-
Notifications
You must be signed in to change notification settings - Fork 5
/
extract_features_fp.py
263 lines (218 loc) · 11.8 KB
/
extract_features_fp.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
"""
Script for extracting features from the raw volumetric data (without saving the patches)
This is useful when low on storage space
If feature extraction from existing patches is desired, use extract_features.py
"""
import argparse
import os
from tqdm import tqdm
import h5py
import pandas as pd
import time
import torch
import torch.nn as nn
from models.feature_extractor import get_extractor_model
from utils.exp_utils import update_config
from utils.feature_utils import extract_patch_features_raw, extract_patch_raw, extract_features_raw
from utils.heatmap_utils import initialize_img
from utils.file_utils import save_hdf5
from data.transforms import get_basic_data_transforms
from preprocess.wsi_core.img_utils import identify_image_thresholds
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
def extract_features(conf,
seg_params={'seg_level': 0, 'sthresh': 8, 'mthresh': 7, 'close': 4, 'use_otsu': False},
filter_params={'a_t': 100, 'a_h': 16, 'max_n_holes': 10},
patch_params={'use_padding': True, 'contour_fn': 'four_pt_easy'},
img_params={'clip_min': 0, 'clip_max': 1, 'black_thresh': 0},
):
if conf['patch_mode'] == '2D':
patch_size = (conf['patch_size'], ) * 2
step_size = (conf['step_size'], ) * 2
elif conf['patch_mode'] == '3D':
patch_size = (conf['patch_size_z'], conf['patch_size'], conf['patch_size'])
step_size = (conf['step_z'], conf['step_size'], conf['step_size'])
print("\n================")
print('Loading model...')
model = get_extractor_model(encoder=conf['encoder'],
mode=conf['patch_mode'],
input=patch_size)
model.load_weights(**conf['pretrained'])
model.eval()
model = model.to(device)
print(model)
channel = model.get_channel_dim()
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model)
aug_suffix = '_aug' if conf['augment_fold'] > 0 else ''
## Define subfolder name
if conf['pretrained']['load_weights']:
subfolder = conf['pretrained']['pretrained_name'] + aug_suffix
else:
subfolder = 'random' + aug_suffix
feats_h5_subdir = os.path.join(conf['extracted_dir'],
'{}_h5_patch_features'.format(conf['encoder']),
subfolder)
os.makedirs(feats_h5_subdir, exist_ok=True)
### Setting Up For-Loop for patch creation/feature extraction
df = pd.read_csv(conf['process_list'])
total = len(df)
pbar_stack = tqdm(range(total))
print(df)
print("=============================")
print("\nBeginning {} patch extraction with {} settings".format(conf['patch_mode'],
conf['data_mode']))
######################
# Feature extraction #
######################
for i in pbar_stack:
idx = df.index[i]
slide_id = df.loc[idx, 'slide_id']
slide_path = os.path.join(conf['data_dir'], slide_id)
# Check whether the slide needs to be processed
IS_PROCESSED = df.loc[idx, 'process'] == 0
TO_SKIP = df.loc[idx, 'process'] == -1
IS_FAILURE = df.loc[idx, 'process'] == -2
if TO_SKIP or IS_FAILURE or IS_PROCESSED:
if TO_SKIP:
df.loc[idx, 'status'] = 'skip'
elif IS_FAILURE:
df.loc[idx, 'status'] = 'failure'
else:
df.loc[idx, 'status'] = 'proccessed'
df.loc[idx, 'bag_size'] = 0
continue
# Load params
current_filter_params = {}
current_seg_params = {}
current_patch_params = {}
current_img_params = {}
for key in filter_params.keys():
current_filter_params.update({key: df.loc[idx, key] if key in df.columns else filter_params[key]})
for key in seg_params.keys():
current_seg_params.update({key: df.loc[idx, key] if key in df.columns else seg_params[key]})
for key in patch_params.keys():
current_patch_params.update({key: df.loc[idx, key] if key in df.columns else patch_params[key]})
for key in img_params.keys():
current_img_params.update({key: df.loc[idx, key] if key in df.columns else img_params[key]})
print('\nSubject: {}'.format(slide_id))
print('Initializing Three Dim object')
if conf['mask_dir'] is None:
mask_file = None
binarize = True
segment = True
else:
mask_file = os.path.join(conf['mask_dir'], slide_id, 'segmentation.pkl')
binarize = False
segment = False
# Load and segment image
ThreeDim_object = initialize_img(slide_path,
patch_mode=conf['patch_mode'],
binarize=binarize,
segment=segment,
seg_mask_path=mask_file,
img_params=current_img_params,
seg_params=current_seg_params,
filter_params=current_filter_params)
s = time.time()
clip_max, clip_min = identify_image_thresholds(ThreeDim_object.wsi.img,
clip_min=current_img_params['clip_min'],
clip_max=current_img_params['clip_max'],
thresh_mode=conf['thresh_mode'])
print("Took {} seconds to find min/max".format(time.time() - s))
# Patch dataset
roi_dataset = extract_patch_raw(img_obj=ThreeDim_object,
patch_size=patch_size,
step_size=step_size,
batch_size=conf['batch_size'],
data_mode=conf['data_mode'],
slice_mode=conf['slice_mode'],
clip_min=clip_min,
clip_max=clip_max
)
######################
# Feature extraction #
######################
print("\nFeature extraction started")
for aug_idx in range(conf['augment_fold'] + 1):
if aug_idx == 0:
aug_suffix = ''
else:
aug_suffix = '_aug{}'.format(aug_idx)
df.to_csv(os.path.join(conf['extracted_dir'], conf['process_list']), index=False)
feats_h5_path = os.path.join(feats_h5_subdir, slide_id + aug_suffix + '.h5')
if os.path.isfile(feats_h5_path):
print(feats_h5_path + " exists!")
