Omar
commited on
Commit
·
904c81d
1
Parent(s):
0fbf7fe
update results
Browse filesThis view is limited to 50 files because it contains too many changes.
See raw diff
- config.json +30 -4
- finetune/boolq/all_results.json +16 -0
- finetune/boolq/config.json +57 -0
- finetune/boolq/eval_results.json +11 -0
- finetune/boolq/merges.txt +0 -0
- finetune/boolq/predict_results.txt +724 -0
- finetune/boolq/pytorch_model.bin +3 -0
- finetune/boolq/special_tokens_map.json +15 -0
- finetune/boolq/structformer_as_hf_no_parser.py +754 -0
- finetune/boolq/tokenizer_config.json +65 -0
- finetune/boolq/train_results.json +8 -0
- finetune/boolq/trainer_state.json +25 -0
- finetune/boolq/training_args.bin +3 -0
- finetune/boolq/vocab.json +0 -0
- finetune/cola/all_results.json +16 -0
- finetune/cola/checkpoint-400/config.json +57 -0
- finetune/cola/checkpoint-400/merges.txt +0 -0
- finetune/cola/checkpoint-400/optimizer.pt +3 -0
- finetune/cola/checkpoint-400/pytorch_model.bin +3 -0
- finetune/cola/checkpoint-400/rng_state.pth +3 -0
- finetune/cola/checkpoint-400/scheduler.pt +3 -0
- finetune/cola/checkpoint-400/special_tokens_map.json +15 -0
- finetune/cola/checkpoint-400/structformer_as_hf_no_parser.py +754 -0
- finetune/cola/checkpoint-400/tokenizer_config.json +65 -0
- finetune/cola/checkpoint-400/trainer_state.json +27 -0
- finetune/cola/checkpoint-400/training_args.bin +3 -0
- finetune/cola/checkpoint-400/vocab.json +0 -0
- finetune/cola/config.json +57 -0
- finetune/cola/eval_results.json +11 -0
- finetune/cola/merges.txt +0 -0
- finetune/cola/predict_results.txt +1020 -0
- finetune/cola/pytorch_model.bin +3 -0
- finetune/cola/special_tokens_map.json +15 -0
- finetune/cola/structformer_as_hf_no_parser.py +754 -0
- finetune/cola/tokenizer_config.json +65 -0
- finetune/cola/train_results.json +8 -0
- finetune/cola/trainer_state.json +42 -0
- finetune/cola/training_args.bin +3 -0
- finetune/cola/vocab.json +0 -0
- finetune/control_raising_control/all_results.json +16 -0
- finetune/control_raising_control/checkpoint-400/config.json +57 -0
- finetune/control_raising_control/checkpoint-400/merges.txt +0 -0
- finetune/control_raising_control/checkpoint-400/optimizer.pt +3 -0
- finetune/control_raising_control/checkpoint-400/pytorch_model.bin +3 -0
- finetune/control_raising_control/checkpoint-400/rng_state.pth +3 -0
- finetune/control_raising_control/checkpoint-400/scheduler.pt +3 -0
- finetune/control_raising_control/checkpoint-400/special_tokens_map.json +15 -0
- finetune/control_raising_control/checkpoint-400/structformer_as_hf_no_parser.py +754 -0
- finetune/control_raising_control/checkpoint-400/tokenizer_config.json +65 -0
- finetune/control_raising_control/checkpoint-400/trainer_state.json +27 -0
config.json
CHANGED
|
@@ -1,11 +1,15 @@
|
|
| 1 |
{
|
| 2 |
"architectures": [
|
| 3 |
-
"StructformerModel"
|
|
|
|
| 4 |
],
|
|
|
|
| 5 |
"auto_map": {
|
| 6 |
"AutoConfig": "structformer_as_hf_no_parser.StructformerConfig",
|
| 7 |
-
"AutoModelForMaskedLM": "structformer_as_hf_no_parser.StructformerModel"
|
|
|
|
| 8 |
},
|
|
|
|
| 9 |
"conv_size": 9,
|
| 10 |
"dropatt": 0.1,
|
| 11 |
"dropout": 0.1,
|
|
@@ -25,5 +29,27 @@
|
|
| 25 |
"relative_bias": false,
|
| 26 |
"torch_dtype": "float32",
|
| 27 |
"transformers_version": "4.18.0",
|
| 28 |
-
"weight_act": "softmax"
|
| 29 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
{
|
| 2 |
"architectures": [
|
| 3 |
+
"StructformerModel",
|
| 4 |
+
"StructformerModelForSequenceClassification"
|
| 5 |
],
|
| 6 |
+
"attention_probs_dropout_prob": 0.1,
|
| 7 |
"auto_map": {
|
| 8 |
"AutoConfig": "structformer_as_hf_no_parser.StructformerConfig",
|
| 9 |
+
"AutoModelForMaskedLM": "structformer_as_hf_no_parser.StructformerModel",
|
| 10 |
+
"AutoModelForSequenceClassification": "structformer_as_hf_no_parser.StructformerModelForSequenceClassification"
|
| 11 |
},
|
| 12 |
+
|
| 13 |
"conv_size": 9,
|
| 14 |
"dropatt": 0.1,
|
| 15 |
"dropout": 0.1,
|
|
|
|
| 29 |
"relative_bias": false,
|
| 30 |
"torch_dtype": "float32",
|
| 31 |
"transformers_version": "4.18.0",
|
| 32 |
+
"weight_act": "softmax",
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
"bos_token_id": 0,
|
| 36 |
+
"classifier_dropout": null,
|
| 37 |
+
|
| 38 |
+
"eos_token_id": 2,
|
| 39 |
+
"hidden_act": "gelu",
|
| 40 |
+
"hidden_dropout_prob": 0.1,
|
| 41 |
+
"initializer_range": 0.02,
|
| 42 |
+
"intermediate_size": 3072,
|
| 43 |
+
"layer_norm_eps": 1e-05,
|
| 44 |
+
"max_position_embeddings": 514,
|
| 45 |
+
|
| 46 |
+
"num_attention_heads": 12,
|
| 47 |
+
"num_hidden_layers": 12,
|
| 48 |
+
"pad_token_id": 1,
|
| 49 |
+
"position_embedding_type": "absolute",
|
| 50 |
+
|
| 51 |
+
"type_vocab_size": 1,
|
| 52 |
+
"use_cache": true,
|
| 53 |
+
"vocab_size": 32000
|
| 54 |
+
|
| 55 |
+
}
|
finetune/boolq/all_results.json
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"epoch": 10.0,
|
| 3 |
+
"eval_accuracy": 0.6293222904205322,
|
| 4 |
+
"eval_f1": 0.7067833698030634,
|
| 5 |
+
"eval_loss": 1.2662478685379028,
|
| 6 |
+
"eval_mcc": 0.20890142408340617,
|
| 7 |
+
"eval_runtime": 1.0109,
|
| 8 |
+
"eval_samples": 723,
|
| 9 |
+
"eval_samples_per_second": 715.225,
|
| 10 |
+
"eval_steps_per_second": 90.021,
|
| 11 |
+
"train_loss": 0.33231258392333984,
|
| 12 |
+
"train_runtime": 67.9575,
|
| 13 |
+
"train_samples": 2072,
|
| 14 |
+
"train_samples_per_second": 304.896,
|
| 15 |
+
"train_steps_per_second": 2.649
|
| 16 |
+
}
|
finetune/boolq/config.json
ADDED
|
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_name_or_path": "final_models/transformer_base_final_2",
|
| 3 |
+
"architectures": [
|
| 4 |
+
"StructformerModelForSequenceClassification"
|
| 5 |
+
],
|
| 6 |
+
"attention_probs_dropout_prob": 0.1,
|
| 7 |
+
"auto_map": {
|
| 8 |
+
"AutoConfig": "structformer_as_hf_no_parser.StructformerConfig",
|
| 9 |
+
"AutoModelForMaskedLM": "structformer_as_hf_no_parser.StructformerModel",
|
| 10 |
+
"AutoModelForSequenceClassification": "structformer_as_hf_no_parser.StructformerModelForSequenceClassification"
|
| 11 |
+
},
|
| 12 |
+
"bos_token_id": 0,
|
| 13 |
+
"classifier_dropout": null,
|
| 14 |
+
"conv_size": 9,
|
| 15 |
+
"dropatt": 0.1,
|
| 16 |
+
"dropout": 0.1,
|
| 17 |
+
"eos_token_id": 2,
|
| 18 |
+
"hidden_act": "gelu",
|
| 19 |
+
"hidden_dropout_prob": 0.1,
|
| 20 |
+
"hidden_size": 768,
|
| 21 |
+
"id2label": {
|
| 22 |
+
"0": 0,
|
| 23 |
+
"1": 1
|
| 24 |
+
},
|
| 25 |
+
"initializer_range": 0.02,
|
| 26 |
+
"intermediate_size": 3072,
|
| 27 |
+
"label2id": {
|
| 28 |
+
"0": 0,
|
| 29 |
+
"1": 1
|
| 30 |
+
},
|
| 31 |
+
"layer_norm_eps": 1e-05,
|
| 32 |
+
"max_position_embeddings": 514,
|
| 33 |
+
"model_type": "structformer",
|
| 34 |
+
"n_context_layers": 0,
|
| 35 |
+
"n_parser_layers": 0,
|
| 36 |
+
"nhead": 12,
|
| 37 |
+
"nlayers": 12,
|
| 38 |
+
"ntokens": 32000,
|
| 39 |
+
"num_attention_heads": 12,
|
| 40 |
+
"num_hidden_layers": 12,
|
| 41 |
+
"pad": 0,
|
| 42 |
+
"pad_token_id": 1,
|
| 43 |
+
"pos_emb": true,
|
| 44 |
+
"position_embedding_type": "absolute",
|
| 45 |
+
"problem_type": "single_label_classification",
|
| 46 |
+
"relations": [
|
| 47 |
+
"head",
|
| 48 |
+
"child"
|
| 49 |
+
],
|
| 50 |
+
"relative_bias": false,
|
| 51 |
+
"torch_dtype": "float32",
|
| 52 |
+
"transformers_version": "4.26.1",
|
| 53 |
+
"type_vocab_size": 1,
|
| 54 |
+
"use_cache": true,
|
| 55 |
+
"vocab_size": 32000,
|
| 56 |
+
"weight_act": "softmax"
|
| 57 |
+
}
|
finetune/boolq/eval_results.json
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"epoch": 10.0,
|
| 3 |
+
"eval_accuracy": 0.6293222904205322,
|
| 4 |
+
"eval_f1": 0.7067833698030634,
|
| 5 |
+
"eval_loss": 1.2662478685379028,
|
| 6 |
+
"eval_mcc": 0.20890142408340617,
|
| 7 |
+
"eval_runtime": 1.0109,
|
| 8 |
+
"eval_samples": 723,
|
| 9 |
+
"eval_samples_per_second": 715.225,
|
| 10 |
+
"eval_steps_per_second": 90.021
|
| 11 |
+
}
|
finetune/boolq/merges.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
finetune/boolq/predict_results.txt
ADDED
|
@@ -0,0 +1,724 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
index prediction
|
| 2 |
+
0 0
|
| 3 |
+
1 0
|
| 4 |
+
2 1
|
| 5 |
+
3 0
|
| 6 |
+
4 1
|
| 7 |
+
5 0
|
| 8 |
+
6 1
|
| 9 |
+
7 0
|
| 10 |
+
8 1
|
| 11 |
+
9 0
|
| 12 |
+
10 0
|
| 13 |
+
11 1
|
| 14 |
+
12 0
|
| 15 |
+
13 1
|
| 16 |
+
14 1
|
| 17 |
+
15 1
|
| 18 |
+
16 0
|
| 19 |
+
17 1
|
| 20 |
+
18 1
|
| 21 |
+
19 1
|
| 22 |
+
20 0
|
| 23 |
+
21 0
|
| 24 |
+
22 1
|
| 25 |
+
23 0
|
| 26 |
+
24 1
|
| 27 |
+
25 0
|
| 28 |
+
26 0
|
| 29 |
+
27 1
|
| 30 |
+
28 1
|
| 31 |
+
29 1
|
| 32 |
+
30 0
|
| 33 |
+
31 0
|
| 34 |
+
32 0
|
| 35 |
+
33 1
|
| 36 |
+
34 0
|
| 37 |
+
35 1
|
| 38 |
+
36 1
|
| 39 |
+
37 0
|
| 40 |
+
38 1
|
| 41 |
+
39 1
|
| 42 |
+
40 0
|
| 43 |
+
41 1
|
| 44 |
+
42 1
|
| 45 |
+
43 1
|
| 46 |
+
44 1
|
| 47 |
+
45 1
|
| 48 |
+
46 1
|
| 49 |
+
47 0
|
| 50 |
+
48 1
|
| 51 |
+
49 1
|
| 52 |
+
50 0
|
| 53 |
+
51 1
|
| 54 |
+
52 1
|
| 55 |
+
53 1
|
| 56 |
+
54 0
|
| 57 |
+
55 0
|
| 58 |
+
56 1
|
| 59 |
+
57 0
|
| 60 |
+
58 0
|
| 61 |
+
59 1
|
| 62 |
+
60 1
|
| 63 |
+
61 0
|
| 64 |
+
62 0
|
| 65 |
+
63 1
|
| 66 |
+
64 1
|
| 67 |
+
65 1
|
| 68 |
+
66 1
|
| 69 |
+
67 1
|
| 70 |
+
68 0
|
| 71 |
+
69 1
|
| 72 |
+
70 0
|
| 73 |
+
71 1
|
| 74 |
+
72 0
|
| 75 |
+
73 1
|
| 76 |
+
74 0
|
| 77 |
+
75 1
|
| 78 |
+
76 0
|
| 79 |
+
77 0
|
| 80 |
+
78 1
|
| 81 |
+
79 0
|
| 82 |
+
80 1
|
| 83 |
+
81 1
|
| 84 |
+
82 0
|
| 85 |
+
83 1
|
| 86 |
+
84 1
|
| 87 |
+
85 1
|
| 88 |
+
86 1
|
| 89 |
+
87 1
|
| 90 |
+
88 1
|
| 91 |
+
89 0
|
| 92 |
+
90 1
|
| 93 |
+
91 1
|
| 94 |
+
92 1
|
| 95 |
+
93 1
|
| 96 |
+
94 0
|
| 97 |
+
95 1
|
| 98 |
+
96 1
|
| 99 |
+
97 1
|
| 100 |
+
98 1
|
| 101 |
+
99 1
|
| 102 |
+
100 1
|
| 103 |
+
101 0
|
| 104 |
+
102 0
|
| 105 |
+
103 0
|
| 106 |
+
104 1
|
| 107 |
+
105 1
|
| 108 |
+
106 1
|
| 109 |
+
107 0
|
| 110 |
+
108 0
|
| 111 |
+
109 1
|
| 112 |
+
110 1
|
| 113 |
+
111 1
|
| 114 |
+
112 0
|
| 115 |
+
113 1
|
| 116 |
+
114 1
|
| 117 |
+
115 1
|
| 118 |
+
116 1
|
| 119 |
+
117 0
|
| 120 |
+
118 1
|
| 121 |
+
119 1
|
| 122 |
+
120 1
|
| 123 |
+
121 0
|
| 124 |
+
122 1
|
| 125 |
+
123 0
|
| 126 |
+
124 0
|
| 127 |
+
125 1
|
| 128 |
+
126 1
|
| 129 |
+
127 0
|
| 130 |
+
128 1
|
| 131 |
+
129 0
|
| 132 |
+
130 0
|
| 133 |
+
131 0
|
| 134 |
+
132 0
|
| 135 |
+
133 1
|
| 136 |
+
134 1
|
| 137 |
+
135 1
|
| 138 |
+
136 1
|
| 139 |
+
137 0
|
| 140 |
+
138 0
|
| 141 |
+
139 1
|
| 142 |
+
140 0
|
| 143 |
+
141 1
|
| 144 |
+
142 1
|
| 145 |
+
143 1
|
| 146 |
+
144 0
|
| 147 |
+
145 1
|
| 148 |
+
146 1
|
| 149 |
+
147 1
|
| 150 |
+
148 1
|
| 151 |
+
149 1
|
| 152 |
+
150 0
|
| 153 |
+
151 1
|
| 154 |
+
152 1
|
| 155 |
+
153 1
|
| 156 |
+
154 1
|
| 157 |
+
155 0
|
| 158 |
+
156 0
|
| 159 |
+
157 0
|
| 160 |
+
158 0
|
| 161 |
+
159 1
|
| 162 |
+
160 1
|
| 163 |
+
161 1
|
| 164 |
+
162 1
|
| 165 |
+
163 1
|
| 166 |
+
164 0
|
| 167 |
+
165 1
|
| 168 |
+
166 1
|
| 169 |
+
167 1
|
| 170 |
+
168 0
|
| 171 |
+
169 0
|
| 172 |
+
170 1
|
| 173 |
+
171 0
|
| 174 |
+
172 0
|
| 175 |
+
173 1
|
| 176 |
+
174 0
|
| 177 |
+
175 1
|
| 178 |
+
176 1
|
| 179 |
+
177 1
|
| 180 |
+
178 0
|
| 181 |
+
179 1
|
| 182 |
+
180 0
|
| 183 |
+
181 1
|
| 184 |
+
182 0
|
| 185 |
+
183 1
|
| 186 |
+
184 0
|
| 187 |
+
185 1
|
| 188 |
+
186 0
|
| 189 |
+
187 0
|
| 190 |
+
188 1
|
| 191 |
+
189 1
|
| 192 |
+
190 1
|
| 193 |
+
191 0
|
| 194 |
+
192 0
|
| 195 |
+
193 0
|
| 196 |
+
194 1
|
| 197 |
+
195 0
|
| 198 |
+
196 1
|
| 199 |
+
197 1
|
| 200 |
+
198 0
|
| 201 |
+
199 1
|
| 202 |
+
200 1
|
| 203 |
+
201 1
|
| 204 |
+
202 1
|
| 205 |
+
203 1
|
| 206 |
+
204 1
|
| 207 |
+
205 1
|
| 208 |
+
206 1
|
| 209 |
+
207 1
|
| 210 |
+
208 0
|
| 211 |
+
209 1
|
| 212 |
+
210 1
|
| 213 |
+
211 1
|
| 214 |
+
212 1
|
| 215 |
+
213 0
|
| 216 |
+
214 1
|
| 217 |
+
215 0
|
| 218 |
+
216 1
|
| 219 |
+
217 0
|
| 220 |
+
218 1
|
| 221 |
+
219 1
|
| 222 |
+
220 1
|
| 223 |
+
221 0
|
| 224 |
+
222 1
|
| 225 |
+
223 0
|
| 226 |
+
224 0
|
| 227 |
+
225 0
|
| 228 |
+
226 1
|
| 229 |
+
227 0
|
| 230 |
+
228 0
|
| 231 |
+
229 1
|
| 232 |
+
230 0
|
| 233 |
+
231 1
|
| 234 |
+
232 1
|
| 235 |
+
233 0
|
| 236 |
+
234 0
|
| 237 |
+
235 1
|
| 238 |
+
236 1
|
| 239 |
+
237 1
|
| 240 |
+
238 1
|
| 241 |
+
239 1
|
| 242 |
+
240 1
|
| 243 |
+
241 1
|
| 244 |
+
242 1
|
| 245 |
+
243 1
|
| 246 |
+
244 0
|
| 247 |
+
245 1
|
| 248 |
+
246 1
|
| 249 |
+
247 0
|
| 250 |
+
248 1
|
| 251 |
+
249 1
|
| 252 |
+
250 1
|
| 253 |
+
251 0
|
| 254 |
+
252 1
|
| 255 |
+
253 1
|
| 256 |
+
254 1
|
| 257 |
+
255 1
|
| 258 |
+
256 1
|
| 259 |
+
257 1
|
| 260 |
+
258 1
|
| 261 |
+
259 0
|
| 262 |
+
260 1
|
| 263 |
+
261 0
|
| 264 |
+
262 1
|
| 265 |
+
263 1
|
| 266 |
+
264 0
|
| 267 |
+
265 1
|
| 268 |
+
266 1
|
| 269 |
+
267 0
|
| 270 |
+
268 1
|
| 271 |
+
269 1
|
| 272 |
+
270 0
|
| 273 |
+
271 1
|
| 274 |
+
272 1
|
| 275 |
+
273 1
|
| 276 |
+
274 1
|
| 277 |
+
275 1
|
| 278 |
+
276 1
|
| 279 |
+
277 0
|
| 280 |
+
278 0
|
| 281 |
+
279 1
|
| 282 |
+
280 1
|
| 283 |
+
281 1
|
| 284 |
+
282 0
|
| 285 |
+
283 0
|
| 286 |
+
284 0
|
| 287 |
+
285 1
|
| 288 |
+
286 1
|
| 289 |
+
287 1
|
| 290 |
+
288 1
|
| 291 |
+
289 0
|
| 292 |
+
290 0
|
| 293 |
+
291 0
|
| 294 |
+
292 1
|
| 295 |
+
293 1
|
| 296 |
+
294 1
|
| 297 |
+
295 1
|
| 298 |
+
296 1
|
| 299 |
+
297 1
|
| 300 |
+
298 1
|
| 301 |
+
299 1
|
| 302 |
+
300 1
|
| 303 |
+
301 1
|
| 304 |
+
302 1
|
| 305 |
+
303 0
|
| 306 |
+
304 1
|
| 307 |
+
305 0
|
| 308 |
+
306 0
|
| 309 |
+
307 0
|
| 310 |
+
308 1
|
| 311 |
+
309 0
|
| 312 |
+
310 0
|
| 313 |
+
311 1
|
| 314 |
+
312 1
|
| 315 |
+
313 1
|
| 316 |
+
314 1
|
| 317 |
+
315 1
|
| 318 |
+
316 1
|
| 319 |
+
317 1
|
| 320 |
+
318 1
|
| 321 |
+
319 1
|
| 322 |
+
320 1
|
| 323 |
+
321 0
|
| 324 |
+
322 1
|
| 325 |
+
323 1
|
| 326 |
+
324 1
|
| 327 |
+
325 0
|
| 328 |
+
326 1
|
| 329 |
+
327 1
|
| 330 |
+
328 1
|
| 331 |
+
329 1
|
| 332 |
+
330 0
|
| 333 |
+
331 0
|
| 334 |
+
332 1
|
| 335 |
+
333 1
|
| 336 |
+
334 0
|
| 337 |
+
335 1
|
| 338 |
+
336 0
|
| 339 |
+
337 1
|
| 340 |
+
338 1
|
| 341 |
+
339 1
|
| 342 |
+
340 1
|
| 343 |
+
341 0
|
| 344 |
+
342 1
|
| 345 |
+
343 0
|
| 346 |
+
344 0
|
| 347 |
+
345 1
|
| 348 |
+
346 1
|
| 349 |
+
347 1
|
| 350 |
+
348 1
|
| 351 |
+
349 1
|
| 352 |
+
350 1
|
| 353 |
+
351 1
|
| 354 |
+
352 1
|
| 355 |
+
353 1
|
| 356 |
+
354 0
|
| 357 |
+
355 1
|
| 358 |
+
356 0
|
| 359 |
+
357 1
|
| 360 |
+
358 1
|
| 361 |
+
359 0
|
| 362 |
+
360 1
|
| 363 |
+
361 1
|
| 364 |
+
362 0
|
| 365 |
+
363 1
|
| 366 |
+
364 0
|
| 367 |
+
365 1
|
| 368 |
+
366 1
|
| 369 |
+
367 1
|
| 370 |
+
368 1
|
| 371 |
+
369 1
|
| 372 |
+
370 0
|
| 373 |
+
371 0
|
| 374 |
+
372 1
|
| 375 |
+
373 1
|
| 376 |
+
374 1
|
| 377 |
+
375 1
|
| 378 |
+
376 1
|
| 379 |
+
377 1
|
| 380 |
+
378 1
|
| 381 |
+
379 0
|
| 382 |
+
380 0
|
| 383 |
+
381 0
|
| 384 |
+
382 1
|
| 385 |
+
383 1
|
| 386 |
+
384 1
|
| 387 |
+
385 0
|
| 388 |
+
386 1
|
| 389 |
+
387 0
|
| 390 |
+
388 1
|
| 391 |
+
389 1
|
| 392 |
+
390 1
|
| 393 |
+
391 0
|
| 394 |
+
392 0
|
| 395 |
+
393 0
|
| 396 |
+
394 1
|
| 397 |
+
395 1
|
| 398 |
+
396 1
|
| 399 |
+
397 1
|
| 400 |
+
398 0
|
| 401 |
+
399 0
|
| 402 |
+
400 1
|
| 403 |
+
401 0
|
| 404 |
+
402 1
|
| 405 |
+
403 1
|
| 406 |
+
404 0
|
| 407 |
+
405 1
|
| 408 |
+
406 0
|
| 409 |
+
407 0
|
| 410 |
+
408 1
|
| 411 |
+
409 1
|
| 412 |
+
410 1
|
| 413 |
+
411 1
|
| 414 |
+
412 0
|
| 415 |
+
413 1
|
| 416 |
+
414 1
|
| 417 |
+
415 0
|
| 418 |
+
416 1
|
| 419 |
+
417 1
|
| 420 |
+
418 0
|
| 421 |
+
419 1
|
| 422 |
+
420 1
|
| 423 |
+
421 1
|
| 424 |
+
422 1
|
| 425 |
+
423 1
|
| 426 |
+
424 1
|
| 427 |
+
425 0
|
| 428 |
+
426 1
|
| 429 |
+
427 1
|
| 430 |
+
428 0
|
| 431 |
+
429 1
|
| 432 |
+
430 1
|
| 433 |
+
431 0
|
| 434 |
+
432 1
|
| 435 |
+
433 1
|
| 436 |
+
434 0
|
| 437 |
+
435 0
|
| 438 |
+
436 1
|
| 439 |
+
437 1
|
| 440 |
+
438 0
|
| 441 |
+
439 0
|
| 442 |
+
440 0
|
| 443 |
+
441 0
|
| 444 |
+
442 1
|
| 445 |
+
443 0
|
| 446 |
+
444 1
|
| 447 |
+
445 1
|
| 448 |
+
446 1
|
| 449 |
+
447 0
|
| 450 |
+
448 1
|
| 451 |
+
449 1
|
| 452 |
+
450 1
|
| 453 |
+
451 0
|
| 454 |
+
452 1
|
| 455 |
+
453 1
|
| 456 |
+
454 1
|
| 457 |
+
455 0
|
| 458 |
+
456 1
|
| 459 |
+
457 0
|
| 460 |
+
458 1
|
| 461 |
+
459 1
|
| 462 |
+
460 1
|
| 463 |
+
461 1
|
| 464 |
+
462 0
|
| 465 |
+
463 0
|
| 466 |
+
464 0
|
| 467 |
+
465 0
|
| 468 |
+
466 1
|
| 469 |
+
467 1
|
| 470 |
+
468 1
|
| 471 |
+
469 0
|
| 472 |
+
470 1
|
| 473 |
+
471 1
|
| 474 |
+
472 1
|
| 475 |
+
473 1
|
| 476 |
+
474 1
|
| 477 |
+
475 0
|
| 478 |
+
476 0
|
| 479 |
+
477 1
|
| 480 |
+
478 0
|
| 481 |
+
479 1
|
| 482 |
+
480 1
|
| 483 |
+
481 1
|
| 484 |
+
482 0
|
| 485 |
+
483 1
|
| 486 |
+
484 1
|
| 487 |
+
485 1
|
| 488 |
+
486 1
|
| 489 |
+
487 1
|
| 490 |
+
488 1
|
| 491 |
+
489 0
|
| 492 |
+
490 1
|
| 493 |
+
491 1
|
| 494 |
+
492 0
|
| 495 |
+
493 1
|
| 496 |
+
494 1
|
| 497 |
+
495 1
|
| 498 |
+
496 1
|
| 499 |
+
497 1
|
| 500 |
+
498 1
|
| 501 |
+
499 1
|
| 502 |
+
500 1
|
| 503 |
+
501 0
|
| 504 |
+
502 0
|
| 505 |
+
503 1
|
| 506 |
+
504 0
|
| 507 |
+
505 1
|
| 508 |
+
506 1
|
| 509 |
+
507 1
|
| 510 |
+
508 1
|
| 511 |
+
509 0
|
| 512 |
+
510 1
|
| 513 |
+
511 1
|
| 514 |
+
512 1
|
| 515 |
+
513 1
|
| 516 |
+
514 0
|
| 517 |
+
515 0
|
| 518 |
+
516 1
|
| 519 |
+
517 1
|
| 520 |
+
518 1
|
| 521 |
+
519 1
|
| 522 |
+
520 1
|
| 523 |
+
521 1
|
| 524 |
+
522 0
|
| 525 |
+
523 1
|
| 526 |
+
524 1
|
| 527 |
+
525 1
|
| 528 |
+
526 1
|
| 529 |
+
527 1
|
| 530 |
+
528 0
|
| 531 |
+
529 0
|
| 532 |
+
530 1
|
| 533 |
+
531 0
|
| 534 |
+
532 0
|
| 535 |
+
533 1
|
| 536 |
+
534 0
|
| 537 |
+
535 0
|
| 538 |
+
536 1
|
| 539 |
+
537 0
|
| 540 |
+
538 0
|
| 541 |
+
539 1
|
| 542 |
+
540 0
|
| 543 |
+
541 1
|
| 544 |
+
542 1
|
| 545 |
+
543 1
|
| 546 |
+
544 0
|
| 547 |
+
545 1
|
| 548 |
+
546 1
|
| 549 |
+
547 1
|
| 550 |
+
548 1
|
| 551 |
+
549 1
|
| 552 |
+
550 1
|
| 553 |
+
551 0
|
| 554 |
+
552 0
|
| 555 |
+
553 1
|
| 556 |
+
554 0
|
| 557 |
+
555 0
|
| 558 |
+
556 1
|
| 559 |
+
557 0
|
| 560 |
+
558 0
|
| 561 |
+
559 1
|
| 562 |
+
560 0
|
| 563 |
+
561 1
|
| 564 |
+
562 1
|
| 565 |
+
563 1
|
| 566 |
+
564 1
|
| 567 |
+
565 0
|
| 568 |
+
566 1
|
| 569 |
+
567 0
|
| 570 |
+
568 1
|
| 571 |
+
569 1
|
| 572 |
+
570 1
|
| 573 |
+
571 1
|
| 574 |
+
572 1
|
| 575 |
+
573 1
|
| 576 |
+
574 0
|
| 577 |
+
575 1
|
| 578 |
+
576 1
|
| 579 |
+
577 1
|
| 580 |
+
578 1
|
| 581 |
+
579 1
|
| 582 |
+
580 0
|
| 583 |
+
581 0
|
| 584 |
+
582 1
|
| 585 |
+
583 1
|
| 586 |
+
584 0
|
| 587 |
+
585 1
|
| 588 |
+
586 1
|
| 589 |
+
587 1
|
| 590 |
+
588 0
|
| 591 |
+
589 1
|
| 592 |
+
590 0
|
| 593 |
+
591 1
|
| 594 |
+
592 1
|
| 595 |
+
593 1
|
| 596 |
+
594 1
|
| 597 |
+
595 0
|
| 598 |
+
596 0
|
| 599 |
+
597 1
|
| 600 |
+
598 0
|
| 601 |
+
599 1
|
| 602 |
+
600 1
|
| 603 |
+
601 0
|
| 604 |
+
602 1
|
| 605 |
+
603 1
|
| 606 |
+
604 1
|
| 607 |
+
605 0
|
| 608 |
+
606 0
|
| 609 |
+
607 1
|
| 610 |
+
608 1
|
| 611 |
+
609 1
|
| 612 |
+
610 1
|
| 613 |
+
611 1
|
| 614 |
+
612 1
|
| 615 |
+
613 1
|
| 616 |
+
614 1
|
| 617 |
+
615 0
|
| 618 |
+
616 1
|
| 619 |
+
617 1
|
| 620 |
+
618 1
|
| 621 |
+
619 1
|
| 622 |
+
620 1
|
| 623 |
+
621 1
|
| 624 |
+
622 1
|
| 625 |
+
623 1
|
| 626 |
+
624 1
|
| 627 |
+
625 0
|
| 628 |
+
626 1
|
| 629 |
+
627 1
|
| 630 |
+
628 1
|
| 631 |
+
629 0
|
| 632 |
+
630 1
|
| 633 |
+
631 1
|
| 634 |
+
632 0
|
| 635 |
+
633 1
|
| 636 |
+
634 1
|
| 637 |
+
635 0
|
| 638 |
+
636 1
|
| 639 |
+
637 1
|
| 640 |
+
638 1
|
| 641 |
+
639 1
|
| 642 |
+
640 1
|
| 643 |
+
641 1
|
| 644 |
+
642 1
|
| 645 |
+
643 0
|
| 646 |
+
644 0
|
| 647 |
+
645 1
|
| 648 |
+
646 1
|
| 649 |
+
647 0
|
| 650 |
+
648 0
|
| 651 |
+
649 1
|
| 652 |
+
650 1
|
| 653 |
+
651 1
|
| 654 |
+
652 1
|
| 655 |
+
653 0
|
| 656 |
+
654 0
|
| 657 |
+
655 1
|
| 658 |
+
656 1
|
| 659 |
+
657 0
|
| 660 |
+
658 1
|
| 661 |
+
659 1
|
| 662 |
+
660 1
|
| 663 |
+
661 0
|
| 664 |
+
662 1
|
| 665 |
+
663 1
|
| 666 |
+
664 1
|
| 667 |
+
665 1
|
| 668 |
+
666 0
|
| 669 |
+
667 1
|
| 670 |
+
668 0
|
| 671 |
+
669 1
|
| 672 |
+
670 1
|
| 673 |
+
671 1
|
| 674 |
+
672 1
|
| 675 |
+
673 0
|
| 676 |
+
674 1
|
| 677 |
+
675 0
|
| 678 |
+
676 1
|
| 679 |
+
677 1
|
| 680 |
+
678 1
|
| 681 |
+
679 1
|
| 682 |
+
680 0
|
| 683 |
+
681 1
|
| 684 |
+
682 1
|
| 685 |
+
683 1
|
| 686 |
+
684 0
|
| 687 |
+
685 1
|
| 688 |
+
686 0
|
| 689 |
+
687 1
|
| 690 |
+
688 0
|
| 691 |
+
689 1
|
| 692 |
+
690 1
|
| 693 |
+
691 0
|
| 694 |
+
692 0
|
| 695 |
+
693 1
|
| 696 |
+
694 1
|
| 697 |
+
695 1
|
| 698 |
+
696 0
|
| 699 |
+
697 1
|
| 700 |
+
698 1
|
| 701 |
+
699 1
|
| 702 |
+
700 0
|
| 703 |
+
701 1
|
| 704 |
+
702 1
|
| 705 |
+
703 1
|
| 706 |
+
704 1
|
| 707 |
+
705 1
|
| 708 |
+
706 1
|
| 709 |
+
707 1
|
| 710 |
+
708 1
|
| 711 |
+
709 1
|
| 712 |
+
710 1
|
| 713 |
+
711 0
|
| 714 |
+
712 0
|
| 715 |
+
713 0
|
| 716 |
+
714 0
|
| 717 |
+
715 1
|
| 718 |
+
716 0
|
| 719 |
+
717 1
|
| 720 |
+
718 0
|
| 721 |
+
719 0
|
| 722 |
+
720 1
|
| 723 |
+
721 1
|
| 724 |
+
722 1
|
finetune/boolq/pytorch_model.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c5483030e47ecf2b2adcbfba33ea2c1741db1d170a0f2b751457e63d49fd38c8
|
| 3 |
+
size 442624943
|
finetune/boolq/special_tokens_map.json
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bos_token": "<s>",
|
| 3 |
+
"cls_token": "<s>",
|
| 4 |
+
"eos_token": "</s>",
|
| 5 |
+
"mask_token": {
|
| 6 |
+
"content": "<mask>",
|
| 7 |
+
"lstrip": true,
|
| 8 |
+
"normalized": false,
|
| 9 |
+
"rstrip": false,
|
| 10 |
+
"single_word": false
|
| 11 |
+
},
|
| 12 |
+
"pad_token": "<pad>",
|
| 13 |
+
"sep_token": "</s>",
|
| 14 |
+
"unk_token": "<unk>"
|
| 15 |
+
}
|
finetune/boolq/structformer_as_hf_no_parser.py
ADDED
|
@@ -0,0 +1,754 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import torch.nn as nn
|
| 3 |
+
import torch.nn.functional as F
|
| 4 |
+
from torch.nn import init
|
| 5 |
+
from transformers import PreTrainedModel
|
| 6 |
+
from transformers import PretrainedConfig
|
| 7 |
+
from transformers.modeling_outputs import MaskedLMOutput
|
| 8 |
+
from typing import List
|
| 9 |
+
from torch.nn import CrossEntropyLoss, MSELoss, BCEWithLogitsLoss
|
| 10 |
+
from transformers.modeling_outputs import (
|
| 11 |
+
BaseModelOutputWithPastAndCrossAttentions,
|
| 12 |
+
BaseModelOutputWithPoolingAndCrossAttentions,
|
| 13 |
+
MaskedLMOutput,
|
| 14 |
+
SequenceClassifierOutput
|
| 15 |
+
)
|
| 16 |
+
|
| 17 |
+
##########################################
|
| 18 |
+
# HuggingFace Config
|
| 19 |
+
##########################################
|
| 20 |
+
class StructformerConfig(PretrainedConfig):
|
| 21 |
+
model_type = "structformer"
|
| 22 |
+
|
| 23 |
+
def __init__(
|
| 24 |
+
self,
|
| 25 |
+
hidden_size=768,
|
| 26 |
+
n_context_layers=2,
|
| 27 |
+
nlayers=6,
|
| 28 |
+
ntokens=32000,
|
| 29 |
+
nhead=8,
|
| 30 |
+
dropout=0.1,
|
| 31 |
+
dropatt=0.1,
|
| 32 |
+
relative_bias=False,
|
| 33 |
+
pos_emb=False,
|
| 34 |
+
pad=0,
|
| 35 |
+
n_parser_layers=4,
|
| 36 |
+
conv_size=9,
|
| 37 |
+
relations=('head', 'child'),
|
| 38 |
+
weight_act='softmax',
|
| 39 |
+
**kwargs,
|
| 40 |
+
):
|
| 41 |
+
self.hidden_size = hidden_size
|
| 42 |
+
self.n_context_layers = n_context_layers
|
| 43 |
+
self.nlayers = nlayers
|
| 44 |
+
self.ntokens = ntokens
|
| 45 |
+
self.nhead = nhead
|
| 46 |
+
self.dropout = dropout
|
| 47 |
+
self.dropatt = dropatt
|
| 48 |
+
self.relative_bias = relative_bias
|
| 49 |
+
self.pos_emb = pos_emb
|
| 50 |
+
self.pad = pad
|
| 51 |
+
self.n_parser_layers = n_parser_layers
|
| 52 |
+
self.conv_size = conv_size
|
| 53 |
+
self.relations = relations
|
| 54 |
+
self.weight_act = weight_act
|
| 55 |
+
super().__init__(**kwargs)
|
| 56 |
+
|
| 57 |
+
##########################################
|
| 58 |
+
# Custom Layers
|
| 59 |
+
##########################################
|
| 60 |
+
def _get_activation_fn(activation):
|
| 61 |
+
"""Get specified activation function."""
