-
Notifications
You must be signed in to change notification settings - Fork 1
/
extract_display_feat.py
218 lines (196 loc) · 8.57 KB
/
extract_display_feat.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
import os
import numpy as np
from PIL import Image
import torch
from torch import nn
from torch.utils.data import DataLoader
import torch.nn.functional as F
from sklearn.decomposition import PCA
from sklearn.manifold import TSNE
import umap
import time
import pipeline.models.ViT as vits
from pipeline.utils import load_model, disp_sar, draw_progress_bar
from pipeline.datasets.datasets import test_data
from pipeline.datasets.load import eval_data
from pipeline.datasets.preprocessing import *
denorm = denormalization(-2.5,12.)
def disp_debug(input_data: torch.Tensor):
im = torch.squeeze(input_data[0,:,:,:]).cpu().data.numpy()
if len(im.shape) == 3 :# multi polarization image
im = im.transpose(1,2,0)
im_d = denorm(im)
print('\nbatch shape : {}'.format(input_data.shape))
print('batch - mean {} - std {}'.format(torch.mean(input_data),torch.std(input_data)))
print('im - mean {} - std {}'.format(np.mean(im),np.std(im)))
print('im denorm - mean {} - std {}'.format(np.mean(im_d),np.std(im_d)))
disp_sar(im_d,tresh=None)
def save_basic(im,fold):
im = np.abs(im)*255
assert len(im.shape)==3 and im.shape[2]==3
im = Image.fromarray(im.astype(np.uint8), 'RGB')
im.save(fold)
def enlarge_pixels(image, p):
enlarged_image = np.repeat(image, p, axis=0)
enlarged_image = np.repeat(enlarged_image, p, axis=1)
return enlarged_image
def display(features,params):
# Number of vector to reduce
nb_token = 0
for im in features:
shape = im['features'].shape
nb_token += shape[0]*shape[1]
token_sz = im['features'].shape[2]
print('Display type : {}\nData : {} vectors of size {}\n'.format(params['display_type'],nb_token,token_sz))
# Concatenating vector
x = np.empty((nb_token,token_sz))
start_id = 0
for im in features:
shape = im['features'].shape
end_id = start_id+shape[0]*shape[1]
x[start_id:end_id] = im['features'].reshape((shape[0]*shape[1],token_sz))
start_id = end_id
# Data reduction
if params['display_type'] == 'pca':
pca_dim = 3
pca = PCA(n_components=pca_dim)
x_reduced = pca.fit_transform(x)
print('Cumulative explained variation for {} principal components: {}'.format(pca_dim,np.sum(pca.explained_variance_ratio_)))
elif params['display_type'] == 'tsne':
tsne_dim = 3
time_start = time.time()
tsne = TSNE(n_components=tsne_dim, verbose=0, perplexity=40, n_iter=500)
x_reduced = tsne.fit_transform(x)
print('t-SNE done with {} components. Time elapsed: {} seconds'.format(tsne_dim,time.time()-time_start))
elif params['display_type'] == 'umap':
umap_dim = 3
umap_model = umap.UMAP(
n_neighbors=15,
min_dist=0.1,
n_components=umap_dim,
metric='euclidean' # or 'cosine'
)
x_reduced = umap_model.fit_transform(x)
elif params['display_type'] == 'pca_tsne':
pca_dim = 50
tsne_dim = 3
# PCA
pca = PCA(n_components=pca_dim)
pca_result = pca.fit_transform(x)
print('Cumulative explained variation for {} principal components: {}'.format(pca_dim,np.sum(pca.explained_variance_ratio_)))
# t-SNE
time_start = time.time()
tsne = TSNE(n_components=tsne_dim, verbose=0, perplexity=40, n_iter=500)
x_reduced = tsne.fit_transform(pca_result)
print('t-SNE done with {} components. Time elapsed: {} seconds'.format(tsne_dim,time.time()-time_start))
# Normalize the 3 components
for ch in range(x_reduced.shape[1]):
x_reduced[:,ch] = x_reduced[:,ch]/np.amax(x_reduced[:,ch])
# Display the 3 components
start_id = 0
for im in features:
shape = im['features'].shape
end_id = start_id+shape[0]*shape[1]
features_im = x_reduced[start_id:end_id]
features_im = features_im.reshape((shape[0],shape[1],3))
start_id = end_id
if params['stride'] != 'full':
features_im = enlarge_pixels(features_im,im['stride'])
save_basic(features_im,'{}/{}_{}_p{}_s{}.png'.format(params['save_fold'],im['name'],params['display_type'],im['patch_size'],im['stride']))
