
|
300 |
+
# set global_step to gobal_step of last saved checkpoint from model path
|
301 |
+
checkpoint_suffix = args.model_name_or_path.split("-")[-1].split("/")[0]
|
302 |
+
global_step = int(checkpoint_suffix)
|
303 |
+
epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
|
304 |
+
steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
|
305 |
+
|
306 |
+
logger.info(" Continuing training from checkpoint, will skip to saved global_step")
|
307 |
+
logger.info(" Continuing training from epoch %d", epochs_trained)
|
308 |
+
logger.info(" Continuing training from global step %d", global_step)
|
309 |
+
logger.info(" Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
|
310 |
+
except ValueError:
|
311 |
+
logger.info(" Starting fine-tuning.")
|
312 |
+
|
313 |
+
tr_loss, logging_loss = 0.0, 0.0
|
314 |
+
|
315 |
+
model.zero_grad()
|
316 |
+
train_iterator = trange(
|
317 |
+
epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]
|
318 |
+
)
|
319 |
+
set_seed(args) # Added here for reproducibility
|
320 |
+
for _ in train_iterator:
|
321 |
+
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
|
322 |
+
for step, batch in enumerate(epoch_iterator):
|
323 |
+
|
324 |
+
# Skip past any already trained steps if resuming training
|
325 |
+
if steps_trained_in_current_epoch > 0:
|
326 |
+
steps_trained_in_current_epoch -= 1
|
327 |
+
continue
|
328 |
+
|
329 |
+
inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch)
|
330 |
+
inputs = inputs.to(args.device)
|
331 |
+
labels = labels.to(args.device)
|
332 |
+
model.train()
|
333 |
+
outputs = model(inputs, masked_lm_labels=labels) if args.mlm else model(inputs, labels=labels)
|
334 |
+
loss = outputs[0] # model outputs are always tuple in transformers (see doc)
|
335 |
+
|
336 |
+
if args.n_gpu > 1:
|
337 |
+
loss = loss.mean() # mean() to average on multi-gpu parallel training
|
338 |
+
if args.gradient_accumulation_steps > 1:
|
339 |
+
loss = loss / args.gradient_accumulation_steps
|
340 |
+
|
341 |
+
if args.fp16:
|
342 |
+
with amp.scale_loss(loss, optimizer) as scaled_loss:
|
343 |
+
scaled_loss.backward()
|
344 |
+
else:
|
345 |
+
loss.backward()
|
346 |
+
|
347 |
+
tr_loss += loss.item()
|
348 |
+
if (step + 1) % args.gradient_accumulation_steps == 0:
|
349 |
+
if args.fp16:
|
350 |
+
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
|
351 |
+
else:
|
352 |
+
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
|
353 |
+
optimizer.step()
|
354 |
+
scheduler.step() # Update learning rate schedule
|
355 |
+
model.zero_grad()
|
356 |
+
global_step += 1
|
357 |
+
|
358 |
+
if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
|
359 |
+
# Log metrics
|
360 |
+
if (
|
361 |
+
args.local_rank == -1 and args.evaluate_during_training
|
362 |
+
): # Only evaluate when single GPU otherwise metrics may not average well
|
363 |
+
results = evaluate(args, model, tokenizer)
|
364 |
+
for key, value in results.items():
|
365 |
+
tb_writer.add_scalar("eval_{}".format(key), value, global_step)
|
366 |
+
tb_writer.add_scalar("lr", scheduler.get_lr()[0], global_step)
|
367 |
+
tb_writer.add_scalar("loss", (tr_loss - logging_loss) / args.logging_steps, global_step)
|
368 |
+
logging_loss = tr_loss
|
369 |
+
|
370 |
+
if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
|
371 |
+
checkpoint_prefix = "checkpoint"
|
372 |
+
# Save model checkpoint
|
373 |
+
output_dir = os.path.join(args.output_dir, "{}-{}".format(checkpoint_prefix, global_step))
|
374 |
+
os.makedirs(output_dir, exist_ok=True)
|
375 |
+
model_to_save = (
|
376 |
+
model.module if hasattr(model, "module") else model
|
377 |
+
) # Take care of distributed/parallel training
|
378 |
+