|
| 62 |
+
if activation == "relu":
|
| 63 |
+
return nn.ReLU()
|
| 64 |
+
elif activation == "gelu":
|
| 65 |
+
return nn.GELU()
|
| 66 |
+
elif activation == "leakyrelu":
|
| 67 |
+
return nn.LeakyReLU()
|
| 68 |
+
|
| 69 |
+
raise RuntimeError(
|
| 70 |
+
"activation should be relu/gelu, not {}".format(activation))
|
| 71 |
+
|
| 72 |
+
class Conv1d(nn.Module):
|
| 73 |
+
"""1D convolution layer."""
|
| 74 |
+
|
| 75 |
+
def __init__(self, hidden_size, kernel_size, dilation=1):
|
| 76 |
+
"""Initialization.
|
| 77 |
+
|
| 78 |
+
Args:
|
| 79 |
+
hidden_size: dimension of input embeddings
|
| 80 |
+
kernel_size: convolution kernel size
|
| 81 |
+
dilation: the spacing between the kernel points
|
| 82 |
+
"""
|
| 83 |
+
super(Conv1d, self).__init__()
|
| 84 |
+
|
| 85 |
+
if kernel_size % 2 == 0:
|
| 86 |
+
padding = (kernel_size // 2) * dilation
|
| 87 |
+
self.shift = True
|
| 88 |
+
else:
|
| 89 |
+
padding = ((kernel_size - 1) // 2) * dilation
|
| 90 |
+
self.shift = False
|
| 91 |
+
self.conv = nn.Conv1d(
|
| 92 |
+
hidden_size,
|
| 93 |
+
hidden_size,
|
| 94 |
+
kernel_size,
|
| 95 |
+
padding=padding,
|
| 96 |
+
dilation=dilation)
|
| 97 |
+
|
| 98 |
+
def forward(self, x):
|
| 99 |
+
"""Compute convolution.
|
| 100 |
+
|
| 101 |
+
Args:
|
| 102 |
+
x: input embeddings
|
| 103 |
+
Returns:
|
| 104 |
+
conv_output: convolution results
|
| 105 |
+
"""
|
| 106 |
+
|
| 107 |
+
if self.shift:
|
| 108 |
+
return self.conv(x.transpose(1, 2)).transpose(1, 2)[:, 1:]
|
| 109 |
+
else:
|
| 110 |
+
return self.conv(x.transpose(1, 2)).transpose(1, 2)
|
| 111 |
+
|
| 112 |
+
class MultiheadAttention(nn.Module):
|
| 113 |
+
"""Multi-head self-attention layer."""
|
| 114 |
+
|
| 115 |
+
def __init__(self,
|
| 116 |
+
embed_dim,
|
| 117 |
+
num_heads,
|
| 118 |
+
dropout=0.,
|
| 119 |
+
bias=True,
|
| 120 |
+
v_proj=True,
|
| 121 |
+
out_proj=True,
|
| 122 |
+
relative_bias=True):
|
| 123 |
+
"""Initialization.
|
| 124 |
+
|
| 125 |
+
Args:
|
| 126 |
+
embed_dim: dimension of input embeddings
|
| 127 |
+
num_heads: number of self-attention heads
|
| 128 |
+
dropout: dropout rate
|
| 129 |
+
bias: bool, indicate whether include bias for linear transformations
|
| 130 |
+
v_proj: bool, indicate whether project inputs to new values
|
| 131 |
+
out_proj: bool, indicate whether project outputs to new values
|
| 132 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 133 |
+
attention bias
|
| 134 |
+
"""
|
| 135 |
+
|
| 136 |
+
super(MultiheadAttention, self).__init__()
|
| 137 |
+
self.embed_dim = embed_dim
|
| 138 |
+
|
| 139 |
+
self.num_heads = num_heads
|
| 140 |
+
self.drop = nn.Dropout(dropout)
|
| 141 |
+
self.head_dim = embed_dim // num_heads
|
| 142 |
+
assert self.head_dim * num_heads == self.embed_dim, ("embed_dim must be "
|
| 143 |
+
"divisible by "
|
| 144 |
+
"num_heads")
|
| 145 |
+
|
| 146 |
+
self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 147 |
+
self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 148 |
+
if v_proj:
|
| 149 |
+
self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 150 |
+
else:
|
| 151 |
+
self.v_proj = nn.Identity()
|
| 152 |
+
|
| 153 |
+
if out_proj:
|
| 154 |
+
self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 155 |
+
else:
|
| 156 |
+
self.out_proj = nn.Identity()
|
| 157 |
+
|
| 158 |
+
if relative_bias:
|
| 159 |
+
self.relative_bias = nn.Parameter(torch.zeros((self.num_heads, 512)))
|
| 160 |
+
else:
|
| 161 |
+
self.relative_bias = None
|
| 162 |
+
|
| 163 |
+
self._reset_parameters()
|
| 164 |
+
|
| 165 |
+
def _reset_parameters(self):
|
| 166 |
+
"""Initialize attention parameters."""
|
| 167 |
+
|
| 168 |
+
init.xavier_uniform_(self.q_proj.weight)
|
| 169 |
+
init.constant_(self.q_proj.bias, 0.)
|
| 170 |
+
|
| 171 |
+
init.xavier_uniform_(self.k_proj.weight)
|
| 172 |
+
init.constant_(self.k_proj.bias, 0.)
|
| 173 |
+
|
| 174 |
+
if isinstance(self.v_proj, nn.Linear):
|
| 175 |
+
init.xavier_uniform_(self.v_proj.weight)
|
| 176 |
+
init.constant_(self.v_proj.bias, 0.)
|
| 177 |
+
|
| 178 |
+
if isinstance(self.out_proj, nn.Linear):
|
| 179 |
+
init.xavier_uniform_(self.out_proj.weight)
|
| 180 |
+
init.constant_(self.out_proj.bias, 0.)
|
| 181 |
+
|
| 182 |
+
def forward(self, query, key_padding_mask=None, attn_mask=None):
|
| 183 |
+
"""Compute multi-head self-attention.
|
| 184 |
+
|
| 185 |
+
Args:
|
| 186 |
+
query: input embeddings
|
| 187 |
+
key_padding_mask: 3D mask that prevents attention to certain positions
|
| 188 |
+
attn_mask: 3D mask that rescale the attention weight at each position
|
| 189 |
+
Returns:
|
| 190 |
+
attn_output: self-attention output
|
| 191 |
+
"""
|
| 192 |
+
|
| 193 |
+
length, bsz, embed_dim = query.size()
|
| 194 |
+
assert embed_dim == self.embed_dim
|
| 195 |
+
|
| 196 |
+
head_dim = embed_dim // self.num_heads
|
| 197 |
+
assert head_dim * self.num_heads == embed_dim, ("embed_dim must be "
|
| 198 |
+
"divisible by num_heads")
|
| 199 |
+
scaling = float(head_dim)**-0.5
|
| 200 |
+
|
| 201 |
+
q = self.q_proj(query)
|
| 202 |
+
k = self.k_proj(query)
|
| 203 |
+
v = self.v_proj(query)
|
| 204 |
+
|
| 205 |
+
q = q * scaling
|
| 206 |
+
|
| 207 |
+
if attn_mask is not None:
|
| 208 |
+
assert list(attn_mask.size()) == [bsz * self.num_heads,
|
| 209 |
+
query.size(0), query.size(0)]
|
| 210 |
+
|
| 211 |
+
q = q.contiguous().view(length, bsz * self.num_heads,
|
| 212 |
+
head_dim).transpose(0, 1)
|
| 213 |
+
k = k.contiguous().view(length, bsz * self.num_heads,
|
| 214 |
+
head_dim).transpose(0, 1)
|
| 215 |
+
v = v.contiguous().view(length, bsz * self.num_heads,
|
| 216 |
+
head_dim).transpose(0, 1)
|
| 217 |
+
|
| 218 |
+
attn_output_weights = torch.bmm(q, k.transpose(1, 2))
|
| 219 |
+
assert list(
|
| 220 |
+
attn_output_weights.size()) == [bsz * self.num_heads, length, length]
|
| 221 |
+
|
| 222 |
+
if self.relative_bias is not None:
|
| 223 |
+
pos = torch.arange(length, device=query.device)
|
| 224 |
+
relative_pos = torch.abs(pos[:, None] - pos[None, :]) + 256
|
| 225 |
+
relative_pos = relative_pos[None, :, :].expand(bsz * self.num_heads, -1,
|
| 226 |
+
-1)
|
| 227 |
+
|
| 228 |
+
relative_bias = self.relative_bias.repeat_interleave(bsz, dim=0)
|
| 229 |
+
relative_bias = relative_bias[:, None, :].expand(-1, length, -1)
|
| 230 |
+
relative_bias = torch.gather(relative_bias, 2, relative_pos)
|
| 231 |
+
attn_output_weights = attn_output_weights + relative_bias
|
| 232 |
+
|
| 233 |
+
if key_padding_mask is not None:
|
| 234 |
+
attn_output_weights = attn_output_weights + key_padding_mask
|
| 235 |
+
|
| 236 |
+
if attn_mask is None:
|
| 237 |
+
attn_output_weights = torch.softmax(attn_output_weights, dim=-1)
|
| 238 |
+
else:
|
| 239 |
+
attn_output_weights = torch.sigmoid(attn_output_weights) * attn_mask
|
| 240 |
+
|
| 241 |
+
attn_output_weights = self.drop(attn_output_weights)
|
| 242 |
+
|
| 243 |
+
attn_output = torch.bmm(attn_output_weights, v)
|
| 244 |
+
|
| 245 |
+
assert list(attn_output.size()) == [bsz * self.num_heads, length, head_dim]
|
| 246 |
+
attn_output = attn_output.transpose(0, 1).contiguous().view(
|
| 247 |
+
length, bsz, embed_dim)
|
| 248 |
+
attn_output = self.out_proj(attn_output)
|
| 249 |
+
|
| 250 |
+
return attn_output
|
| 251 |
+
|
| 252 |
+
class TransformerLayer(nn.Module):
|
| 253 |
+
"""TransformerEncoderLayer is made up of self-attn and feedforward network."""
|
| 254 |
+
|
| 255 |
+
def __init__(self,
|
| 256 |
+
d_model,
|
| 257 |
+
nhead,
|
| 258 |
+
dim_feedforward=2048,
|
| 259 |
+
dropout=0.1,
|
| 260 |
+
dropatt=0.1,
|
| 261 |
+
activation="leakyrelu",
|
| 262 |
+
relative_bias=True):
|
| 263 |
+
"""Initialization.
|
| 264 |
+
|
| 265 |
+
Args:
|
| 266 |
+
d_model: dimension of inputs
|
| 267 |
+
nhead: number of self-attention heads
|
| 268 |
+
dim_feedforward: dimension of hidden layer in feedforward layer
|
| 269 |
+
dropout: dropout rate
|
| 270 |
+
dropatt: drop attention rate
|
| 271 |
+
activation: activation function
|
| 272 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 273 |
+
attention bias
|
| 274 |
+
"""
|
| 275 |
+
|
| 276 |
+
super(TransformerLayer, self).__init__()
|
| 277 |
+
|
| 278 |
+
self.self_attn = MultiheadAttention(
|
| 279 |
+
d_model, nhead, dropout=dropatt, relative_bias=relative_bias)
|
| 280 |
+
|
| 281 |
+
# Implementation of Feedforward model
|
| 282 |
+
self.feedforward = nn.Sequential(
|
| 283 |
+
nn.LayerNorm(d_model), nn.Linear(d_model, dim_feedforward),
|
| 284 |
+
_get_activation_fn(activation), nn.Dropout(dropout),
|
| 285 |
+
nn.Linear(dim_feedforward, d_model))
|
| 286 |
+
|
| 287 |
+
self.norm = nn.LayerNorm(d_model)
|
| 288 |
+
self.dropout1 = nn.Dropout(dropout)
|
| 289 |
+
self.dropout2 = nn.Dropout(dropout)
|
| 290 |
+
|
| 291 |
+
self.nhead = nhead
|
| 292 |
+
|
| 293 |
+
def forward(self, src, attn_mask=None, key_padding_mask=None):
|
| 294 |
+
"""Pass the input through the encoder layer.
|
| 295 |
+
|
| 296 |
+
Args:
|
| 297 |
+
src: the sequence to the encoder layer (required).
|
| 298 |
+
attn_mask: the mask for the src sequence (optional).
|
| 299 |
+
key_padding_mask: the mask for the src keys per batch (optional).
|
| 300 |
+
Returns:
|
| 301 |
+
src3: the output of transformer layer, share the same shape as src.
|
| 302 |
+
"""
|
| 303 |
+
src2 = self.self_attn(
|
| 304 |
+
self.norm(src), attn_mask=attn_mask, key_padding_mask=key_padding_mask)
|
| 305 |
+
src2 = src + self.dropout1(src2)
|
| 306 |
+
src3 = self.feedforward(src2)
|
| 307 |
+
src3 = src2 + self.dropout2(src3)
|
| 308 |
+
|
| 309 |
+
return src3
|
| 310 |
+
|
| 311 |
+
|
| 312 |
+
|
| 313 |
+
class RobertaClassificationHead(nn.Module):
|
| 314 |
+
"""Head for sentence-level classification tasks."""
|
| 315 |
+
|
| 316 |
+
def __init__(self, config):
|
| 317 |
+
super().__init__()
|
| 318 |
+
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
|
| 319 |
+
classifier_dropout = (
|
| 320 |
+
config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
|
| 321 |
+
)
|
| 322 |
+
self.dropout = nn.Dropout(classifier_dropout)
|
| 323 |
+
self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
|
| 324 |
+
|
| 325 |
+
def forward(self, features, **kwargs):
|
| 326 |
+
x = features[:, 0, :] # take <s> token (equiv. to [CLS])
|
| 327 |
+
x = self.dropout(x)
|
| 328 |
+
x = self.dense(x)
|
| 329 |
+
x = torch.tanh(x)
|
| 330 |
+
x = self.dropout(x)
|
| 331 |
+
x = self.out_proj(x)
|
| 332 |
+
return x
|
| 333 |
+
|
| 334 |
+
|
| 335 |
+
##########################################
|
| 336 |
+
# Custom Models
|
| 337 |
+
##########################################
|
| 338 |
+
def cumprod(x, reverse=False, exclusive=False):
|
| 339 |
+
"""cumulative product."""
|
| 340 |
+
if reverse:
|
| 341 |
+
x = x.flip([-1])
|
| 342 |
+
|
| 343 |
+
if exclusive:
|
| 344 |
+
x = F.pad(x[:, :, :-1], (1, 0), value=1)
|
| 345 |
+
|
| 346 |
+
cx = x.cumprod(-1)
|
| 347 |
+
|
| 348 |
+
if reverse:
|
| 349 |
+
cx = cx.flip([-1])
|
| 350 |
+
return cx
|
| 351 |
+
|
| 352 |
+
def cumsum(x, reverse=False, exclusive=False):
|
| 353 |
+
"""cumulative sum."""
|
| 354 |
+
bsz, _, length = x.size()
|
| 355 |
+
device = x.device
|
| 356 |
+
if reverse:
|
| 357 |
+
if exclusive:
|
| 358 |
+
w = torch.ones([bsz, length, length], device=device).tril(-1)
|
| 359 |
+
else:
|
| 360 |
+
w = torch.ones([bsz, length, length], device=device).tril(0)
|
| 361 |
+
cx = torch.bmm(x, w)
|
| 362 |
+
else:
|
| 363 |
+
if exclusive:
|
| 364 |
+
w = torch.ones([bsz, length, length], device=device).triu(1)
|
| 365 |
+
else:
|
| 366 |
+
w = torch.ones([bsz, length, length], device=device).triu(0)
|
| 367 |
+
cx = torch.bmm(x, w)
|
| 368 |
+
return cx
|
| 369 |
+
|
| 370 |
+
def cummin(x, reverse=False, exclusive=False, max_value=1e9):
|
| 371 |
+
"""cumulative min."""
|
| 372 |
+
if reverse:
|
| 373 |
+
if exclusive:
|
| 374 |
+
x = F.pad(x[:, :, 1:], (0, 1), value=max_value)
|
| 375 |
+
x = x.flip([-1]).cummin(-1)[0].flip([-1])
|
| 376 |
+
else:
|
| 377 |
+
if exclusive:
|
| 378 |
+
x = F.pad(x[:, :, :-1], (1, 0), value=max_value)
|
| 379 |
+
x = x.cummin(-1)[0]
|
| 380 |
+
return x
|
| 381 |
+
|
| 382 |
+
class Transformer(nn.Module):
|
| 383 |
+
"""Transformer model."""
|
| 384 |
+
|
| 385 |
+
def __init__(self,
|
| 386 |
+
hidden_size,
|
| 387 |
+
nlayers,
|
| 388 |
+
ntokens,
|
| 389 |
+
nhead=8,
|
| 390 |
+
dropout=0.1,
|
| 391 |
+
dropatt=0.1,
|
| 392 |
+
relative_bias=True,
|
| 393 |
+
pos_emb=False,
|
| 394 |
+
pad=0):
|
| 395 |
+
"""Initialization.
|
| 396 |
+
|
| 397 |
+
Args:
|
| 398 |
+
hidden_size: dimension of inputs and hidden states
|
| 399 |
+
nlayers: number of layers
|
| 400 |
+
ntokens: number of output categories
|
| 401 |
+
nhead: number of self-attention heads
|
| 402 |
+
dropout: dropout rate
|
| 403 |
+
dropatt: drop attention rate
|
| 404 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 405 |
+
attention bias
|
| 406 |
+
pos_emb: bool, indicate whether use a learnable positional embedding
|
| 407 |
+
pad: pad token index
|
| 408 |
+
"""
|
| 409 |
+
|
| 410 |
+
super(Transformer, self).__init__()
|
| 411 |
+
|
| 412 |
+
self.drop = nn.Dropout(dropout)
|
| 413 |
+
|
| 414 |
+
self.emb = nn.Embedding(ntokens, hidden_size)
|
| 415 |
+
if pos_emb:
|
| 416 |
+
self.pos_emb = nn.Embedding(500, hidden_size)
|
| 417 |
+
|
| 418 |
+
self.layers = nn.ModuleList([
|
| 419 |
+
TransformerLayer(hidden_size, nhead, hidden_size * 4, dropout,
|
| 420 |
+
dropatt=dropatt, relative_bias=relative_bias)
|
| 421 |
+
for _ in range(nlayers)])
|
| 422 |
+
|
| 423 |
+
self.norm = nn.LayerNorm(hidden_size)
|
| 424 |
+
|
| 425 |
+
self.output_layer = nn.Linear(hidden_size, ntokens)
|
| 426 |
+
self.output_layer.weight = self.emb.weight
|
| 427 |
+
|
| 428 |
+
self.init_weights()
|
| 429 |
+
|
| 430 |
+
self.nlayers = nlayers
|
| 431 |
+
self.nhead = nhead
|
| 432 |
+
self.ntokens = ntokens
|
| 433 |
+
self.hidden_size = hidden_size
|
| 434 |
+
self.pad = pad
|
| 435 |
+
|
| 436 |
+
def init_weights(self):
|
| 437 |
+
"""Initialize token embedding and output bias."""
|
| 438 |
+
initrange = 0.1
|
| 439 |
+
self.emb.weight.data.uniform_(-initrange, initrange)
|
| 440 |
+
if hasattr(self, 'pos_emb'):
|
| 441 |
+
self.pos_emb.weight.data.uniform_(-initrange, initrange)
|
| 442 |
+
self.output_layer.bias.data.fill_(0)
|
| 443 |
+
|
| 444 |
+
def visibility(self, x, device):
|
| 445 |
+
"""Mask pad tokens."""
|
| 446 |
+
visibility = (x != self.pad).float()
|
| 447 |
+
visibility = visibility[:, None, :].expand(-1, x.size(1), -1)
|
| 448 |
+
visibility = torch.repeat_interleave(visibility, self.nhead, dim=0)
|
| 449 |
+
return visibility.log()
|
| 450 |
+
|
| 451 |
+
def encode(self, x, pos):
|
| 452 |
+
"""Standard transformer encode process."""
|
| 453 |
+
h = self.emb(x)
|
| 454 |
+
if hasattr(self, 'pos_emb'):
|
| 455 |
+
h = h + self.pos_emb(pos)
|
| 456 |
+
h_list = []
|
| 457 |
+
visibility = self.visibility(x, x.device)
|
| 458 |
+
|
| 459 |
+
for i in range(self.nlayers):
|
| 460 |
+
h_list.append(h)
|
| 461 |
+
h = self.layers[i](
|
| 462 |
+
h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
|
| 463 |
+
|
| 464 |
+
output = h
|
| 465 |
+
h_array = torch.stack(h_list, dim=2)
|
| 466 |
+
|
| 467 |
+
return output, h_array
|
| 468 |
+
|
| 469 |
+
def forward(self, x, pos):
|
| 470 |
+
"""Pass the input through the encoder layer.
|
| 471 |
+
|
| 472 |
+
Args:
|
| 473 |
+
x: input tokens (required).
|
| 474 |
+
pos: position for each token (optional).
|
| 475 |
+
Returns:
|
| 476 |
+
output: probability distributions for missing tokens.
|
| 477 |
+
state_dict: parsing results and raw output
|
| 478 |
+
"""
|
| 479 |
+
|
| 480 |
+
batch_size, length = x.size()
|
| 481 |
+
|
| 482 |
+
raw_output, _ = self.encode(x, pos)
|
| 483 |
+
raw_output = self.norm(raw_output)
|
| 484 |
+
raw_output = self.drop(raw_output)
|
| 485 |
+
|
| 486 |
+
output = self.output_layer(raw_output)
|
| 487 |
+
return output.view(batch_size * length, -1), {'raw_output': raw_output,}
|
| 488 |
+
|
| 489 |
+
class StructFormer(Transformer):
|
| 490 |
+
"""StructFormer model."""
|
| 491 |
+
|
| 492 |
+
def __init__(self,
|
| 493 |
+
hidden_size,
|
| 494 |
+
n_context_layers,
|
| 495 |
+
nlayers,
|
| 496 |
+
ntokens,
|
| 497 |
+
nhead=8,
|
| 498 |
+
dropout=0.1,
|
| 499 |
+
dropatt=0.1,
|
| 500 |
+
relative_bias=False,
|
| 501 |
+
pos_emb=False,
|
| 502 |
+
pad=0,
|
| 503 |
+
n_parser_layers=4,
|
| 504 |
+
conv_size=9,
|
| 505 |
+
relations=('head', 'child'),
|
| 506 |
+
weight_act='softmax'):
|
| 507 |
+
"""Initialization.
