-
Notifications
You must be signed in to change notification settings - Fork 1
/
tsne.py
220 lines (166 loc) · 7.09 KB
/
tsne.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
import os
import torch
import numpy as np
from sklearn.manifold import TSNE
import seaborn as sns
import pandas as pd
import matplotlib.pyplot as plt
import time
from sklearn.decomposition import PCA
from datasets import data_loader
from config.m_parser import load_config, arg_parser
from models.model_utils import (model_selector, multipathway_input,
load_pretrained_model, save_checkpoint, load_checkpoint,
AverageMeter, accuracy, create_output_dirs)
from evaluate import evaluate
import pickle as pkl
from models.s3d.select_backbone import select_backbone
from coclr_utils.utils import neq_load_customized
from coclr_utils.classifier import LinearClassifier
import torch.nn.functional as F
if __name__ == '__main__':
args = arg_parser().parse_args()
cfg = load_config(args)
print('GPU:', args.gpu)
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu)
# device = torch.device('cuda:0' if torch.cuda.is_available() else "cpu")
device = torch.cuda.current_device()
print('device:', device)
cuda=True
###############
train_loader, (train_data, _) = data_loader.build_data_loader('train', cfg,
triplets=False, req_train_shuffle=False, drop_last=False)
val_loader, (val_data, _) = data_loader.build_data_loader('val', cfg,
triplets=False, val_sample=None, req_train_shuffle=False,
drop_last=False)
label_to_class_map = train_data.get_label_to_class_map()
#print(label_to_class_map)
NUM_CLASSES = 101
split = 'train'
################
feature_dir = 'tsne_proj_feature_dir'
if not os.path.exists(feature_dir):
os.makedirs(feature_dir)
embeddings_pkl = os.path.join(feature_dir, '{}_embeddings.pkl'.format(split))
idxs_pkl = os.path.join(feature_dir, '{}_idxs.pkl'.format(split))
labels_pkl = os.path.join(feature_dir, '{}_labels.pkl'.format(split))
COMPUTE_EMBEDDINGS = False
if COMPUTE_EMBEDDINGS:
if cfg.MODEL.ARCH == 'coclr':
print('getting coclr embeddings')
#model, param = select_backbone('s3d')
#feature_size = param['feature_size']
model = LinearClassifier(
network='s3d',
num_class=101,
dropout=0.9,
use_dropout=False,
use_final_bn=True,
use_l2_norm=True)
model = model.cuda(device=device)
model = torch.nn.DataParallel(model)
model_without_dp = model.module
checkpoint = torch.load(args.checkpoint_path, map_location=torch.device('cpu'))
epoch = checkpoint['epoch']
state_dict = checkpoint['state_dict']
#if args.retrieval: # if directly test on pretrained network
new_dict = {}
for k,v in state_dict.items():
k = k.replace('encoder_q.0.', 'backbone.')
new_dict[k] = v
state_dict = new_dict
try: model_without_dp.load_state_dict(state_dict)
except: neq_load_customized(model_without_dp, state_dict, verbose=True)
#print(model_without_dp)
else:
model=model_selector(cfg, projection_head=True, classifier=False)
#print(model)
#exit()
start_epoch, best_acc = load_checkpoint(model, args.checkpoint_path, classifier=False)
model = model.cuda(device=device)
print("start_epoch:{}, best_acc:{}".format(start_epoch, best_acc))
if split == 'train':
print('getting train embeddings')
embeddings, labels, idxs = evaluate(cfg, model, cuda, device,
train_loader, split='train', is_master_proc=True)
elif split == 'val':
print('getting val embeddings')
embeddings, labels, idxs = evaluate(cfg, model, cuda, device,
val_loader, split='val', is_master_proc=True)
with open(embeddings_pkl, 'wb') as handle:
torch.save(embeddings, handle, pickle_protocol=pkl.HIGHEST_PROTOCOL)
with open(labels_pkl, 'wb') as handle:
torch.save(labels, handle, pickle_protocol=pkl.HIGHEST_PROTOCOL)
with open(idxs_pkl, 'wb') as handle:
torch.save(idxs, handle, pickle_protocol=pkl.HIGHEST_PROTOCOL)
print('saved {}_embeddings'.format(split), embeddings.size(), 'labels', len(labels))
exit()
if os.path.exists(embeddings_pkl) and os.path.exists(labels_pkl) and os.path.exists(idxs_pkl):
with open(embeddings_pkl, 'rb') as handle:
embeddings = torch.load(handle)
with open(labels_pkl, 'rb') as handle:
labels = torch.load(handle)
with open(idxs_pkl, 'rb') as handle:
idxs = torch.load(handle)
labels = np.asarray(labels)
print(embeddings.shape)
print(labels.shape)
print(len(set(labels)))
#######
if cfg.MODEL.ARCH == 'coclr':
print('normalizing embeddings')
embeddings = F.normalize(embeddings, p=2, dim=1)
#####
X = embeddings
y = labels
labels = [label_to_class_map[i] for i in y]
feat_cols = [ 'pixel'+str(i) for i in range(X.shape[1]) ]
df = pd.DataFrame(X,columns=feat_cols)
df['y'] = y
#df['style'] = y%2
#df['label'] = df['y'].apply(lambda i: str(i))
#df['label'] = df['y'].apply(lambda i: label_to_class_map[i])
df['label'] = labels
X, y = None, None
print('Size of the dataframe: {}'.format(df.shape))
# For reproducability of the results
np.random.seed(42)
#rndperm = np.random.permutation(df.shape[0])
#N = 10000
#df_subset = df.loc[rndperm[:N],:].copy()
rndperm = np.random.permutation(NUM_CLASSES)
N_CLASS_SUBSET = 20
class_subset = rndperm[:N_CLASS_SUBSET]
#df = df.sample(n=10,axis='y')
df_subset = df.loc[df['y'].isin(class_subset)].copy()
#df_subset = df
data_subset = df_subset[feat_cols].values
print(df_subset.shape)
######### PCA ########
pca_50 = PCA(n_components=50)
pca_result_50 = pca_50.fit_transform(data_subset)
print('Cumulative explained variation for 50 principal components:{}'.format(np.sum(pca_50.explained_variance_ratio_)))
#####################
time_start = time.time()
#tsne = TSNE(n_components=2, verbose=1, perplexity=40, n_iter=300)
#tsne_results = tsne.fit_transform(data_subset)
tsne_results = TSNE(n_components=2, perplexity=30).fit_transform(pca_result_50)
print('t-SNE done! Time elapsed: {} seconds'.format(time.time()-time_start))
df_subset['tsne-2d-one'] = tsne_results[:,0]
df_subset['tsne-2d-two'] = tsne_results[:,1]
plt.figure(figsize=(16,10))
#plt.figure()
sns.scatterplot(
x="tsne-2d-one", y="tsne-2d-two",
hue="label",
style="label",
s=100,
palette=sns.color_palette("hls", N_CLASS_SUBSET),
#palette=sns.color_palette("husl", N_CLASS_SUBSET),
data=df_subset,
legend="brief",
alpha=1.0
)
plt.savefig('tsne_proj_{}_tight.png'.format(split), dpi=300, bbox_inches ='tight')
#plt.show()
#print(X_embedded.shape)