-
Notifications
You must be signed in to change notification settings - Fork 0
/
finetunedXLM.py
288 lines (226 loc) · 10.7 KB
/
finetunedXLM.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
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
import json, time
import numpy as np
import pandas as pd
import re
from transformers import AutoTokenizer, AutoModel
from sklearn.preprocessing import StandardScaler
import torch
from torch.utils.data import TensorDataset, DataLoader
from torch.nn.utils.clip_grad import clip_grad_norm
from transformers import get_linear_schedule_with_warmup
from transformers import AdamW
tokenizer = AutoTokenizer.from_pretrained('xlm-roberta-base')
tokenizer_name = "xlm"
batch_size = 16
import torch
import torch.nn as nn
if torch.cuda.is_available():
device = torch.device("cuda")
print("Using GPU.")
else:
print("No GPU available, using the CPU instead.")
device = torch.device("cpu")
# Data Processing
def text_preprocessing(text):
text = re.sub(r'(@.*?)[\s]', ' ', text)
text = re.sub(r'&', '&', text)
text = re.sub(r'\s+', ' ', text).strip()
return text
def load_data(input_data, input_dir):
files_present,files_absent = 0,0
inputs1, inputs2 = list(), list()
masks1, masks2 = list(), list()
targets = list()
for i in range(len(input_data)):
row = input_data.iloc[i]
pair = row['pair_id']
y_target = row['Overall']
# y_scaler = StandardScaler()
# y_target = y_scaler.transform(y_target.reshape(-1, 1))
f1, f2 = pair.split("_")
folder1 = f1[-2:]
folder2 = f2[-2:]
try:
with open(input_dir+"/"+folder1+"/"+f1+".json","r") as f:
d1 = json.load(f)
with open(input_dir+"/"+folder2+"/"+f2+".json","r") as f:
d2 = json.load(f)
except Exception as E:
files_absent = files_absent + 1
continue
files_present = files_present+1
s1 = text_preprocessing(d1['text'])
s2 = text_preprocessing(d2['text'])
encoded_input1 = tokenizer.encode_plus(s1, return_tensors='pt',return_attention_mask=True, pad_to_max_length=True,max_length=256)
encoded_input2 = tokenizer.encode_plus(s2, return_tensors='pt',return_attention_mask=True, pad_to_max_length=True,max_length=256)
input_ids1 = encoded_input1['input_ids']
input_ids2 = encoded_input2['input_ids']
attention_masks1 = encoded_input1['attention_mask']
attention_masks2 = encoded_input2['attention_mask']
inputs1.append(input_ids1)
inputs2.append(input_ids2)
masks1.append(attention_masks1)
masks2.append(attention_masks2)
targets.append(y_target)
print(f"Files present: {files_present} and Files absent: {files_absent}. % missing = {(files_absent*100)/(files_present+files_absent)}%")
return inputs1, inputs2, masks1, masks2, targets
def create_dataloaders(inputs1, inputs2, masks1, masks2, targets, batch_size):
input_tensor1 = torch.tensor([t.numpy() for t in inputs1])
input_tensor2 = torch.tensor([t.numpy() for t in inputs2])
mask_tensor1 = torch.tensor([t.numpy() for t in masks1])
mask_tensor2 = torch.tensor([t.numpy() for t in masks2])
targets_tensor = torch.tensor(targets)
dataset = TensorDataset(input_tensor1, input_tensor2, mask_tensor1, mask_tensor2, targets_tensor)
dataloader = DataLoader(dataset, batch_size=batch_size,shuffle=True)
return dataloader
############################################################# Modeling Baselines #################################################################
#Model Architecure
#Baseline 1: mBERT
class mBERTRegressor(nn.Module):
def __init__(self, drop_rate=0.2, freeze_camembert=False):
super(mBERTRegressor, self).__init__()
D_in, D_out = 1536, 1
self.mbert = AutoModel.from_pretrained("bert-base-multilingual-cased")
self.regressor = nn.Sequential(
nn.Dropout(drop_rate),
nn.Linear(D_in, D_out))
def forward(self, input_ids1, input_ids2, attention_masks1, attention_masks2):
outputs1 = self.mbert(input_ids1, attention_masks1)
outputs2 = self.mbert(input_ids2, attention_masks2)
#outputs = torch.dot(outputs1[1],outputs2[1])
#outputs = outputs1[1]+outputs2[1]
outputs = torch.cat((outputs1[1], outputs2[1]), 1)
# last_hidden_state_cls = outputs[0][:, 0, :]
last_hidden_state_cls = outputs
final_outputs = self.regressor(last_hidden_state_cls)
return final_outputs
#Baseline 2: XLM-Roberta
class XLMRegressor(nn.Module):
def __init__(self, drop_rate=0.2, freeze_camembert=False):
super(XLMRegressor, self).__init__()
D_in, D_out = 1536, 1
self.xlm = AutoModel.from_pretrained("xlm-roberta-base")
self.regressor = nn.Sequential(
nn.Dropout(drop_rate),
nn.Linear(D_in, D_out))
def forward(self, input_ids1, input_ids2, attention_masks1, attention_masks2):
outputs1 = self.xlm(input_ids1, attention_masks1)
outputs2 = self.xlm(input_ids2, attention_masks2)
#outputs = torch.dot(outputs1[1],outputs2[1])
#outputs = outputs1[1]+outputs2[1]
outputs = torch.cat((outputs1[1], outputs2[1]), 1)