|
| 508 |
+
|
| 509 |
+
Args:
|
| 510 |
+
hidden_size: dimension of inputs and hidden states
|
| 511 |
+
nlayers: number of layers
|
| 512 |
+
ntokens: number of output categories
|
| 513 |
+
nhead: number of self-attention heads
|
| 514 |
+
dropout: dropout rate
|
| 515 |
+
dropatt: drop attention rate
|
| 516 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 517 |
+
attention bias
|
| 518 |
+
pos_emb: bool, indicate whether use a learnable positional embedding
|
| 519 |
+
pad: pad token index
|
| 520 |
+
n_parser_layers: number of parsing layers
|
| 521 |
+
conv_size: convolution kernel size for parser
|
| 522 |
+
relations: relations that are used to compute self attention
|
| 523 |
+
weight_act: relations distribution activation function
|
| 524 |
+
"""
|
| 525 |
+
|
| 526 |
+
super(StructFormer, self).__init__(
|
| 527 |
+
hidden_size,
|
| 528 |
+
nlayers,
|
| 529 |
+
ntokens,
|
| 530 |
+
nhead=nhead,
|
| 531 |
+
dropout=dropout,
|
| 532 |
+
dropatt=dropatt,
|
| 533 |
+
relative_bias=relative_bias,
|
| 534 |
+
pos_emb=pos_emb,
|
| 535 |
+
pad=pad)
|
| 536 |
+
|
| 537 |
+
|
| 538 |
+
def encode(self, x, pos):
|
| 539 |
+
h = self.emb(x)
|
| 540 |
+
if hasattr(self, 'pos_emb'):
|
| 541 |
+
h = h + self.pos_emb(pos)
|
| 542 |
+
h_list = []
|
| 543 |
+
visibility = self.visibility(x, x.device)
|
| 544 |
+
|
| 545 |
+
for i in range(self.nlayers):
|
| 546 |
+
h_list.append(h)
|
| 547 |
+
h = self.layers[i](
|
| 548 |
+
h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
|
| 549 |
+
|
| 550 |
+
output = h
|
| 551 |
+
h_array = torch.stack(h_list, dim=2)
|
| 552 |
+
|
| 553 |
+
return output
|
| 554 |
+
|
| 555 |
+
|
| 556 |
+
def forward(self, input_ids, labels=None, position_ids=None, **kwargs):
|
| 557 |
+
|
| 558 |
+
x = input_ids
|
| 559 |
+
batch_size, length = x.size()
|
| 560 |
+
|
| 561 |
+
if position_ids is None:
|
| 562 |
+
pos = torch.arange(length, device=x.device).expand(batch_size, length)
|
| 563 |
+
|
| 564 |
+
raw_output = self.encode(x, pos)
|
| 565 |
+
raw_output = self.norm(raw_output)
|
| 566 |
+
raw_output = self.drop(raw_output)
|
| 567 |
+
|
| 568 |
+
output = self.output_layer(raw_output)
|
| 569 |
+
|
| 570 |
+
loss = None
|
| 571 |
+
if labels is not None:
|
| 572 |
+
loss_fct = nn.CrossEntropyLoss()
|
| 573 |
+
loss = loss_fct(output.view(batch_size * length, -1), labels.reshape(-1))
|
| 574 |
+
|
| 575 |
+
return MaskedLMOutput(
|
| 576 |
+
loss=loss, # shape: 1
|
| 577 |
+
logits=output, # shape: (batch_size * length, ntokens)
|
| 578 |
+
hidden_states=None,
|
| 579 |
+
attentions=None,
|
| 580 |
+
)
|
| 581 |
+
|
| 582 |
+
##########################################
|
| 583 |
+
# HuggingFace Model
|
| 584 |
+
##########################################
|
| 585 |
+
class StructformerModel(PreTrainedModel):
|
| 586 |
+
config_class = StructformerConfig
|
| 587 |
+
|
| 588 |
+
def __init__(self, config):
|
| 589 |
+
super().__init__(config)
|
| 590 |
+
self.model = StructFormer(
|
| 591 |
+
hidden_size=config.hidden_size,
|
| 592 |
+
n_context_layers=config.n_context_layers,
|
| 593 |
+
nlayers=config.nlayers,
|
| 594 |
+
ntokens=config.ntokens,
|
| 595 |
+
nhead=config.nhead,
|
| 596 |
+
dropout=config.dropout,
|
| 597 |
+
dropatt=config.dropatt,
|
| 598 |
+
relative_bias=config.relative_bias,
|
| 599 |
+
pos_emb=config.pos_emb,
|
| 600 |
+
pad=config.pad,
|
| 601 |
+
n_parser_layers=config.n_parser_layers,
|
| 602 |
+
conv_size=config.conv_size,
|
| 603 |
+
relations=config.relations,
|
| 604 |
+
weight_act=config.weight_act
|
| 605 |
+
)
|
| 606 |
+
|
| 607 |
+
def forward(self, input_ids, labels=None, **kwargs):
|
| 608 |
+
return self.model(input_ids, labels=labels, **kwargs)
|
| 609 |
+
|
| 610 |
+
|
| 611 |
+
class StructFormerClassification(Transformer):
|
| 612 |
+
"""StructFormer model."""
|
| 613 |
+
|
| 614 |
+
def __init__(self,
|
| 615 |
+
hidden_size,
|
| 616 |
+
n_context_layers,
|
| 617 |
+
nlayers,
|
| 618 |
+
ntokens,
|
| 619 |
+
nhead=8,
|
| 620 |
+
dropout=0.1,
|
| 621 |
+
dropatt=0.1,
|
| 622 |
+
relative_bias=False,
|
| 623 |
+
pos_emb=False,
|
| 624 |
+
pad=0,
|
| 625 |
+
n_parser_layers=4,
|
| 626 |
+
conv_size=9,
|
| 627 |
+
relations=('head', 'child'),
|
| 628 |
+
weight_act='softmax',
|
| 629 |
+
config=None,
|
| 630 |
+
):
|
| 631 |
+
|
| 632 |
+
|
| 633 |
+
super(StructFormerClassification, self).__init__(
|
| 634 |
+
hidden_size,
|
| 635 |
+
nlayers,
|
| 636 |
+
ntokens,
|
| 637 |
+
nhead=nhead,
|
| 638 |
+
dropout=dropout,
|
| 639 |
+
dropatt=dropatt,
|
| 640 |
+
relative_bias=relative_bias,
|
| 641 |
+
pos_emb=pos_emb,
|
| 642 |
+
pad=pad)
|
| 643 |
+
|
| 644 |
+
self.num_labels = config.num_labels
|
| 645 |
+
self.config = config
|
| 646 |
+
|
| 647 |
+
self.classifier = RobertaClassificationHead(config)
|
| 648 |
+
|
| 649 |
+
def encode(self, x, pos):
|
| 650 |
+
h = self.emb(x)
|
| 651 |
+
if hasattr(self, 'pos_emb'):
|
| 652 |
+
h = h + self.pos_emb(pos)
|
| 653 |
+
h_list = []
|
| 654 |
+
visibility = self.visibility(x, x.device)
|
| 655 |
+
|
| 656 |
+
for i in range(self.nlayers):
|
| 657 |
+
h_list.append(h)
|
| 658 |
+
h = self.layers[i](
|
| 659 |
+
h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
|
| 660 |
+
|
| 661 |
+
output = h
|
| 662 |
+
h_array = torch.stack(h_list, dim=2)
|
| 663 |
+
|
| 664 |
+
return output
|
| 665 |
+
|
| 666 |
+
|
| 667 |
+
def forward(self, input_ids, labels=None, position_ids=None, **kwargs):
|
| 668 |
+
|
| 669 |
+
x = input_ids
|
| 670 |
+
batch_size, length = x.size()
|
| 671 |
+
|
| 672 |
+
if position_ids is None:
|
| 673 |
+
pos = torch.arange(length, device=x.device).expand(batch_size, length)
|
| 674 |
+
|
| 675 |
+
raw_output = self.encode(x, pos)
|
| 676 |
+
raw_output = self.norm(raw_output)
|
| 677 |
+
raw_output = self.drop(raw_output)
|
| 678 |
+
|
| 679 |
+
#output = self.output_layer(raw_output)
|
| 680 |
+
logits = self.classifier(raw_output)
|
| 681 |
+
|
| 682 |
+
loss = None
|
| 683 |
+
if labels is not None:
|
| 684 |
+
if self.config.problem_type is None:
|
| 685 |
+
if self.num_labels == 1:
|
| 686 |
+
self.config.problem_type = "regression"
|
| 687 |
+
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
| 688 |
+
self.config.problem_type = "single_label_classification"
|
| 689 |
+
else:
|
| 690 |
+
self.config.problem_type = "multi_label_classification"
|
| 691 |
+
|
| 692 |
+
if self.config.problem_type == "regression":
|
| 693 |
+
loss_fct = MSELoss()
|
| 694 |
+
if self.num_labels == 1:
|
| 695 |
+
loss = loss_fct(logits.squeeze(), labels.squeeze())
|
| 696 |
+
else:
|
| 697 |
+
loss = loss_fct(logits, labels)
|
| 698 |
+
elif self.config.problem_type == "single_label_classification":
|
| 699 |
+
loss_fct = CrossEntropyLoss()
|
| 700 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
| 701 |
+
elif self.config.problem_type == "multi_label_classification":
|
| 702 |
+
loss_fct = BCEWithLogitsLoss()
|
| 703 |
+
loss = loss_fct(logits, labels)
|
| 704 |
+
|
| 705 |
+
|
| 706 |
+
return SequenceClassifierOutput(
|
| 707 |
+
loss=loss,
|
| 708 |
+
logits=logits,
|
| 709 |
+
hidden_states=None,
|
| 710 |
+
attentions=None,
|
| 711 |
+
)
|
| 712 |
+
|
| 713 |
+
|
| 714 |
+
class StructformerModelForSequenceClassification(PreTrainedModel):
|
| 715 |
+
config_class = StructformerConfig
|
| 716 |
+
def __init__(self, config):
|
| 717 |
+
super().__init__(config)
|
| 718 |
+
self.model = StructFormerClassification(
|
| 719 |
+
hidden_size=config.hidden_size,
|
| 720 |
+
n_context_layers=config.n_context_layers,
|
| 721 |
+
nlayers=config.nlayers,
|
| 722 |
+
ntokens=config.ntokens,
|
| 723 |
+
nhead=config.nhead,
|
| 724 |
+
dropout=config.dropout,
|
| 725 |
+
dropatt=config.dropatt,
|
| 726 |
+
relative_bias=config.relative_bias,
|
| 727 |
+
pos_emb=config.pos_emb,
|
| 728 |
+
pad=config.pad,
|
| 729 |
+
n_parser_layers=config.n_parser_layers,
|
| 730 |
+
conv_size=config.conv_size,
|
| 731 |
+
relations=config.relations,
|
| 732 |
+
weight_act=config.weight_act,
|
| 733 |
+
config=config)
|
| 734 |
+
|
| 735 |
+
def _init_weights(self, module):
|
| 736 |
+
"""Initialize the weights"""
|
| 737 |
+
if isinstance(module, nn.Linear):
|
| 738 |
+
# Slightly different from the TF version which uses truncated_normal for initialization
|
| 739 |
+
# cf https://github.com/pytorch/pytorch/pull/5617
|
| 740 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 741 |
+
if module.bias is not None:
|
| 742 |
+
module.bias.data.zero_()
|
| 743 |
+
elif isinstance(module, nn.Embedding):
|
| 744 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 745 |
+
if module.padding_idx is not None:
|
| 746 |
+
module.weight.data[module.padding_idx].zero_()
|
| 747 |
+
elif isinstance(module, nn.LayerNorm):
|
| 748 |
+
if module.bias is not None:
|
| 749 |
+
module.bias.data.zero_()
|
| 750 |
+
module.weight.data.fill_(1.0)
|
| 751 |
+
|
| 752 |
+
|
| 753 |
+
def forward(self, input_ids, labels=None, **kwargs):
|
| 754 |
+
return self.model(input_ids, labels=labels, **kwargs)
|
finetune/boolq/tokenizer_config.json
ADDED
|
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_prefix_space": false,
|
| 3 |
+
"bos_token": {
|
| 4 |
+
"__type": "AddedToken",
|
| 5 |
+
"content": "<s>",
|
| 6 |
+
"lstrip": false,
|
| 7 |
+
"normalized": true,
|
| 8 |
+
"rstrip": false,
|
| 9 |
+
"single_word": false
|
| 10 |
+
},
|
| 11 |
+
"cls_token": {
|
| 12 |
+
"__type": "AddedToken",
|
| 13 |
+
"content": "<s>",
|
| 14 |
+
"lstrip": false,
|
| 15 |
+
"normalized": true,
|
| 16 |
+
"rstrip": false,
|
| 17 |
+
"single_word": false
|
| 18 |
+
},
|
| 19 |
+
"eos_token": {
|
| 20 |
+
"__type": "AddedToken",
|
| 21 |
+
"content": "</s>",
|
| 22 |
+
"lstrip": false,
|
| 23 |
+
"normalized": true,
|
| 24 |
+
"rstrip": false,
|
| 25 |
+
"single_word": false
|
| 26 |
+
},
|
| 27 |
+
"errors": "replace",
|
| 28 |
+
"mask_token": {
|
| 29 |
+
"__type": "AddedToken",
|
| 30 |
+
"content": "<mask>",
|
| 31 |
+
"lstrip": true,
|
| 32 |
+
"normalized": false,
|
| 33 |
+
"rstrip": false,
|
| 34 |
+
"single_word": false
|
| 35 |
+
},
|
| 36 |
+
"model_max_length": 512,
|
| 37 |
+
"name_or_path": "final_models/transformer_base_final_2",
|
| 38 |
+
"pad_token": {
|
| 39 |
+
"__type": "AddedToken",
|
| 40 |
+
"content": "<pad>",
|
| 41 |
+
"lstrip": false,
|
| 42 |
+
"normalized": true,
|
| 43 |
+
"rstrip": false,
|
| 44 |
+
"single_word": false
|
| 45 |
+
},
|
| 46 |
+
"sep_token": {
|
| 47 |
+
"__type": "AddedToken",
|
| 48 |
+
"content": "</s>",
|
| 49 |
+
"lstrip": false,
|
| 50 |
+
"normalized": true,
|
| 51 |
+
"rstrip": false,
|
| 52 |
+
"single_word": false
|
| 53 |
+
},
|
| 54 |
+
"special_tokens_map_file": null,
|
| 55 |
+
"tokenizer_class": "RobertaTokenizer",
|
| 56 |
+
"trim_offsets": true,
|
| 57 |
+
"unk_token": {
|
| 58 |
+
"__type": "AddedToken",
|
| 59 |
+
"content": "<unk>",
|
| 60 |
+
"lstrip": false,
|
| 61 |
+
"normalized": true,
|
| 62 |
+
"rstrip": false,
|
| 63 |
+
"single_word": false
|
| 64 |
+
}
|
| 65 |
+
}
|
finetune/boolq/train_results.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"epoch": 10.0,
|
| 3 |
+
"train_loss": 0.33231258392333984,
|
| 4 |
+
"train_runtime": 67.9575,
|
| 5 |
+
"train_samples": 2072,
|
| 6 |
+
"train_samples_per_second": 304.896,
|
| 7 |
+
"train_steps_per_second": 2.649
|
| 8 |
+
}
|
finetune/boolq/trainer_state.json
ADDED
|
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"best_metric": null,
|
| 3 |
+
"best_model_checkpoint": null,
|
| 4 |
+
"epoch": 10.0,
|
| 5 |
+
"global_step": 180,
|
| 6 |
+
"is_hyper_param_search": false,
|
| 7 |
+
"is_local_process_zero": true,
|
| 8 |
+
"is_world_process_zero": true,
|
| 9 |
+
"log_history": [
|
| 10 |
+
{
|
| 11 |
+
"epoch": 10.0,
|
| 12 |
+
"step": 180,
|
| 13 |
+
"total_flos": 1363424481484800.0,
|
| 14 |
+
"train_loss": 0.33231258392333984,
|
| 15 |
+
"train_runtime": 67.9575,
|
| 16 |
+
"train_samples_per_second": 304.896,
|
| 17 |
+
"train_steps_per_second": 2.649
|
| 18 |
+
}
|
| 19 |
+
],
|
| 20 |
+
"max_steps": 180,
|
| 21 |
+
"num_train_epochs": 10,
|
| 22 |
+
"total_flos": 1363424481484800.0,
|
| 23 |
+
"trial_name": null,
|
| 24 |
+
"trial_params": null
|
| 25 |
+
}
|
finetune/boolq/training_args.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5c49e448b89277e79f0f1ccaf10e6cae9a79ed2754e23dd1d3fde17f45d1945b
|
| 3 |
+
size 3503
|
finetune/boolq/vocab.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
finetune/cola/all_results.json
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"epoch": 10.0,
|
| 3 |
+
"eval_accuracy": 0.6712462902069092,
|
| 4 |
+
"eval_f1": 0.7777040477770405,
|
| 5 |
+
"eval_loss": 0.7399011254310608,
|
| 6 |
+
"eval_mcc": 0.15981907397227785,
|
| 7 |
+
"eval_runtime": 1.4236,
|
| 8 |
+
"eval_samples": 1019,
|
| 9 |
+
"eval_samples_per_second": 715.815,
|
| 10 |
+
"eval_steps_per_second": 89.916,
|
| 11 |
+
"train_loss": 0.42006420190783517,
|
| 12 |
+
"train_runtime": 267.2668,
|
| 13 |
+
"train_samples": 8164,
|
| 14 |
+
"train_samples_per_second": 305.463,
|
| 15 |
+
"train_steps_per_second": 2.582
|
| 16 |
+
}
|
finetune/cola/checkpoint-400/config.json
ADDED
|
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_name_or_path": "final_models/transformer_base_final_2",
|
| 3 |
+
"architectures": [
|
| 4 |
+
"StructformerModelForSequenceClassification"
|
| 5 |
+
],
|
| 6 |
+
"attention_probs_dropout_prob": 0.1,
|
| 7 |
+
"auto_map": {
|
| 8 |
+
"AutoConfig": "structformer_as_hf_no_parser.StructformerConfig",
|
| 9 |
+
"AutoModelForMaskedLM": "structformer_as_hf_no_parser.StructformerModel",
|
| 10 |
+
"AutoModelForSequenceClassification": "structformer_as_hf_no_parser.StructformerModelForSequenceClassification"
|
| 11 |
+
},
|
| 12 |
+
"bos_token_id": 0,
|
| 13 |
+
"classifier_dropout": null,
|
| 14 |
+
"conv_size": 9,
|
| 15 |
+
"dropatt": 0.1,
|
| 16 |
+
"dropout": 0.1,
|
| 17 |
+
"eos_token_id": 2,
|
| 18 |
+
"hidden_act": "gelu",
|
| 19 |
+
"hidden_dropout_prob": 0.1,
|
| 20 |
+
"hidden_size": 768,
|
| 21 |
+
"id2label": {
|
| 22 |
+
"0": 0,
|
| 23 |
+
"1": 1
|
| 24 |
+
},
|
| 25 |
+
"initializer_range": 0.02,
|
| 26 |
+
"intermediate_size": 3072,
|
| 27 |
+
"label2id": {
|
| 28 |
+
"0": 0,
|
| 29 |
+
"1": 1
|
| 30 |
+
},
|
| 31 |
+
"layer_norm_eps": 1e-05,
|
| 32 |
+
"max_position_embeddings": 514,
|
| 33 |
+
"model_type": "structformer",
|
| 34 |
+
"n_context_layers": 0,
|
| 35 |
+
"n_parser_layers": 0,
|
| 36 |
+
"nhead": 12,
|
| 37 |
+
"nlayers": 12,
|
| 38 |
+
"ntokens": 32000,
|
| 39 |
+
"num_attention_heads": 12,
|
| 40 |
+
"num_hidden_layers": 12,
|
| 41 |
+
"pad": 0,
|
| 42 |
+
"pad_token_id": 1,
|
| 43 |
+
"pos_emb": true,
|
| 44 |
+
"position_embedding_type": "absolute",
|
| 45 |
+
"problem_type": "single_label_classification",
|
| 46 |
+
"relations": [
|
| 47 |
+
"head",
|
| 48 |
+
"child"
|
| 49 |
+
],
|
| 50 |
+
"relative_bias": false,
|
| 51 |
+
"torch_dtype": "float32",
|
| 52 |
+
"transformers_version": "4.26.1",
|
| 53 |
+
"type_vocab_size": 1,
|
| 54 |
+
"use_cache": true,
|
| 55 |
+
"vocab_size": 32000,
|
| 56 |
+
"weight_act": "softmax"
|
| 57 |
+
}
|
finetune/cola/checkpoint-400/merges.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
finetune/cola/checkpoint-400/optimizer.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3ac0a55ab34318598e03256dc8115b7cdbe039ebcd7ae1e842a5c7b2add862cd
|
| 3 |
+
size 884981025
|
finetune/cola/checkpoint-400/pytorch_model.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8e347d72000b0f5a8771aa62eedeb5a78b9aa79669b5796edb1e66f8e21e832c
|
| 3 |
+
size 442624943
|
finetune/cola/checkpoint-400/rng_state.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5769fda57687e02db326bee3281d32f541cdc788f737f37c47c3a91239b699cc
|
| 3 |
+
size 14503
|
finetune/cola/checkpoint-400/scheduler.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2b09db5c9b9264697e44562c70af05a12a67180f480992d419fa287214c7be7a
|
| 3 |
+
size 623
|
finetune/cola/checkpoint-400/special_tokens_map.json
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bos_token": "<s>",
|
| 3 |
+
"cls_token": "<s>",
|
| 4 |
+
"eos_token": "</s>",
|
| 5 |
+
"mask_token": {
|
| 6 |
+
"content": "<mask>",
|
| 7 |
+
"lstrip": true,
|
| 8 |
+
"normalized": false,
|
| 9 |
+
"rstrip": false,
|
| 10 |
+
"single_word": false
|
| 11 |
+
},
|
| 12 |
+
"pad_token": "<pad>",
|
| 13 |
+
"sep_token": "</s>",
|
| 14 |
+
"unk_token": "<unk>"
|
| 15 |
+
}
|
finetune/cola/checkpoint-400/structformer_as_hf_no_parser.py
ADDED
|
@@ -0,0 +1,754 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import torch.nn as nn
|
| 3 |
+
import torch.nn.functional as F
|
| 4 |
+
from torch.nn import init
|
| 5 |
+
from transformers import PreTrainedModel
|
| 6 |
+
from transformers import PretrainedConfig
|
| 7 |
+
from transformers.modeling_outputs import MaskedLMOutput
|
| 8 |
+
from typing import List
|
| 9 |
+
from torch.nn import CrossEntropyLoss, MSELoss, BCEWithLogitsLoss
|
| 10 |
+
from transformers.modeling_outputs import (
|
| 11 |
+
BaseModelOutputWithPastAndCrossAttentions,
|
| 12 |
+
BaseModelOutputWithPoolingAndCrossAttentions,
|
| 13 |
+
MaskedLMOutput,
|
| 14 |
+
SequenceClassifierOutput
|
| 15 |
+
)
|
| 16 |
+
|
| 17 |
+
##########################################
|
| 18 |
+
# HuggingFace Config
|
| 19 |
+
##########################################
|
| 20 |
+
class StructformerConfig(PretrainedConfig):
|
| 21 |
+
model_type = "structformer"
|
| 22 |
+
|
| 23 |
+
def __init__(
|
| 24 |
+
self,
|
| 25 |
+
hidden_size=768,
|
| 26 |
+
n_context_layers=2,
|
| 27 |
+
nlayers=6,
|
| 28 |
+
ntokens=32000,
|
| 29 |
+
nhead=8,
|
| 30 |
+
dropout=0.1,
|
| 31 |
+
dropatt=0.1,
|
| 32 |
+
relative_bias=False,
|
| 33 |
+
pos_emb=False,
|
| 34 |
+
pad=0,
|
| 35 |
+
n_parser_layers=4,
|
| 36 |
+
conv_size=9,
|
| 37 |
+
relations=('head', 'child'),
|
| 38 |
+
weight_act='softmax',
|
| 39 |
+
**kwargs,
|
| 40 |
+
):
|
| 41 |
+
self.hidden_size = hidden_size
|
| 42 |
+
self.n_context_layers = n_context_layers
|
| 43 |
+
self.nlayers = nlayers
|
| 44 |
+
self.ntokens = ntokens
|
| 45 |
+
self.nhead = nhead
|
| 46 |
+
self.dropout = dropout
|
| 47 |
+
self.dropatt = dropatt
|
| 48 |
+
self.relative_bias = relative_bias
|
| 49 |
+
self.pos_emb = pos_emb
|
| 50 |
+
self.pad = pad
|
| 51 |
+
self.n_parser_layers = n_parser_layers
|
| 52 |
+
self.conv_size = conv_size
|
| 53 |
+
self.relations = relations
|
| 54 |
+
self.weight_act = weight_act
|
| 55 |
+
super().__init__(**kwargs)
|
| 56 |
+
|
| 57 |
+
##########################################
|
| 58 |
+
# Custom Layers
|
| 59 |
+
##########################################
|
| 60 |
+
def _get_activation_fn(activation):
|
| 61 |
+
"""Get specified activation function."""
|
| 62 |
+
if activation == "relu":
|
| 63 |
+
return nn.ReLU()
|
| 64 |
+
elif activation == "gelu":
|
| 65 |
+
return nn.GELU()
|
| 66 |
+
elif activation == "leakyrelu":
|
| 67 |
+
return nn.LeakyReLU()
|
| 68 |
+
|
| 69 |
+
raise RuntimeError(
|
| 70 |
+
"activation should be relu/gelu, not {}".format(activation))
|
| 71 |
+
|
| 72 |
+
class Conv1d(nn.Module):
|
| 73 |
+
"""1D convolution layer."""
|
| 74 |
+
|
| 75 |
+
def __init__(self, hidden_size, kernel_size, dilation=1):
|
| 76 |
+
"""Initialization.
|
| 77 |
+
|
| 78 |
+
Args:
|
| 79 |
+
hidden_size: dimension of input embeddings
|
| 80 |
+
kernel_size: convolution kernel size
|
| 81 |
+
dilation: the spacing between the kernel points
|
| 82 |
+
"""
|
| 83 |
+
super(Conv1d, self).__init__()
|
| 84 |
+
|
| 85 |
+
if kernel_size % 2 == 0:
|
| 86 |
+
padding = (kernel_size // 2) * dilation
|
| 87 |
+
self.shift = True
|
| 88 |
+
else:
|
| 89 |
+
padding = ((kernel_size - 1) // 2) * dilation
|
| 90 |
+
self.shift = False
|
| 91 |
+
self.conv = nn.Conv1d(
|
| 92 |
+
hidden_size,
|
| 93 |
+
hidden_size,
|
| 94 |
+
kernel_size,
|
| 95 |
+
padding=padding,
|
| 96 |
+
dilation=dilation)
|
| 97 |
+
|
| 98 |
+
def forward(self, x):
|
| 99 |
+
"""Compute convolution.
|
| 100 |
+
|
| 101 |
+
Args:
|
| 102 |
+
x: input embeddings
|
| 103 |
+
Returns:
|
| 104 |
+
conv_output: convolution results
|
| 105 |
+
"""
|
| 106 |
+
|
| 107 |
+
if self.shift:
|
| 108 |
+
return self.conv(x.transpose(1, 2)).transpose(1, 2)[:, 1:]
|
| 109 |
+
else:
|
| 110 |
+
return self.conv(x.transpose(1, 2)).transpose(1, 2)
|
| 111 |
+
|
| 112 |
+
class MultiheadAttention(nn.Module):
|
| 113 |
+
"""Multi-head self-attention layer."""
|
| 114 |
+
|
| 115 |
+
def __init__(self,
|
| 116 |
+
embed_dim,
|
| 117 |
+
num_heads,
|
| 118 |
+
dropout=0.,
|
| 119 |
+
bias=True,
|
| 120 |
+
v_proj=True,
|
| 121 |
+
out_proj=True,
|
| 122 |
+
relative_bias=True):
|
| 123 |
+
"""Initialization.
|
| 124 |
+
|
| 125 |
+
Args:
|
| 126 |
+
embed_dim: dimension of input embeddings
|
| 127 |
+
num_heads: number of self-attention heads
|
| 128 |
+
dropout: dropout rate
|
| 129 |
+
bias: bool, indicate whether include bias for linear transformations
|
| 130 |
+
v_proj: bool, indicate whether project inputs to new values
|
| 131 |
+
out_proj: bool, indicate whether project outputs to new values
|
| 132 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 133 |
+
attention bias
|
| 134 |
+
"""
|
| 135 |
+
|
| 136 |
+
super(MultiheadAttention, self).__init__()
|
| 137 |
+
self.embed_dim = embed_dim
|
| 138 |
+
|
| 139 |
+
self.num_heads = num_heads
|
| 140 |
+
self.drop = nn.Dropout(dropout)
|
| 141 |
+
self.head_dim = embed_dim // num_heads
|
| 142 |
+
assert self.head_dim * num_heads == self.embed_dim, ("embed_dim must be "
|
| 143 |
+
"divisible by "
|
| 144 |
+
"num_heads")
|
| 145 |
+
|
| 146 |
+
self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 147 |
+
self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 148 |
+
if v_proj:
|
| 149 |
+
self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 150 |
+
else:
|
| 151 |
+
self.v_proj = nn.Identity()
|
| 152 |
+
|
| 153 |
+
if out_proj:
|
| 154 |
+
self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 155 |
+
else:
|
| 156 |
+
self.out_proj = nn.Identity()
|
| 157 |
+
|
| 158 |
+
if relative_bias:
|
| 159 |
+
self.relative_bias = nn.Parameter(torch.zeros((self.num_heads, 512)))
|
| 160 |
+
else:
|
| 161 |
+
self.relative_bias = None
|
| 162 |
+
|
| 163 |
+
self._reset_parameters()
|
| 164 |
+
|
| 165 |
+
def _reset_parameters(self):
|
| 166 |
+
"""Initialize attention parameters."""
|
| 167 |
+
|
| 168 |
+
init.xavier_uniform_(self.q_proj.weight)
|
| 169 |
+
init.constant_(self.q_proj.bias, 0.)
|
| 170 |
+
|
| 171 |
+
init.xavier_uniform_(self.k_proj.weight)
|
| 172 |
+
init.constant_(self.k_proj.bias, 0.)
|
| 173 |
+
|
| 174 |
+
if isinstance(self.v_proj, nn.Linear):
|
| 175 |
+
init.xavier_uniform_(self.v_proj.weight)
|
| 176 |
+
init.constant_(self.v_proj.bias, 0.)
|
| 177 |
+
|
| 178 |
+
if isinstance(self.out_proj, nn.Linear):
|
| 179 |
+
init.xavier_uniform_(self.out_proj.weight)
|
| 180 |
+
init.constant_(self.out_proj.bias, 0.)
|
| 181 |
+
|
| 182 |
+
def forward(self, query, key_padding_mask=None, attn_mask=None):
|
| 183 |
+
"""Compute multi-head self-attention.
|
| 184 |
+
|
| 185 |
+
Args:
|
| 186 |
+
query: input embeddings
|
| 187 |
+
key_padding_mask: 3D mask that prevents attention to certain positions
|
| 188 |
+
attn_mask: 3D mask that rescale the attention weight at each position
|
| 189 |
+
Returns:
|
| 190 |
+
attn_output: self-attention output
|
| 191 |
+
"""
|
| 192 |
+
|
| 193 |
+
length, bsz, embed_dim = query.size()
|
| 194 |
+
assert embed_dim == self.embed_dim
|
| 195 |
+
|
| 196 |
+
head_dim = embed_dim // self.num_heads
|
| 197 |
+
assert head_dim * self.num_heads == embed_dim, ("embed_dim must be "
|
| 198 |
+
"divisible by num_heads")
|
| 199 |
+
scaling = float(head_dim)**-0.5
|
| 200 |
+
|
| 201 |
+
q = self.q_proj(query)
|
| 202 |
+
k = self.k_proj(query)
|
| 203 |
+
v = self.v_proj(query)
|
| 204 |
+
|
| 205 |
+
q = q * scaling
|
| 206 |
+
|
| 207 |
+
if attn_mask is not None:
|
| 208 |
+
assert list(attn_mask.size()) == [bsz * self.num_heads,
|
| 209 |
+
query.size(0), query.size(0)]
|
| 210 |
+
|
| 211 |
+
q = q.contiguous().view(length, bsz * self.num_heads,
|
| 212 |
+
head_dim).transpose(0, 1)
|
| 213 |
+
k = k.contiguous().view(length, bsz * self.num_heads,
|
| 214 |
+
head_dim).transpose(0, 1)
|
| 215 |
+
v = v.contiguous().view(length, bsz * self.num_heads,
|
| 216 |
+
head_dim).transpose(0, 1)
|
| 217 |
+
|
| 218 |
+
attn_output_weights = torch.bmm(q, k.transpose(1, 2))
|
| 219 |
+
assert list(
|
| 220 |
+
attn_output_weights.size()) == [bsz * self.num_heads, length, length]
|
| 221 |
+
|
| 222 |
+
if self.relative_bias is not None:
|
| 223 |
+
pos = torch.arange(length, device=query.device)
|
| 224 |
+
relative_pos = torch.abs(pos[:, None] - pos[None, :]) + 256
|
| 225 |
+
relative_pos = relative_pos[None, :, :].expand(bsz * self.num_heads, -1,
|
| 226 |
+
-1)
|
| 227 |
+
|
| 228 |
+
relative_bias = self.relative_bias.repeat_interleave(bsz, dim=0)
|
| 229 |
+
relative_bias = relative_bias[:, None, :].expand(-1, length, -1)
|
| 230 |
+
relative_bias = torch.gather(relative_bias, 2, relative_pos)
|
| 231 |
+
attn_output_weights = attn_output_weights + relative_bias
|
| 232 |
+
|
| 233 |
+
if key_padding_mask is not None:
|
| 234 |
+
attn_output_weights = attn_output_weights + key_padding_mask
|
| 235 |
+
|
| 236 |
+
if attn_mask is None:
|
| 237 |
+
attn_output_weights = torch.softmax(attn_output_weights, dim=-1)
|
| 238 |
+
else:
|
| 239 |
+
attn_output_weights = torch.sigmoid(attn_output_weights) * attn_mask
|
| 240 |
+
|
| 241 |
+
attn_output_weights = self.drop(attn_output_weights)
|
| 242 |
+
|
| 243 |
+
attn_output = torch.bmm(attn_output_weights, v)
|
| 244 |
+
|
| 245 |
+
assert list(attn_output.size()) == [bsz * self.num_heads, length, head_dim]
|
| 246 |
+
attn_output = attn_output.transpose(0, 1).contiguous().view(
|
| 247 |
+
length, bsz, embed_dim)
|
| 248 |
+
attn_output = self.out_proj(attn_output)
|
| 249 |
+
|
| 250 |
+
return attn_output
|
| 251 |
+
|
| 252 |
+
class TransformerLayer(nn.Module):
|
| 253 |
+
"""TransformerEncoderLayer is made up of self-attn and feedforward network."""
