-
Notifications
You must be signed in to change notification settings - Fork 2
/
latent_analyse.py
executable file
·179 lines (159 loc) · 7.5 KB
/
latent_analyse.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
# 隐空间可视化
import torch
import os
import pickle
import csv
import json
from tqdm import tqdm
import random
from models.CVAE_cont.vae_framework import VAE_Framework
from models.Hidden_analyzer.style_classifier import Style_Classifier
from data_load import data_load
from config import config
if config.vis_mode == 'vis':
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
from sklearn.manifold import TSNE
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
seed = config.seed
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
random.seed(seed)
torch.backends.cudnn.deterministic = True
with open(config.vocab, 'rb') as f:
vocab = pickle.load(f)
"""
train_loader = data_load(config, 'train', './data/dataset_train.json')
# 观察整个训练集中的object类别
with torch.no_grad():
counter = {}
for step, (cap, cap_len, obj, obj_num, cap_style, cap_style_len, style_label, feat) in enumerate(
tqdm(train_loader)):
if int(style_label[0]) != 4:
continue
for noun in obj[0]:
counter[int(noun)] = counter.get(int(noun), 0) + 1
freq_words = sorted(counter, key=counter.__getitem__, reverse=True)
for k in freq_words:
print(str(vocab.idList_to_sent([1, k, 2])) + ": " + str(counter[k]))
object_categories = {0: ["man", "men"], 1: ["woman", "women"], 2: ["people", "crowd"], 3: ["boy", "boys"],
4: ["girl", "girls"], 5: ["child", "children"], 6: ["dog", "dogs"], 7: ["cat", "cats"],
8: ["water"], 9: ["street"], 10: ["beach"], 11: ["field"], 12: ["food"]}
input()
"""
if config.vis_mode == 'save':
step = config.step
model = VAE_Framework(config).to(device)
model_sc = Style_Classifier(config).to(device)
log_path = config.log_dir.format(config.id)
best_model_path = log_path + '/model/model_' + str(step) + '.pt'
best_model_sc_path = log_path + '/model/model_sc_' + str(step) + '.pt'
model.load_state_dict(torch.load(best_model_path))
model_sc.load_state_dict(torch.load(best_model_sc_path))
model.eval()
model_sc.eval()
"""
# 观察重要的维度
print(model_sc.state_dict())
weights = model_sc.state_dict()['classifier.weight']
k = 10
for i in range(5):
weight = list(weights[i])
weight = [abs(float(item)) for item in weight]
index_k = []
for j in range(k):
index_j = weight.index(max(weight))
index_k.append(index_j)
weight[index_j] = float('-inf')
print(index_k)
input()
"""
# 生成不同风格对应的隐变量(保存所有维度)
important_dims = torch.tensor(list(range(100))).to(device)
print("loading data...")
train_loader = data_load(config, 'train', './data/dataset_train_finetune.json')
model.train()
num_step = 50
data = torch.zeros(num_step*config.batch_size, 100)
label = torch.zeros(num_step*config.batch_size)
label_sp = torch.zeros(num_step*config.batch_size)
with torch.no_grad():
for step, (cap, cap_len, obj, obj_num, cap_style, cap_style_len, style_label, feat) in enumerate(tqdm(train_loader)):
# if int(style_label[0]) == 2:
if step >= num_step:
break
cap = cap.to(device)
cap_len = cap_len.to(device)
obj = obj.to(device)
obj_num = obj_num.to(device)
cap_style = cap_style.to(device)
cap_style_len = cap_style_len.to(device)
style_label = style_label.to(device)
feat = feat.to(device)
cap_len = cap_len + 2 # 开始符结束符
obj_num = obj_num + 2
cap_style_len = cap_style_len + 2
logit, mu, sigma2, mu_pri, sigma2_pri, latent_vec, obj_vec, res_vec, feat_vec = model(cap, cap_len, obj, obj_num, cap_style,
cap_style_len, feat, style_label)
for i in range(cap.size(0)):
if vocab.word_to_id("to") in cap_style[i].tolist() \
and vocab.word_to_id("meet") in cap_style[i].tolist() \
and vocab.word_to_id("his") in cap_style[i].tolist() \
and vocab.word_to_id("lover") in cap_style[i].tolist():
label_sp[step*config.batch_size+i] = 1
if vocab.word_to_id("with") in cap_style[i].tolist() \
and vocab.word_to_id("full") in cap_style[i].tolist() \
and vocab.word_to_id("joy") in cap_style[i].tolist():
label_sp[step*config.batch_size+i] = 2
if vocab.word_to_id("in") in cap_style[i].tolist()\
and vocab.word_to_id("love") in cap_style[i].tolist():
print(cap_style[i].tolist())
input()
label_sp[step*config.batch_size+i] = 3
if vocab.word_to_id("determined") in cap_style[i].tolist() \
and vocab.word_to_id("to") in cap_style[i].tolist() \
and vocab.word_to_id("win") in cap_style[i].tolist():
label_sp[step*config.batch_size+i] = 4
data[step*config.batch_size:step*config.batch_size+config.batch_size] = torch.index_select(latent_vec, 1, important_dims)
label[step*config.batch_size:step*config.batch_size+config.batch_size] = style_label
log_path = config.log_dir.format(config.id)
torch.save(data, log_path+'/latent_vec.pt', _use_new_zipfile_serialization=False)
torch.save(label, log_path+'/label_latent.pt', _use_new_zipfile_serialization=False)
torch.save(label_sp, log_path+'/label_latent_sp.pt', _use_new_zipfile_serialization=False)
print("save success")
elif config.vis_mode == 'vis':
# important_dims = torch.tensor([34, 68, 84, 77, 7, 85, 31, 70, 16, 76, 15])
important_dims = torch.tensor(list(range(100)))
data = torch.load('./vis/latent_vec.pt')
label = torch.load('./vis/label_latent.pt')
label_sp = torch.load('./vis/label_latent_sp.pt')
num_samples = data.size(0)
data = torch.index_select(data, 1, important_dims)
print("num of samples: "+str(num_samples))
tsne = TSNE(n_components=2, init='pca', random_state=0)
result = tsne.fit_transform(data.detach().numpy())
styles = ["ro", "fu", "pos", "neg"]
for i, style in enumerate(styles):
path = './vis/'+style+'_result.csv'
with open(path, 'w', encoding='utf-8') as f:
csv_writer = csv.writer(f)
for j, item in enumerate(result):
if label[j] == i:
csv_writer.writerow(list(item)+list([int(label_sp[j])]))
f.close()
plt.rcParams['font.sans-serif'] = ['Arial Unicode MS']
for i in tqdm(range(num_samples)):
if label[i] == 0:
s1 = plt.scatter(result[i][0], result[i][1], s=20, color='g', marker='+')
elif label[i] == 1:
s2 = plt.scatter(result[i][0], result[i][1], s=20, color='r', marker='+')
elif label[i] == 2:
s3 = plt.scatter(result[i][0], result[i][1], s=20, color='b', marker='+')
elif label[i] == 3:
s4 = plt.scatter(result[i][0], result[i][1], s=20, color='brown', marker='+')
#elif label[i] == 4:
# s5 = plt.scatter(result[i][0], result[i][1], s=15, color='pink')
plt.legend([s1, s2, s3, s4], ['romantic', 'humorous', 'positive', 'negative', 'factual'], loc='upper left') # lower right
plt.axis('off')
plt.title("latent space")
plt.show()