# last_hidden_state_cls = outputs[0][:, 0, :]
last_hidden_state_cls = outputs
final_outputs = self.regressor(last_hidden_state_cls)
return final_outputs
freeze = False #set to True for 3rd baseline
epochs = 10
total_steps = len(train_dataloader) * epochs
if freeze:
optimizer = AdamW(model.regressor.parameters(), lr=5e-5, eps=1e-8)
else:
optimizer = AdamW(model.parameters(), lr=5e-5, eps=1e-8)
scheduler = get_linear_schedule_with_warmup(optimizer,num_warmup_steps=0, num_training_steps=total_steps)
loss_function = nn.MSELoss()
def train(model, optimizer, scheduler, loss_function, epochs, train_dataloader, device, clip_value=2):
for epoch_i in range(epochs):
print(f"Epoch {epoch_i}")
print("------------------------------------------------------------")
best_loss = 1e10
t0_epoch, t0_batch = time.time(), time.time()
total_loss, batch_loss, batch_counts = 0, 0, 0
model.train()
for step, batch in enumerate(train_dataloader):
batch_counts +=1
batch_inputs1, batch_inputs2, batch_masks1, batch_masks2, batch_targets = tuple(b.to(device) for b in batch)
batch_inputs1 = batch_inputs1.squeeze(1)
batch_inputs2 = batch_inputs2.squeeze(1)
batch_masks1 = batch_masks1.squeeze(1)
batch_masks2 = batch_masks2.squeeze(1)
model.zero_grad()
outputs = model(batch_inputs1, batch_inputs2, batch_masks1, batch_masks2)
loss = loss_function(outputs.squeeze(), batch_targets.float().squeeze())
batch_loss += loss.item()
total_loss += loss.item()
loss.backward()
if freeze:
clip_grad_norm(model.regressor.parameters(), clip_value)
else:
clip_grad_norm(model.parameters(), clip_value)
optimizer.step()
scheduler.step()
if (step % 20 == 0 and step != 0) or (step == len(train_dataloader) - 1):
# Calculate time elapsed for 20 batches
time_elapsed = time.time() - t0_batch
# Print training results
print(f"{epoch_i + 1:^7} | {step:^7} | {batch_loss / batch_counts:^12.6f} | {'-':^10} | {'-':^9} | {time_elapsed:^9.2f}")
# Reset batch tracking variables
batch_loss, batch_counts = 0, 0
t0_batch = time.time()
# with torch.no_grad():
# del outputs
# torch.cuda.empty_cache()
# show_gpu(f'{epoch_i}: GPU memory usage after training model:')
del loss, outputs
torch.cuda.empty_cache()
torch.cuda.synchronize()
# show_gpu(f'{epoch_i}: GPU memory usage after clearing cache:')
avg_train_loss = total_loss / len(train_dataloader)
print("-"*70)
print(f"{epoch_i + 1:^7} | {'-':^7} | {avg_train_loss:^12.6f}")
print();
return model
######################################################################### Evaluation #############################################################33
def pearson(X,Y):
print(X,Y)
xm = torch.mean(X.float())
ym = torch.mean(Y.float())
print(xm, ym)
num = 0.0
deno1 = 0.0
deno2 = 0.0
deno = 0.0
n = X.shape[0]
for i in range(n):
num += (X[i]-xm)*(Y[i]-ym)
deno1 += (X[i]-xm)*(X[i]-xm)
deno2 += (Y[i]-ym)*(Y[i]-ym)
deno = deno1*deno2
deno = torch.sqrt(deno)
return (num/deno).item()
def evaluate(model, loss_function, test_dataloader, device):
model.eval()
test_loss, test_pearson, test_mse = [], [], []
for batch in test_dataloader:
batch_inputs1, batch_inputs2, batch_masks1, batch_masks2, batch_targets = tuple(b.to(device) for b in batch)
batch_inputs1 = batch_inputs1.squeeze(1)
batch_inputs2 = batch_inputs2.squeeze(1)
batch_masks1 = batch_masks1.squeeze(1)
batch_masks2 = batch_masks2.squeeze(1)
with torch.no_grad():
outputs = model(batch_inputs1, batch_inputs2, batch_masks1, batch_masks2)
loss = loss_function(outputs, batch_targets)
test_loss.append(loss.item())
preds = outputs.cpu()
targs = batch_targets.cpu()
# print(preds, targs)
pearson_score = pearson(preds, targs)
MSE = np.square(np.subtract(preds, targs)).mean()
test_pearson.append(pearson_score)
test_mse.append(MSE)
return test_loss, test_pearson, test_mse
#Loading Train and Test Dataset Files
train_data = pd.read_csv("train_v2.csv")
test_data = pd.read_csv("eval_with_result.csv")
#Train Set
train_inputs1, train_inputs2, train_masks1, train_masks2, train_targets = load_data(train_data, "articles")
train_dataloader = create_dataloaders(train_inputs1, train_inputs2, train_masks1, train_masks2, train_targets, batch_size)
#Test Set
test_inputs1, test_inputs2, test_masks1, test_masks2, test_targets = load_data(test_data, "eval_data")
test_dataloader = create_dataloaders(test_inputs1, test_inputs2, test_masks1, test_masks2, test_targets, batch_size)
model = XLMRegressor(drop_rate=0.2).to(device)
# %%time
model = train(model, optimizer, scheduler, loss_function, epochs,train_dataloader, device, clip_value=2)
test_loss, test_pearson, test_mse = evaluate(model, loss_function, test_dataloader, device)
print(f"Test Pearson Score is {np.sum(test_pearson)/len(test_pearson)}")
print(f"Test MSE Score is {np.sum(test_mse)/len(test_mse)}")