np.save('{}/{}_{}_p{}_s{}.npy'.format(params['save_fold'],im['name'],params['display_type'],im['patch_size'],im['stride']),features_im)
@torch.no_grad()
def extract_feature_pipeline(params):
# ============ preparing data ... ============
device = params['device']
patch_sz = params['patch_sz']
patch_drop = params['drop_rate']
data_ = eval_data(params['load_fold'])
transform = TransformEvalSar()
data = test_data(data_,transform)
feature_data = DataLoader(data,batch_size=1,shuffle=False)
# ============ building network ... ============
print("Loading model : {}\nweights path : {}\nchannel(s) : {} \ndrop rate : {} ".format(params['arch'],params['weights_path'],params['channels'],params['drop_rate']))
if params['arch'] in vits.__dict__.keys():
model = vits.__dict__[params['arch']](
patch_size = params['arch_patch_sz'],
in_chans = params['channels'])
model.set_patch_drop(patch_drop)
model.to(device)
load_model(model, params['weights_path'])
model.eval()
# ============ Choosing stride option ... ============
if params['stride'] == 'auto':
stride = patch_sz
elif type(params['stride']) == int and params['stride'] >= 1:
stride = params['stride']
else :
print('Stride parameter must be \'auto\' or a positive integer, not {}'.format(params['stride']))
# ============ extract features ... ============
nb_im = len(data)
print("Extracting features - \n number of images : {}\npatch size : {}\ stride : {}".format(nb_im,params['patch_sz'],stride))
outputs = []
count = 1
for im,name,_ in feature_data :
draw_progress_bar(count,nb_im,'Feature exctraction - {}'.format(name))
name = name[0]
b,c,h,w = im.shape
# Number of patches along each dimension
num_patches_H = h // stride
num_patches_W = w // stride
# Total length required to fit all patches exactly
total_length_H = patch_sz + (num_patches_H - 1) * stride
total_length_W = patch_sz + (num_patches_W - 1) * stride
# Padding needed: difference between required total length and original dimension
pad_h = total_length_H - h
pad_w = total_length_W - w
pad_top = pad_h // 2
pad_bottom = pad_h - pad_top
pad_left = pad_w // 2
pad_right = pad_w - pad_left
im = F.pad(im, (pad_left, pad_right, pad_top, pad_bottom), mode='reflect')
# Slicing the image in patches
im = im.unfold(2, patch_sz, stride).unfold(3, patch_sz, stride)
im = im.contiguous().view(b, c, -1, patch_sz, patch_sz)
im = im.permute(0, 2, 1, 3, 4).reshape(-1, c, patch_sz, patch_sz)
# Extracing features
patch_outputs = []
for i in range(0, im.size(0), params['batch_sz']):
output = model(im[i:i + params['batch_sz']].to(device))
patch_outputs.append(output.cpu())
output = torch.cat(patch_outputs, dim=0)
# Resizing to keep the spatial dimension
output_h = num_patches_H
output_w = num_patches_W
output = output.reshape(b, output_h, output_w, output.shape[-1])
output = torch.squeeze(output.permute(0, 1, 2, 3)) # H,W,D
output = nn.functional.normalize(output, dim=2, p=2)
output = output.data.numpy().astype(np.float32)
# Data storing
outputs.append({'name' : name,
'stride' : stride,
'patch_size' : patch_sz,
'features' : output })
count += 1
return outputs
if __name__ == '__main__':
params = {}
# Save/load folders
params['save_fold'] = '/data/display_features/results'
params['load_fold'] = '/data/display_features/eval'
# Dataset parameters
params['device'] = 'cuda'
params['patch_sz'] = 64
params['stride'] = 4 # auto to set stride = patch_size, or a positive int to set a custom stride
# Network parameters
params['weights_path'] = 'pipeline/out/encoder_1ch'
params['arch'] = 'vit_tiny'
params['arch_patch_sz'] = 8
params['channels'] = 1
params['drop_rate'] = 0.
params['batch_sz'] = 256
# Display parameters
params['display_type'] = 'umap'
assert params['channels'] in [1,4]
os.makedirs(params['save_fold'],exist_ok=True)
features = extract_feature_pipeline(params)
display(features,params)