Spaces:
Runtime error
Runtime error
Create model.py
Browse files
model.py
ADDED
@@ -0,0 +1,252 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gc
|
2 |
+
|
3 |
+
import numpy as np
|
4 |
+
import pandas as pd
|
5 |
+
from tqdm.notebook import tqdm, trange
|
6 |
+
|
7 |
+
import torch
|
8 |
+
from torch import nn
|
9 |
+
import transformers
|
10 |
+
from transformers import AutoModel, AutoTokenizer, AutoConfig
|
11 |
+
|
12 |
+
|
13 |
+
config = dict(
|
14 |
+
# basic
|
15 |
+
seed = 3407,
|
16 |
+
num_jobs=1,
|
17 |
+
num_labels=2,
|
18 |
+
|
19 |
+
# model info
|
20 |
+
tokenizer_path = 'allenai/biomed_roberta_base', # 'roberta-base',
|
21 |
+
model_checkpoint = '../input/biomed-roberta', # 'roberta-base',
|
22 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu',
|
23 |
+
|
24 |
+
# training paramters
|
25 |
+
max_length = 512,
|
26 |
+
batch_size=16,
|
27 |
+
|
28 |
+
# for this notebook
|
29 |
+
debug = False,
|
30 |
+
)
|
31 |
+
|
32 |
+
|
33 |
+
def create_sample_test():
|
34 |
+
feats = pd.read_csv(f"../input/nbme-score-clinical-patient-notes/features.csv")
|
35 |
+
feats.loc[27, 'feature_text'] = "Last-Pap-smear-1-year-ago"
|
36 |
+
|
37 |
+
notes = pd.read_csv(f"../input/nbme-score-clinical-patient-notes/patient_notes.csv")
|
38 |
+
test = pd.read_csv(f"../input/nbme-score-clinical-patient-notes/test.csv")
|
39 |
+
|
40 |
+
merged = test.merge(notes, how = "left")
|
41 |
+
merged = merged.merge(feats, how = "left")
|
42 |
+
|
43 |
+
def process_feature_text(text):
|
44 |
+
return text.replace("-OR-", ";-").replace("-", " ")
|
45 |
+
merged["feature_text"] = [process_feature_text(x) for x in merged["feature_text"]]
|
46 |
+
|
47 |
+
return merged.sample(1).reset_index(drop=True)
|
48 |
+
|
49 |
+
class NBMETestData(torch.utils.data.Dataset):
|
50 |
+
def __init__(self, feature_text, pn_history, tokenizer):
|
51 |
+
self.feature_text = feature_text
|
52 |
+
self.pn_history = pn_history
|
53 |
+
self.tokenizer = tokenizer
|
54 |
+
|
55 |
+
def __len__(self):
|
56 |
+
return len(self.feature_text)
|
57 |
+
|
58 |
+
def __getitem__(self, idx):
|
59 |
+
tokenized = self.tokenizer(
|
60 |
+
self.feature_text[idx],
|
61 |
+
self.pn_history[idx],
|
62 |
+
truncation = "only_second",
|
63 |
+
max_length = config['max_length'],
|
64 |
+
padding = "max_length",
|
65 |
+
return_offsets_mapping = True
|
66 |
+
)
|
67 |
+
tokenized["sequence_ids"] = tokenized.sequence_ids()
|
68 |
+
|
69 |
+
input_ids = np.array(tokenized["input_ids"])
|
70 |
+
attention_mask = np.array(tokenized["attention_mask"])
|
71 |
+
offset_mapping = np.array(tokenized["offset_mapping"])
|
72 |
+
sequence_ids = np.array(tokenized["sequence_ids"]).astype("float16")
|
73 |
+
|
74 |
+
return {
|
75 |
+
'input_ids': input_ids,
|
76 |
+
'attention_mask': attention_mask,
|
77 |
+
'offset_mapping': offset_mapping,
|
78 |
+
'sequence_ids': sequence_ids,
|
79 |
+
}
|
80 |
+
|
81 |
+
class NBMEModel(nn.Module):
|
82 |
+
def __init__(self, num_labels=1, path=None):
|