try:
_ = h5py.File(feats_h5_path, "r")
df.loc[idx, 'status'] = 'processed'
continue
except OSError:
print('Error Opening %s' % feats_h5_path)
os.system('rm %s' % feats_h5_path)
df.loc[idx, 'process'] = -2
df.loc[idx, 'status'] = 'failure'
# Feature extraction
augment = False if aug_idx == 0 else True
trns = get_basic_data_transforms(augment=augment, data_mode=conf['data_mode'], invert=conf['invert'])
roi_dataset.set_transforms(trns)
print("Extracting from ", slide_id,
" for aug ", aug_idx, " with ", clip_min, clip_max,
" (aug: {})".format(augment))
asset_dict = extract_features_raw(roi_dataset,
feature_extractor=model,
batch_size=conf['batch_size'],
channel=channel,
device=device)
save_hdf5(feats_h5_path, asset_dict, mode='w') # Save extracted features/coords
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Feature Extraction')
parser.add_argument('--data_dir', type=str, help='The root project folder directory. \
We assume that for most projects, you would want your extracted features to live in the same directory as your WSIs.')
parser.add_argument('--save_dir', type=str, help='Folder to save extracted results for patching, tissue segmentation, and stitching. \
By default, we assume args.extracted_dir is a directory within args.datroot. However, passing an absolute path into args.extracted_dir (by checking \
if "/" is in args.extracted_dir) will override using args.dataroot as a root path.')
parser.add_argument('--mask_dir', type=str, help="If segmentation file already exists, use it")
parser.add_argument('--process_list', type=str,
help='name of list of images to process with parameters (.csv)')
parser.add_argument('--config', type=str, default='.',
help='Config files that contain default parameters')
parser.add_argument('--batch_size', type=int, default=50)
parser.add_argument('--patch_size', type=int)
parser.add_argument('--patch_size_z', type=int)
parser.add_argument('--step_size', type=int)
parser.add_argument('--step_z', type=int,
help='Step size along z-direction. This allows for non-isotropic step size along z direction. If None, same as step_size')
parser.add_argument('--patch_mode', choices=['2D', '3D'], type=str,
help='2D patching or 3D patching')
parser.add_argument('--clip_min', type=int)
parser.add_argument('--clip_max', type=int)
parser.add_argument('--encoder', type=str,
help='cnn feature extractor to use')
parser.add_argument('--augment_fold', default=5, type=int,
help='Number of augmentations to perform')
parser.add_argument('--data_mode', type=str,
help='The input device mode, e.g., CT, OTLS')
parser.add_argument('--thresh_mode', type=str, choices=['fixed', 'global'])
parser.add_argument('--slice_mode', type=str)
parser.add_argument('--invert', action='store_true', default=False,
help='Whether to invert intesinty or not')
args = parser.parse_args()
# Update args namespace with parameters in config file
conf = update_config(args)
print("\nPARAMETERS: ", conf)
if conf['patch_size_z'] is None:
conf['patch_size_z'] = conf['patch_size']
if conf['step_z'] is None:
conf['step_z'] = conf['step_size']
conf['extracted_dir'] = os.path.join(conf['save_dir'], 'patch_{}-{}_step_{}-{}_{}_{}_{}'.format(conf['patch_size'],
conf['patch_size_z'],
conf['step_size'],
conf['step_z'],
conf['patch_mode'],
conf['slice_mode'],
conf['thresh_mode']))
os.makedirs(conf['extracted_dir'], exist_ok=True)
if conf['patch_mode'] == '3D' and conf['batch_size'] >= 100:
print("*************************************")
print("WARNING: Make sure you are using pytorch 2.0 for large batch size of 3D inputs!")
extract_features(conf)