model_to_save.save_pretrained(output_dir)
|
379 |
+
tokenizer.save_pretrained(output_dir)
|
380 |
+
|
381 |
+
torch.save(args, os.path.join(output_dir, "training_args.bin"))
|
382 |
+
logger.info("Saving model checkpoint to %s", output_dir)
|
383 |
+
|
384 |
+
_rotate_checkpoints(args, checkpoint_prefix)
|
385 |
+
|
386 |
+
torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
|
387 |
+
torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
|
388 |
+
logger.info("Saving optimizer and scheduler states to %s", output_dir)
|
389 |
+
|
390 |
+
if args.max_steps > 0 and global_step > args.max_steps:
|
391 |
+
epoch_iterator.close()
|
392 |
+
break
|
393 |
+
if args.max_steps > 0 and global_step > args.max_steps:
|
394 |
+
train_iterator.close()
|
395 |
+
break
|
396 |
+
|
397 |
+
if args.local_rank in [-1, 0]:
|
398 |
+
tb_writer.close()
|
399 |
+
|
400 |
+
return global_step, tr_loss / global_step
|
401 |
+
|
402 |
+
|
403 |
+
def evaluate(args, model: PreTrainedModel, tokenizer: PreTrainedTokenizer, prefix="") -> Dict:
|
404 |
+
# Loop to handle MNLI double evaluation (matched, mis-matched)
|
405 |
+
eval_output_dir = args.output_dir
|
406 |
+
|
407 |
+
eval_dataset = load_and_cache_examples(args, tokenizer, evaluate=True)
|
408 |
+
|
409 |
+
if args.local_rank in [-1, 0]:
|
410 |
+
os.makedirs(eval_output_dir, exist_ok=True)
|
411 |
+
|
412 |
+
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
|
413 |
+
# Note that DistributedSampler samples randomly
|
414 |
+
|
415 |
+
def collate(examples: List[torch.Tensor]):
|
416 |
+
if tokenizer._pad_token is None:
|
417 |
+
return pad_sequence(examples, batch_first=True)
|
418 |
+
return pad_sequence(examples, batch_first=True, padding_value=tokenizer.pad_token_id)
|
419 |
+
|
420 |
+
eval_sampler = SequentialSampler(eval_dataset)
|
421 |
+
eval_dataloader = DataLoader(
|
422 |
+
eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size, collate_fn=collate
|
423 |
+
)
|
424 |
+
|
425 |
+
# multi-gpu evaluate
|
426 |
+
if args.n_gpu > 1:
|
427 |
+
model = torch.nn.DataParallel(model)
|
428 |
+
|
429 |
+
# Eval!
|
430 |
+
logger.info("***** Running evaluation {} *****".format(prefix))
|
431 |
+
logger.info(" Num examples = %d", len(eval_dataset))
|
432 |
+
logger.info(" Batch size = %d", args.eval_batch_size)
|
433 |
+
eval_loss = 0.0
|
434 |
+
nb_eval_steps = 0
|
435 |
+
model.eval()
|
436 |
+
|
437 |
+
for batch in tqdm(eval_dataloader, desc="Evaluating"):
|
438 |
+
inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch)
|
439 |
+
inputs = inputs.to(args.device)
|
440 |
+
labels = labels.to(args.device)
|
441 |
+
|
442 |
+
with torch.no_grad():
|
443 |
+
outputs = model(inputs, masked_lm_labels=labels) if args.mlm else model(inputs, labels=labels)
|
444 |
+
lm_loss = outputs[0]
|
445 |
+
eval_loss += lm_loss.mean().item()
|
446 |
+
nb_eval_steps += 1
|
447 |
+
|
448 |
+
eval_loss = eval_loss / nb_eval_steps
|
449 |
+
perplexity = torch.exp(torch.tensor(eval_loss))
|
450 |
+
|
451 |
+
result = {"perplexity": perplexity}
|
452 |
+
|
453 |
+
output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
|
454 |
+
with open(output_eval_file, "w") as writer:
|
455 |
+
logger.info("***** Eval results {} *****".format(prefix))
|
456 |
+
for key in sorted(result.keys()):
|
457 |
+
logger.info(" %s = %s", key, str(result[key]))
|
458 |
+
writer.write("%s = %s\n" % (key, str(result[key])))
|
459 |
+
|
460 |
+
return result
|
461 |
+
|
462 |
+
|
463 |
+
def main():
|
464 |
+
parser = argparse.ArgumentParser()
|
465 |
+
|
466 |
+
# Required parameters
|
467 |
+
parser.add_argument(
|
468 |
+
"--train_data_file", default=None, type=str, required=True, help="The input training data file (a text file)."
