forked from ChangwenXu98/TransPolymer
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Attention_vis.py
165 lines (129 loc) · 6.39 KB
/
Attention_vis.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
import pandas as pd
import numpy as np
import sys
import yaml
from tqdm.auto import tqdm
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
from torch.nn.utils.clip_grad import clip_grad_norm
from transformers import AdamW, get_linear_schedule_with_warmup, RobertaModel, RobertaConfig, RobertaTokenizer
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report, multilabel_confusion_matrix
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import KFold
from rdkit import Chem
import seaborn as sns
from pylab import rcParams
import matplotlib.pyplot as plt
from matplotlib import rc
from packaging import version
import torchmetrics
from torchmetrics import R2Score
from PolymerSmilesTokenization import PolymerSmilesTokenizer
from copy import deepcopy
from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter()
import pdb
from colorama import init, Fore, Style
class DownstreamRegression(nn.Module):
def __init__(self, drop_rate=0.1):
super(DownstreamRegression, self).__init__()
self.PretrainedModel = deepcopy(PretrainedModel)
self.PretrainedModel.resize_token_embeddings(len(tokenizer))
self.Regressor = nn.Sequential(
nn.Dropout(drop_rate),
nn.Linear(self.PretrainedModel.config.hidden_size, self.PretrainedModel.config.hidden_size),
nn.SiLU(),
nn.Linear(self.PretrainedModel.config.hidden_size, 1)
)
def forward(self, input_ids, attention_mask):
outputs = self.PretrainedModel(input_ids=input_ids, attention_mask=attention_mask)
logits = outputs.last_hidden_state[:, 0, :]
output = self.Regressor(logits)
return output
def main(attention_config):
if attention_config['task'] == 'pretrain':
smiles = attention_config['smiles']
else:
data = pd.read_csv(attention_config['file_path'])
smiles = data.values[attention_config['index'],0]
# pdb.set_trace()
if attention_config['add_vocab_flag']:
vocab_sup = pd.read_csv(attention_config['vocab_sup_file'], header=None).values.flatten().tolist()
tokenizer.add_tokens(vocab_sup)
encoding = tokenizer(
str(smiles),
add_special_tokens=True,
max_length=attention_config['blocksize'],
return_token_type_ids=False,
padding="max_length",
truncation=True,
return_attention_mask=True,
return_tensors='pt',
)
input_ids = encoding["input_ids"].to(device)
attention_mask = encoding["attention_mask"].to(device)
if attention_config['task'] == 'pretrain':
outputs = PretrainedModel(input_ids=input_ids, attention_mask=attention_mask, output_attentions=True)
else:
model = DownstreamRegression(drop_rate=0).to(device)
checkpoint = torch.load(attention_config['model_path'])
# model.load_state_dict(checkpoint['model'])
model = model.double()
model.eval()
with torch.no_grad():
outputs = model.PretrainedModel(input_ids=input_ids, attention_mask=attention_mask, output_attentions=True)
attention = outputs[-1]
fig, axes = plt.subplots(3,4, figsize=(attention_config['figsize_x'],attention_config['figsize_y']))
xticklabels = tokenizer.convert_ids_to_tokens(input_ids.squeeze())
if attention_config['task'] == 'pretrain':
for i in range(3):
for j in range(4):
sns.heatmap(attention[attention_config['layer']][0,4*i+j,:,:].cpu().detach().numpy(), ax = axes[i,j], xticklabels=xticklabels, yticklabels=xticklabels)
axes[i,j].set_title(label="Attention Head %s" % str(4*i+j+1), fontsize=attention_config['fontsize'])
axes[i,j].tick_params(labelsize=attention_config['labelsize'])
cbar = axes[i,j].collections[0].colorbar
# here set the labelsize by 20
cbar.ax.tick_params(labelsize=attention_config['labelsize'])
else:
# pdb.set_trace()
yticklabels = ['Layer 1','Layer 2','Layer 3','Layer 4','Layer 5','Layer 6']
for i in range(3):
for j in range(4):
for layer in range(6):
attention_sub = attention[layer][0,4*i+j,0,:].cpu().detach().numpy().reshape(1,-1)
if layer == 0:
attention_CLS = attention_sub
else:
attention_CLS = np.vstack((attention_CLS, attention_sub))
max_attention_CLS = attention_CLS.max(axis = 0)
max_attention= np.vstack(max_attention_CLS.reshape(1,max_attention_CLS.shape[0]))
# pdb.set_trace()
sns.heatmap(attention_CLS, ax = axes[i,j], xticklabels=False)
axes[i, j].set_title(label="Attention Head %s" % str(4 * i + j + 1), fontsize=attention_config['fontsize'])
axes[i, j].set_yticklabels(rotation=attention_config['rotation'], labels=yticklabels)
axes[i, j].tick_params(labelsize=attention_config['labelsize'])
cbar = axes[i, j].collections[0].colorbar
# here set the labelsize by 20
cbar.ax.tick_params(labelsize=attention_config['labelsize'])
plt.savefig(attention_config['save_path'], bbox_inches='tight')
# max_array = max_attention.max(axis = 0)
# # Define color mappings based on float ranges (you can customize these)
# def get_color(float_value):
# if float_value < 0.3:
# return Fore.RED
# elif float_value < 0.7:
# return Fore.GREEN
# else:
# return Fore.BLUE
# colorized_string = ''.join([get_color(value) + char for value, char in zip(max_array, smiles)])
# print(colorized_string)
# pdb.set_trace()
if __name__ == "__main__":
attention_config = yaml.load(open("config_attention.yaml", "r"), Loader=yaml.FullLoader)
"""Device"""
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
PretrainedModel = RobertaModel.from_pretrained(attention_config['pretrain_path']).to(device)
tokenizer = PolymerSmilesTokenizer.from_pretrained("/project/rcc/hyadav/roberta-base", max_len=attention_config['blocksize'])
main(attention_config)