|
| 254 |
+
|
| 255 |
+
def __init__(self,
|
| 256 |
+
d_model,
|
| 257 |
+
nhead,
|
| 258 |
+
dim_feedforward=2048,
|
| 259 |
+
dropout=0.1,
|
| 260 |
+
dropatt=0.1,
|
| 261 |
+
activation="leakyrelu",
|
| 262 |
+
relative_bias=True):
|
| 263 |
+
"""Initialization.
|
| 264 |
+
|
| 265 |
+
Args:
|
| 266 |
+
d_model: dimension of inputs
|
| 267 |
+
nhead: number of self-attention heads
|
| 268 |
+
dim_feedforward: dimension of hidden layer in feedforward layer
|
| 269 |
+
dropout: dropout rate
|
| 270 |
+
dropatt: drop attention rate
|
| 271 |
+
activation: activation function
|
| 272 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 273 |
+
attention bias
|
| 274 |
+
"""
|
| 275 |
+
|
| 276 |
+
super(TransformerLayer, self).__init__()
|
| 277 |
+
|
| 278 |
+
self.self_attn = MultiheadAttention(
|
| 279 |
+
d_model, nhead, dropout=dropatt, relative_bias=relative_bias)
|
| 280 |
+
|
| 281 |
+
# Implementation of Feedforward model
|
| 282 |
+
self.feedforward = nn.Sequential(
|
| 283 |
+
nn.LayerNorm(d_model), nn.Linear(d_model, dim_feedforward),
|
| 284 |
+
_get_activation_fn(activation), nn.Dropout(dropout),
|
| 285 |
+
nn.Linear(dim_feedforward, d_model))
|
| 286 |
+
|
| 287 |
+
self.norm = nn.LayerNorm(d_model)
|
| 288 |
+
self.dropout1 = nn.Dropout(dropout)
|
| 289 |
+
self.dropout2 = nn.Dropout(dropout)
|
| 290 |
+
|
| 291 |
+
self.nhead = nhead
|
| 292 |
+
|
| 293 |
+
def forward(self, src, attn_mask=None, key_padding_mask=None):
|
| 294 |
+
"""Pass the input through the encoder layer.
|
| 295 |
+
|
| 296 |
+
Args:
|
| 297 |
+
src: the sequence to the encoder layer (required).
|
| 298 |
+
attn_mask: the mask for the src sequence (optional).
|
| 299 |
+
key_padding_mask: the mask for the src keys per batch (optional).
|
| 300 |
+
Returns:
|
| 301 |
+
src3: the output of transformer layer, share the same shape as src.
|
| 302 |
+
"""
|
| 303 |
+
src2 = self.self_attn(
|
| 304 |
+
self.norm(src), attn_mask=attn_mask, key_padding_mask=key_padding_mask)
|
| 305 |
+
src2 = src + self.dropout1(src2)
|
| 306 |
+
src3 = self.feedforward(src2)
|
| 307 |
+
src3 = src2 + self.dropout2(src3)
|
| 308 |
+
|
| 309 |
+
return src3
|
| 310 |
+
|
| 311 |
+
|
| 312 |
+
|
| 313 |
+
class RobertaClassificationHead(nn.Module):
|
| 314 |
+
"""Head for sentence-level classification tasks."""
|
| 315 |
+
|
| 316 |
+
def __init__(self, config):
|
| 317 |
+
super().__init__()
|
| 318 |
+
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
|
| 319 |
+
classifier_dropout = (
|
| 320 |
+
config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
|
| 321 |
+
)
|
| 322 |
+
self.dropout = nn.Dropout(classifier_dropout)
|
| 323 |
+
self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
|
| 324 |
+
|
| 325 |
+
def forward(self, features, **kwargs):
|
| 326 |
+
x = features[:, 0, :] # take <s> token (equiv. to [CLS])
|
| 327 |
+
x = self.dropout(x)
|
| 328 |
+
x = self.dense(x)
|
| 329 |
+
x = torch.tanh(x)
|
| 330 |
+
x = self.dropout(x)
|
| 331 |
+
x = self.out_proj(x)
|
| 332 |
+
return x
|
| 333 |
+
|
| 334 |
+
|
| 335 |
+
##########################################
|
| 336 |
+
# Custom Models
|
| 337 |
+
##########################################
|
| 338 |
+
def cumprod(x, reverse=False, exclusive=False):
|
| 339 |
+
"""cumulative product."""
|
| 340 |
+
if reverse:
|
| 341 |
+
x = x.flip([-1])
|
| 342 |
+
|
| 343 |
+
if exclusive:
|
| 344 |
+
x = F.pad(x[:, :, :-1], (1, 0), value=1)
|
| 345 |
+
|
| 346 |
+
cx = x.cumprod(-1)
|
| 347 |
+
|
| 348 |
+
if reverse:
|
| 349 |
+
cx = cx.flip([-1])
|
| 350 |
+
return cx
|
| 351 |
+
|
| 352 |
+
def cumsum(x, reverse=False, exclusive=False):
|
| 353 |
+
"""cumulative sum."""
|
| 354 |
+
bsz, _, length = x.size()
|
| 355 |
+
device = x.device
|
| 356 |
+
if reverse:
|
| 357 |
+
if exclusive:
|
| 358 |
+
w = torch.ones([bsz, length, length], device=device).tril(-1)
|
| 359 |
+
else:
|
| 360 |
+
w = torch.ones([bsz, length, length], device=device).tril(0)
|
| 361 |
+
cx = torch.bmm(x, w)
|
| 362 |
+
else:
|
| 363 |
+
if exclusive:
|
| 364 |
+
w = torch.ones([bsz, length, length], device=device).triu(1)
|
| 365 |
+
else:
|
| 366 |
+
w = torch.ones([bsz, length, length], device=device).triu(0)
|
| 367 |
+
cx = torch.bmm(x, w)
|
| 368 |
+
return cx
|
| 369 |
+
|
| 370 |
+
def cummin(x, reverse=False, exclusive=False, max_value=1e9):
|
| 371 |
+
"""cumulative min."""
|
| 372 |
+
if reverse:
|
| 373 |
+
if exclusive:
|
| 374 |
+
x = F.pad(x[:, :, 1:], (0, 1), value=max_value)
|
| 375 |
+
x = x.flip([-1]).cummin(-1)[0].flip([-1])
|
| 376 |
+
else:
|
| 377 |
+
if exclusive:
|
| 378 |
+
x = F.pad(x[:, :, :-1], (1, 0), value=max_value)
|
| 379 |
+
x = x.cummin(-1)[0]
|
| 380 |
+
return x
|
| 381 |
+
|
| 382 |
+
class Transformer(nn.Module):
|
| 383 |
+
"""Transformer model."""
|
| 384 |
+
|
| 385 |
+
def __init__(self,
|
| 386 |
+
hidden_size,
|
| 387 |
+
nlayers,
|
| 388 |
+
ntokens,
|
| 389 |
+
nhead=8,
|
| 390 |
+
dropout=0.1,
|
| 391 |
+
dropatt=0.1,
|
| 392 |
+
relative_bias=True,
|
| 393 |
+
pos_emb=False,
|
| 394 |
+
pad=0):
|
| 395 |
+
"""Initialization.
|
| 396 |
+
|
| 397 |
+
Args:
|
| 398 |
+
hidden_size: dimension of inputs and hidden states
|
| 399 |
+
nlayers: number of layers
|
| 400 |
+
ntokens: number of output categories
|
| 401 |
+
nhead: number of self-attention heads
|
| 402 |
+
dropout: dropout rate
|
| 403 |
+
dropatt: drop attention rate
|
| 404 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 405 |
+
attention bias
|
| 406 |
+
pos_emb: bool, indicate whether use a learnable positional embedding
|
| 407 |
+
pad: pad token index
|
| 408 |
+
"""
|
| 409 |
+
|
| 410 |
+
super(Transformer, self).__init__()
|
| 411 |
+
|
| 412 |
+
self.drop = nn.Dropout(dropout)
|
| 413 |
+
|
| 414 |
+
self.emb = nn.Embedding(ntokens, hidden_size)
|
| 415 |
+
if pos_emb:
|
| 416 |
+
self.pos_emb = nn.Embedding(500, hidden_size)
|
| 417 |
+
|
| 418 |
+
self.layers = nn.ModuleList([
|
| 419 |
+
TransformerLayer(hidden_size, nhead, hidden_size * 4, dropout,
|
| 420 |
+
dropatt=dropatt, relative_bias=relative_bias)
|
| 421 |
+
for _ in range(nlayers)])
|
| 422 |
+
|
| 423 |
+
self.norm = nn.LayerNorm(hidden_size)
|
| 424 |
+
|
| 425 |
+
self.output_layer = nn.Linear(hidden_size, ntokens)
|
| 426 |
+
self.output_layer.weight = self.emb.weight
|
| 427 |
+
|
| 428 |
+
self.init_weights()
|
| 429 |
+
|
| 430 |
+
self.nlayers = nlayers
|
| 431 |
+
self.nhead = nhead
|
| 432 |
+
self.ntokens = ntokens
|
| 433 |
+
self.hidden_size = hidden_size
|
| 434 |
+
self.pad = pad
|
| 435 |
+
|
| 436 |
+
def init_weights(self):
|
| 437 |
+
"""Initialize token embedding and output bias."""
|
| 438 |
+
initrange = 0.1
|
| 439 |
+
self.emb.weight.data.uniform_(-initrange, initrange)
|
| 440 |
+
if hasattr(self, 'pos_emb'):
|
| 441 |
+
self.pos_emb.weight.data.uniform_(-initrange, initrange)
|
| 442 |
+
self.output_layer.bias.data.fill_(0)
|
| 443 |
+
|
| 444 |
+
def visibility(self, x, device):
|
| 445 |
+
"""Mask pad tokens."""
|
| 446 |
+
visibility = (x != self.pad).float()
|
| 447 |
+
visibility = visibility[:, None, :].expand(-1, x.size(1), -1)
|
| 448 |
+
visibility = torch.repeat_interleave(visibility, self.nhead, dim=0)
|
| 449 |
+
return visibility.log()
|
| 450 |
+
|
| 451 |
+
def encode(self, x, pos):
|
| 452 |
+
"""Standard transformer encode process."""
|
| 453 |
+
h = self.emb(x)
|
| 454 |
+
if hasattr(self, 'pos_emb'):
|
| 455 |
+
h = h + self.pos_emb(pos)
|
| 456 |
+
h_list = []
|
| 457 |
+
visibility = self.visibility(x, x.device)
|
| 458 |
+
|
| 459 |
+
for i in range(self.nlayers):
|
| 460 |
+
h_list.append(h)
|
| 461 |
+
h = self.layers[i](
|
| 462 |
+
h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
|
| 463 |
+
|
| 464 |
+
output = h
|
| 465 |
+
h_array = torch.stack(h_list, dim=2)
|
| 466 |
+
|
| 467 |
+
return output, h_array
|
| 468 |
+
|
| 469 |
+
def forward(self, x, pos):
|
| 470 |
+
"""Pass the input through the encoder layer.
|
| 471 |
+
|
| 472 |
+
Args:
|
| 473 |
+
x: input tokens (required).
|
| 474 |
+
pos: position for each token (optional).
|
| 475 |
+
Returns:
|
| 476 |
+
output: probability distributions for missing tokens.
|
| 477 |
+
state_dict: parsing results and raw output
|
| 478 |
+
"""
|
| 479 |
+
|
| 480 |
+
batch_size, length = x.size()
|
| 481 |
+
|
| 482 |
+
raw_output, _ = self.encode(x, pos)
|
| 483 |
+
raw_output = self.norm(raw_output)
|
| 484 |
+
raw_output = self.drop(raw_output)
|
| 485 |
+
|
| 486 |
+
output = self.output_layer(raw_output)
|
| 487 |
+
return output.view(batch_size * length, -1), {'raw_output': raw_output,}
|
| 488 |
+
|
| 489 |
+
class StructFormer(Transformer):
|
| 490 |
+
"""StructFormer model."""
|
| 491 |
+
|
| 492 |
+
def __init__(self,
|
| 493 |
+
hidden_size,
|
| 494 |
+
n_context_layers,
|
| 495 |
+
nlayers,
|
| 496 |
+
ntokens,
|
| 497 |
+
nhead=8,
|
| 498 |
+
dropout=0.1,
|
| 499 |
+
dropatt=0.1,
|
| 500 |
+
relative_bias=False,
|
| 501 |
+
pos_emb=False,
|
| 502 |
+
pad=0,
|
| 503 |
+
n_parser_layers=4,
|
| 504 |
+
conv_size=9,
|
| 505 |
+
relations=('head', 'child'),
|
| 506 |
+
weight_act='softmax'):
|
| 507 |
+
"""Initialization.
|
| 508 |
+
|
| 509 |
+
Args:
|
| 510 |
+
hidden_size: dimension of inputs and hidden states
|
| 511 |
+
nlayers: number of layers
|
| 512 |
+
ntokens: number of output categories
|
| 513 |
+
nhead: number of self-attention heads
|
| 514 |
+
dropout: dropout rate
|
| 515 |
+
dropatt: drop attention rate
|
| 516 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 517 |
+
attention bias
|
| 518 |
+
pos_emb: bool, indicate whether use a learnable positional embedding
|
| 519 |
+
pad: pad token index
|
| 520 |
+
n_parser_layers: number of parsing layers
|
| 521 |
+
conv_size: convolution kernel size for parser
|
| 522 |
+
relations: relations that are used to compute self attention
|
| 523 |
+
weight_act: relations distribution activation function
|
| 524 |
+
"""
|
| 525 |
+
|
| 526 |
+
super(StructFormer, self).__init__(
|
| 527 |
+
hidden_size,
|
| 528 |
+
nlayers,
|
| 529 |
+
ntokens,
|
| 530 |
+
nhead=nhead,
|
| 531 |
+
dropout=dropout,
|
| 532 |
+
dropatt=dropatt,
|
| 533 |
+
relative_bias=relative_bias,
|
| 534 |
+
pos_emb=pos_emb,
|
| 535 |
+
pad=pad)
|
| 536 |
+
|
| 537 |
+
|
| 538 |
+
def encode(self, x, pos):
|
| 539 |
+
h = self.emb(x)
|
| 540 |
+
if hasattr(self, 'pos_emb'):
|
| 541 |
+
h = h + self.pos_emb(pos)
|
| 542 |
+
h_list = []
|
| 543 |
+
visibility = self.visibility(x, x.device)
|
| 544 |
+
|
| 545 |
+
for i in range(self.nlayers):
|
| 546 |
+
h_list.append(h)
|
| 547 |
+
h = self.layers[i](
|
| 548 |
+
h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
|
| 549 |
+
|
| 550 |
+
output = h
|
| 551 |
+
h_array = torch.stack(h_list, dim=2)
|
| 552 |
+
|
| 553 |
+
return output
|
| 554 |
+
|
| 555 |
+
|
| 556 |
+
def forward(self, input_ids, labels=None, position_ids=None, **kwargs):
|
| 557 |
+
|
| 558 |
+
x = input_ids
|
| 559 |
+
batch_size, length = x.size()
|
| 560 |
+
|
| 561 |
+
if position_ids is None:
|
| 562 |
+
pos = torch.arange(length, device=x.device).expand(batch_size, length)
|
| 563 |
+
|
| 564 |
+
raw_output = self.encode(x, pos)
|
| 565 |
+
raw_output = self.norm(raw_output)
|
| 566 |
+
raw_output = self.drop(raw_output)
|
| 567 |
+
|
| 568 |
+
output = self.output_layer(raw_output)
|
| 569 |
+
|
| 570 |
+
loss = None
|
| 571 |
+
if labels is not None:
|
| 572 |
+
loss_fct = nn.CrossEntropyLoss()
|
| 573 |
+
loss = loss_fct(output.view(batch_size * length, -1), labels.reshape(-1))
|
| 574 |
+
|
| 575 |
+
return MaskedLMOutput(
|
| 576 |
+
loss=loss, # shape: 1
|
| 577 |
+
logits=output, # shape: (batch_size * length, ntokens)
|
| 578 |
+
hidden_states=None,
|
| 579 |
+
attentions=None,
|
| 580 |
+
)
|
| 581 |
+
|
| 582 |
+
##########################################
|
| 583 |
+
# HuggingFace Model
|
| 584 |
+
##########################################
|
| 585 |
+
class StructformerModel(PreTrainedModel):
|
| 586 |
+
config_class = StructformerConfig
|
| 587 |
+
|
| 588 |
+
def __init__(self, config):
|
| 589 |
+
super().__init__(config)
|
| 590 |
+
self.model = StructFormer(
|
| 591 |
+
hidden_size=config.hidden_size,
|
| 592 |
+
n_context_layers=config.n_context_layers,
|
| 593 |
+
nlayers=config.nlayers,
|
| 594 |
+
ntokens=config.ntokens,
|
| 595 |
+
nhead=config.nhead,
|
| 596 |
+
dropout=config.dropout,
|
| 597 |
+
dropatt=config.dropatt,
|
| 598 |
+
relative_bias=config.relative_bias,
|
| 599 |
+
pos_emb=config.pos_emb,
|
| 600 |
+
pad=config.pad,
|
| 601 |
+
n_parser_layers=config.n_parser_layers,
|
| 602 |
+
conv_size=config.conv_size,
|
| 603 |
+
relations=config.relations,
|
| 604 |
+
weight_act=config.weight_act
|
| 605 |
+
)
|
| 606 |
+
|
| 607 |
+
def forward(self, input_ids, labels=None, **kwargs):
|
| 608 |
+
return self.model(input_ids, labels=labels, **kwargs)
|
| 609 |
+
|
| 610 |
+
|
| 611 |
+
class StructFormerClassification(Transformer):
|
| 612 |
+
"""StructFormer model."""
|
| 613 |
+
|
| 614 |
+
def __init__(self,
|
| 615 |
+
hidden_size,
|
| 616 |
+
n_context_layers,
|
| 617 |
+
nlayers,
|
| 618 |
+
ntokens,
|
| 619 |
+
nhead=8,
|
| 620 |
+
dropout=0.1,
|
| 621 |
+
dropatt=0.1,
|
| 622 |
+
relative_bias=False,
|
| 623 |
+
pos_emb=False,
|
| 624 |
+
pad=0,
|
| 625 |
+
n_parser_layers=4,
|
| 626 |
+
conv_size=9,
|
| 627 |
+
relations=('head', 'child'),
|
| 628 |
+
weight_act='softmax',
|
| 629 |
+
config=None,
|
| 630 |
+
):
|
| 631 |
+
|
| 632 |
+
|
| 633 |
+
super(StructFormerClassification, self).__init__(
|
| 634 |
+
hidden_size,
|
| 635 |
+
nlayers,
|
| 636 |
+
ntokens,
|
| 637 |
+
nhead=nhead,
|
| 638 |
+
dropout=dropout,
|
| 639 |
+
dropatt=dropatt,
|
| 640 |
+
relative_bias=relative_bias,
|
| 641 |
+
pos_emb=pos_emb,
|
| 642 |
+
pad=pad)
|
| 643 |
+
|
| 644 |
+
self.num_labels = config.num_labels
|
| 645 |
+
self.config = config
|
| 646 |
+
|
| 647 |
+
self.classifier = RobertaClassificationHead(config)
|
| 648 |
+
|
| 649 |
+
def encode(self, x, pos):
|
| 650 |
+
h = self.emb(x)
|
| 651 |
+
if hasattr(self, 'pos_emb'):
|
| 652 |
+
h = h + self.pos_emb(pos)
|
| 653 |
+
h_list = []
|
| 654 |
+
visibility = self.visibility(x, x.device)
|
| 655 |
+
|
| 656 |
+
for i in range(self.nlayers):
|
| 657 |
+
h_list.append(h)
|
| 658 |
+
h = self.layers[i](
|
| 659 |
+
h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
|
| 660 |
+
|
| 661 |
+
output = h
|
| 662 |
+
h_array = torch.stack(h_list, dim=2)
|
| 663 |
+
|
| 664 |
+
return output
|
| 665 |
+
|
| 666 |
+
|
| 667 |
+
def forward(self, input_ids, labels=None, position_ids=None, **kwargs):
|
| 668 |
+
|
| 669 |
+
x = input_ids
|
| 670 |
+
batch_size, length = x.size()
|
| 671 |
+
|
| 672 |
+
if position_ids is None:
|
| 673 |
+
pos = torch.arange(length, device=x.device).expand(batch_size, length)
|
| 674 |
+
|
| 675 |
+
raw_output = self.encode(x, pos)
|
| 676 |
+
raw_output = self.norm(raw_output)
|
| 677 |
+
raw_output = self.drop(raw_output)
|
| 678 |
+
|
| 679 |
+
#output = self.output_layer(raw_output)
|
| 680 |
+
logits = self.classifier(raw_output)
|
| 681 |
+
|
| 682 |
+
loss = None
|
| 683 |
+
if labels is not None:
|
| 684 |
+
if self.config.problem_type is None:
|
| 685 |
+
if self.num_labels == 1:
|
| 686 |
+
self.config.problem_type = "regression"
|
| 687 |
+
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
| 688 |
+
self.config.problem_type = "single_label_classification"
|
| 689 |
+
else:
|
| 690 |
+
self.config.problem_type = "multi_label_classification"
|
| 691 |
+
|
| 692 |
+
if self.config.problem_type == "regression":
|
| 693 |
+
loss_fct = MSELoss()
|
| 694 |
+
if self.num_labels == 1:
|
| 695 |
+
loss = loss_fct(logits.squeeze(), labels.squeeze())
|
| 696 |
+
else:
|
| 697 |
+
loss = loss_fct(logits, labels)
|
| 698 |
+
elif self.config.problem_type == "single_label_classification":
|
| 699 |
+
loss_fct = CrossEntropyLoss()
|
| 700 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
| 701 |
+
elif self.config.problem_type == "multi_label_classification":
|
| 702 |
+
loss_fct = BCEWithLogitsLoss()
|
| 703 |
+
loss = loss_fct(logits, labels)
|
| 704 |
+
|
| 705 |
+
|
| 706 |
+
return SequenceClassifierOutput(
|
| 707 |
+
loss=loss,
|
| 708 |
+
logits=logits,
|
| 709 |
+
hidden_states=None,
|
| 710 |
+
attentions=None,
|
| 711 |
+
)
|
| 712 |
+
|
| 713 |
+
|
| 714 |
+
class StructformerModelForSequenceClassification(PreTrainedModel):
|
| 715 |
+
config_class = StructformerConfig
|
| 716 |
+
def __init__(self, config):
|
| 717 |
+
super().__init__(config)
|
| 718 |
+
self.model = StructFormerClassification(
|
| 719 |
+
hidden_size=config.hidden_size,
|
| 720 |
+
n_context_layers=config.n_context_layers,
|
| 721 |
+
nlayers=config.nlayers,
|
| 722 |
+
ntokens=config.ntokens,
|
| 723 |
+
nhead=config.nhead,
|
| 724 |
+
dropout=config.dropout,
|
| 725 |
+
dropatt=config.dropatt,
|
| 726 |
+
relative_bias=config.relative_bias,
|
| 727 |
+
pos_emb=config.pos_emb,
|
| 728 |
+
pad=config.pad,
|
| 729 |
+
n_parser_layers=config.n_parser_layers,
|
| 730 |
+
conv_size=config.conv_size,
|
| 731 |
+
relations=config.relations,
|
| 732 |
+
weight_act=config.weight_act,
|
| 733 |
+
config=config)
|
| 734 |
+
|
| 735 |
+
def _init_weights(self, module):
|
| 736 |
+
"""Initialize the weights"""
|
| 737 |
+
if isinstance(module, nn.Linear):
|
| 738 |
+
# Slightly different from the TF version which uses truncated_normal for initialization
|
| 739 |
+
# cf https://github.com/pytorch/pytorch/pull/5617
|
| 740 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 741 |
+
if module.bias is not None:
|
| 742 |
+
module.bias.data.zero_()
|
| 743 |
+
elif isinstance(module, nn.Embedding):
|
| 744 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 745 |
+
if module.padding_idx is not None:
|
| 746 |
+
module.weight.data[module.padding_idx].zero_()
|
| 747 |
+
elif isinstance(module, nn.LayerNorm):
|
| 748 |
+
if module.bias is not None:
|
| 749 |
+
module.bias.data.zero_()
|
| 750 |
+
module.weight.data.fill_(1.0)
|
| 751 |
+
|
| 752 |
+
|
| 753 |
+
def forward(self, input_ids, labels=None, **kwargs):
|
| 754 |
+
return self.model(input_ids, labels=labels, **kwargs)
|
finetune/cola/checkpoint-400/tokenizer_config.json
ADDED
|
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_prefix_space": false,
|
| 3 |
+
"bos_token": {
|
| 4 |
+
"__type": "AddedToken",
|
| 5 |
+
"content": "<s>",
|
| 6 |
+
"lstrip": false,
|
| 7 |
+
"normalized": true,
|
| 8 |
+
"rstrip": false,
|
| 9 |
+
"single_word": false
|
| 10 |
+
},
|
| 11 |
+
"cls_token": {
|
| 12 |
+
"__type": "AddedToken",
|
| 13 |
+
"content": "<s>",
|
| 14 |
+
"lstrip": false,
|
| 15 |
+
"normalized": true,
|
| 16 |
+
"rstrip": false,
|
| 17 |
+
"single_word": false
|
| 18 |
+
},
|
| 19 |
+
"eos_token": {
|
| 20 |
+
"__type": "AddedToken",
|
| 21 |
+
"content": "</s>",
|
| 22 |
+
"lstrip": false,
|
| 23 |
+
"normalized": true,
|
| 24 |
+
"rstrip": false,
|
| 25 |
+
"single_word": false
|
| 26 |
+
},
|
| 27 |
+
"errors": "replace",
|
| 28 |
+
"mask_token": {
|
| 29 |
+
"__type": "AddedToken",
|
| 30 |
+
"content": "<mask>",
|
| 31 |
+
"lstrip": true,
|
| 32 |
+
"normalized": false,
|
| 33 |
+
"rstrip": false,
|
| 34 |
+
"single_word": false
|
| 35 |
+
},
|
| 36 |
+
"model_max_length": 512,
|
| 37 |
+
"name_or_path": "final_models/transformer_base_final_2",
|
| 38 |
+
"pad_token": {
|
| 39 |
+
"__type": "AddedToken",
|
| 40 |
+
"content": "<pad>",
|
| 41 |
+
"lstrip": false,
|
| 42 |
+
"normalized": true,
|
| 43 |
+
"rstrip": false,
|
| 44 |
+
"single_word": false
|
| 45 |
+
},
|
| 46 |
+
"sep_token": {
|
| 47 |
+
"__type": "AddedToken",
|
| 48 |
+
"content": "</s>",
|
| 49 |
+
"lstrip": false,
|
| 50 |
+
"normalized": true,
|
| 51 |
+
"rstrip": false,
|
| 52 |
+
"single_word": false
|
| 53 |
+
},
|
| 54 |
+
"special_tokens_map_file": null,
|
| 55 |
+
"tokenizer_class": "RobertaTokenizer",
|
| 56 |
+
"trim_offsets": true,
|
| 57 |
+
"unk_token": {
|
| 58 |
+
"__type": "AddedToken",
|
| 59 |
+
"content": "<unk>",
|
| 60 |
+
"lstrip": false,
|
| 61 |
+
"normalized": true,
|
| 62 |
+
"rstrip": false,
|
| 63 |
+
"single_word": false
|
| 64 |
+
}
|
| 65 |
+
}
|
finetune/cola/checkpoint-400/trainer_state.json
ADDED
|
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"best_metric": 0.7777040477770405,
|
| 3 |
+
"best_model_checkpoint": "final_models/transformer_base_final_2/finetune/cola/checkpoint-400",
|
| 4 |
+
"epoch": 5.797101449275362,
|
| 5 |
+
"global_step": 400,
|
| 6 |
+
"is_hyper_param_search": false,
|
| 7 |
+
"is_local_process_zero": true,
|
| 8 |
+
"is_world_process_zero": true,
|
| 9 |
+
"log_history": [
|
| 10 |
+
{
|
| 11 |
+
"epoch": 5.8,
|
| 12 |
+
"eval_accuracy": 0.6712462902069092,
|
| 13 |
+
"eval_f1": 0.7777040477770405,
|
| 14 |
+
"eval_loss": 0.7399011254310608,
|
| 15 |
+
"eval_mcc": 0.15981907397227785,
|
| 16 |
+
"eval_runtime": 1.4063,
|
| 17 |
+
"eval_samples_per_second": 724.62,
|
| 18 |
+
"eval_steps_per_second": 91.022,
|
| 19 |
+
"step": 400
|
| 20 |
+
}
|
| 21 |
+
],
|
| 22 |
+
"max_steps": 690,
|
| 23 |
+
"num_train_epochs": 10,
|
| 24 |
+
"total_flos": 3120346955212800.0,
|
| 25 |
+
"trial_name": null,
|
| 26 |
+
"trial_params": null
|
| 27 |
+
}
|
finetune/cola/checkpoint-400/training_args.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:42f03300332c3f519b9e94f52c9e382ae77f93b0d20616ae0c3858d1e25c21dc
|
| 3 |
+
size 3503
|
finetune/cola/checkpoint-400/vocab.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
finetune/cola/config.json
ADDED
|
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_name_or_path": "final_models/transformer_base_final_2",
|
| 3 |
+
"architectures": [
|
| 4 |
+
"StructformerModelForSequenceClassification"
|
| 5 |
+
],
|
| 6 |
+
"attention_probs_dropout_prob": 0.1,
|
| 7 |
+
"auto_map": {
|
| 8 |
+
"AutoConfig": "structformer_as_hf_no_parser.StructformerConfig",
|
| 9 |
+
"AutoModelForMaskedLM": "structformer_as_hf_no_parser.StructformerModel",
|
| 10 |
+
"AutoModelForSequenceClassification": "structformer_as_hf_no_parser.StructformerModelForSequenceClassification"
|
| 11 |
+
},
|
| 12 |
+
"bos_token_id": 0,
|
| 13 |
+
"classifier_dropout": null,
|
| 14 |
+
"conv_size": 9,
|
| 15 |
+
"dropatt": 0.1,
|
| 16 |
+
"dropout": 0.1,
|
| 17 |
+
"eos_token_id": 2,
|
| 18 |
+
"hidden_act": "gelu",
|
| 19 |
+
"hidden_dropout_prob": 0.1,
|
| 20 |
+
"hidden_size": 768,
|
| 21 |
+
"id2label": {
|
| 22 |
+
"0": 0,
|
| 23 |
+
"1": 1
|
| 24 |
+
},
|
| 25 |
+
"initializer_range": 0.02,
|
| 26 |
+
"intermediate_size": 3072,
|
| 27 |
+
"label2id": {
|
| 28 |
+
"0": 0,
|
| 29 |
+
"1": 1
|
| 30 |
+
},
|
| 31 |
+
"layer_norm_eps": 1e-05,
|
| 32 |
+
"max_position_embeddings": 514,