|
469 |
+
)
|
470 |
+
parser.add_argument(
|
471 |
+
"--output_dir",
|
472 |
+
type=str,
|
473 |
+
required=True,
|
474 |
+
help="The output directory where the model predictions and checkpoints will be written.",
|
475 |
+
)
|
476 |
+
parser.add_argument(
|
477 |
+
"--model_type", type=str, required=True, help="The model architecture to be trained or fine-tuned.",
|
478 |
+
)
|
479 |
+
|
480 |
+
# Other parameters
|
481 |
+
parser.add_argument(
|
482 |
+
"--eval_data_file",
|
483 |
+
default=None,
|
484 |
+
type=str,
|
485 |
+
help="An optional input evaluation data file to evaluate the perplexity on (a text file).",
|
486 |
+
)
|
487 |
+
parser.add_argument(
|
488 |
+
"--line_by_line",
|
489 |
+
action="store_true",
|
490 |
+
help="Whether distinct lines of text in the dataset are to be handled as distinct sequences.",
|
491 |
+
)
|
492 |
+
parser.add_argument(
|
493 |
+
"--should_continue", action="store_true", help="Whether to continue from latest checkpoint in output_dir"
|
494 |
+
)
|
495 |
+
parser.add_argument(
|
496 |
+
"--model_name_or_path",
|
497 |
+
default=None,
|
498 |
+
type=str,
|
499 |
+
help="The model checkpoint for weights initialization. Leave None if you want to train a model from scratch.",
|
500 |
+
)
|
501 |
+
|
502 |
+
parser.add_argument(
|
503 |
+
"--mlm", action="store_true", help="Train with masked-language modeling loss instead of language modeling."
|
504 |
+
)
|
505 |
+
parser.add_argument(
|
506 |
+
"--mlm_probability", type=float, default=0.15, help="Ratio of tokens to mask for masked language modeling loss"
|
507 |
+
)
|
508 |
+
|
509 |
+
parser.add_argument(
|
510 |
+
"--config_name",
|
511 |
+
default=None,
|
512 |
+
type=str,
|
513 |
+
help="Optional pretrained config name or path if not the same as model_name_or_path. If both are None, initialize a new config.",
|
514 |
+
)
|
515 |
+
parser.add_argument(
|
516 |
+
"--tokenizer_name",
|
517 |
+
default=None,
|
518 |
+
type=str,
|
519 |
+
help="Optional pretrained tokenizer name or path if not the same as model_name_or_path. If both are None, initialize a new tokenizer.",
|
520 |
+
)
|
521 |
+
parser.add_argument(
|
522 |
+
"--cache_dir",
|
523 |
+
default=None,
|
524 |
+
type=str,
|
525 |
+
help="Optional directory to store the pre-trained models downloaded from s3 (instead of the default one)",
|
526 |
+
)
|
527 |
+
parser.add_argument(
|
528 |
+
"--block_size",
|
529 |
+
default=-1,
|
530 |
+
type=int,
|
531 |
+
help="Optional input sequence length after tokenization."
|
532 |
+
"The training dataset will be truncated in block of this size for training."
|
533 |
+
"Default to the model max input length for single sentence inputs (take into account special tokens).",
|
534 |
+
)
|
535 |
+
parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
|
536 |
+
parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
|
537 |
+
parser.add_argument(
|
538 |
+
"--evaluate_during_training", action="store_true", help="Run evaluation during training at each logging step."
