-
Notifications
You must be signed in to change notification settings - Fork 0
/
extract_features_fp.py
146 lines (124 loc) · 5.41 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
from cmath import exp
import torch
import torch.nn as nn
from math import floor
import os
import random
import numpy as np
import pdb,glob
import time
from datasets.dataset_h5 import Dataset_All_Bags, Whole_Slide_Bag_FP
from torch.utils.data import DataLoader
from models.resnet_custom import resnet50_baseline
import argparse
from utils.utils import print_network, collate_features
from utils.file_utils import save_hdf5
from PIL import Image
import h5py
import pandas as pd
import openslide
# os.environ["CUDA_VISIBLE_DEVICES"] = "0"
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
def compute_w_loader(file_path, output_path, wsi, model,
batch_size = 8, verbose = 0, print_every=20, pretrained=True,
custom_downsample=1, target_patch_size=-1):
"""
args:
file_path: directory of bag (.h5 file)
output_path: directory to save computed features (.h5 file)
model: pytorch model
batch_size: batch_size for computing features in batches
verbose: level of feedback
pretrained: use weights pretrained on imagenet
custom_downsample: custom defined downscale factor of image patches
target_patch_size: custom defined, rescaled image size before embedding
"""
dataset = Whole_Slide_Bag_FP(file_path=file_path, wsi=wsi, pretrained=pretrained,
custom_downsample=custom_downsample, target_patch_size=target_patch_size)
x, y = dataset[0]
kwargs = {'num_workers': 4, 'pin_memory': True} if device.type == "cuda" else {}
loader = DataLoader(dataset=dataset, batch_size=batch_size, **kwargs, collate_fn=collate_features)
if verbose > 0:
print('processing {}: total of {} batches'.format(file_path,len(loader)))
mode = 'w'
for count, (batch, coords) in enumerate(loader):
with torch.no_grad():
if count % print_every == 0:
print('batch {}/{}, {} files processed'.format(count, len(loader), count * batch_size))
batch = batch.to(device, non_blocking=True)
features = model(batch)
features = features.cpu().numpy()
asset_dict = {'features': features, 'coords': coords}
save_hdf5(output_path, asset_dict, attr_dict= None, mode=mode)
mode = 'a'
return output_path
parser = argparse.ArgumentParser(description='Feature Extraction')
parser.add_argument('--data_h5_dir', type=str, default='/data_local2/ljjdata/TCGA/CLAM_preprocessing/size_512/LUSC/patches/')
parser.add_argument('--highest_resolution', type=str, default='/data_local3/ljjdata/TCGA/clinical/LUSC/tcga_lusc_all_clean.csv')
parser.add_argument('--slide_ext', type=str, default= '.svs')
parser.add_argument('--csv_path', type=str, default='/data_local2/ljjdata/TCGA/CLAM_preprocessing/size_512/LUSC/process_list_autogen.csv')
parser.add_argument('--feat_dir', type=str, default='/data_local2/ljjdata/TCGA/CLAM_preprocessing/size_512/LUSC/feat_dir/')
parser.add_argument('--batch_size', type=int, default=256)
parser.add_argument('--no_auto_skip', default=False, action='store_true')
# parser.add_argument('--custom_downsample', type=int, default=1)
parser.add_argument('--target_patch_size', type=int, default=512)
args = parser.parse_args()
if __name__ == '__main__':
print('initializing dataset')
csv_path = args.csv_path
if csv_path is None:
raise NotImplementedError
bags_dataset = Dataset_All_Bags(csv_path)
os.makedirs(args.feat_dir, exist_ok=True)
os.makedirs(os.path.join(args.feat_dir, 'pt_files'), exist_ok=True)
os.makedirs(os.path.join(args.feat_dir, 'h5_files'), exist_ok=True)
dest_files = os.listdir(os.path.join(args.feat_dir, 'pt_files'))
print('loading model checkpoint')
model = resnet50_baseline(pretrained=True)
model = model.to(device)
# print_network(model)
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model)
model.eval()
total = len(bags_dataset)
sur_data = pd.read_csv(args.highest_resolution)
pid_list = sur_data['slide_id'].to_list()
sur_data.index = pid_list
for bag_candidate_idx in range(total):
slide_file_path = bags_dataset[bag_candidate_idx]#.split(args.slide_ext)[0]
slide_id = slide_file_path.split('/')[-1][:-4]
bag_name = slide_id+'.h5'
h5_file_path = os.path.join(args.data_h5_dir, bag_name)
if not os.path.exists(h5_file_path):
continue
# slide_file_path = os.path.join(args.data_slide_dir,'*', slide_id+args.slide_ext)
# slide_file_path = glob.glob(slide_file_path)[0]
print('\nprogress: {}/{}'.format(bag_candidate_idx, total))
# print(slide_file_path)
print(slide_id)
if not args.no_auto_skip and slide_id+'.pt' in dest_files:
print('skipped {}'.format(slide_id))
continue
try:
custom_downsample = sur_data.loc[slide_id+'.svs','highest_resolution']
except Exception as e:
continue
if custom_downsample==40:
c= 2
else:
c=1
output_path = os.path.join(args.feat_dir, 'h5_files', bag_name)
time_start = time.time()
wsi = openslide.open_slide(slide_file_path)
output_file_path = compute_w_loader(h5_file_path, output_path, wsi,
model = model, batch_size = args.batch_size, verbose = 1, print_every = 20,
custom_downsample=c, target_patch_size=args.target_patch_size)
time_elapsed = time.time() - time_start
print('\ncomputing features for {} took {} s'.format(output_file_path, time_elapsed))
file = h5py.File(output_file_path, "r")
features = file['features'][:]
print('features size: ', features.shape)
print('coordinates size: ', file['coords'].shape)
features = torch.from_numpy(features)
bag_base, _ = os.path.splitext(bag_name)
torch.save(features, os.path.join(args.feat_dir, 'pt_files', bag_base+'.pt'))