|
| 33 |
+
"model_type": "structformer",
|
| 34 |
+
"n_context_layers": 0,
|
| 35 |
+
"n_parser_layers": 0,
|
| 36 |
+
"nhead": 12,
|
| 37 |
+
"nlayers": 12,
|
| 38 |
+
"ntokens": 32000,
|
| 39 |
+
"num_attention_heads": 12,
|
| 40 |
+
"num_hidden_layers": 12,
|
| 41 |
+
"pad": 0,
|
| 42 |
+
"pad_token_id": 1,
|
| 43 |
+
"pos_emb": true,
|
| 44 |
+
"position_embedding_type": "absolute",
|
| 45 |
+
"problem_type": "single_label_classification",
|
| 46 |
+
"relations": [
|
| 47 |
+
"head",
|
| 48 |
+
"child"
|
| 49 |
+
],
|
| 50 |
+
"relative_bias": false,
|
| 51 |
+
"torch_dtype": "float32",
|
| 52 |
+
"transformers_version": "4.26.1",
|
| 53 |
+
"type_vocab_size": 1,
|
| 54 |
+
"use_cache": true,
|
| 55 |
+
"vocab_size": 32000,
|
| 56 |
+
"weight_act": "softmax"
|
| 57 |
+
}
|
finetune/cola/eval_results.json
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"epoch": 10.0,
|
| 3 |
+
"eval_accuracy": 0.6712462902069092,
|
| 4 |
+
"eval_f1": 0.7777040477770405,
|
| 5 |
+
"eval_loss": 0.7399011254310608,
|
| 6 |
+
"eval_mcc": 0.15981907397227785,
|
| 7 |
+
"eval_runtime": 1.4236,
|
| 8 |
+
"eval_samples": 1019,
|
| 9 |
+
"eval_samples_per_second": 715.815,
|
| 10 |
+
"eval_steps_per_second": 89.916
|
| 11 |
+
}
|
finetune/cola/merges.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
finetune/cola/predict_results.txt
ADDED
|
@@ -0,0 +1,1020 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
index prediction
|
| 2 |
+
0 1
|
| 3 |
+
1 1
|
| 4 |
+
2 1
|
| 5 |
+
3 1
|
| 6 |
+
4 1
|
| 7 |
+
5 1
|
| 8 |
+
6 0
|
| 9 |
+
7 1
|
| 10 |
+
8 1
|
| 11 |
+
9 1
|
| 12 |
+
10 1
|
| 13 |
+
11 1
|
| 14 |
+
12 1
|
| 15 |
+
13 0
|
| 16 |
+
14 1
|
| 17 |
+
15 1
|
| 18 |
+
16 0
|
| 19 |
+
17 1
|
| 20 |
+
18 1
|
| 21 |
+
19 1
|
| 22 |
+
20 1
|
| 23 |
+
21 1
|
| 24 |
+
22 1
|
| 25 |
+
23 1
|
| 26 |
+
24 0
|
| 27 |
+
25 0
|
| 28 |
+
26 1
|
| 29 |
+
27 1
|
| 30 |
+
28 0
|
| 31 |
+
29 1
|
| 32 |
+
30 1
|
| 33 |
+
31 1
|
| 34 |
+
32 1
|
| 35 |
+
33 1
|
| 36 |
+
34 1
|
| 37 |
+
35 1
|
| 38 |
+
36 1
|
| 39 |
+
37 1
|
| 40 |
+
38 1
|
| 41 |
+
39 1
|
| 42 |
+
40 0
|
| 43 |
+
41 0
|
| 44 |
+
42 1
|
| 45 |
+
43 1
|
| 46 |
+
44 0
|
| 47 |
+
45 0
|
| 48 |
+
46 1
|
| 49 |
+
47 0
|
| 50 |
+
48 1
|
| 51 |
+
49 1
|
| 52 |
+
50 1
|
| 53 |
+
51 0
|
| 54 |
+
52 0
|
| 55 |
+
53 0
|
| 56 |
+
54 0
|
| 57 |
+
55 1
|
| 58 |
+
56 1
|
| 59 |
+
57 1
|
| 60 |
+
58 1
|
| 61 |
+
59 0
|
| 62 |
+
60 0
|
| 63 |
+
61 1
|
| 64 |
+
62 1
|
| 65 |
+
63 1
|
| 66 |
+
64 1
|
| 67 |
+
65 0
|
| 68 |
+
66 1
|
| 69 |
+
67 0
|
| 70 |
+
68 1
|
| 71 |
+
69 1
|
| 72 |
+
70 1
|
| 73 |
+
71 1
|
| 74 |
+
72 1
|
| 75 |
+
73 0
|
| 76 |
+
74 0
|
| 77 |
+
75 0
|
| 78 |
+
76 1
|
| 79 |
+
77 1
|
| 80 |
+
78 1
|
| 81 |
+
79 1
|
| 82 |
+
80 0
|
| 83 |
+
81 1
|
| 84 |
+
82 1
|
| 85 |
+
83 0
|
| 86 |
+
84 1
|
| 87 |
+
85 0
|
| 88 |
+
86 0
|
| 89 |
+
87 1
|
| 90 |
+
88 0
|
| 91 |
+
89 0
|
| 92 |
+
90 1
|
| 93 |
+
91 1
|
| 94 |
+
92 1
|
| 95 |
+
93 1
|
| 96 |
+
94 1
|
| 97 |
+
95 1
|
| 98 |
+
96 0
|
| 99 |
+
97 1
|
| 100 |
+
98 1
|
| 101 |
+
99 1
|
| 102 |
+
100 1
|
| 103 |
+
101 0
|
| 104 |
+
102 1
|
| 105 |
+
103 1
|
| 106 |
+
104 1
|
| 107 |
+
105 1
|
| 108 |
+
106 0
|
| 109 |
+
107 1
|
| 110 |
+
108 0
|
| 111 |
+
109 0
|
| 112 |
+
110 1
|
| 113 |
+
111 0
|
| 114 |
+
112 1
|
| 115 |
+
113 1
|
| 116 |
+
114 0
|
| 117 |
+
115 1
|
| 118 |
+
116 1
|
| 119 |
+
117 1
|
| 120 |
+
118 0
|
| 121 |
+
119 1
|
| 122 |
+
120 1
|
| 123 |
+
121 1
|
| 124 |
+
122 1
|
| 125 |
+
123 1
|
| 126 |
+
124 1
|
| 127 |
+
125 1
|
| 128 |
+
126 1
|
| 129 |
+
127 1
|
| 130 |
+
128 0
|
| 131 |
+
129 1
|
| 132 |
+
130 1
|
| 133 |
+
131 1
|
| 134 |
+
132 1
|
| 135 |
+
133 1
|
| 136 |
+
134 1
|
| 137 |
+
135 1
|
| 138 |
+
136 1
|
| 139 |
+
137 1
|
| 140 |
+
138 1
|
| 141 |
+
139 1
|
| 142 |
+
140 1
|
| 143 |
+
141 1
|
| 144 |
+
142 1
|
| 145 |
+
143 1
|
| 146 |
+
144 1
|
| 147 |
+
145 1
|
| 148 |
+
146 1
|
| 149 |
+
147 1
|
| 150 |
+
148 1
|
| 151 |
+
149 1
|
| 152 |
+
150 1
|
| 153 |
+
151 0
|
| 154 |
+
152 1
|
| 155 |
+
153 1
|
| 156 |
+
154 1
|
| 157 |
+
155 1
|
| 158 |
+
156 1
|
| 159 |
+
157 0
|
| 160 |
+
158 1
|
| 161 |
+
159 0
|
| 162 |
+
160 1
|
| 163 |
+
161 1
|
| 164 |
+
162 1
|
| 165 |
+
163 1
|
| 166 |
+
164 0
|
| 167 |
+
165 1
|
| 168 |
+
166 1
|
| 169 |
+
167 1
|
| 170 |
+
168 0
|
| 171 |
+
169 1
|
| 172 |
+
170 0
|
| 173 |
+
171 0
|
| 174 |
+
172 1
|
| 175 |
+
173 1
|
| 176 |
+
174 0
|
| 177 |
+
175 1
|
| 178 |
+
176 1
|
| 179 |
+
177 1
|
| 180 |
+
178 0
|
| 181 |
+
179 1
|
| 182 |
+
180 0
|
| 183 |
+
181 0
|
| 184 |
+
182 1
|
| 185 |
+
183 1
|
| 186 |
+
184 1
|
| 187 |
+
185 1
|
| 188 |
+
186 0
|
| 189 |
+
187 0
|
| 190 |
+
188 0
|
| 191 |
+
189 1
|
| 192 |
+
190 1
|
| 193 |
+
191 1
|
| 194 |
+
192 0
|
| 195 |
+
193 1
|
| 196 |
+
194 1
|
| 197 |
+
195 1
|
| 198 |
+
196 1
|
| 199 |
+
197 0
|
| 200 |
+
198 0
|
| 201 |
+
199 1
|
| 202 |
+
200 1
|
| 203 |
+
201 1
|
| 204 |
+
202 1
|
| 205 |
+
203 1
|
| 206 |
+
204 1
|
| 207 |
+
205 1
|
| 208 |
+
206 1
|
| 209 |
+
207 1
|
| 210 |
+
208 1
|
| 211 |
+
209 1
|
| 212 |
+
210 1
|
| 213 |
+
211 1
|
| 214 |
+
212 1
|
| 215 |
+
213 1
|
| 216 |
+
214 1
|
| 217 |
+
215 1
|
| 218 |
+
216 0
|
| 219 |
+
217 1
|
| 220 |
+
218 0
|
| 221 |
+
219 1
|
| 222 |
+
220 0
|
| 223 |
+
221 1
|
| 224 |
+
222 1
|
| 225 |
+
223 1
|
| 226 |
+
224 1
|
| 227 |
+
225 0
|
| 228 |
+
226 0
|
| 229 |
+
227 1
|
| 230 |
+
228 1
|
| 231 |
+
229 1
|
| 232 |
+
230 1
|
| 233 |
+
231 1
|
| 234 |
+
232 0
|
| 235 |
+
233 1
|
| 236 |
+
234 1
|
| 237 |
+
235 1
|
| 238 |
+
236 1
|
| 239 |
+
237 1
|
| 240 |
+
238 1
|
| 241 |
+
239 1
|
| 242 |
+
240 1
|
| 243 |
+
241 1
|
| 244 |
+
242 0
|
| 245 |
+
243 0
|
| 246 |
+
244 1
|
| 247 |
+
245 1
|
| 248 |
+
246 1
|
| 249 |
+
247 1
|
| 250 |
+
248 0
|
| 251 |
+
249 1
|
| 252 |
+
250 1
|
| 253 |
+
251 1
|
| 254 |
+
252 1
|
| 255 |
+
253 1
|
| 256 |
+
254 0
|
| 257 |
+
255 0
|
| 258 |
+
256 1
|
| 259 |
+
257 0
|
| 260 |
+
258 1
|
| 261 |
+
259 1
|
| 262 |
+
260 1
|
| 263 |
+
261 1
|
| 264 |
+
262 0
|
| 265 |
+
263 1
|
| 266 |
+
264 1
|
| 267 |
+
265 1
|
| 268 |
+
266 0
|
| 269 |
+
267 1
|
| 270 |
+
268 1
|
| 271 |
+
269 0
|
| 272 |
+
270 1
|
| 273 |
+
271 1
|
| 274 |
+
272 1
|
| 275 |
+
273 1
|
| 276 |
+
274 1
|
| 277 |
+
275 1
|
| 278 |
+
276 0
|
| 279 |
+
277 1
|
| 280 |
+
278 1
|
| 281 |
+
279 1
|
| 282 |
+
280 1
|
| 283 |
+
281 1
|
| 284 |
+
282 1
|
| 285 |
+
283 0
|
| 286 |
+
284 1
|
| 287 |
+
285 0
|
| 288 |
+
286 1
|
| 289 |
+
287 1
|
| 290 |
+
288 1
|
| 291 |
+
289 1
|
| 292 |
+
290 1
|
| 293 |
+
291 0
|
| 294 |
+
292 0
|
| 295 |
+
293 0
|
| 296 |
+
294 1
|
| 297 |
+
295 1
|
| 298 |
+
296 0
|
| 299 |
+
297 1
|
| 300 |
+
298 1
|
| 301 |
+
299 1
|
| 302 |
+
300 0
|
| 303 |
+
301 1
|
| 304 |
+
302 1
|
| 305 |
+
303 1
|
| 306 |
+
304 0
|
| 307 |
+
305 1
|
| 308 |
+
306 1
|
| 309 |
+
307 1
|
| 310 |
+
308 0
|
| 311 |
+
309 1
|
| 312 |
+
310 1
|
| 313 |
+
311 1
|
| 314 |
+
312 0
|
| 315 |
+
313 0
|
| 316 |
+
314 1
|
| 317 |
+
315 1
|
| 318 |
+
316 1
|
| 319 |
+
317 1
|
| 320 |
+
318 1
|
| 321 |
+
319 1
|
| 322 |
+
320 1
|
| 323 |
+
321 1
|
| 324 |
+
322 0
|
| 325 |
+
323 0
|
| 326 |
+
324 1
|
| 327 |
+
325 1
|
| 328 |
+
326 1
|
| 329 |
+
327 1
|
| 330 |
+
328 1
|
| 331 |
+
329 1
|
| 332 |
+
330 1
|
| 333 |
+
331 1
|
| 334 |
+
332 1
|
| 335 |
+
333 1
|
| 336 |
+
334 1
|
| 337 |
+
335 1
|
| 338 |
+
336 1
|
| 339 |
+
337 1
|
| 340 |
+
338 1
|
| 341 |
+
339 1
|
| 342 |
+
340 1
|
| 343 |
+
341 1
|
| 344 |
+
342 1
|
| 345 |
+
343 0
|
| 346 |
+
344 1
|
| 347 |
+
345 0
|
| 348 |
+
346 1
|
| 349 |
+
347 1
|
| 350 |
+
348 1
|
| 351 |
+
349 1
|
| 352 |
+
350 0
|
| 353 |
+
351 1
|
| 354 |
+
352 1
|
| 355 |
+
353 1
|
| 356 |
+
354 1
|
| 357 |
+
355 1
|
| 358 |
+
356 1
|
| 359 |
+
357 1
|
| 360 |
+
358 1
|
| 361 |
+
359 1
|
| 362 |
+
360 1
|
| 363 |
+
361 1
|
| 364 |
+
362 1
|
| 365 |
+
363 1
|
| 366 |
+
364 1
|
| 367 |
+
365 1
|
| 368 |
+
366 1
|
| 369 |
+
367 1
|
| 370 |
+
368 1
|
| 371 |
+
369 1
|
| 372 |
+
370 1
|
| 373 |
+
371 1
|
| 374 |
+
372 1
|
| 375 |
+
373 1
|
| 376 |
+
374 1
|
| 377 |
+
375 1
|
| 378 |
+
376 1
|
| 379 |
+
377 1
|
| 380 |
+
378 1
|
| 381 |
+
379 1
|
| 382 |
+
380 1
|
| 383 |
+
381 0
|
| 384 |
+
382 1
|
| 385 |
+
383 1
|
| 386 |
+
384 1
|
| 387 |
+
385 1
|
| 388 |
+
386 1
|
| 389 |
+
387 1
|
| 390 |
+
388 0
|
| 391 |
+
389 1
|
| 392 |
+
390 1
|
| 393 |
+
391 1
|
| 394 |
+
392 1
|
| 395 |
+
393 1
|
| 396 |
+
394 1
|
| 397 |
+
395 1
|
| 398 |
+
396 1
|
| 399 |
+
397 1
|
| 400 |
+
398 1
|
| 401 |
+
399 1
|
| 402 |
+
400 1
|
| 403 |
+
401 0
|
| 404 |
+
402 0
|
| 405 |
+
403 1
|
| 406 |
+
404 1
|
| 407 |
+
405 1
|
| 408 |
+
406 1
|
| 409 |
+
407 1
|
| 410 |
+
408 1
|
| 411 |
+
409 0
|
| 412 |
+
410 1
|
| 413 |
+
411 1
|
| 414 |
+
412 1
|
| 415 |
+
413 1
|
| 416 |
+
414 1
|
| 417 |
+
415 0
|
| 418 |
+
416 1
|
| 419 |
+
417 1
|
| 420 |
+
418 1
|
| 421 |
+
419 1
|
| 422 |
+
420 1
|
| 423 |
+
421 0
|
| 424 |
+
422 1
|
| 425 |
+
423 1
|
| 426 |
+
424 1
|
| 427 |
+
425 1
|
| 428 |
+
426 1
|
| 429 |
+
427 1
|
| 430 |
+
428 0
|
| 431 |
+
429 1
|
| 432 |
+
430 0
|
| 433 |
+
431 0
|
| 434 |
+
432 1
|
| 435 |
+
433 1
|
| 436 |
+
434 1
|
| 437 |
+
435 1
|
| 438 |
+
436 1
|
| 439 |
+
437 1
|
| 440 |
+
438 0
|
| 441 |
+
439 1
|
| 442 |
+
440 1
|
| 443 |
+
441 1
|
| 444 |
+
442 1
|
| 445 |
+
443 1
|
| 446 |
+
444 1
|
| 447 |
+
445 1
|
| 448 |
+
446 1
|
| 449 |
+
447 1
|
| 450 |
+
448 1
|
| 451 |
+
449 1
|
| 452 |
+
450 1
|
| 453 |
+
451 1
|
| 454 |
+
452 1
|
| 455 |
+
453 0
|
| 456 |
+
454 1
|
| 457 |
+
455 1
|
| 458 |
+
456 1
|
| 459 |
+
457 1
|
| 460 |
+
458 1
|
| 461 |
+
459 1
|
| 462 |
+
460 1
|
| 463 |
+
461 1
|
| 464 |
+
462 0
|
| 465 |
+
463 1
|
| 466 |
+
464 0
|
| 467 |
+
465 1
|
| 468 |
+
466 0
|
| 469 |
+
467 0
|
| 470 |
+
468 1
|
| 471 |
+
469 1
|
| 472 |
+
470 1
|
| 473 |
+
471 1
|
| 474 |
+
472 0
|
| 475 |
+
473 1
|
| 476 |
+
474 1
|
| 477 |
+
475 1
|
| 478 |
+
476 0
|
| 479 |
+
477 1
|
| 480 |
+
478 1
|
| 481 |
+
479 1
|
| 482 |
+
480 0
|
| 483 |
+
481 1
|
| 484 |
+
482 1
|
| 485 |
+
483 1
|
| 486 |
+
484 1
|
| 487 |
+
485 1
|
| 488 |
+
486 1
|
| 489 |
+
487 1
|
| 490 |
+
488 1
|
| 491 |
+
489 1
|
| 492 |
+
490 1
|
| 493 |
+
491 1
|
| 494 |
+
492 1
|
| 495 |
+
493 0
|
| 496 |
+
494 1
|
| 497 |
+
495 1
|
| 498 |
+
496 1
|
| 499 |
+
497 1
|
| 500 |
+
498 1
|
| 501 |
+
499 1
|
| 502 |
+
500 1
|
| 503 |
+
501 1
|
| 504 |
+
502 0
|
| 505 |
+
503 0
|
| 506 |
+
504 0
|
| 507 |
+
505 0
|
| 508 |
+
506 0
|
| 509 |
+
507 1
|
| 510 |
+
508 0
|
| 511 |
+
509 0
|
| 512 |
+
510 1
|
| 513 |
+
511 1
|
| 514 |
+
512 1
|
| 515 |
+
513 1
|
| 516 |
+
514 1
|
| 517 |
+
515 1
|
| 518 |
+
516 1
|
| 519 |
+
517 1
|
| 520 |
+
518 1
|
| 521 |
+
519 1
|
| 522 |
+
520 1
|
| 523 |
+
521 0
|
| 524 |
+
522 1
|
| 525 |
+
523 1
|
| 526 |
+
524 1
|
| 527 |
+
525 1
|
| 528 |
+
526 0
|
| 529 |
+
527 1
|
| 530 |
+
528 1
|
| 531 |
+
529 1
|
| 532 |
+
530 1
|
| 533 |
+
531 1
|
| 534 |
+
532 1
|
| 535 |
+
533 0
|
| 536 |
+
534 1
|
| 537 |
+
535 1
|
| 538 |
+
536 1
|
| 539 |
+
537 1
|
| 540 |
+
538 1
|
| 541 |
+
539 1
|
| 542 |
+
540 1
|
| 543 |
+
541 1
|
| 544 |
+
542 1
|
| 545 |
+
543 1
|
| 546 |
+
544 1
|
| 547 |
+
545 0
|
| 548 |
+
546 0
|
| 549 |
+
547 1
|
| 550 |
+
548 1
|
| 551 |
+
549 1
|
| 552 |
+
550 0
|
| 553 |
+
551 1
|
| 554 |
+
552 1
|
| 555 |
+
553 1
|
| 556 |
+
554 0
|
| 557 |
+
555 0
|
| 558 |
+
556 0
|
| 559 |
+
557 1
|
| 560 |
+
558 1
|
| 561 |
+
559 1
|
| 562 |
+
560 1
|
| 563 |
+
561 1
|
| 564 |
+
562 1
|
| 565 |
+
563 1
|
| 566 |
+
564 0
|
| 567 |
+
565 1
|
| 568 |
+
566 1
|
| 569 |
+
567 1
|
| 570 |
+
568 1
|
| 571 |
+
569 0
|
| 572 |
+
570 1
|
| 573 |
+
571 1
|
| 574 |
+
572 1
|
| 575 |
+
573 1
|
| 576 |
+
574 1
|
| 577 |
+
575 1
|
| 578 |
+
576 1
|
| 579 |
+
577 1
|
| 580 |
+
578 1
|
| 581 |
+
579 1
|
| 582 |
+
580 1
|
| 583 |
+
581 1
|
| 584 |
+
582 1
|
| 585 |
+
583 1
|
| 586 |
+
584 1
|
| 587 |
+
585 1
|
| 588 |
+
586 0
|
| 589 |
+
587 1
|
| 590 |
+
588 1
|
| 591 |
+
589 1
|
| 592 |
+
590 0
|
| 593 |
+
591 1
|
| 594 |
+
592 1
|
| 595 |
+
593 0
|
| 596 |
+
594 0
|
| 597 |
+
595 1
|
| 598 |
+
596 1
|
| 599 |
+
597 0
|
| 600 |
+
598 0
|
| 601 |
+
599 0
|
| 602 |
+
600 0
|
| 603 |
+
601 1
|
| 604 |
+
602 0
|
| 605 |
+
603 0
|
| 606 |
+
604 0
|
| 607 |
+
605 1
|
| 608 |
+
606 1
|
| 609 |
+
607 1
|
| 610 |
+
608 1
|
| 611 |
+
609 1
|
| 612 |
+
610 0
|
| 613 |
+
611 1
|
| 614 |
+
612 0
|
| 615 |
+
613 1
|
| 616 |
+
614 1
|
| 617 |
+
615 1
|
| 618 |
+
616 1
|
| 619 |
+
617 1
|
| 620 |
+
618 1
|
| 621 |
+
619 1
|
| 622 |
+
620 1
|
| 623 |
+
621 0
|
| 624 |
+
622 1
|
| 625 |
+
623 1
|
| 626 |
+
624 1
|
| 627 |
+
625 1
|
| 628 |
+
626 1
|
| 629 |
+
627 1
|
| 630 |
+
628 1
|
| 631 |
+
629 1
|
| 632 |
+
630 1
|
| 633 |
+
631 0
|
| 634 |
+
632 1
|
| 635 |
+
633 1
|
| 636 |
+
634 0
|
| 637 |
+
635 1
|
| 638 |
+
636 1
|
| 639 |
+
637 1
|
| 640 |
+
638 1
|
| 641 |
+
639 1
|
| 642 |
+
640 0
|
| 643 |
+
641 1
|
| 644 |
+
642 1
|
| 645 |
+
643 1
|
| 646 |
+
644 0
|
| 647 |
+
645 0
|
| 648 |
+
646 1
|
| 649 |
+
647 1
|
| 650 |
+
648 0
|
| 651 |
+
649 1
|
| 652 |
+
650 1
|
| 653 |
+
651 1
|
| 654 |
+
652 1
|
| 655 |
+
653 1
|
| 656 |
+
654 1
|
| 657 |
+
655 1
|
| 658 |
+
656 1
|
| 659 |
+
657 1
|
| 660 |
+
658 1
|
| 661 |
+
659 1
|
| 662 |
+
660 1
|
| 663 |
+
661 1
|
| 664 |
+
662 1
|
| 665 |
+
663 1
|
| 666 |
+
664 1
|
| 667 |
+
665 0
|
| 668 |
+
666 1
|
| 669 |
+
667 1
|
| 670 |
+
668 1
|
| 671 |
+
669 1
|
| 672 |
+
670 0
|
| 673 |
+
671 1
|
| 674 |
+
672 1
|
| 675 |
+
673 0
|
| 676 |
+
674 1
|
| 677 |
+
675 1
|
| 678 |
+
676 0
|
| 679 |
+
677 1
|
| 680 |
+
678 0
|
| 681 |
+
679 0
|
| 682 |
+
680 0
|
| 683 |
+
681 1
|
| 684 |
+
682 0
|
| 685 |
+
683 1
|
| 686 |
+
684 1
|
| 687 |
+
685 1
|
| 688 |
+
686 0
|
| 689 |
+
687 1
|
| 690 |
+
688 0
|
| 691 |
+
689 1
|
| 692 |
+
690 1
|
| 693 |
+
691 1
|
| 694 |
+
692 0
|
| 695 |
+
693 0
|
| 696 |
+
694 0
|
| 697 |
+
695 0
|
| 698 |
+
696 1
|
| 699 |
+
697 1
|
| 700 |
+
698 1
|
| 701 |
+
699 1
|
| 702 |
+
700 1
|
| 703 |
+
701 0
|
| 704 |
+
702 1
|
| 705 |
+
703 1
|
| 706 |
+
704 1
|
| 707 |
+
705 0
|
| 708 |
+
706 0
|
| 709 |
+
707 0
|
| 710 |
+
708 0
|
| 711 |
+
709 0
|
| 712 |
+
710 1
|
| 713 |
+
711 0
|
| 714 |
+
712 1
|
| 715 |
+
713 1
|
| 716 |
+
714 0
|
| 717 |
+
715 1
|
| 718 |
+
716 1
|
| 719 |
+
717 1
|
| 720 |
+
718 1
|
| 721 |
+
719 1
|
| 722 |
+
720 0
|
| 723 |
+
721 1
|
| 724 |
+
722 1
|
| 725 |
+
723 1
|
| 726 |
+
724 1
|
| 727 |
+
725 1
|
| 728 |
+
726 1
|
| 729 |
+
727 1
|
| 730 |
+
728 0
|
| 731 |
+
729 1
|
| 732 |
+
730 1
|
| 733 |
+
731 1
|
| 734 |
+
732 0
|
| 735 |
+
733 1
|
| 736 |
+
734 1
|
| 737 |
+
735 0
|
| 738 |
+
736 1
|
| 739 |
+
737 1
|
| 740 |
+
738 1
|
| 741 |
+
739 0
|
| 742 |
+
740 0
|
| 743 |
+
741 1
|
| 744 |
+
742 1
|
| 745 |
+
743 1
|
| 746 |
+
744 1
|
| 747 |
+
745 1
|
| 748 |
+
746 1
|
| 749 |
+
747 1
|
| 750 |
+
748 1
|
| 751 |
+
749 1
|
| 752 |
+
750 1
|
| 753 |
+
751 1
|
| 754 |
+
752 1
|
| 755 |
+
753 1
|
| 756 |
+
754 1
|
| 757 |
+
755 0
|
| 758 |
+
756 1
|
| 759 |
+
757 1
|
| 760 |
+
758 1
|
| 761 |
+
759 1
|
| 762 |
+
760 1
|
| 763 |
+
761 1
|
| 764 |
+
762 1
|
| 765 |
+
763 1
|
| 766 |
+
764 1
|
| 767 |
+
765 0
|
| 768 |
+
766 0
|
| 769 |
+
767 1
|
| 770 |
+
768 1
|
| 771 |
+
769 1
|
| 772 |
+
770 1
|
| 773 |
+
771 1
|
| 774 |
+
772 1
|
| 775 |
+
773 1
|
| 776 |
+
774 1
|
| 777 |
+
775 1
|
| 778 |
+
776 1
|
| 779 |
+
777 1
|
| 780 |
+
778 1
|
| 781 |
+
779 1
|
| 782 |
+
780 1
|
| 783 |
+
781 1
|
| 784 |
+
782 1
|
| 785 |
+
783 1
|
| 786 |
+
784 1
|
| 787 |
+
785 1
|
| 788 |
+
786 0
|
| 789 |
+
787 1
|
| 790 |
+
788 1
|
| 791 |
+
789 1
|
| 792 |
+
790 1
|
| 793 |
+
791 1
|
| 794 |
+
792 1
|
| 795 |
+
793 1
|
| 796 |
+
794 0
|
| 797 |
+
795 0
|
| 798 |
+
796 0
|
| 799 |
+
797 1
|
| 800 |
+
798 1
|
| 801 |
+
799 1
|
| 802 |
+
800 1
|
| 803 |
+
801 1
|
| 804 |
+
802 1
|
| 805 |
+
803 0
|
| 806 |
+
804 1
|
| 807 |
+
805 1
|
| 808 |
+
806 1
|
| 809 |
+
807 1
|
| 810 |
+
808 1
|
| 811 |
+
809 0
|
| 812 |
+
810 1
|
| 813 |
+
811 1
|
| 814 |
+
812 0
|
| 815 |
+
813 0
|
| 816 |
+
814 0
|
| 817 |
+
815 0
|
| 818 |
+
816 1
|
| 819 |
+
817 0
|
| 820 |
+
818 0
|
| 821 |
+
819 1
|
| 822 |
+
820 1
|
| 823 |
+
821 1
|
| 824 |
+
822 1
|
| 825 |
+
823 1
|
| 826 |
+
824 1
|
| 827 |
+
825 1
|
| 828 |
+
826 1
|
| 829 |
+
827 0
|
| 830 |
+
828 0
|
| 831 |
+
829 1
|
| 832 |
+
830 1
|
| 833 |
+
831 1
|
| 834 |
+
832 1
|
| 835 |
+
833 1
|
| 836 |
+
834 0
|
| 837 |
+
835 1
|
| 838 |
+
836 0
|
| 839 |
+
837 1
|
| 840 |
+
838 1
|
| 841 |
+
839 1
|
| 842 |
+
840 1
|
| 843 |
+
841 1
|
| 844 |
+
842 1
|
| 845 |
+
843 1
|
| 846 |
+
844 1
|
| 847 |
+
845 1
|
| 848 |
+
846 1
|
| 849 |
+
847 1
|
| 850 |
+
848 1
|
| 851 |
+
849 1
|
| 852 |
+
850 1
|
| 853 |
+
851 1
|
| 854 |
+
852 1
|
| 855 |
+
853 0
|
| 856 |
+
854 1
|
| 857 |
+
855 1
|
| 858 |
+
856 1
|
| 859 |
+
857 1
|
| 860 |
+
858 1
|
| 861 |
+
859 1
|
| 862 |
+
860 1
|
| 863 |
+
861 1
|
| 864 |
+
862 0
|
| 865 |
+
863 1
|
| 866 |
+
864 0
|
| 867 |
+
865 1
|
| 868 |
+
866 1
|
| 869 |
+
867 1
|
| 870 |
+
868 1
|
| 871 |
+
869 1
|
| 872 |
+
870 1
|
| 873 |
+
871 1
|
| 874 |
+
872 0
|
| 875 |
+
873 1
|
| 876 |
+
874 1
|
| 877 |
+
875 1
|
| 878 |
+
876 1
|
| 879 |
+
877 1
|
| 880 |
+
878 1
|
| 881 |
+
879 1
|
| 882 |
+
880 1
|
| 883 |
+
881 1
|
| 884 |
+
882 1
|
| 885 |
+
883 1
|
| 886 |
+
884 1
|
| 887 |
+
885 1
|
| 888 |
+
886 0
|
| 889 |
+
887 1
|
| 890 |
+
888 1
|
| 891 |
+
889 1
|
| 892 |
+
890 1
|
| 893 |
+
891 1
|
| 894 |
+
892 1
|
| 895 |
+
893 1
|
| 896 |
+
894 1
|
| 897 |
+
895 1
|
| 898 |
+
896 1
|
| 899 |
+
897 0
|
| 900 |
+
898 0
|
| 901 |
+
899 1
|
| 902 |
+
900 1
|
| 903 |
+
901 1
|
| 904 |
+
902 1
|
| 905 |
+
903 1
|
| 906 |
+
904 1
|
| 907 |
+
905 1
|
| 908 |
+
906 1
|
| 909 |
+
907 1
|
| 910 |
+
908 1
|
| 911 |
+
909 1
|
| 912 |
+
910 0
|
| 913 |
+
911 1
|
| 914 |
+
912 1
|
| 915 |
+
913 1
|
| 916 |
+
914 0
|
| 917 |
+
915 0
|
| 918 |
+
916 0
|
| 919 |
+
917 1
|
| 920 |
+
918 1
|
| 921 |
+
919 1
|
| 922 |
+
920 1
|
| 923 |
+
921 1
|
| 924 |
+
922 1
|
| 925 |
+
923 1
|
| 926 |
+
924 1
|
| 927 |
+
925 1
|
| 928 |
+
926 1
|
| 929 |
+
927 1
|
| 930 |
+
928 1
|
| 931 |
+
929 1
|
| 932 |
+
930 1
|
| 933 |
+
931 1
|
| 934 |
+
932 1
|
| 935 |
+
933 1
|
| 936 |
+
934 1
|
| 937 |
+
935 1
|
| 938 |
+
936 1
|
| 939 |
+
937 1
|
| 940 |
+
938 1
|
| 941 |
+
939 1
|
| 942 |
+
940 1
|
| 943 |
+
941 1
|
| 944 |
+
942 1
|
| 945 |
+
943 1
|
| 946 |
+
944 1
|
| 947 |
+
945 1
|
| 948 |
+
946 1
|
| 949 |
+
947 1
|
| 950 |
+
948 0
|
| 951 |
+
949 1
|
| 952 |
+
950 0
|
| 953 |
+
951 0
|
| 954 |
+
952 1
|
| 955 |
+
953 1
|
| 956 |
+
954 1
|
| 957 |
+
955 0
|
| 958 |
+
956 0
|
| 959 |
+
957 1
|
| 960 |
+
958 1
|
| 961 |
+
959 1
|
| 962 |
+
960 1
|
| 963 |
+
961 0
|
| 964 |
+
962 1
|
| 965 |
+
963 1
|
| 966 |
+
964 1
|
| 967 |
+
965 1
|
| 968 |
+
966 1
|
| 969 |
+
967 1
|
| 970 |
+
968 1
|
| 971 |
+
969 1
|
| 972 |
+
970 1
|
| 973 |
+
971 0
|
| 974 |
+
972 1
|
| 975 |
+
973 1
|
| 976 |
+
974 1
|
| 977 |
+
975 1
|
| 978 |
+
976 1
|
| 979 |
+
977 1
|
| 980 |
+
978 0
|
| 981 |
+
979 1
|
| 982 |
+
980 1
|
| 983 |
+
981 1
|
| 984 |
+
982 1
|
| 985 |
+
983 1
|
| 986 |
+
984 0
|
| 987 |
+
985 1
|
| 988 |
+
986 1
|
| 989 |
+
987 1
|
| 990 |
+
988 1
|
| 991 |
+
989 1
|
| 992 |
+
990 1
|
| 993 |
+
991 1
|
| 994 |
+
992 1
|
| 995 |
+
993 0
|
| 996 |
+
994 1
|
| 997 |
+
995 0
|
| 998 |
+
996 0
|
| 999 |
+
997 1
|
| 1000 |
+
998 1
|
| 1001 |
+
999 1
|
| 1002 |
+
1000 1
|
| 1003 |
+
1001 0
|
| 1004 |
+
1002 1
|
| 1005 |
+
1003 1
|
| 1006 |
+
1004 1
|
| 1007 |
+
1005 0
|
| 1008 |
+
1006 1
|
| 1009 |
+
1007 1
|
| 1010 |
+
1008 1
|
| 1011 |
+
1009 1
|
| 1012 |
+
1010 1
|
| 1013 |
+
1011 0
|
| 1014 |
+
1012 0
|
| 1015 |
+
1013 0
|
| 1016 |
+
1014 1
|
| 1017 |
+
1015 1
|
| 1018 |
+
1016 1
|
| 1019 |
+
1017 1
|
| 1020 |
+
1018 0
|
finetune/cola/pytorch_model.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8e347d72000b0f5a8771aa62eedeb5a78b9aa79669b5796edb1e66f8e21e832c
|
| 3 |
+
size 442624943
|
finetune/cola/special_tokens_map.json
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bos_token": "<s>",
|
| 3 |
+
"cls_token": "<s>",
|
| 4 |
+
"eos_token": "</s>",
|
| 5 |
+
"mask_token": {
|
| 6 |
+
"content": "<mask>",
|
| 7 |
+
"lstrip": true,
|
| 8 |
+
"normalized": false,
|
| 9 |
+
"rstrip": false,
|
| 10 |
+
"single_word": false
|
| 11 |
+
},
|
| 12 |
+
"pad_token": "<pad>",
|
| 13 |
+
"sep_token": "</s>",
|
| 14 |
+
"unk_token": "<unk>"
|
| 15 |
+
}
|
finetune/cola/structformer_as_hf_no_parser.py
ADDED
|
@@ -0,0 +1,754 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import torch.nn as nn
|
| 3 |
+
import torch.nn.functional as F
|
| 4 |
+
from torch.nn import init
|
| 5 |
+
from transformers import PreTrainedModel
|
| 6 |
+
from transformers import PretrainedConfig
|
| 7 |
+
from transformers.modeling_outputs import MaskedLMOutput
|
| 8 |
+
from typing import List
|
| 9 |
+
from torch.nn import CrossEntropyLoss, MSELoss, BCEWithLogitsLoss
|
| 10 |
+
from transformers.modeling_outputs import (
|
| 11 |
+
BaseModelOutputWithPastAndCrossAttentions,
|
| 12 |
+
BaseModelOutputWithPoolingAndCrossAttentions,
|
| 13 |
+
MaskedLMOutput,
|
| 14 |
+
SequenceClassifierOutput
|
| 15 |
+
)
|
| 16 |
+
|
| 17 |
+
##########################################
|
| 18 |
+
# HuggingFace Config
|
| 19 |
+
##########################################
|
| 20 |
+
class StructformerConfig(PretrainedConfig):
|
| 21 |
+
model_type = "structformer"
|
| 22 |
+
|
| 23 |
+
def __init__(
|
| 24 |
+
self,
|
| 25 |
+
hidden_size=768,
|
| 26 |
+
n_context_layers=2,
|
| 27 |
+
nlayers=6,
|
| 28 |
+
ntokens=32000,
|
| 29 |
+
nhead=8,
|
| 30 |
+
dropout=0.1,
|
| 31 |
+
dropatt=0.1,
|
| 32 |
+
relative_bias=False,
|
| 33 |
+
pos_emb=False,
|
| 34 |
+
pad=0,
|
| 35 |
+
n_parser_layers=4,
|
| 36 |
+
conv_size=9,
|
| 37 |
+
relations=('head', 'child'),
|
| 38 |
+
weight_act='softmax',
|
| 39 |
+
**kwargs,
|
| 40 |
+
):
|
| 41 |
+
self.hidden_size = hidden_size
|
| 42 |
+
self.n_context_layers = n_context_layers
|
| 43 |
+
self.nlayers = nlayers
|
| 44 |
+
self.ntokens = ntokens
|
| 45 |
+
self.nhead = nhead
|
| 46 |
+
self.dropout = dropout
|
| 47 |
+
self.dropatt = dropatt
|
| 48 |
+
self.relative_bias = relative_bias
|
| 49 |
+
self.pos_emb = pos_emb
|
| 50 |
+
self.pad = pad
|
| 51 |
+
self.n_parser_layers = n_parser_layers
|
| 52 |
+
self.conv_size = conv_size
|
| 53 |
+
self.relations = relations
|
| 54 |
+
self.weight_act = weight_act
|
| 55 |
+
super().__init__(**kwargs)
|
| 56 |
+
|
| 57 |
+
##########################################
|
| 58 |
+
# Custom Layers
|
| 59 |
+
##########################################
|
| 60 |
+
def _get_activation_fn(activation):
|
| 61 |
+
"""Get specified activation function."""