|
539 |
+
)
|
540 |
+
|
541 |
+
parser.add_argument("--per_gpu_train_batch_size", default=4, type=int, help="Batch size per GPU/CPU for training.")
|
542 |
+
parser.add_argument(
|
543 |
+
"--per_gpu_eval_batch_size", default=4, type=int, help="Batch size per GPU/CPU for evaluation."
|
544 |
+
)
|
545 |
+
parser.add_argument(
|
546 |
+
"--gradient_accumulation_steps",
|
547 |
+
type=int,
|
548 |
+
default=1,
|
549 |
+
help="Number of updates steps to accumulate before performing a backward/update pass.",
|
550 |
+
)
|
551 |
+
parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
|
552 |
+
parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
|
553 |
+
parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
|
554 |
+
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
|
555 |
+
parser.add_argument(
|
556 |
+
"--num_train_epochs", default=1.0, type=float, help="Total number of training epochs to perform."
|
557 |
+
)
|
558 |
+
parser.add_argument(
|
559 |
+
"--max_steps",
|
560 |
+
default=-1,
|
561 |
+
type=int,
|
562 |
+
help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
|
563 |
+
)
|
564 |
+
parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
|
565 |
+
|
566 |
+
parser.add_argument("--logging_steps", type=int, default=500, help="Log every X updates steps.")
|
567 |
+
parser.add_argument("--save_steps", type=int, default=500, help="Save checkpoint every X updates steps.")
|
568 |
+
parser.add_argument(
|
569 |
+
"--save_total_limit",
|
570 |
+
type=int,
|
571 |
+
default=None,
|
572 |
+
help="Limit the total amount of checkpoints, delete the older checkpoints in the output_dir, does not delete by default",
|
573 |
+
)
|
574 |
+
parser.add_argument(
|
575 |
+
"--eval_all_checkpoints",
|
576 |
+
action="store_true",
|
577 |
+
help="Evaluate all checkpoints starting with the same prefix as model_name_or_path ending and ending with step number",
|
578 |
+
)
|
579 |
+
parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
|
580 |
+
parser.add_argument(
|
581 |
+
"--overwrite_output_dir", action="store_true", help="Overwrite the content of the output directory"
|
582 |
+
)
|
583 |
+
parser.add_argument(
|
584 |
+
"--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets"
|
585 |
+
)
|
586 |
+
parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
|
587 |
+
|
588 |
+
parser.add_argument(
|
589 |
+
"--fp16",
|
590 |
+
action="store_true",
|
591 |
+
help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
|
592 |
+
)
|
593 |
+
parser.add_argument(
|
594 |
+
"--fp16_opt_level",
|
595 |
+
type=str,
|
596 |
+
default="O1",
|
597 |
+
help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
|
598 |
+
"See details at https://nvidia.github.io/apex/amp.html",
|
599 |
+
)
|
600 |
+
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
|
601 |
+
parser.add_argument("--server_ip", type=str, default="", help="For distant debugging.")
|
602 |
+
parser.add_argument("--server_port", type=str, default="", help="For distant debugging.")
|
603 |
+
args = parser.parse_args()
|
604 |
+
|
605 |
+
if args.model_type in ["bert", "roberta", "distilbert", "camembert"] and not args.mlm:
|
606 |
+
raise ValueError(
|
607 |
+
"BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the --mlm "
|
608 |
+
"flag (masked language modeling)."
|
609 |
+
)
|
610 |
+
if args.eval_data_file is None and args.do_eval:
|
611 |
+
raise ValueError(
|
612 |
+
"Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file "
|
613 |
+
"or remove the --do_eval argument."
|
614 |
+
)
|
615 |
+
if args.should_continue:
|
616 |
+
sorted_checkpoints = _sorted_checkpoints(args)
|
617 |
+
if len(sorted_checkpoints) == 0:
|
618 |
+
raise ValueError("Used --should_continue but no checkpoint was found in --output_dir.")