83 |
+
super().__init__()
|
84 |
+
|
85 |
+
layer_norm_eps: float = 1e-6
|
86 |
+
|
87 |
+
self.path = path
|
88 |
+
self.num_labels = num_labels
|
89 |
+
self.config = transformers.AutoConfig.from_pretrained(config['model_checkpoint'])
|
90 |
+
|
91 |
+
self.config.update(
|
92 |
+
{
|
93 |
+
"layer_norm_eps": layer_norm_eps,
|
94 |
+
}
|
95 |
+
)
|
96 |
+
self.transformer = transformers.AutoModel.from_pretrained(config['model_checkpoint'], config=self.config)
|
97 |
+
self.dropout = nn.Dropout(0.2)
|
98 |
+
self.output = nn.Linear(self.config.hidden_size, 1)
|
99 |
+
|
100 |
+
if self.path is not None:
|
101 |
+
self.load_state_dict(torch.load(self.path)['model'])
|
102 |
+
|
103 |
+
def forward(self, data):
|
104 |
+
|
105 |
+
ids = data['input_ids']
|
106 |
+
mask = data['attention_mask']
|
107 |
+
try:
|
108 |
+
target = data['targets']
|
109 |
+
except:
|
110 |
+
target = None
|
111 |
+
|
112 |
+
transformer_out = self.transformer(ids, mask)
|
113 |
+
sequence_output = transformer_out[0]
|
114 |
+
sequence_output = self.dropout(sequence_output)
|
115 |
+
logits = self.output(sequence_output)
|
116 |
+
|
117 |
+
ret = {
|
118 |
+
"logits": torch.sigmoid(logits),
|
119 |
+
}
|
120 |
+
|
121 |
+
if target is not None:
|
122 |
+
loss = self.get_loss(logits, target)
|
123 |
+
ret['loss'] = loss
|
124 |
+
ret['targets'] = target
|
125 |
+
|
126 |
+
return ret
|
127 |
+
|
128 |
+
|
129 |
+
def get_optimizer(self, learning_rate, weigth_decay):
|
130 |
+
optimizer = torch.optim.AdamW(
|
131 |
+
self.parameters(),
|
132 |
+
lr=learning_rate,
|
133 |
+
weight_decay=weigth_decay,
|
134 |
+
)
|
135 |
+
if self.path is not None:
|
136 |
+
optimizer.load_state_dict(torch.load(self.path)['optimizer'])
|
137 |
+
|
138 |
+
return optimizer
|
139 |
+
|
140 |
+
def get_scheduler(self, optimizer, num_warmup_steps, num_training_steps):
|
141 |
+
scheduler = transformers.get_linear_schedule_with_warmup(
|
142 |
+
optimizer,
|
143 |
+
num_warmup_steps=num_warmup_steps,
|
144 |
+
num_training_steps=num_training_steps,
|
145 |
+
)
|
146 |
+
if self.path is not None:
|
147 |
+
scheduler.load_state_dict(torch.load(self.path)['scheduler'])
|
148 |
+
|
149 |
+
return scheduler
|
150 |
+
|
151 |
+
def get_loss(self, output, target):
|
152 |
+
loss_fn = nn.BCEWithLogitsLoss(reduction="none")
|
153 |
+
loss = loss_fn(output.view(-1, 1), target.view(-1, 1))
|
154 |
+
loss = torch.masked_select(loss, target.view(-1, 1) != -100).mean()
|
155 |
+
return loss
|
156 |
+
|
157 |
+
def get_location_predictions(preds, offset_mapping, sequence_ids, test=False):
|
158 |
+
all_predictions = []
|
159 |
+
for pred, offsets, seq_ids in zip(preds, offset_mapping, sequence_ids):
|
160 |
+
start_idx = None
|
161 |
+
current_preds = []
|
162 |
+
for p, o, s_id in zip(pred, offsets, seq_ids):
|
163 |
+
if s_id is None or s_id == 0:
|
164 |
+
continue
|
165 |
+
if p > 0.5:
|
166 |
+
if start_idx is None:
|
167 |
+
start_idx = o[0]
|
168 |
+
end_idx = o[1]
|
169 |
+
elif start_idx is not None:
|
170 |
+
if test:
|
171 |
+
current_preds.append(f"{start_idx} {end_idx}")
|
172 |
+
else:
|
173 |
+
current_preds.append((start_idx, end_idx))
|
174 |
+
start_idx = None
|
175 |
+
if test:
|
176 |
+
all_predictions.append("; ".join(current_preds))
|
177 |
+
else:
|
178 |
+
all_predictions.append(current_preds)
|
179 |
+
return all_predictions
|
180 |
+
|
181 |
+
|
182 |
+
|
183 |
+
def predict_location_preds(tokenizer, model, feature_text, pn_history):
|
184 |
+
|
185 |
+
test_ds = NBMETestData(feature_text, pn_history, tokenizer)
|
186 |
+
test_dl = torch.utils.data.DataLoader(
|
187 |
+
test_ds,
|
188 |
+
batch_size=config['batch_size'],
|
189 |
+
pin_memory=True,
|
190 |
+
shuffle=False,
|
191 |
+
drop_last=False
|
192 |
+
)
|
193 |
+
|
194 |
+
all_preds = None
|
195 |
+
offsets = []
|
196 |
+
seq_ids = []
|
197 |
+
|
198 |
+
preds = []
|
199 |
+
|
200 |
+
with torch.no_grad():
|
201 |
+
for batch in tqdm(test_dl):
|
202 |
+
|
203 |
+
for k, v in batch.items():
|
204 |
+
if k not in ['offset_mapping', 'sequence_id']:
|
205 |
+
batch[k] = v.to(config['device'])
|
206 |
+
|
207 |
+
logits = model(batch)['logits']
|
208 |
+
preds.append(logits.cpu().numpy())
|
209 |
+
|
210 |
+
offset_mapping = batch['offset_mapping']
|
211 |
+
sequence_ids = batch['sequence_ids']
|
212 |
+
offsets.append(offset_mapping.cpu().numpy())
|
213 |
+
seq_ids.append(sequence_ids.cpu().numpy())
|
214 |
+
|
215 |
+
preds = np.concatenate(preds, axis=0)
|
216 |
+
if all_preds is None:
|
217 |
+
all_preds = np.array(preds).astype(np.float32)
|
218 |
+
else:
|
219 |
+
all_preds += np.array(preds).astype(np.float32)
|
220 |
+
torch.cuda.empty_cache()
|
221 |
+
|
222 |
+
all_preds = all_preds.squeeze()
|
223 |
+
|
224 |
+
offsets = np.concatenate(offsets, axis=0)
|
225 |
+
seq_ids = np.concatenate(seq_ids, axis=0)
|
226 |
+
|
227 |
+
# print(all_preds.shape, offsets.shape, seq_ids.shape)
|
228 |
+
|
229 |
+
location_preds = get_location_predictions([all_preds], offsets, seq_ids, test=False)[0]
|
230 |
+
|
231 |
+
x = []
|
232 |
+
|
233 |
+
for location in location_preds:
|
234 |
+
x.append(pn_history[0][location[0]: location[1]])
|
235 |
+
|
236 |
+
return location_preds, ', '.join(x)
|
237 |
+
|
238 |
+
def get_predictions(feature_text, pn_history):
|
239 |
+
location_preds, pred_string = predict_location_preds(tokenizer, model, [feature_text], [pn_history])
|
240 |
+
print(pred_string)
|
241 |
+
|
242 |
+
tokenizer = AutoTokenizer.from_pretrained(config['tokenizer_path'])
|
243 |
+
path = '../input/nbme-training-biomed-roberta-base/best_model_0.bin'
|
244 |
+
|
245 |
+
model = NBMEModel().to(config['device'])
|
246 |
+
model.load_state_dict(torch.load(path, map_location=torch.device(config['device']))['model'])
|
247 |
+
model.eval();
|
248 |
+
|
249 |
+
# input_text = create_sample_test()
|
250 |
+
# feature_text = input_text.feature_text[0]
|
251 |
+
# pn_history = input_text.pn_history[0]
|
252 |
+
# get_predictions(feature_text, pn_history)
|