-
Notifications
You must be signed in to change notification settings - Fork 7
/
metric_analysis.py
192 lines (153 loc) · 6.94 KB
/
metric_analysis.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
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
from transformers import pipeline
from clinicgen.nli import BERTScorer
import pandas as pd
import numpy as np
import re
from tqdm import tqdm
import torch
# Set visible gpu
# import os
# os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID" # see issue #152
# os.environ["CUDA_VISIBLE_DEVICES"]=2
# Questions for QA model
QUESTIONS = [
'Is there pneumonia?',
'Is there edema?',
'Is there thorax?',
'Are there devices?',
'Is there opacity?',
'Is there atelectasis?',
'Is there cardiomegaly?',
'Is there lung lesion?',
'Is there consolidation?',
'Is there fracture?',
]
# Model for BERTScore
bert_model = 'distilbert-base-uncased'
# Read original reports
reports_df = pd.read_csv('/home/otabek.nazarov/Downloads/thesis/ifcc/labeled_reports_test.csv')
# Batch size configuration for model
samples_cnt = 1500#3854
batch_size = 50#47 # or 41
batch_count = int(samples_cnt / batch_size)
# Load QA model
device_id = 2 # -1 for cpu
qa_model = pipeline("question-answering",
model='franklu/pubmed_bert_squadv2',
framework='pt',
device=device_id)
qa_model.model.to(torch.device('cuda:2'))
QA_THRESHOLD = 0.25
# Load BERTScore model
bert_score_qa_model = BERTScorer(model_type=bert_model, batch_size=batch_size,
nthreads=2, lang='en', rescale_with_baseline=True,
penalty=False)
# bert_score_qa_model.cuda()#.model.to(torch.device('cuda:2'))
# Dictionary for final dataframe
data_dict = {
'mask_prob' : [],
'f1_full' : [],
'f1_qa' : [],
'prec_full' : [],
'prec_qa' : [],
'recall_full' : [],
'recall_qa' : [],
}
bert_scores_detailed = []
mask_reports_dict = {}
for percent in tqdm(range(0, 100, 4)):
# Mask out all the reports
orig_reports = reports_df['Report Impression'].values[:samples_cnt]
mask_reports = []
masking_prob = percent / 100
data_dict['mask_prob'].append(masking_prob)
for cur_report in orig_reports:
# Split report into list of words
words = cur_report.split()
words_array = np.array(words)
length = len(words_array)
# Mask out the words with masking_prob
mask = np.random.choice([0, 1], size=length, replace=True, p=[1-masking_prob, masking_prob]).astype(bool)
mask_vals = np.array([''] * length)
words_array[mask] = mask_vals[mask]
# Append masked report to the list
masked_report = ' '.join(words_array.tolist())
masked_report = re.sub(' +', ' ', masked_report)
mask_reports.append(masked_report)
# Save masked reports for dataframe
mask_reports_dict[f'masked_{percent}'] = mask_reports
# Turn into batches for fast processing
orig_reports = np.reshape(orig_reports, (batch_count, batch_size))
mask_reports = np.reshape(np.array(mask_reports), (batch_count, batch_size))
f1_score_means = []
f1_score_means_orig = []
prec_means = []
prec_means_orig = []
recall_means = []
recall_means_orig = []
bert_scores = []
for idx in range(batch_count):
refs_l = orig_reports[idx,:].tolist()
hypos_l = mask_reports[idx,:].tolist()
f1_scores = np.empty((len(refs_l), len(QUESTIONS)))
f1_scores.fill(np.nan)
f1_scores_orig = np.empty((len(refs_l), len(QUESTIONS)))
f1_scores_orig.fill(np.nan)
prec_scores = np.empty((len(refs_l), len(QUESTIONS)))
prec_scores.fill(np.nan)
prec_scores_orig = np.empty((len(refs_l), len(QUESTIONS)))
prec_scores_orig.fill(np.nan)
recall_scores = np.empty((len(refs_l), len(QUESTIONS)))
recall_scores.fill(np.nan)
recall_scores_orig = np.empty((len(refs_l), len(QUESTIONS)))
recall_scores_orig.fill(np.nan)
for q_idx, cur_question in enumerate(QUESTIONS):
# Copy questions for batch forwarding to the model
question_batch = [cur_question] * len(hypos_l)
# Get results from QA model
refs_cur_results = qa_model(question=question_batch, context=refs_l)
hypo_cur_results = qa_model(question=question_batch, context=hypos_l)
# Get bert scores for given answers
bert_score_refs = []
bert_score_hypo = []
for sample_idx, (cur_ref_res, cur_hypo_res) in enumerate(zip(refs_cur_results, hypo_cur_results)):
bert_score_refs.append(cur_ref_res['answer'])
bert_score_hypo.append(cur_hypo_res['answer'])
b_prec, b_recall, b_f1 = bert_score_qa_model.score(bert_score_hypo, bert_score_refs)
b_prec, b_recall, b_f1 = b_prec.numpy(), b_recall.numpy(), b_f1.numpy()
full_prec, full_recall, full_f1 = bert_score_qa_model.score(hypos_l, refs_l)
full_prec, full_recall, full_f1 = full_prec.numpy(), full_recall.numpy(), full_f1.numpy()
# Select scores for loss based on threshold
for sample_idx, (cur_ref_res, cur_hypo_res) in enumerate(zip(refs_cur_results, hypo_cur_results)):
if cur_ref_res['score'] > QA_THRESHOLD or cur_hypo_res['score'] > QA_THRESHOLD:
f1_scores[sample_idx, q_idx] = b_f1[sample_idx]
f1_scores_orig[sample_idx, q_idx] = full_f1[sample_idx]
prec_scores[sample_idx, q_idx] = b_prec[sample_idx]
prec_scores_orig[sample_idx, q_idx] = full_prec[sample_idx]
recall_scores[sample_idx, q_idx] = b_recall[sample_idx]
recall_scores_orig[sample_idx, q_idx] = full_recall[sample_idx]
bert_scores.append(np.nanmean(f1_scores, axis=0))
f1_score_means.append(np.nanmean(f1_scores))
f1_score_means_orig.append(np.nanmean(f1_scores_orig))
prec_means.append(np.nanmean(prec_scores))
prec_means_orig.append(np.nanmean(prec_scores_orig))
recall_means.append(np.nanmean(recall_scores))
recall_means_orig.append(np.nanmean(recall_scores_orig))
# Save data for final dataframe
bert_scores_detailed.append(np.array(bert_scores).mean(axis=0))
data_dict['f1_full'].append(np.array(f1_score_means_orig).mean())
data_dict['f1_qa'].append(np.array(f1_score_means).mean())
data_dict['prec_full'].append(np.array(prec_means_orig).mean())
data_dict['prec_qa'].append(np.array(prec_means).mean())
data_dict['recall_full'].append(np.array(recall_means_orig).mean())
data_dict['recall_qa'].append(np.array(recall_means).mean())
# Save metrics dataframe
save_df = pd.DataFrame(data_dict)
save_df.to_csv(f'metric_experiments_{QA_THRESHOLD}.csv', index=False)
# Save masked reports dataframe
save_df = pd.DataFrame(mask_reports_dict)
save_df.to_csv(f'masked_reports_{QA_THRESHOLD}.csv', index=False)
# Save detailed bert scores
bert_scores_np = np.array(bert_scores_detailed)
save_df = pd.DataFrame(bert_scores_np, columns=QUESTIONS)
save_df.to_csv(f'bert_scores_{QA_THRESHOLD}.csv', index=False)