|
| 62 |
+
if activation == "relu":
|
| 63 |
+
return nn.ReLU()
|
| 64 |
+
elif activation == "gelu":
|
| 65 |
+
return nn.GELU()
|
| 66 |
+
elif activation == "leakyrelu":
|
| 67 |
+
return nn.LeakyReLU()
|
| 68 |
+
|
| 69 |
+
raise RuntimeError(
|
| 70 |
+
"activation should be relu/gelu, not {}".format(activation))
|
| 71 |
+
|
| 72 |
+
class Conv1d(nn.Module):
|
| 73 |
+
"""1D convolution layer."""
|
| 74 |
+
|
| 75 |
+
def __init__(self, hidden_size, kernel_size, dilation=1):
|
| 76 |
+
"""Initialization.
|
| 77 |
+
|
| 78 |
+
Args:
|
| 79 |
+
hidden_size: dimension of input embeddings
|
| 80 |
+
kernel_size: convolution kernel size
|
| 81 |
+
dilation: the spacing between the kernel points
|
| 82 |
+
"""
|
| 83 |
+
super(Conv1d, self).__init__()
|
| 84 |
+
|
| 85 |
+
if kernel_size % 2 == 0:
|
| 86 |
+
padding = (kernel_size // 2) * dilation
|
| 87 |
+
self.shift = True
|
| 88 |
+
else:
|
| 89 |
+
padding = ((kernel_size - 1) // 2) * dilation
|
| 90 |
+
self.shift = False
|
| 91 |
+
self.conv = nn.Conv1d(
|
| 92 |
+
hidden_size,
|
| 93 |
+
hidden_size,
|
| 94 |
+
kernel_size,
|
| 95 |
+
padding=padding,
|
| 96 |
+
dilation=dilation)
|
| 97 |
+
|
| 98 |
+
def forward(self, x):
|
| 99 |
+
"""Compute convolution.
|
| 100 |
+
|
| 101 |
+
Args:
|
| 102 |
+
x: input embeddings
|
| 103 |
+
Returns:
|
| 104 |
+
conv_output: convolution results
|
| 105 |
+
"""
|
| 106 |
+
|
| 107 |
+
if self.shift:
|
| 108 |
+
return self.conv(x.transpose(1, 2)).transpose(1, 2)[:, 1:]
|
| 109 |
+
else:
|
| 110 |
+
return self.conv(x.transpose(1, 2)).transpose(1, 2)
|
| 111 |
+
|
| 112 |
+
class MultiheadAttention(nn.Module):
|
| 113 |
+
"""Multi-head self-attention layer."""
|
| 114 |
+
|
| 115 |
+
def __init__(self,
|
| 116 |
+
embed_dim,
|
| 117 |
+
num_heads,
|
| 118 |
+
dropout=0.,
|
| 119 |
+
bias=True,
|
| 120 |
+
v_proj=True,
|
| 121 |
+
out_proj=True,
|
| 122 |
+
relative_bias=True):
|
| 123 |
+
"""Initialization.
|
| 124 |
+
|
| 125 |
+
Args:
|
| 126 |
+
embed_dim: dimension of input embeddings
|
| 127 |
+
num_heads: number of self-attention heads
|
| 128 |
+
dropout: dropout rate
|
| 129 |
+
bias: bool, indicate whether include bias for linear transformations
|
| 130 |
+
v_proj: bool, indicate whether project inputs to new values
|
| 131 |
+
out_proj: bool, indicate whether project outputs to new values
|
| 132 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 133 |
+
attention bias
|
| 134 |
+
"""
|
| 135 |
+
|
| 136 |
+
super(MultiheadAttention, self).__init__()
|
| 137 |
+
self.embed_dim = embed_dim
|
| 138 |
+
|
| 139 |
+
self.num_heads = num_heads
|
| 140 |
+
self.drop = nn.Dropout(dropout)
|
| 141 |
+
self.head_dim = embed_dim // num_heads
|
| 142 |
+
assert self.head_dim * num_heads == self.embed_dim, ("embed_dim must be "
|
| 143 |
+
"divisible by "
|
| 144 |
+
"num_heads")
|
| 145 |
+
|
| 146 |
+
self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 147 |
+
self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 148 |
+
if v_proj:
|
| 149 |
+
self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 150 |
+
else:
|
| 151 |
+
self.v_proj = nn.Identity()
|
| 152 |
+
|
| 153 |
+
if out_proj:
|
| 154 |
+
self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 155 |
+
else:
|
| 156 |
+
self.out_proj = nn.Identity()
|
| 157 |
+
|
| 158 |
+
if relative_bias:
|
| 159 |
+
self.relative_bias = nn.Parameter(torch.zeros((self.num_heads, 512)))
|
| 160 |
+
else:
|
| 161 |
+
self.relative_bias = None
|
| 162 |
+
|
| 163 |
+
self._reset_parameters()
|
| 164 |
+
|
| 165 |
+
def _reset_parameters(self):
|
| 166 |
+
"""Initialize attention parameters."""
|
| 167 |
+
|
| 168 |
+
init.xavier_uniform_(self.q_proj.weight)
|
| 169 |
+
init.constant_(self.q_proj.bias, 0.)
|
| 170 |
+
|
| 171 |
+
init.xavier_uniform_(self.k_proj.weight)
|
| 172 |
+
init.constant_(self.k_proj.bias, 0.)
|
| 173 |
+
|
| 174 |
+
if isinstance(self.v_proj, nn.Linear):
|
| 175 |
+
init.xavier_uniform_(self.v_proj.weight)
|
| 176 |
+
init.constant_(self.v_proj.bias, 0.)
|
| 177 |
+
|
| 178 |
+
if isinstance(self.out_proj, nn.Linear):
|
| 179 |
+
init.xavier_uniform_(self.out_proj.weight)
|
| 180 |
+
init.constant_(self.out_proj.bias, 0.)
|
| 181 |
+
|
| 182 |
+
def forward(self, query, key_padding_mask=None, attn_mask=None):
|
| 183 |
+
"""Compute multi-head self-attention.
|
| 184 |
+
|
| 185 |
+
Args:
|
| 186 |
+
query: input embeddings
|
| 187 |
+
key_padding_mask: 3D mask that prevents attention to certain positions
|
| 188 |
+
attn_mask: 3D mask that rescale the attention weight at each position
|
| 189 |
+
Returns:
|
| 190 |
+
attn_output: self-attention output
|
| 191 |
+
"""
|
| 192 |
+
|
| 193 |
+
length, bsz, embed_dim = query.size()
|
| 194 |
+
assert embed_dim == self.embed_dim
|
| 195 |
+
|
| 196 |
+
head_dim = embed_dim // self.num_heads
|
| 197 |
+
assert head_dim * self.num_heads == embed_dim, ("embed_dim must be "
|
| 198 |
+
"divisible by num_heads")
|
| 199 |
+
scaling = float(head_dim)**-0.5
|
| 200 |
+
|
| 201 |
+
q = self.q_proj(query)
|
| 202 |
+
k = self.k_proj(query)
|
| 203 |
+
v = self.v_proj(query)
|
| 204 |
+
|
| 205 |
+
q = q * scaling
|
| 206 |
+
|
| 207 |
+
if attn_mask is not None:
|
| 208 |
+
assert list(attn_mask.size()) == [bsz * self.num_heads,
|
| 209 |
+
query.size(0), query.size(0)]
|
| 210 |
+
|
| 211 |
+
q = q.contiguous().view(length, bsz * self.num_heads,
|
| 212 |
+
head_dim).transpose(0, 1)
|
| 213 |
+
k = k.contiguous().view(length, bsz * self.num_heads,
|
| 214 |
+
head_dim).transpose(0, 1)
|
| 215 |
+
v = v.contiguous().view(length, bsz * self.num_heads,
|
| 216 |
+
head_dim).transpose(0, 1)
|
| 217 |
+
|
| 218 |
+
attn_output_weights = torch.bmm(q, k.transpose(1, 2))
|
| 219 |
+
assert list(
|
| 220 |
+
attn_output_weights.size()) == [bsz * self.num_heads, length, length]
|
| 221 |
+
|
| 222 |
+
if self.relative_bias is not None:
|
| 223 |
+
pos = torch.arange(length, device=query.device)
|
| 224 |
+
relative_pos = torch.abs(pos[:, None] - pos[None, :]) + 256
|
| 225 |
+
relative_pos = relative_pos[None, :, :].expand(bsz * self.num_heads, -1,
|
| 226 |
+
-1)
|
| 227 |
+
|
| 228 |
+
relative_bias = self.relative_bias.repeat_interleave(bsz, dim=0)
|
| 229 |
+
relative_bias = relative_bias[:, None, :].expand(-1, length, -1)
|
| 230 |
+
relative_bias = torch.gather(relative_bias, 2, relative_pos)
|
| 231 |
+
attn_output_weights = attn_output_weights + relative_bias
|
| 232 |
+
|
| 233 |
+
if key_padding_mask is not None:
|
| 234 |
+
attn_output_weights = attn_output_weights + key_padding_mask
|
| 235 |
+
|
| 236 |
+
if attn_mask is None:
|
| 237 |
+
attn_output_weights = torch.softmax(attn_output_weights, dim=-1)
|
| 238 |
+
else:
|
| 239 |
+
attn_output_weights = torch.sigmoid(attn_output_weights) * attn_mask
|
| 240 |
+
|
| 241 |
+
attn_output_weights = self.drop(attn_output_weights)
|
| 242 |
+
|
| 243 |
+
attn_output = torch.bmm(attn_output_weights, v)
|
| 244 |
+
|
| 245 |
+
assert list(attn_output.size()) == [bsz * self.num_heads, length, head_dim]
|
| 246 |
+
attn_output = attn_output.transpose(0, 1).contiguous().view(
|
| 247 |
+
length, bsz, embed_dim)
|
| 248 |
+
attn_output = self.out_proj(attn_output)
|
| 249 |
+
|
| 250 |
+
return attn_output
|
| 251 |
+
|
| 252 |
+
class TransformerLayer(nn.Module):
|
| 253 |
+
"""TransformerEncoderLayer is made up of self-attn and feedforward network."""
|
| 254 |
+
|
| 255 |
+
def __init__(self,
|
| 256 |
+
d_model,
|
| 257 |
+
nhead,
|
| 258 |
+
dim_feedforward=2048,
|
| 259 |
+
dropout=0.1,
|
| 260 |
+
dropatt=0.1,
|
| 261 |
+
activation="leakyrelu",
|
| 262 |
+
relative_bias=True):
|
| 263 |
+
"""Initialization.
|
| 264 |
+
|
| 265 |
+
Args:
|
| 266 |
+
d_model: dimension of inputs
|
| 267 |
+
nhead: number of self-attention heads
|
| 268 |
+
dim_feedforward: dimension of hidden layer in feedforward layer
|
| 269 |
+
dropout: dropout rate
|
| 270 |
+
dropatt: drop attention rate
|
| 271 |
+
activation: activation function
|
| 272 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 273 |
+
attention bias
|
| 274 |
+
"""
|
| 275 |
+
|
| 276 |
+
super(TransformerLayer, self).__init__()
|
| 277 |
+
|
| 278 |
+
self.self_attn = MultiheadAttention(
|
| 279 |
+
d_model, nhead, dropout=dropatt, relative_bias=relative_bias)
|
| 280 |
+
|
| 281 |
+
# Implementation of Feedforward model
|
| 282 |
+
self.feedforward = nn.Sequential(
|
| 283 |
+
nn.LayerNorm(d_model), nn.Linear(d_model, dim_feedforward),
|
| 284 |
+
_get_activation_fn(activation), nn.Dropout(dropout),
|
| 285 |
+
nn.Linear(dim_feedforward, d_model))
|
| 286 |
+
|
| 287 |
+
self.norm = nn.LayerNorm(d_model)
|
| 288 |
+
self.dropout1 = nn.Dropout(dropout)
|
| 289 |
+
self.dropout2 = nn.Dropout(dropout)
|
| 290 |
+
|
| 291 |
+
self.nhead = nhead
|
| 292 |
+
|
| 293 |
+
def forward(self, src, attn_mask=None, key_padding_mask=None):
|
| 294 |
+
"""Pass the input through the encoder layer.
|
| 295 |
+
|
| 296 |
+
Args:
|
| 297 |
+
src: the sequence to the encoder layer (required).
|
| 298 |
+
attn_mask: the mask for the src sequence (optional).
|
| 299 |
+
key_padding_mask: the mask for the src keys per batch (optional).
|
| 300 |
+
Returns:
|
| 301 |
+
src3: the output of transformer layer, share the same shape as src.
|
| 302 |
+
"""
|
| 303 |
+
src2 = self.self_attn(
|
| 304 |
+
self.norm(src), attn_mask=attn_mask, key_padding_mask=key_padding_mask)
|
| 305 |
+
src2 = src + self.dropout1(src2)
|
| 306 |
+
src3 = self.feedforward(src2)
|
| 307 |
+
src3 = src2 + self.dropout2(src3)
|
| 308 |
+
|
| 309 |
+
return src3
|
| 310 |
+
|
| 311 |
+
|
| 312 |
+
|
| 313 |
+
class RobertaClassificationHead(nn.Module):
|
| 314 |
+
"""Head for sentence-level classification tasks."""
|
| 315 |
+
|
| 316 |
+
def __init__(self, config):
|
| 317 |
+
super().__init__()
|
| 318 |
+
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
|
| 319 |
+
classifier_dropout = (
|
| 320 |
+
config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
|
| 321 |
+
)
|
| 322 |
+
self.dropout = nn.Dropout(classifier_dropout)
|
| 323 |
+
self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
|
| 324 |
+
|
| 325 |
+
def forward(self, features, **kwargs):
|
| 326 |
+
x = features[:, 0, :] # take <s> token (equiv. to [CLS])
|
| 327 |
+
x = self.dropout(x)
|
| 328 |
+
x = self.dense(x)
|
| 329 |
+
x = torch.tanh(x)
|
| 330 |
+
x = self.dropout(x)
|
| 331 |
+
x = self.out_proj(x)
|
| 332 |
+
return x
|
| 333 |
+
|
| 334 |
+
|
| 335 |
+
##########################################
|
| 336 |
+
# Custom Models
|
| 337 |
+
##########################################
|
| 338 |
+
def cumprod(x, reverse=False, exclusive=False):
|
| 339 |
+
"""cumulative product."""
|
| 340 |
+
if reverse:
|
| 341 |
+
x = x.flip([-1])
|
| 342 |
+
|
| 343 |
+
if exclusive:
|
| 344 |
+
x = F.pad(x[:, :, :-1], (1, 0), value=1)
|
| 345 |
+
|
| 346 |
+
cx = x.cumprod(-1)
|
| 347 |
+
|
| 348 |
+
if reverse:
|
| 349 |
+
cx = cx.flip([-1])
|
| 350 |
+
return cx
|
| 351 |
+
|
| 352 |
+
def cumsum(x, reverse=False, exclusive=False):
|
| 353 |
+
"""cumulative sum."""
|
| 354 |
+
bsz, _, length = x.size()
|
| 355 |
+
device = x.device
|
| 356 |
+
if reverse:
|
| 357 |
+
if exclusive:
|
| 358 |
+
w = torch.ones([bsz, length, length], device=device).tril(-1)
|
| 359 |
+
else:
|
| 360 |
+
w = torch.ones([bsz, length, length], device=device).tril(0)
|
| 361 |
+
cx = torch.bmm(x, w)
|
| 362 |
+
else:
|
| 363 |
+
if exclusive:
|
| 364 |
+
w = torch.ones([bsz, length, length], device=device).triu(1)
|
| 365 |
+
else:
|
| 366 |
+
w = torch.ones([bsz, length, length], device=device).triu(0)
|
| 367 |
+
cx = torch.bmm(x, w)
|
| 368 |
+
return cx
|
| 369 |
+
|
| 370 |
+
def cummin(x, reverse=False, exclusive=False, max_value=1e9):
|
| 371 |
+
"""cumulative min."""
|
| 372 |
+
if reverse:
|
| 373 |
+
if exclusive:
|
| 374 |
+
x = F.pad(x[:, :, 1:], (0, 1), value=max_value)
|
| 375 |
+
x = x.flip([-1]).cummin(-1)[0].flip([-1])
|
| 376 |
+
else:
|
| 377 |
+
if exclusive:
|
| 378 |
+
x = F.pad(x[:, :, :-1], (1, 0), value=max_value)
|
| 379 |
+
x = x.cummin(-1)[0]
|
| 380 |
+
return x
|
| 381 |
+
|
| 382 |
+
class Transformer(nn.Module):
|
| 383 |
+
"""Transformer model."""
|
| 384 |
+
|
| 385 |
+
def __init__(self,
|
| 386 |
+
hidden_size,
|
| 387 |
+
nlayers,
|
| 388 |
+
ntokens,
|
| 389 |
+
nhead=8,
|
| 390 |
+
dropout=0.1,
|
| 391 |
+
dropatt=0.1,
|
| 392 |
+
relative_bias=True,
|
| 393 |
+
pos_emb=False,
|
| 394 |
+
pad=0):
|
| 395 |
+
"""Initialization.
|
| 396 |
+
|
| 397 |
+
Args:
|
| 398 |
+
hidden_size: dimension of inputs and hidden states
|
| 399 |
+
nlayers: number of layers
|
| 400 |
+
ntokens: number of output categories
|
| 401 |
+
nhead: number of self-attention heads
|
| 402 |
+
dropout: dropout rate
|
| 403 |
+
dropatt: drop attention rate
|
| 404 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 405 |
+
attention bias
|
| 406 |
+
pos_emb: bool, indicate whether use a learnable positional embedding
|
| 407 |
+
pad: pad token index
|
| 408 |
+
"""
|
| 409 |
+
|
| 410 |
+
super(Transformer, self).__init__()
|
| 411 |
+
|
| 412 |
+
self.drop = nn.Dropout(dropout)
|
| 413 |
+
|
| 414 |
+
self.emb = nn.Embedding(ntokens, hidden_size)
|
| 415 |
+
if pos_emb:
|
| 416 |
+
self.pos_emb = nn.Embedding(500, hidden_size)
|
| 417 |
+
|
| 418 |
+
self.layers = nn.ModuleList([
|
| 419 |
+
TransformerLayer(hidden_size, nhead, hidden_size * 4, dropout,
|
| 420 |
+
dropatt=dropatt, relative_bias=relative_bias)
|
| 421 |
+
for _ in range(nlayers)])
|
| 422 |
+
|
| 423 |
+
self.norm = nn.LayerNorm(hidden_size)
|
| 424 |
+
|
| 425 |
+
self.output_layer = nn.Linear(hidden_size, ntokens)
|
| 426 |
+
self.output_layer.weight = self.emb.weight
|
| 427 |
+
|
| 428 |
+
self.init_weights()
|
| 429 |
+
|
| 430 |
+
self.nlayers = nlayers
|
| 431 |
+
self.nhead = nhead
|
| 432 |
+
self.ntokens = ntokens
|
| 433 |
+
self.hidden_size = hidden_size
|
| 434 |
+
self.pad = pad
|
| 435 |
+
|
| 436 |
+
def init_weights(self):
|
| 437 |
+
"""Initialize token embedding and output bias."""
|
| 438 |
+
initrange = 0.1
|
| 439 |
+
self.emb.weight.data.uniform_(-initrange, initrange)
|
| 440 |
+
if hasattr(self, 'pos_emb'):
|
| 441 |
+
self.pos_emb.weight.data.uniform_(-initrange, initrange)
|
| 442 |
+
self.output_layer.bias.data.fill_(0)
|
| 443 |
+
|
| 444 |
+
def visibility(self, x, device):
|
| 445 |
+
"""Mask pad tokens."""
|
| 446 |
+
visibility = (x != self.pad).float()
|
| 447 |
+
visibility = visibility[:, None, :].expand(-1, x.size(1), -1)
|
| 448 |
+
visibility = torch.repeat_interleave(visibility, self.nhead, dim=0)
|
| 449 |
+
return visibility.log()
|
| 450 |
+
|
| 451 |
+
def encode(self, x, pos):
|
| 452 |
+
"""Standard transformer encode process."""
|
| 453 |
+
h = self.emb(x)
|
| 454 |
+
if hasattr(self, 'pos_emb'):
|
| 455 |
+
h = h + self.pos_emb(pos)
|
| 456 |
+
h_list = []
|
| 457 |
+
visibility = self.visibility(x, x.device)
|
| 458 |
+
|
| 459 |
+
for i in range(self.nlayers):
|
| 460 |
+
h_list.append(h)
|
| 461 |
+
h = self.layers[i](
|
| 462 |
+
h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
|
| 463 |
+
|
| 464 |
+
output = h
|
| 465 |
+
h_array = torch.stack(h_list, dim=2)
|
| 466 |
+
|
| 467 |
+
return output, h_array
|
| 468 |
+
|
| 469 |
+
def forward(self, x, pos):
|
| 470 |
+
"""Pass the input through the encoder layer.
|
| 471 |
+
|
| 472 |
+
Args:
|
| 473 |
+
x: input tokens (required).
|
| 474 |
+
pos: position for each token (optional).
|
| 475 |
+
Returns:
|
| 476 |
+
output: probability distributions for missing tokens.
|
| 477 |
+
state_dict: parsing results and raw output
|
| 478 |
+
"""
|
| 479 |
+
|
| 480 |
+
batch_size, length = x.size()
|
| 481 |
+
|
| 482 |
+
raw_output, _ = self.encode(x, pos)
|
| 483 |
+
raw_output = self.norm(raw_output)
|
| 484 |
+
raw_output = self.drop(raw_output)
|
| 485 |
+
|
| 486 |
+
output = self.output_layer(raw_output)
|
| 487 |
+
return output.view(batch_size * length, -1), {'raw_output': raw_output,}
|
| 488 |
+
|
| 489 |
+
class StructFormer(Transformer):
|
| 490 |
+
"""StructFormer model."""
|
| 491 |
+
|
| 492 |
+
def __init__(self,
|
| 493 |
+
hidden_size,
|
| 494 |
+
n_context_layers,
|
| 495 |
+
nlayers,
|
| 496 |
+
ntokens,
|
| 497 |
+
nhead=8,
|
| 498 |
+
dropout=0.1,
|
| 499 |
+
dropatt=0.1,
|
| 500 |
+
relative_bias=False,
|
| 501 |
+
pos_emb=False,
|
| 502 |
+
pad=0,
|
| 503 |
+
n_parser_layers=4,
|
| 504 |
+
conv_size=9,
|
| 505 |
+
relations=('head', 'child'),
|
| 506 |
+
weight_act='softmax'):
|
| 507 |
+
"""Initialization.
