- Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathscript.py
150 lines (119 loc) · 6.63 KB
/
script.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
importmath
importcollections
importjson
importpprint
# import simplejson # TODO: check if needed
importtorch
fromtorchimportnn, optim
fromtorch.utils.dataimport (DataLoader,
SequentialSampler,
TensorDataset)
fromtqdmimporttqdm
fromtermcolorimportcolored
frompytorch_pretrained_bert.tokenizationimportBertTokenizer
frompytorch_pretrained_bert.modelingimportBertForQuestionAnswering, BertConfig
fromautoencoderimportEncoderRNN, DecoderRNN, train_autoencoder
fromsquadimportread_squad_examples, convert_examples_to_features, predict
frombert_explainerimportBertExplainer
device=torch.device("cuda"iftorch.cuda.is_available() else"cpu")
n_gpu=torch.cuda.device_count()
RawResult=collections.namedtuple("RawResult",
["unique_id", "start_logits", "end_logits"])
# para_file = "../Input_file.txt"
para_file="/content/drive/My Drive/train-v2.0.json"# TODO: use proper file path
model_path="/content/drive/My Drive/pytorch_model.bin"# TODO: use proper file path
tokenizer=BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True)
### Loading Pretrained model for QnA
config=BertConfig("../Results/bert_config.json")
model=BertForQuestionAnswering(config)
model.load_state_dict(torch.load(model_path, map_location='cpu'))
# model = BertForQuestionAnswering.from_pretrained('bert-base-uncased')
model.to(device)
print()
### initializing the autoencoder
hidden_size=384
encoder1=EncoderRNN(384, config.hidden_size, hidden_size).to(device)
decoder1=DecoderRNN(384, config.hidden_size, hidden_size).to(device)
encoder_optimizer=optim.Adam(encoder1.parameters())
decoder_optimizer=optim.Adam(decoder1.parameters())
criterion=nn.MSELoss()
pp=pprint.PrettyPrinter(indent=4)
# input_data is a list of dictionary which has a paragraph and questions
withopen("/content/drive/My Drive/train-v2.0.json") asf:
squad=json.load(f)
forarticleinsquad["data"]:
# input_data = []
# i = 1
forcontext_questionsinarticle["paragraphs"]:
input_data= []
i=1
paragraphs= {"id": i, "text": context_questions["context"]}
paragraphs["ques"] = [(x["question"], x["is_impossible"]) forxincontext_questions["qas"]]
input_data.append(paragraphs)
i+=1
examples=read_squad_examples(input_data)
eval_features=convert_examples_to_features(
examples=examples,
tokenizer=tokenizer,
max_seq_length=384,
doc_stride=128,
max_query_length=64)
all_input_ids=torch.tensor([f.input_idsforfineval_features], dtype=torch.long)
all_input_mask=torch.tensor([f.input_maskforfineval_features], dtype=torch.long)
all_segment_ids=torch.tensor([f.segment_idsforfineval_features], dtype=torch.long)
all_example_index=torch.arange(all_input_ids.size(0), dtype=torch.long)
pred_data=TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_example_index)
# Run prediction for full data
pred_sampler=SequentialSampler(pred_data)
pred_dataloader=DataLoader(pred_data, sampler=pred_sampler, batch_size=9)
predictions= []
forinput_ids, input_mask, segment_ids, example_indicesintqdm(pred_dataloader):
input_ids=input_ids.to(device)
input_mask=input_mask.to(device)
segment_ids=segment_ids.to(device)
# explainer = shap.DeepExplainer(model, [input_ids, segment_ids, input_mask])
withtorch.no_grad():
# tensor_output = model(input_ids, segment_ids, input_mask)
batch_start_logits, batch_end_logits=model(input_ids, segment_ids, input_mask)
# batch_start_logits, batch_end_logits = torch.split(tensor_output, int(tensor_output.shape[1]/2), dim=1)
# shap_values = explainer.shap_values([input_ids, segment_ids, input_mask])
features= []
examples_batch= []
all_results= []
print(len(examples), example_indices.max())
fori, example_indexinenumerate(example_indices):
start_logits=batch_start_logits[i].detach().cpu().tolist()
end_logits=batch_end_logits[i].detach().cpu().tolist()
feature=eval_features[example_index.item()]
unique_id=int(feature.unique_id)
features.append(feature)
examples_batch.append(examples[example_index.item()])
all_results.append(RawResult(unique_id=unique_id,
start_logits=start_logits,
end_logits=end_logits))
output_indices=predict(examples_batch, features, all_results, 30)
predictions.append(output_indices)
explainer=BertExplainer(model)
relevance, attentions, self_attentions=explainer.explain(input_ids, segment_ids, input_mask,
[o["span"] foroinoutput_indices.values()])
input_tensor=torch.stack(
[r.sum(-1).unsqueeze(-1) *explainer.layer_values_global["bert.encoder"]["input"][0] forrin
relevance], 0)
target_tensor=torch.stack(relevance, 0).sum(-1)
loss=train_autoencoder(input_tensor, target_tensor, encoder1,
decoder1, encoder_optimizer, decoder_optimizer, criterion, max_length=13)
print('Encoder loss: %.4f'%loss)
# For printing the results ####
index=None
forexampleinexamples:
ifindex!=example.example_id:
pp.pprint(example.para_text)
index=example.example_id
print('\n')
print(colored('***********Question and Answers *************', 'red'))
ques_text=colored(example.question_text+" Unanswerable: "+str(example.unanswerable), 'blue')
print(ques_text)
prediction=colored(predictions[math.floor(example.unique_id/9)][example]['text'], 'green',
attrs=['reverse', 'blink'])
print(prediction)
print('\n')