|
619 |
+
else:
|
620 |
+
args.model_name_or_path = sorted_checkpoints[-1]
|
621 |
+
|
622 |
+
if (
|
623 |
+
os.path.exists(args.output_dir)
|
624 |
+
and os.listdir(args.output_dir)
|
625 |
+
and args.do_train
|
626 |
+
and not args.overwrite_output_dir
|
627 |
+
and not args.should_continue
|
628 |
+
):
|
629 |
+
raise ValueError(
|
630 |
+
"Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
|
631 |
+
args.output_dir
|
632 |
+
)
|
633 |
+
)
|
634 |
+
|
635 |
+
# Setup distant debugging if needed
|
636 |
+
if args.server_ip and args.server_port:
|
637 |
+
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
|
638 |
+
import ptvsd
|
639 |
+
|
640 |
+
print("Waiting for debugger attach")
|
641 |
+
ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
|
642 |
+
ptvsd.wait_for_attach()
|
643 |
+
|
644 |
+
# Setup CUDA, GPU & distributed training
|
645 |
+
if args.local_rank == -1 or args.no_cuda:
|
646 |
+
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
|
647 |
+
args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
|
648 |
+
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
|
649 |
+
torch.cuda.set_device(args.local_rank)
|
650 |
+
device = torch.device("cuda", args.local_rank)
|
651 |
+
torch.distributed.init_process_group(backend="nccl")
|
652 |
+
args.n_gpu = 1
|
653 |
+
args.device = device
|
654 |
+
|
655 |
+
# Setup logging
|
656 |
+
logging.basicConfig(
|
657 |
+
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
|
658 |
+
datefmt="%m/%d/%Y %H:%M:%S",
|
659 |
+
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
|
660 |
+
)
|
661 |
+
logger.warning(
|
662 |
+
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
|
663 |
+
args.local_rank,
|
664 |
+
device,
|
665 |
+
args.n_gpu,
|
666 |
+
bool(args.local_rank != -1),
|
667 |
+
args.fp16,
|
668 |
+
)
|
669 |
+
|
670 |
+
# Set seed
|
671 |
+
set_seed(args)
|
672 |
+
|
673 |
+
# Load pretrained model and tokenizer
|
674 |
+
if args.local_rank not in [-1, 0]:
|
675 |
+
torch.distributed.barrier() # Barrier to make sure only the first process in distributed training download model & vocab
|
676 |
+
|
677 |
+
if args.config_name:
|
678 |
+
config = AutoConfig.from_pretrained(args.config_name, cache_dir=args.cache_dir)
|
679 |
+
elif args.model_name_or_path:
|
680 |
+
config = AutoConfig.from_pretrained(args.model_name_or_path, cache_dir=args.cache_dir)
|
681 |
+
else:
|
682 |
+
# When we release a pip version exposing CONFIG_MAPPING,
|
683 |
+
# we can do `config = CONFIG_MAPPING[args.model_type]()`.
|
684 |
+
raise ValueError(
|
685 |
+
"You are instantiating a new config instance from scratch. This is not supported, but you can do it from another script, save it,"
|
686 |
+
"and load it from here, using --config_name"
|
687 |
+
)
|
688 |
+
|
689 |
+
if args.tokenizer_name:
|
690 |
+
tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, cache_dir=args.cache_dir)
|
691 |
+
elif args.model_name_or_path:
|
692 |
+
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, cache_dir=args.cache_dir)
|
693 |
+
else:
|
694 |
+
raise ValueError(
|
695 |
+
"You are instantiating a new tokenizer from scratch. This is not supported, but you can do it from another script, save it,"
|
696 |
+
"and load it from here, using --tokenizer_name"
|
697 |
+
)
|
698 |
+
|
699 |
+
if args.block_size <= 0:
|
700 |
+
args.block_size = tokenizer.max_len
|
701 |
+
# Our input block size will be the max possible for the model
|
702 |
+
else:
|
703 |
+
args.block_size = min(args.block_size, tokenizer.max_len)
|
704 |
+
|
705 |
+
if args.model_name_or_path:
|
706 |
+
model = AutoModelWithLMHead.from_pretrained(
|
707 |
+
args.model_name_or_path,
|
708 |
+
from_tf=bool(".ckpt" in args.model_name_or_path),
|
709 |
+
config=config,
|
710 |
+
cache_dir=args.cache_dir,
|
711 |
+
)
|
712 |
+
else:
|
713 |
+
logger.info("Training new model from scratch")
|
714 |
+
model = AutoModelWithLMHead.from_config(config)
|
715 |
+
|
716 |
+
model.to(args.device)
|
717 |
+
|
718 |
+
if args.local_rank == 0:
|
719 |
+
torch.distributed.barrier() # End of barrier to make sure only the first process in distributed training download model & vocab
|
720 |
+
|
721 |
+
logger.info("Training/evaluation parameters %s", args)
|
722 |
+
|
723 |
+
# Training
|
724 |
+
if args.do_train:
|
725 |
+
if args.local_rank not in [-1, 0]:
|
726 |
+
torch.distributed.barrier() # Barrier to make sure only the first process in distributed training process the dataset, and the others will use the cache
|
727 |
+
|
728 |
+
train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False)
|
729 |
+
|
730 |
+
if args.local_rank == 0:
|
731 |
+
torch.distributed.barrier()
|
732 |
+
|
733 |
+
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
|
734 |
+
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
|
735 |
+
|
736 |
+
# Saving best-practices: if you use save_pretrained for the model and tokenizer, you can reload them using from_pretrained()
|
737 |
+
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
|
738 |
+
# Create output directory if needed
|
739 |
+
if args.local_rank in [-1, 0]:
|
740 |
+
os.makedirs(args.output_dir, exist_ok=True)
|
741 |
+
|
742 |
+
logger.info("Saving model checkpoint to %s", args.output_dir)
|
743 |
+
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
|
744 |
+
# They can then be reloaded using `from_pretrained()`
|
745 |
+
model_to_save = (
|
746 |
+
model.module if hasattr(model, "module") else model
|
747 |
+
) # Take care of distributed/parallel training
|
748 |
+
model_to_save.save_pretrained(args.output_dir)
|
749 |
+
tokenizer.save_pretrained(args.output_dir)
|
750 |
+
|
751 |
+
# Good practice: save your training arguments together with the trained model
|
752 |
+
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
|
753 |
+
|
754 |
+
# Load a trained model and vocabulary that you have fine-tuned
|
755 |
+
model = AutoModelWithLMHead.from_pretrained(args.output_dir)
|
756 |
+
tokenizer = AutoTokenizer.from_pretrained(args.output_dir)
|
757 |
+
model.to(args.device)
|
758 |
+
|
759 |
+
# Evaluation
|
760 |
+
results = {}
|
761 |
+
if args.do_eval and args.local_rank in [-1, 0]:
|
762 |
+
checkpoints = [args.output_dir]
|
763 |
+
if args.eval_all_checkpoints:
|
764 |
+
checkpoints = list(
|
765 |
+
os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
|
766 |
+
)
|
767 |
+
logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging
|
768 |
+
logger.info("Evaluate the following checkpoints: %s", checkpoints)
|
769 |
+
for checkpoint in checkpoints:
|
770 |
+
global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
|
771 |
+
prefix = checkpoint.split("/")[-1] if checkpoint.find("checkpoint") != -1 else ""
|
772 |
+
|
773 |
+
model = AutoModelWithLMHead.from_pretrained(checkpoint)
|
774 |
+
model.to(args.device)
|
775 |
+
result = evaluate(args, model, tokenizer, prefix=prefix)
|
776 |
+
result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
|
777 |
+
results.update(result)
|
778 |
+
|
779 |
+
return results
|
780 |
+
|
781 |
+
|
782 |
+
if __name__ == "__main__":
|
783 |
+
main()
|
runs/Dec15_14-18-19_2c94adf95c33/events.out.tfevents.1608041899.2c94adf95c33.522.0
ADDED
Binary file (1.11 kB). View file
|