|
| 508 |
+
|
| 509 |
+
Args:
|
| 510 |
+
hidden_size: dimension of inputs and hidden states
|
| 511 |
+
nlayers: number of layers
|
| 512 |
+
ntokens: number of output categories
|
| 513 |
+
nhead: number of self-attention heads
|
| 514 |
+
dropout: dropout rate
|
| 515 |
+
dropatt: drop attention rate
|
| 516 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 517 |
+
attention bias
|
| 518 |
+
pos_emb: bool, indicate whether use a learnable positional embedding
|
| 519 |
+
pad: pad token index
|
| 520 |
+
n_parser_layers: number of parsing layers
|
| 521 |
+
conv_size: convolution kernel size for parser
|
| 522 |
+
relations: relations that are used to compute self attention
|
| 523 |
+
weight_act: relations distribution activation function
|
| 524 |
+
"""
|
| 525 |
+
|
| 526 |
+
super(StructFormer, self).__init__(
|
| 527 |
+
hidden_size,
|
| 528 |
+
nlayers,
|
| 529 |
+
ntokens,
|
| 530 |
+
nhead=nhead,
|
| 531 |
+
dropout=dropout,
|
| 532 |
+
dropatt=dropatt,
|
| 533 |
+
relative_bias=relative_bias,
|
| 534 |
+
pos_emb=pos_emb,
|
| 535 |
+
pad=pad)
|
| 536 |
+
|
| 537 |
+
|
| 538 |
+
def encode(self, x, pos):
|
| 539 |
+
h = self.emb(x)
|
| 540 |
+
if hasattr(self, 'pos_emb'):
|
| 541 |
+
h = h + self.pos_emb(pos)
|
| 542 |
+
h_list = []
|
| 543 |
+
visibility = self.visibility(x, x.device)
|
| 544 |
+
|
| 545 |
+
for i in range(self.nlayers):
|
| 546 |
+
h_list.append(h)
|
| 547 |
+
h = self.layers[i](
|
| 548 |
+
h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
|
| 549 |
+
|
| 550 |
+
output = h
|
| 551 |
+
h_array = torch.stack(h_list, dim=2)
|
| 552 |
+
|
| 553 |
+
return output
|
| 554 |
+
|
| 555 |
+
|
| 556 |
+
def forward(self, input_ids, labels=None, position_ids=None, **kwargs):
|
| 557 |
+
|
| 558 |
+
x = input_ids
|
| 559 |
+
batch_size, length = x.size()
|
| 560 |
+
|
| 561 |
+
if position_ids is None:
|
| 562 |
+
pos = torch.arange(length, device=x.device).expand(batch_size, length)
|
| 563 |
+
|
| 564 |
+
raw_output = self.encode(x, pos)
|
| 565 |
+
raw_output = self.norm(raw_output)
|
| 566 |
+
raw_output = self.drop(raw_output)
|
| 567 |
+
|
| 568 |
+
output = self.output_layer(raw_output)
|
| 569 |
+
|
| 570 |
+
loss = None
|
| 571 |
+
if labels is not None:
|
| 572 |
+
loss_fct = nn.CrossEntropyLoss()
|
| 573 |
+
loss = loss_fct(output.view(batch_size * length, -1), labels.reshape(-1))
|
| 574 |
+
|
| 575 |
+
return MaskedLMOutput(
|
| 576 |
+
loss=loss, # shape: 1
|
| 577 |
+
logits=output, # shape: (batch_size * length, ntokens)
|
| 578 |
+
hidden_states=None,
|
| 579 |
+
attentions=None,
|
| 580 |
+
)
|
| 581 |
+
|
| 582 |
+
##########################################
|
| 583 |
+
# HuggingFace Model
|
| 584 |
+
##########################################
|
| 585 |
+
class StructformerModel(PreTrainedModel):
|
| 586 |
+
config_class = StructformerConfig
|
| 587 |
+
|
| 588 |
+
def __init__(self, config):
|
| 589 |
+
super().__init__(config)
|
| 590 |
+
self.model = StructFormer(
|
| 591 |
+
hidden_size=config.hidden_size,
|
| 592 |
+
n_context_layers=config.n_context_layers,
|
| 593 |
+
nlayers=config.nlayers,
|
| 594 |
+
ntokens=config.ntokens,
|
| 595 |
+
nhead=config.nhead,
|
| 596 |
+
dropout=config.dropout,
|
| 597 |
+
dropatt=config.dropatt,
|
| 598 |
+
relative_bias=config.relative_bias,
|
| 599 |
+
pos_emb=config.pos_emb,
|
| 600 |
+
pad=config.pad,
|
| 601 |
+
n_parser_layers=config.n_parser_layers,
|
| 602 |
+
conv_size=config.conv_size,
|
| 603 |
+
relations=config.relations,
|
| 604 |
+
weight_act=config.weight_act
|
| 605 |
+
)
|
| 606 |
+
|
| 607 |
+
def forward(self, input_ids, labels=None, **kwargs):
|
| 608 |
+
return self.model(input_ids, labels=labels, **kwargs)
|
| 609 |
+
|
| 610 |
+
|
| 611 |
+
class StructFormerClassification(Transformer):
|
| 612 |
+
"""StructFormer model."""
|
| 613 |
+
|
| 614 |
+
def __init__(self,
|
| 615 |
+
hidden_size,
|
| 616 |
+
n_context_layers,
|
| 617 |
+
nlayers,
|
| 618 |
+
ntokens,
|
| 619 |
+
nhead=8,
|
| 620 |
+
dropout=0.1,
|
| 621 |
+
dropatt=0.1,
|
| 622 |
+
relative_bias=False,
|
| 623 |
+
pos_emb=False,
|
| 624 |
+
pad=0,
|
| 625 |
+
n_parser_layers=4,
|
| 626 |
+
conv_size=9,
|
| 627 |
+
relations=('head', 'child'),
|
| 628 |
+
weight_act='softmax',
|
| 629 |
+
config=None,
|
| 630 |
+
):
|
| 631 |
+
|
| 632 |
+
|
| 633 |
+
super(StructFormerClassification, self).__init__(
|
| 634 |
+
hidden_size,
|
| 635 |
+
nlayers,
|
| 636 |
+
ntokens,
|
| 637 |
+
nhead=nhead,
|
| 638 |
+
dropout=dropout,
|
| 639 |
+
dropatt=dropatt,
|
| 640 |
+
relative_bias=relative_bias,
|
| 641 |
+
pos_emb=pos_emb,
|
| 642 |
+
pad=pad)
|
| 643 |
+
|
| 644 |
+
self.num_labels = config.num_labels
|
| 645 |
+
self.config = config
|
| 646 |
+
|
| 647 |
+
self.classifier = RobertaClassificationHead(config)
|
| 648 |
+
|
| 649 |
+
def encode(self, x, pos):
|
| 650 |
+
h = self.emb(x)
|
| 651 |
+
if hasattr(self, 'pos_emb'):
|
| 652 |
+
h = h + self.pos_emb(pos)
|
| 653 |
+
h_list = []
|
| 654 |
+
visibility = self.visibility(x, x.device)
|
| 655 |
+
|
| 656 |
+
for i in range(self.nlayers):
|
| 657 |
+
h_list.append(h)
|
| 658 |
+
h = self.layers[i](
|
| 659 |
+
h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
|
| 660 |
+
|
| 661 |
+
output = h
|
| 662 |
+
h_array = torch.stack(h_list, dim=2)
|
| 663 |
+
|
| 664 |
+
return output
|
| 665 |
+
|
| 666 |
+
|
| 667 |
+
def forward(self, input_ids, labels=None, position_ids=None, **kwargs):
|
| 668 |
+
|
| 669 |
+
x = input_ids
|
| 670 |
+
batch_size, length = x.size()
|
| 671 |
+
|
| 672 |
+
if position_ids is None:
|
| 673 |
+
pos = torch.arange(length, device=x.device).expand(batch_size, length)
|
| 674 |
+
|
| 675 |
+
raw_output = self.encode(x, pos)
|
| 676 |
+
raw_output = self.norm(raw_output)
|
| 677 |
+
raw_output = self.drop(raw_output)
|
| 678 |
+
|
| 679 |
+
#output = self.output_layer(raw_output)
|
| 680 |
+
logits = self.classifier(raw_output)
|
| 681 |
+
|
| 682 |
+
loss = None
|
| 683 |
+
if labels is not None:
|
| 684 |
+
if self.config.problem_type is None:
|
| 685 |
+
if self.num_labels == 1:
|
| 686 |
+
self.config.problem_type = "regression"
|
| 687 |
+
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
| 688 |
+
self.config.problem_type = "single_label_classification"
|
| 689 |
+
else:
|
| 690 |
+
self.config.problem_type = "multi_label_classification"
|
| 691 |
+
|
| 692 |
+
if self.config.problem_type == "regression":
|
| 693 |
+
loss_fct = MSELoss()
|
| 694 |
+
if self.num_labels == 1:
|
| 695 |
+
loss = loss_fct(logits.squeeze(), labels.squeeze())
|
| 696 |
+
else:
|
| 697 |
+
loss = loss_fct(logits, labels)
|
| 698 |
+
elif self.config.problem_type == "single_label_classification":
|
| 699 |
+
loss_fct = CrossEntropyLoss()
|
| 700 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
| 701 |
+
elif self.config.problem_type == "multi_label_classification":
|
| 702 |
+
loss_fct = BCEWithLogitsLoss()
|
| 703 |
+
loss = loss_fct(logits, labels)
|
| 704 |
+
|
| 705 |
+
|
| 706 |
+
return SequenceClassifierOutput(
|
| 707 |
+
loss=loss,
|
| 708 |
+
logits=logits,
|
| 709 |
+
hidden_states=None,
|
| 710 |
+
attentions=None,
|
| 711 |
+
)
|
| 712 |
+
|
| 713 |
+
|
| 714 |
+
class StructformerModelForSequenceClassification(PreTrainedModel):
|
| 715 |
+
config_class = StructformerConfig
|
| 716 |
+
def __init__(self, config):
|
| 717 |
+
super().__init__(config)
|
| 718 |
+
self.model = StructFormerClassification(
|
| 719 |
+
hidden_size=config.hidden_size,
|
| 720 |
+
n_context_layers=config.n_context_layers,
|
| 721 |
+
nlayers=config.nlayers,
|
| 722 |
+
ntokens=config.ntokens,
|
| 723 |
+
nhead=config.nhead,
|
| 724 |
+
dropout=config.dropout,
|
| 725 |
+
dropatt=config.dropatt,
|
| 726 |
+
relative_bias=config.relative_bias,
|
| 727 |
+
pos_emb=config.pos_emb,
|
| 728 |
+
pad=config.pad,
|
| 729 |
+
n_parser_layers=config.n_parser_layers,
|
| 730 |
+
conv_size=config.conv_size,
|
| 731 |
+
relations=config.relations,
|
| 732 |
+
weight_act=config.weight_act,
|
| 733 |
+
config=config)
|
| 734 |
+
|
| 735 |
+
def _init_weights(self, module):
|
| 736 |
+
"""Initialize the weights"""
|
| 737 |
+
if isinstance(module, nn.Linear):
|
| 738 |
+
# Slightly different from the TF version which uses truncated_normal for initialization
|
| 739 |
+
# cf https://github.com/pytorch/pytorch/pull/5617
|
| 740 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 741 |
+
if module.bias is not None:
|
| 742 |
+
module.bias.data.zero_()
|
| 743 |
+
elif isinstance(module, nn.Embedding):
|
| 744 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 745 |
+
if module.padding_idx is not None:
|
| 746 |
+
module.weight.data[module.padding_idx].zero_()
|
| 747 |
+
elif isinstance(module, nn.LayerNorm):
|
| 748 |
+
if module.bias is not None:
|
| 749 |
+
module.bias.data.zero_()
|
| 750 |
+
module.weight.data.fill_(1.0)
|
| 751 |
+
|
| 752 |
+
|
| 753 |
+
def forward(self, input_ids, labels=None, **kwargs):
|
| 754 |
+
return self.model(input_ids, labels=labels, **kwargs)
|
finetune/cola/tokenizer_config.json
ADDED
|
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_prefix_space": false,
|
| 3 |
+
"bos_token": {
|
| 4 |
+
"__type": "AddedToken",
|
| 5 |
+
"content": "<s>",
|
| 6 |
+
"lstrip": false,
|
| 7 |
+
"normalized": true,
|
| 8 |
+
"rstrip": false,
|
| 9 |
+
"single_word": false
|
| 10 |
+
},
|
| 11 |
+
"cls_token": {
|
| 12 |
+
"__type": "AddedToken",
|
| 13 |
+
"content": "<s>",
|
| 14 |
+
"lstrip": false,
|
| 15 |
+
"normalized": true,
|
| 16 |
+
"rstrip": false,
|
| 17 |
+
"single_word": false
|
| 18 |
+
},
|
| 19 |
+
"eos_token": {
|
| 20 |
+
"__type": "AddedToken",
|
| 21 |
+
"content": "</s>",
|
| 22 |
+
"lstrip": false,
|
| 23 |
+
"normalized": true,
|
| 24 |
+
"rstrip": false,
|
| 25 |
+
"single_word": false
|
| 26 |
+
},
|
| 27 |
+
"errors": "replace",
|
| 28 |
+
"mask_token": {
|
| 29 |
+
"__type": "AddedToken",
|
| 30 |
+
"content": "<mask>",
|
| 31 |
+
"lstrip": true,
|
| 32 |
+
"normalized": false,
|
| 33 |
+
"rstrip": false,
|
| 34 |
+
"single_word": false
|
| 35 |
+
},
|
| 36 |
+
"model_max_length": 512,
|
| 37 |
+
"name_or_path": "final_models/transformer_base_final_2",
|
| 38 |
+
"pad_token": {
|
| 39 |
+
"__type": "AddedToken",
|
| 40 |
+
"content": "<pad>",
|
| 41 |
+
"lstrip": false,
|
| 42 |
+
"normalized": true,
|
| 43 |
+
"rstrip": false,
|
| 44 |
+
"single_word": false
|
| 45 |
+
},
|
| 46 |
+
"sep_token": {
|
| 47 |
+
"__type": "AddedToken",
|
| 48 |
+
"content": "</s>",
|
| 49 |
+
"lstrip": false,
|
| 50 |
+
"normalized": true,
|
| 51 |
+
"rstrip": false,
|
| 52 |
+
"single_word": false
|
| 53 |
+
},
|
| 54 |
+
"special_tokens_map_file": null,
|
| 55 |
+
"tokenizer_class": "RobertaTokenizer",
|
| 56 |
+
"trim_offsets": true,
|
| 57 |
+
"unk_token": {
|
| 58 |
+
"__type": "AddedToken",
|
| 59 |
+
"content": "<unk>",
|
| 60 |
+
"lstrip": false,
|
| 61 |
+
"normalized": true,
|
| 62 |
+
"rstrip": false,
|
| 63 |
+
"single_word": false
|
| 64 |
+
}
|
| 65 |
+
}
|
finetune/cola/train_results.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"epoch": 10.0,
|
| 3 |
+
"train_loss": 0.42006420190783517,
|
| 4 |
+
"train_runtime": 267.2668,
|
| 5 |
+
"train_samples": 8164,
|
| 6 |
+
"train_samples_per_second": 305.463,
|
| 7 |
+
"train_steps_per_second": 2.582
|
| 8 |
+
}
|
finetune/cola/trainer_state.json
ADDED
|
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"best_metric": 0.7777040477770405,
|
| 3 |
+
"best_model_checkpoint": "final_models/transformer_base_final_2/finetune/cola/checkpoint-400",
|
| 4 |
+
"epoch": 10.0,
|
| 5 |
+
"global_step": 690,
|
| 6 |
+
"is_hyper_param_search": false,
|
| 7 |
+
"is_local_process_zero": true,
|
| 8 |
+
"is_world_process_zero": true,
|
| 9 |
+
"log_history": [
|
| 10 |
+
{
|
| 11 |
+
"epoch": 5.8,
|
| 12 |
+
"eval_accuracy": 0.6712462902069092,
|
| 13 |
+
"eval_f1": 0.7777040477770405,
|
| 14 |
+
"eval_loss": 0.7399011254310608,
|
| 15 |
+
"eval_mcc": 0.15981907397227785,
|
| 16 |
+
"eval_runtime": 1.4063,
|
| 17 |
+
"eval_samples_per_second": 724.62,
|
| 18 |
+
"eval_steps_per_second": 91.022,
|
| 19 |
+
"step": 400
|
| 20 |
+
},
|
| 21 |
+
{
|
| 22 |
+
"epoch": 7.25,
|
| 23 |
+
"learning_rate": 1.3768115942028985e-05,
|
| 24 |
+
"loss": 0.4934,
|
| 25 |
+
"step": 500
|
| 26 |
+
},
|
| 27 |
+
{
|
| 28 |
+
"epoch": 10.0,
|
| 29 |
+
"step": 690,
|
| 30 |
+
"total_flos": 5372103024537600.0,
|
| 31 |
+
"train_loss": 0.42006420190783517,
|
| 32 |
+
"train_runtime": 267.2668,
|
| 33 |
+
"train_samples_per_second": 305.463,
|
| 34 |
+
"train_steps_per_second": 2.582
|
| 35 |
+
}
|
| 36 |
+
],
|
| 37 |
+
"max_steps": 690,
|
| 38 |
+
"num_train_epochs": 10,
|
| 39 |
+
"total_flos": 5372103024537600.0,
|
| 40 |
+
"trial_name": null,
|
| 41 |
+
"trial_params": null
|
| 42 |
+
}
|
finetune/cola/training_args.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:42f03300332c3f519b9e94f52c9e382ae77f93b0d20616ae0c3858d1e25c21dc
|
| 3 |
+
size 3503
|
finetune/cola/vocab.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
finetune/control_raising_control/all_results.json
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"epoch": 10.0,
|
| 3 |
+
"eval_accuracy": 0.8665371537208557,
|
| 4 |
+
"eval_f1": 0.8797792137856758,
|
| 5 |
+
"eval_loss": 0.8952956199645996,
|
| 6 |
+
"eval_mcc": 0.7560620169097876,
|
| 7 |
+
"eval_runtime": 18.4787,
|
| 8 |
+
"eval_samples": 13382,
|
| 9 |
+
"eval_samples_per_second": 724.185,
|
| 10 |
+
"eval_steps_per_second": 90.537,
|
| 11 |
+
"train_loss": 0.04565125611526045,
|
| 12 |
+
"train_runtime": 234.6588,
|
| 13 |
+
"train_samples": 6570,
|
| 14 |
+
"train_samples_per_second": 279.981,
|
| 15 |
+
"train_steps_per_second": 2.344
|
| 16 |
+
}
|
finetune/control_raising_control/checkpoint-400/config.json
ADDED
|
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_name_or_path": "final_models/transformer_base_final_2",
|
| 3 |
+
"architectures": [
|
| 4 |
+
"StructformerModelForSequenceClassification"
|
| 5 |
+
],
|
| 6 |
+
"attention_probs_dropout_prob": 0.1,
|
| 7 |
+
"auto_map": {
|
| 8 |
+
"AutoConfig": "structformer_as_hf_no_parser.StructformerConfig",
|
| 9 |
+
"AutoModelForMaskedLM": "structformer_as_hf_no_parser.StructformerModel",
|
| 10 |
+
"AutoModelForSequenceClassification": "structformer_as_hf_no_parser.StructformerModelForSequenceClassification"
|
| 11 |
+
},
|
| 12 |
+
"bos_token_id": 0,
|
| 13 |
+
"classifier_dropout": null,
|
| 14 |
+
"conv_size": 9,
|
| 15 |
+
"dropatt": 0.1,
|
| 16 |
+
"dropout": 0.1,
|
| 17 |
+
"eos_token_id": 2,
|
| 18 |
+
"hidden_act": "gelu",
|
| 19 |
+
"hidden_dropout_prob": 0.1,
|
| 20 |
+
"hidden_size": 768,
|
| 21 |
+
"id2label": {
|
| 22 |
+
"0": 0,
|
| 23 |
+
"1": 1
|
| 24 |
+
},
|
| 25 |
+
"initializer_range": 0.02,
|
| 26 |
+
"intermediate_size": 3072,
|
| 27 |
+
"label2id": {
|
| 28 |
+
"0": 0,
|
| 29 |
+
"1": 1
|
| 30 |
+
},
|
| 31 |
+
"layer_norm_eps": 1e-05,
|
| 32 |
+
"max_position_embeddings": 514,
|
| 33 |
+
"model_type": "structformer",
|
| 34 |
+
"n_context_layers": 0,
|
| 35 |
+
"n_parser_layers": 0,
|
| 36 |
+
"nhead": 12,
|
| 37 |
+
"nlayers": 12,
|
| 38 |
+
"ntokens": 32000,
|
| 39 |
+
"num_attention_heads": 12,
|
| 40 |
+
"num_hidden_layers": 12,
|
| 41 |
+
"pad": 0,
|
| 42 |
+
"pad_token_id": 1,
|
| 43 |
+
"pos_emb": true,
|
| 44 |
+
"position_embedding_type": "absolute",
|
| 45 |
+
"problem_type": "single_label_classification",
|
| 46 |
+
"relations": [
|
| 47 |
+
"head",
|
| 48 |
+
"child"
|
| 49 |
+
],
|
| 50 |
+
"relative_bias": false,
|
| 51 |
+
"torch_dtype": "float32",
|
| 52 |
+
"transformers_version": "4.26.1",
|
| 53 |
+
"type_vocab_size": 1,
|
| 54 |
+
"use_cache": true,
|
| 55 |
+
"vocab_size": 32000,
|
| 56 |
+
"weight_act": "softmax"
|
| 57 |
+
}
|
finetune/control_raising_control/checkpoint-400/merges.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
finetune/control_raising_control/checkpoint-400/optimizer.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:eb5abc7df98c1236620c1f8058c2ccdd643b47bf1e8151cda6472497757b7033
|
| 3 |
+
size 884981025
|
finetune/control_raising_control/checkpoint-400/pytorch_model.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a17588df7b02a3563a195e3933f14f28429949d3cad166f41883869da4ce57a5
|
| 3 |
+
size 442624943
|
finetune/control_raising_control/checkpoint-400/rng_state.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:02576a68480d1bd556bf2545dbb91a6fcd2d751faf798e19f80054fd04bf34a7
|
| 3 |
+
size 14503
|
finetune/control_raising_control/checkpoint-400/scheduler.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0c2403f14641b3caeb1b4d17bf70ec776358494ec9059cbe53a4c9c5a18c4c15
|
| 3 |
+
size 623
|
finetune/control_raising_control/checkpoint-400/special_tokens_map.json
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bos_token": "<s>",
|
| 3 |
+
"cls_token": "<s>",
|
| 4 |
+
"eos_token": "</s>",
|
| 5 |
+
"mask_token": {
|
| 6 |
+
"content": "<mask>",
|
| 7 |
+
"lstrip": true,
|
| 8 |
+
"normalized": false,
|
| 9 |
+
"rstrip": false,
|
| 10 |
+
"single_word": false
|
| 11 |
+
},
|
| 12 |
+
"pad_token": "<pad>",
|
| 13 |
+
"sep_token": "</s>",
|
| 14 |
+
"unk_token": "<unk>"
|
| 15 |
+
}
|
finetune/control_raising_control/checkpoint-400/structformer_as_hf_no_parser.py
ADDED
|
@@ -0,0 +1,754 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import torch.nn as nn
|
| 3 |
+
import torch.nn.functional as F
|
| 4 |
+
from torch.nn import init
|
| 5 |
+
from transformers import PreTrainedModel
|
| 6 |
+
from transformers import PretrainedConfig
|
| 7 |
+
from transformers.modeling_outputs import MaskedLMOutput
|
| 8 |
+
from typing import List
|
| 9 |
+
from torch.nn import CrossEntropyLoss, MSELoss, BCEWithLogitsLoss
|
| 10 |
+
from transformers.modeling_outputs import (
|
| 11 |
+
BaseModelOutputWithPastAndCrossAttentions,
|
| 12 |
+
BaseModelOutputWithPoolingAndCrossAttentions,
|
| 13 |
+
MaskedLMOutput,
|
| 14 |
+
SequenceClassifierOutput
|
| 15 |
+
)
|
| 16 |
+
|
| 17 |
+
##########################################
|
| 18 |
+
# HuggingFace Config
|
| 19 |
+
##########################################
|
| 20 |
+
class StructformerConfig(PretrainedConfig):
|
| 21 |
+
model_type = "structformer"
|
| 22 |
+
|
| 23 |
+
def __init__(
|
| 24 |
+
self,
|
| 25 |
+
hidden_size=768,
|
| 26 |
+
n_context_layers=2,
|
| 27 |
+
nlayers=6,
|
| 28 |
+
ntokens=32000,
|
| 29 |
+
nhead=8,
|
| 30 |
+
dropout=0.1,
|
| 31 |
+
dropatt=0.1,
|
| 32 |
+
relative_bias=False,
|
| 33 |
+
pos_emb=False,
|
| 34 |
+
pad=0,
|
| 35 |
+
n_parser_layers=4,
|
| 36 |
+
conv_size=9,
|
| 37 |
+
relations=('head', 'child'),
|
| 38 |
+
weight_act='softmax',
|
| 39 |
+
**kwargs,
|
| 40 |
+
):
|
| 41 |
+
self.hidden_size = hidden_size
|
| 42 |
+
self.n_context_layers = n_context_layers
|
| 43 |
+
self.nlayers = nlayers
|
| 44 |
+
self.ntokens = ntokens
|
| 45 |
+
self.nhead = nhead
|
| 46 |
+
self.dropout = dropout
|
| 47 |
+
self.dropatt = dropatt
|
| 48 |
+
self.relative_bias = relative_bias
|
| 49 |
+
self.pos_emb = pos_emb
|
| 50 |
+
self.pad = pad
|
| 51 |
+
self.n_parser_layers = n_parser_layers
|
| 52 |
+
self.conv_size = conv_size
|
| 53 |
+
self.relations = relations
|
| 54 |
+
self.weight_act = weight_act
|
| 55 |
+
super().__init__(**kwargs)
|
| 56 |
+
|
| 57 |
+
##########################################
|
| 58 |
+
# Custom Layers
|
| 59 |
+
##########################################
|
| 60 |
+
def _get_activation_fn(activation):
|
| 61 |
+
"""Get specified activation function."""
|
| 62 |
+
if activation == "relu":
|
| 63 |
+
return nn.ReLU()
|
| 64 |
+
elif activation == "gelu":
|
| 65 |
+
return nn.GELU()
|
| 66 |
+
elif activation == "leakyrelu":
|
| 67 |
+
return nn.LeakyReLU()
|
| 68 |
+
|
| 69 |
+
raise RuntimeError(
|
| 70 |
+
"activation should be relu/gelu, not {}".format(activation))
|
| 71 |
+
|
| 72 |
+
class Conv1d(nn.Module):
|
| 73 |
+
"""1D convolution layer."""
|
| 74 |
+
|
| 75 |
+
def __init__(self, hidden_size, kernel_size, dilation=1):
|
| 76 |
+
"""Initialization.
|
| 77 |
+
|
| 78 |
+
Args:
|
| 79 |
+
hidden_size: dimension of input embeddings
|
| 80 |
+
kernel_size: convolution kernel size
|
| 81 |
+
dilation: the spacing between the kernel points
|
| 82 |
+
"""
|
| 83 |
+
super(Conv1d, self).__init__()
|
| 84 |
+
|
| 85 |
+
if kernel_size % 2 == 0:
|
| 86 |
+
padding = (kernel_size // 2) * dilation
|
| 87 |
+
self.shift = True
|
| 88 |
+
else:
|
| 89 |
+
padding = ((kernel_size - 1) // 2) * dilation
|
| 90 |
+
self.shift = False
|
| 91 |
+
self.conv = nn.Conv1d(
|
| 92 |
+
hidden_size,
|
| 93 |
+
hidden_size,
|
| 94 |
+
kernel_size,
|
| 95 |
+
padding=padding,
|
| 96 |
+
dilation=dilation)
|
| 97 |
+
|
| 98 |
+
def forward(self, x):
|
| 99 |
+
"""Compute convolution.
|
| 100 |
+
|
| 101 |
+
Args:
|
| 102 |
+
x: input embeddings
|
| 103 |
+
Returns:
|
| 104 |
+
conv_output: convolution results
|
| 105 |
+
"""
|
| 106 |
+
|
| 107 |
+
if self.shift:
|
| 108 |
+
return self.conv(x.transpose(1, 2)).transpose(1, 2)[:, 1:]
|
| 109 |
+
else:
|
| 110 |
+
return self.conv(x.transpose(1, 2)).transpose(1, 2)
|
| 111 |
+
|
| 112 |
+
class MultiheadAttention(nn.Module):
|
| 113 |
+
"""Multi-head self-attention layer."""
|
| 114 |
+
|
| 115 |
+
def __init__(self,
|
| 116 |
+
embed_dim,
|
| 117 |
+
num_heads,
|
| 118 |
+
dropout=0.,
|
| 119 |
+
bias=True,
|
| 120 |
+
v_proj=True,
|
| 121 |
+
out_proj=True,
|
| 122 |
+
relative_bias=True):
|
| 123 |
+
"""Initialization.
|
| 124 |
+
|
| 125 |
+
Args:
|
| 126 |
+
embed_dim: dimension of input embeddings
|
| 127 |
+
num_heads: number of self-attention heads
|
| 128 |
+
dropout: dropout rate
|
| 129 |
+
bias: bool, indicate whether include bias for linear transformations
|
| 130 |
+
v_proj: bool, indicate whether project inputs to new values
|
| 131 |
+
out_proj: bool, indicate whether project outputs to new values
|
| 132 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 133 |
+
attention bias
|
| 134 |
+
"""
|
| 135 |
+
|
| 136 |
+
super(MultiheadAttention, self).__init__()
|
| 137 |
+
self.embed_dim = embed_dim
|
| 138 |
+
|
| 139 |
+
self.num_heads = num_heads
|
| 140 |
+
self.drop = nn.Dropout(dropout)
|
| 141 |
+
self.head_dim = embed_dim // num_heads
|
| 142 |
+
assert self.head_dim * num_heads == self.embed_dim, ("embed_dim must be "
|
| 143 |
+
"divisible by "
|
| 144 |
+
"num_heads")
|
| 145 |
+
|
| 146 |
+
self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 147 |
+
self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 148 |
+
if v_proj:
|
| 149 |
+
self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 150 |
+
else:
|
| 151 |
+
self.v_proj = nn.Identity()
|
| 152 |
+
|
| 153 |
+
if out_proj:
|
| 154 |
+
self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
| 155 |
+
else:
|
| 156 |
+
self.out_proj = nn.Identity()
|
| 157 |
+
|
| 158 |
+
if relative_bias:
|
| 159 |
+
self.relative_bias = nn.Parameter(torch.zeros((self.num_heads, 512)))
|
| 160 |
+
else:
|
| 161 |
+
self.relative_bias = None
|
| 162 |
+
|
| 163 |
+
self._reset_parameters()
|
| 164 |
+
|
| 165 |
+
def _reset_parameters(self):
|
| 166 |
+
"""Initialize attention parameters."""
|
| 167 |
+
|
| 168 |
+
init.xavier_uniform_(self.q_proj.weight)
|
| 169 |
+
init.constant_(self.q_proj.bias, 0.)
|
| 170 |
+
|
| 171 |
+
init.xavier_uniform_(self.k_proj.weight)
|
| 172 |
+
init.constant_(self.k_proj.bias, 0.)
|
| 173 |
+
|
| 174 |
+
if isinstance(self.v_proj, nn.Linear):
|
| 175 |
+
init.xavier_uniform_(self.v_proj.weight)
|
| 176 |
+
init.constant_(self.v_proj.bias, 0.)
|
| 177 |
+
|
| 178 |
+
if isinstance(self.out_proj, nn.Linear):
|
| 179 |
+
init.xavier_uniform_(self.out_proj.weight)
|
| 180 |
+
init.constant_(self.out_proj.bias, 0.)
|
| 181 |
+
|
| 182 |
+
def forward(self, query, key_padding_mask=None, attn_mask=None):
|
| 183 |
+
"""Compute multi-head self-attention.
|
| 184 |
+
|
| 185 |
+
Args:
|
| 186 |
+
query: input embeddings
|
| 187 |
+
key_padding_mask: 3D mask that prevents attention to certain positions
|
| 188 |
+
attn_mask: 3D mask that rescale the attention weight at each position
|
| 189 |
+
Returns:
|
| 190 |
+
attn_output: self-attention output
|
| 191 |
+
"""
|
| 192 |
+
|
| 193 |
+
length, bsz, embed_dim = query.size()
|
| 194 |
+
assert embed_dim == self.embed_dim
|
| 195 |
+
|
| 196 |
+
head_dim = embed_dim // self.num_heads
|
| 197 |
+
assert head_dim * self.num_heads == embed_dim, ("embed_dim must be "
|
| 198 |
+
"divisible by num_heads")
|
| 199 |
+
scaling = float(head_dim)**-0.5
|
| 200 |
+
|
| 201 |
+
q = self.q_proj(query)
|
| 202 |
+
k = self.k_proj(query)
|
| 203 |
+
v = self.v_proj(query)
|
| 204 |
+
|
| 205 |
+
q = q * scaling
|
| 206 |
+
|
| 207 |
+
if attn_mask is not None:
|
| 208 |
+
assert list(attn_mask.size()) == [bsz * self.num_heads,
|
| 209 |
+
query.size(0), query.size(0)]
|
| 210 |
+
|
| 211 |
+
q = q.contiguous().view(length, bsz * self.num_heads,
|
| 212 |
+
head_dim).transpose(0, 1)
|
| 213 |
+
k = k.contiguous().view(length, bsz * self.num_heads,
|
| 214 |
+
head_dim).transpose(0, 1)
|
| 215 |
+
v = v.contiguous().view(length, bsz * self.num_heads,
|
| 216 |
+
head_dim).transpose(0, 1)
|
| 217 |
+
|
| 218 |
+
attn_output_weights = torch.bmm(q, k.transpose(1, 2))
|
| 219 |
+
assert list(
|
| 220 |
+
attn_output_weights.size()) == [bsz * self.num_heads, length, length]
|
| 221 |
+
|
| 222 |
+
if self.relative_bias is not None:
|
| 223 |
+
pos = torch.arange(length, device=query.device)
|
| 224 |
+
relative_pos = torch.abs(pos[:, None] - pos[None, :]) + 256
|
| 225 |
+
relative_pos = relative_pos[None, :, :].expand(bsz * self.num_heads, -1,
|
| 226 |
+
-1)
|
| 227 |
+
|
| 228 |
+
relative_bias = self.relative_bias.repeat_interleave(bsz, dim=0)
|
| 229 |
+
relative_bias = relative_bias[:, None, :].expand(-1, length, -1)
|
| 230 |
+
relative_bias = torch.gather(relative_bias, 2, relative_pos)
|
| 231 |
+
attn_output_weights = attn_output_weights + relative_bias
|
| 232 |
+
|
| 233 |
+
if key_padding_mask is not None:
|
| 234 |
+
attn_output_weights = attn_output_weights + key_padding_mask
|
| 235 |
+
|
| 236 |
+
if attn_mask is None:
|
| 237 |
+
attn_output_weights = torch.softmax(attn_output_weights, dim=-1)
|
| 238 |
+
else:
|
| 239 |
+
attn_output_weights = torch.sigmoid(attn_output_weights) * attn_mask
|
| 240 |
+
|
| 241 |
+
attn_output_weights = self.drop(attn_output_weights)
|
| 242 |
+
|
| 243 |
+
attn_output = torch.bmm(attn_output_weights, v)
|
| 244 |
+
|
| 245 |
+
assert list(attn_output.size()) == [bsz * self.num_heads, length, head_dim]
|
| 246 |
+
attn_output = attn_output.transpose(0, 1).contiguous().view(
|
| 247 |
+
length, bsz, embed_dim)
|
| 248 |
+
attn_output = self.out_proj(attn_output)
|
| 249 |
+
|
| 250 |
+
return attn_output
|
| 251 |
+
|
| 252 |
+
class TransformerLayer(nn.Module):
|
| 253 |
+
"""TransformerEncoderLayer is made up of self-attn and feedforward network."""
|
| 254 |
+
|
| 255 |
+
def __init__(self,
|
| 256 |
+
d_model,
|
| 257 |
+
nhead,
|
| 258 |
+
dim_feedforward=2048,
|
| 259 |
+
dropout=0.1,
|
| 260 |
+
dropatt=0.1,
|
| 261 |
+
activation="leakyrelu",
|
| 262 |
+
relative_bias=True):
|
| 263 |
+
"""Initialization.
|
| 264 |
+
|
| 265 |
+
Args:
|
| 266 |
+
d_model: dimension of inputs
|
| 267 |
+
nhead: number of self-attention heads
|
| 268 |
+
dim_feedforward: dimension of hidden layer in feedforward layer
|
| 269 |
+
dropout: dropout rate
|
| 270 |
+
dropatt: drop attention rate
|
| 271 |
+
activation: activation function
|
| 272 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 273 |
+
attention bias
|
| 274 |
+
"""
|
| 275 |
+
|
| 276 |
+
super(TransformerLayer, self).__init__()
|
| 277 |
+
|
| 278 |
+
self.self_attn = MultiheadAttention(
|
| 279 |
+
d_model, nhead, dropout=dropatt, relative_bias=relative_bias)
|
| 280 |
+
|
| 281 |
+
# Implementation of Feedforward model
|
| 282 |
+
self.feedforward = nn.Sequential(
|
| 283 |
+
nn.LayerNorm(d_model), nn.Linear(d_model, dim_feedforward),
|
| 284 |
+
_get_activation_fn(activation), nn.Dropout(dropout),
|
| 285 |
+
nn.Linear(dim_feedforward, d_model))
|
| 286 |
+
|
| 287 |
+
self.norm = nn.LayerNorm(d_model)
|
| 288 |
+
self.dropout1 = nn.Dropout(dropout)
|
| 289 |
+
self.dropout2 = nn.Dropout(dropout)
|
| 290 |
+
|
| 291 |
+
self.nhead = nhead
|
| 292 |
+
|
| 293 |
+
def forward(self, src, attn_mask=None, key_padding_mask=None):
|
| 294 |
+
"""Pass the input through the encoder layer.
|
| 295 |
+
|
| 296 |
+
Args:
|
| 297 |
+
src: the sequence to the encoder layer (required).
|
| 298 |
+
attn_mask: the mask for the src sequence (optional).
|
| 299 |
+
key_padding_mask: the mask for the src keys per batch (optional).
|
| 300 |
+
Returns:
|
| 301 |
+
src3: the output of transformer layer, share the same shape as src.
|
| 302 |
+
"""
|
| 303 |
+
src2 = self.self_attn(
|
| 304 |
+
self.norm(src), attn_mask=attn_mask, key_padding_mask=key_padding_mask)
|
| 305 |
+
src2 = src + self.dropout1(src2)
|
| 306 |
+
src3 = self.feedforward(src2)
|
| 307 |
+
src3 = src2 + self.dropout2(src3)
|
| 308 |
+
|
| 309 |
+
return src3
|
| 310 |
+
|
| 311 |
+
|
| 312 |
+
|
| 313 |
+
class RobertaClassificationHead(nn.Module):
|
| 314 |
+
"""Head for sentence-level classification tasks."""
|
| 315 |
+
|
| 316 |
+
def __init__(self, config):
|
| 317 |
+
super().__init__()
|
| 318 |
+
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
|
| 319 |
+
classifier_dropout = (
|
| 320 |
+
config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
|
| 321 |
+
)
|
| 322 |
+
self.dropout = nn.Dropout(classifier_dropout)
|
| 323 |
+
self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
|
| 324 |
+
|
| 325 |
+
def forward(self, features, **kwargs):
|
| 326 |
+
x = features[:, 0, :] # take <s> token (equiv. to [CLS])
|
| 327 |
+
x = self.dropout(x)
|
| 328 |
+
x = self.dense(x)
|
| 329 |
+
x = torch.tanh(x)
|
| 330 |
+
x = self.dropout(x)
|
| 331 |
+
x = self.out_proj(x)
|
| 332 |
+
return x
|
| 333 |
+
|
| 334 |
+
|
| 335 |
+
##########################################
|
| 336 |
+
# Custom Models
|
| 337 |
+
##########################################
|
| 338 |
+
def cumprod(x, reverse=False, exclusive=False):
|
| 339 |
+
"""cumulative product."""
|
| 340 |
+
if reverse:
|
| 341 |
+
x = x.flip([-1])
|
| 342 |
+
|
| 343 |
+
if exclusive:
|
| 344 |
+
x = F.pad(x[:, :, :-1], (1, 0), value=1)
|
| 345 |
+
|
| 346 |
+
cx = x.cumprod(-1)
|
| 347 |
+
|
| 348 |
+
if reverse:
|
| 349 |
+
cx = cx.flip([-1])
|
| 350 |
+
return cx
|
| 351 |
+
|
| 352 |
+
def cumsum(x, reverse=False, exclusive=False):
|
| 353 |
+
"""cumulative sum."""
|
| 354 |
+
bsz, _, length = x.size()
|
| 355 |
+
device = x.device
|
| 356 |
+
if reverse:
|
| 357 |
+
if exclusive:
|
| 358 |
+
w = torch.ones([bsz, length, length], device=device).tril(-1)
|
| 359 |
+
else:
|
| 360 |
+
w = torch.ones([bsz, length, length], device=device).tril(0)
|
| 361 |
+
cx = torch.bmm(x, w)
|
| 362 |
+
else:
|
| 363 |
+
if exclusive:
|
| 364 |
+
w = torch.ones([bsz, length, length], device=device).triu(1)
|
| 365 |
+
else:
|
| 366 |
+
w = torch.ones([bsz, length, length], device=device).triu(0)
|
| 367 |
+
cx = torch.bmm(x, w)
|
| 368 |
+
return cx
|
| 369 |
+
|
| 370 |
+
def cummin(x, reverse=False, exclusive=False, max_value=1e9):
|
| 371 |
+
"""cumulative min."""
|
| 372 |
+
if reverse:
|
| 373 |
+
if exclusive:
|
| 374 |
+
x = F.pad(x[:, :, 1:], (0, 1), value=max_value)
|
| 375 |
+
x = x.flip([-1]).cummin(-1)[0].flip([-1])
|
| 376 |
+
else:
|
| 377 |
+
if exclusive:
|
| 378 |
+
x = F.pad(x[:, :, :-1], (1, 0), value=max_value)
|
| 379 |
+
x = x.cummin(-1)[0]
|
| 380 |
+
return x
|
| 381 |
+
|
| 382 |
+
class Transformer(nn.Module):
|
| 383 |
+
"""Transformer model."""
|
| 384 |
+
|
| 385 |
+
def __init__(self,
|
| 386 |
+
hidden_size,
|
| 387 |
+
nlayers,
|
| 388 |
+
ntokens,
|
| 389 |
+
nhead=8,
|
| 390 |
+
dropout=0.1,
|
| 391 |
+
dropatt=0.1,
|
| 392 |
+
relative_bias=True,
|
| 393 |
+
pos_emb=False,
|
| 394 |
+
pad=0):
|
| 395 |
+
"""Initialization.
|
| 396 |
+
|
| 397 |
+
Args:
|
| 398 |
+
hidden_size: dimension of inputs and hidden states
|
| 399 |
+
nlayers: number of layers
|
| 400 |
+
ntokens: number of output categories
|
| 401 |
+
nhead: number of self-attention heads
|
| 402 |
+
dropout: dropout rate
|
| 403 |
+
dropatt: drop attention rate
|
| 404 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 405 |
+
attention bias
|
| 406 |
+
pos_emb: bool, indicate whether use a learnable positional embedding
|
| 407 |
+
pad: pad token index
|
| 408 |
+
"""
|
| 409 |
+
|
| 410 |
+
super(Transformer, self).__init__()
|
| 411 |
+
|
| 412 |
+
self.drop = nn.Dropout(dropout)
|
| 413 |
+
|
| 414 |
+
self.emb = nn.Embedding(ntokens, hidden_size)
|
| 415 |
+
if pos_emb:
|
| 416 |
+
self.pos_emb = nn.Embedding(500, hidden_size)
|
| 417 |
+
|
| 418 |
+
self.layers = nn.ModuleList([
|
| 419 |
+
TransformerLayer(hidden_size, nhead, hidden_size * 4, dropout,
|
| 420 |
+
dropatt=dropatt, relative_bias=relative_bias)
|
| 421 |
+
for _ in range(nlayers)])
|
| 422 |
+
|
| 423 |
+
self.norm = nn.LayerNorm(hidden_size)
|
| 424 |
+
|
| 425 |
+
self.output_layer = nn.Linear(hidden_size, ntokens)
|
| 426 |
+
self.output_layer.weight = self.emb.weight
|
| 427 |
+
|
| 428 |
+
self.init_weights()
|
| 429 |
+
|
| 430 |
+
self.nlayers = nlayers
|
| 431 |
+
self.nhead = nhead
|
| 432 |
+
self.ntokens = ntokens
|
| 433 |
+
self.hidden_size = hidden_size
|
| 434 |
+
self.pad = pad
|
| 435 |
+
|
| 436 |
+
def init_weights(self):
|
| 437 |
+
"""Initialize token embedding and output bias."""
|
| 438 |
+
initrange = 0.1
|
| 439 |
+
self.emb.weight.data.uniform_(-initrange, initrange)
|
| 440 |
+
if hasattr(self, 'pos_emb'):
|
| 441 |
+
self.pos_emb.weight.data.uniform_(-initrange, initrange)
|
| 442 |
+
self.output_layer.bias.data.fill_(0)
|
| 443 |
+
|
| 444 |
+
def visibility(self, x, device):
|
| 445 |
+
"""Mask pad tokens."""
|
| 446 |
+
visibility = (x != self.pad).float()
|
| 447 |
+
visibility = visibility[:, None, :].expand(-1, x.size(1), -1)
|
| 448 |
+
visibility = torch.repeat_interleave(visibility, self.nhead, dim=0)
|
| 449 |
+
return visibility.log()
|
| 450 |
+
|
| 451 |
+
def encode(self, x, pos):
|
| 452 |
+
"""Standard transformer encode process."""
|
| 453 |
+
h = self.emb(x)
|
| 454 |
+
if hasattr(self, 'pos_emb'):
|
| 455 |
+
h = h + self.pos_emb(pos)
|
| 456 |
+
h_list = []
|
| 457 |
+
visibility = self.visibility(x, x.device)
|
| 458 |
+
|
| 459 |
+
for i in range(self.nlayers):
|
| 460 |
+
h_list.append(h)
|
| 461 |
+
h = self.layers[i](
|
| 462 |
+
h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
|
| 463 |
+
|
| 464 |
+
output = h
|
| 465 |
+
h_array = torch.stack(h_list, dim=2)
|
| 466 |
+
|
| 467 |
+
return output, h_array
|
| 468 |
+
|
| 469 |
+
def forward(self, x, pos):
|
| 470 |
+
"""Pass the input through the encoder layer.
|
| 471 |
+
|
| 472 |
+
Args:
|
| 473 |
+
x: input tokens (required).
|
| 474 |
+
pos: position for each token (optional).
|
| 475 |
+
Returns:
|
| 476 |
+
output: probability distributions for missing tokens.
|
| 477 |
+
state_dict: parsing results and raw output
|
| 478 |
+
"""
|
| 479 |
+
|
| 480 |
+
batch_size, length = x.size()
|
| 481 |
+
|
| 482 |
+
raw_output, _ = self.encode(x, pos)
|
| 483 |
+
raw_output = self.norm(raw_output)
|
| 484 |
+
raw_output = self.drop(raw_output)
|
| 485 |
+
|
| 486 |
+
output = self.output_layer(raw_output)
|
| 487 |
+
return output.view(batch_size * length, -1), {'raw_output': raw_output,}
|
| 488 |
+
|
| 489 |
+
class StructFormer(Transformer):
|
| 490 |
+
"""StructFormer model."""
|
| 491 |
+
|
| 492 |
+
def __init__(self,
|
| 493 |
+
hidden_size,
|
| 494 |
+
n_context_layers,
|
| 495 |
+
nlayers,
|
| 496 |
+
ntokens,
|
| 497 |
+
nhead=8,
|
| 498 |
+
dropout=0.1,
|
| 499 |
+
dropatt=0.1,
|
| 500 |
+
relative_bias=False,
|
| 501 |
+
pos_emb=False,
|
| 502 |
+
pad=0,
|
| 503 |
+
n_parser_layers=4,
|
| 504 |
+
conv_size=9,
|
| 505 |
+
relations=('head', 'child'),
|
| 506 |
+
weight_act='softmax'):
|
| 507 |
+
"""Initialization.
|
| 508 |
+
|
| 509 |
+
Args:
|
| 510 |
+
hidden_size: dimension of inputs and hidden states
|
| 511 |
+
nlayers: number of layers
|
| 512 |
+
ntokens: number of output categories
|
| 513 |
+
nhead: number of self-attention heads
|
| 514 |
+
dropout: dropout rate
|
| 515 |
+
dropatt: drop attention rate
|
| 516 |
+
relative_bias: bool, indicate whether use a relative position based
|
| 517 |
+
attention bias
|
| 518 |
+
pos_emb: bool, indicate whether use a learnable positional embedding
|
| 519 |
+
pad: pad token index
|
| 520 |
+
n_parser_layers: number of parsing layers
|
| 521 |
+
conv_size: convolution kernel size for parser
|
| 522 |
+
relations: relations that are used to compute self attention
|
| 523 |
+
weight_act: relations distribution activation function
|
| 524 |
+
"""
|
| 525 |
+
|
| 526 |
+
super(StructFormer, self).__init__(
|
| 527 |
+
hidden_size,
|
| 528 |
+
nlayers,
|
| 529 |
+
ntokens,
|
| 530 |
+
nhead=nhead,
|
| 531 |
+
dropout=dropout,
|
| 532 |
+
dropatt=dropatt,
|
| 533 |
+
relative_bias=relative_bias,
|
| 534 |
+
pos_emb=pos_emb,
|
| 535 |
+
pad=pad)
|
| 536 |
+
|
| 537 |
+
|
| 538 |
+
def encode(self, x, pos):
|
| 539 |
+
h = self.emb(x)
|
| 540 |
+
if hasattr(self, 'pos_emb'):
|
| 541 |
+
h = h + self.pos_emb(pos)
|
| 542 |
+
h_list = []
|
| 543 |
+
visibility = self.visibility(x, x.device)
|
| 544 |
+
|
| 545 |
+
for i in range(self.nlayers):
|
| 546 |
+
h_list.append(h)
|
| 547 |
+
h = self.layers[i](
|
| 548 |
+
h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
|
| 549 |
+
|
| 550 |
+
output = h
|
| 551 |
+
h_array = torch.stack(h_list, dim=2)
|
| 552 |
+
|
| 553 |
+
return output
|
| 554 |
+
|
| 555 |
+
|
| 556 |
+
def forward(self, input_ids, labels=None, position_ids=None, **kwargs):
|
| 557 |
+
|
| 558 |
+
x = input_ids
|
| 559 |
+
batch_size, length = x.size()
|
| 560 |
+
|
| 561 |
+
if position_ids is None:
|
| 562 |
+
pos = torch.arange(length, device=x.device).expand(batch_size, length)
|
| 563 |
+
|
| 564 |
+
raw_output = self.encode(x, pos)
|
| 565 |
+
raw_output = self.norm(raw_output)
|
| 566 |
+
raw_output = self.drop(raw_output)
|
| 567 |
+
|
| 568 |
+
output = self.output_layer(raw_output)
|
| 569 |
+
|
| 570 |
+
loss = None
|
| 571 |
+
if labels is not None:
|
| 572 |
+
loss_fct = nn.CrossEntropyLoss()
|
| 573 |
+
loss = loss_fct(output.view(batch_size * length, -1), labels.reshape(-1))
|
| 574 |
+
|
| 575 |
+
return MaskedLMOutput(
|
| 576 |
+
loss=loss, # shape: 1
|
| 577 |
+
logits=output, # shape: (batch_size * length, ntokens)
|
| 578 |
+
hidden_states=None,
|
| 579 |
+
attentions=None,
|
| 580 |
+
)
|
| 581 |
+
|
| 582 |
+
##########################################
|
| 583 |
+
# HuggingFace Model
|
| 584 |
+
##########################################
|
| 585 |
+
class StructformerModel(PreTrainedModel):
|
| 586 |
+
config_class = StructformerConfig
|
| 587 |
+
|
| 588 |
+
def __init__(self, config):
|
| 589 |
+
super().__init__(config)
|
| 590 |
+
self.model = StructFormer(
|
| 591 |
+
hidden_size=config.hidden_size,
|
| 592 |
+
n_context_layers=config.n_context_layers,
|
| 593 |
+
nlayers=config.nlayers,
|
| 594 |
+
ntokens=config.ntokens,
|
| 595 |
+
nhead=config.nhead,
|
| 596 |
+
dropout=config.dropout,
|
| 597 |
+
dropatt=config.dropatt,
|
| 598 |
+
relative_bias=config.relative_bias,
|
| 599 |
+
pos_emb=config.pos_emb,
|
| 600 |
+
pad=config.pad,
|
| 601 |
+
n_parser_layers=config.n_parser_layers,
|
| 602 |
+
conv_size=config.conv_size,
|
| 603 |
+
relations=config.relations,
|
| 604 |
+
weight_act=config.weight_act
|
| 605 |
+
)
|
| 606 |
+
|
| 607 |
+
def forward(self, input_ids, labels=None, **kwargs):
|
| 608 |
+
return self.model(input_ids, labels=labels, **kwargs)
|
| 609 |
+
|
| 610 |
+
|
| 611 |
+
class StructFormerClassification(Transformer):
|
| 612 |
+
"""StructFormer model."""
|
| 613 |
+
|
| 614 |
+
def __init__(self,
|
| 615 |
+
hidden_size,
|
| 616 |
+
n_context_layers,
|
| 617 |
+
nlayers,
|
| 618 |
+
ntokens,
|
| 619 |
+
nhead=8,
|
| 620 |
+
dropout=0.1,
|
| 621 |
+
dropatt=0.1,
|
| 622 |
+
relative_bias=False,
|
| 623 |
+
pos_emb=False,
|
| 624 |
+
pad=0,
|
| 625 |
+
n_parser_layers=4,
|
| 626 |
+
conv_size=9,
|
| 627 |
+
relations=('head', 'child'),
|
| 628 |
+
weight_act='softmax',
|
| 629 |
+
config=None,
|
| 630 |
+
):
|
| 631 |
+
|
| 632 |
+
|
| 633 |
+
super(StructFormerClassification, self).__init__(
|
| 634 |
+
hidden_size,
|
| 635 |
+
nlayers,
|
| 636 |
+
ntokens,
|
| 637 |
+
nhead=nhead,
|
| 638 |
+
dropout=dropout,
|
| 639 |
+
dropatt=dropatt,
|
| 640 |
+
relative_bias=relative_bias,
|
| 641 |
+
pos_emb=pos_emb,
|
| 642 |
+
pad=pad)
|
| 643 |
+
|
| 644 |
+
self.num_labels = config.num_labels
|
| 645 |
+
self.config = config
|
| 646 |
+
|
| 647 |
+
self.classifier = RobertaClassificationHead(config)
|
| 648 |
+
|
| 649 |
+
def encode(self, x, pos):
|
| 650 |
+
h = self.emb(x)
|
| 651 |
+
if hasattr(self, 'pos_emb'):
|
| 652 |
+
h = h + self.pos_emb(pos)
|
| 653 |
+
h_list = []
|
| 654 |
+
visibility = self.visibility(x, x.device)
|
| 655 |
+
|
| 656 |
+
for i in range(self.nlayers):
|
| 657 |
+
h_list.append(h)
|
| 658 |
+
h = self.layers[i](
|
| 659 |
+
h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
|
| 660 |
+
|
| 661 |
+
output = h
|
| 662 |
+
h_array = torch.stack(h_list, dim=2)
|
| 663 |
+
|
| 664 |
+
return output
|
| 665 |
+
|
| 666 |
+
|
| 667 |
+
def forward(self, input_ids, labels=None, position_ids=None, **kwargs):
|
| 668 |
+
|
| 669 |
+
x = input_ids
|
| 670 |
+
batch_size, length = x.size()
|
| 671 |
+
|
| 672 |
+
if position_ids is None:
|
| 673 |
+
pos = torch.arange(length, device=x.device).expand(batch_size, length)
|
| 674 |
+
|
| 675 |
+
raw_output = self.encode(x, pos)
|
| 676 |
+
raw_output = self.norm(raw_output)
|
| 677 |
+
raw_output = self.drop(raw_output)
|
| 678 |
+
|
| 679 |
+
#output = self.output_layer(raw_output)
|
| 680 |
+
logits = self.classifier(raw_output)
|
| 681 |
+
|
| 682 |
+
loss = None
|
| 683 |
+
if labels is not None:
|
| 684 |
+
if self.config.problem_type is None:
|
| 685 |
+
if self.num_labels == 1:
|
| 686 |
+
self.config.problem_type = "regression"
|
| 687 |
+
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
| 688 |
+
self.config.problem_type = "single_label_classification"
|
| 689 |
+
else:
|
| 690 |
+
self.config.problem_type = "multi_label_classification"
|
| 691 |
+
|
| 692 |
+
if self.config.problem_type == "regression":
|
| 693 |
+
loss_fct = MSELoss()
|
| 694 |
+
if self.num_labels == 1:
|
| 695 |
+
loss = loss_fct(logits.squeeze(), labels.squeeze())
|
| 696 |
+
else:
|
| 697 |
+
loss = loss_fct(logits, labels)
|
| 698 |
+
elif self.config.problem_type == "single_label_classification":
|
| 699 |
+
loss_fct = CrossEntropyLoss()
|
| 700 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
| 701 |
+
elif self.config.problem_type == "multi_label_classification":
|
| 702 |
+
loss_fct = BCEWithLogitsLoss()
|
| 703 |
+
loss = loss_fct(logits, labels)
|
| 704 |
+
|
| 705 |
+
|
| 706 |
+
return SequenceClassifierOutput(
|
| 707 |
+
loss=loss,
|
| 708 |
+
logits=logits,
|
| 709 |
+
hidden_states=None,
|
| 710 |
+
attentions=None,
|
| 711 |
+
)
|
| 712 |
+
|
| 713 |
+
|
| 714 |
+
class StructformerModelForSequenceClassification(PreTrainedModel):
|
| 715 |
+
config_class = StructformerConfig
|
| 716 |
+
def __init__(self, config):
|
| 717 |
+
super().__init__(config)
|
| 718 |
+
self.model = StructFormerClassification(
|
| 719 |
+
hidden_size=config.hidden_size,
|
| 720 |
+
n_context_layers=config.n_context_layers,
|
| 721 |
+
nlayers=config.nlayers,
|
| 722 |
+
ntokens=config.ntokens,
|
| 723 |
+
nhead=config.nhead,
|
| 724 |
+
dropout=config.dropout,
|
| 725 |
+
dropatt=config.dropatt,
|
| 726 |
+
relative_bias=config.relative_bias,
|
| 727 |
+
pos_emb=config.pos_emb,
|
| 728 |
+
pad=config.pad,
|
| 729 |
+
n_parser_layers=config.n_parser_layers,
|
| 730 |
+
conv_size=config.conv_size,
|
| 731 |
+
relations=config.relations,
|
| 732 |
+
weight_act=config.weight_act,
|
| 733 |
+
config=config)
|
| 734 |
+
|
| 735 |
+
def _init_weights(self, module):
|
| 736 |
+
"""Initialize the weights"""
|
| 737 |
+
if isinstance(module, nn.Linear):
|
| 738 |
+
# Slightly different from the TF version which uses truncated_normal for initialization
|
| 739 |
+
# cf https://github.com/pytorch/pytorch/pull/5617
|
| 740 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 741 |
+
if module.bias is not None:
|
| 742 |
+
module.bias.data.zero_()
|
| 743 |
+
elif isinstance(module, nn.Embedding):
|
| 744 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 745 |
+
if module.padding_idx is not None:
|
| 746 |
+
module.weight.data[module.padding_idx].zero_()
|
| 747 |
+
elif isinstance(module, nn.LayerNorm):
|
| 748 |
+
if module.bias is not None:
|
| 749 |
+
module.bias.data.zero_()
|
| 750 |
+
module.weight.data.fill_(1.0)
|
| 751 |
+
|
| 752 |
+
|
| 753 |
+
def forward(self, input_ids, labels=None, **kwargs):
|
| 754 |
+
return self.model(input_ids, labels=labels, **kwargs)
|
finetune/control_raising_control/checkpoint-400/tokenizer_config.json
ADDED
|
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_prefix_space": false,
|
| 3 |
+
"bos_token": {
|
| 4 |
+
"__type": "AddedToken",
|
| 5 |
+
"content": "<s>",
|
| 6 |
+
"lstrip": false,
|
| 7 |
+
"normalized": true,
|
| 8 |
+
"rstrip": false,
|
| 9 |
+
"single_word": false
|
| 10 |
+
},
|
| 11 |
+
"cls_token": {
|
| 12 |
+
"__type": "AddedToken",
|
| 13 |
+
"content": "<s>",
|
| 14 |
+
"lstrip": false,
|
| 15 |
+
"normalized": true,
|
| 16 |
+
"rstrip": false,
|
| 17 |
+
"single_word": false
|
| 18 |
+
},
|
| 19 |
+
"eos_token": {
|
| 20 |
+
"__type": "AddedToken",
|
| 21 |
+
"content": "</s>",
|
| 22 |
+
"lstrip": false,
|
| 23 |
+
"normalized": true,
|
| 24 |
+
"rstrip": false,
|
| 25 |
+
"single_word": false
|
| 26 |
+
},
|
| 27 |
+
"errors": "replace",
|
| 28 |
+
"mask_token": {
|
| 29 |
+
"__type": "AddedToken",
|
| 30 |
+
"content": "<mask>",
|
| 31 |
+
"lstrip": true,
|
| 32 |
+
"normalized": false,
|
| 33 |
+
"rstrip": false,
|
| 34 |
+
"single_word": false
|
| 35 |
+
},
|
| 36 |
+
"model_max_length": 512,
|
| 37 |
+
"name_or_path": "final_models/transformer_base_final_2",
|
| 38 |
+
"pad_token": {
|
| 39 |
+
"__type": "AddedToken",
|
| 40 |
+
"content": "<pad>",
|
| 41 |
+
"lstrip": false,
|
| 42 |
+
"normalized": true,
|
| 43 |
+
"rstrip": false,
|
| 44 |
+
"single_word": false
|
| 45 |
+
},
|
| 46 |
+
"sep_token": {
|
| 47 |
+
"__type": "AddedToken",
|
| 48 |
+
"content": "</s>",
|
| 49 |
+
"lstrip": false,
|
| 50 |
+
"normalized": true,
|
| 51 |
+
"rstrip": false,
|
| 52 |
+
"single_word": false
|
| 53 |
+
},
|
| 54 |
+
"special_tokens_map_file": null,
|
| 55 |
+
"tokenizer_class": "RobertaTokenizer",
|
| 56 |
+
"trim_offsets": true,
|
| 57 |
+
"unk_token": {
|
| 58 |
+
"__type": "AddedToken",
|
| 59 |
+
"content": "<unk>",
|
| 60 |
+
"lstrip": false,
|
| 61 |
+
"normalized": true,
|
| 62 |
+
"rstrip": false,
|
| 63 |
+
"single_word": false
|
| 64 |
+
}
|
| 65 |
+
}
|
finetune/control_raising_control/checkpoint-400/trainer_state.json
ADDED
|
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"best_metric": 0.8797792137856758,
|
| 3 |
+
"best_model_checkpoint": "final_models/transformer_base_final_2/finetune/control_raising_control/checkpoint-400",
|
| 4 |
+
"epoch": 7.2727272727272725,
|
| 5 |
+
"global_step": 400,
|
| 6 |
+
"is_hyper_param_search": false,
|
| 7 |
+
"is_local_process_zero": true,
|
| 8 |
+
"is_world_process_zero": true,
|
| 9 |
+
"log_history": [
|
| 10 |
+
{
|
| 11 |
+
"epoch": 7.27,
|
| 12 |
+
"eval_accuracy": 0.8665371537208557,
|
| 13 |
+
"eval_f1": 0.8797792137856758,
|
| 14 |
+
"eval_loss": 0.8952956199645996,
|
| 15 |
+
"eval_mcc": 0.7560620169097876,
|
| 16 |
+
"eval_runtime": 18.3898,
|
| 17 |
+
"eval_samples_per_second": 727.687,
|
| 18 |
+
"eval_steps_per_second": 90.974,
|
| 19 |
+
"step": 400
|
| 20 |
+
}
|
| 21 |
+
],
|
| 22 |
+
"max_steps": 550,
|
| 23 |
+
"num_train_epochs": 10,
|
| 24 |
+
"total_flos": 3144693820953600.0,
|
| 25 |
+
"trial_name": null,
|
| 26 |
+
"trial_params": null
|
| 27 |
+
}
|