-
Notifications
You must be signed in to change notification settings - Fork 0
/
submission.py
204 lines (161 loc) · 6.54 KB
/
submission.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
import os
import json
import tqdm
import glob
import argparse
from pathlib import Path
from collections import defaultdict
from typing import Any, Union, Dict, Literal
from trimesh.voxel.runlength import dense_to_brle
import h5py
import pandas as pd
import cv2
import numpy as np
from numpy.typing import NDArray
import torch
import albumentations as A
from skimage import measure
from models import get_model
from utils.engine_hub import weight_and_experiment
def _stretch_8bit(band, lower_percent=0, higher_percent=98):
"""Serves `stretch_bands` by clipping extreme numbers and stretching 1 band.
Parameters
----------
band : np.ndarray
A single band (h,w)
lower_percent : integer
Lower percentile to clip from image
higher_percent : type
Higher percentile to clip from image
Returns
-------
np.ndarray
(H, W) stretched band with same dimensions as input band
"""
a = 0
b = 255
real_values = band.flatten()
# real_values = real_values[real_values > 0]
c = np.percentile(real_values, lower_percent)
d = np.percentile(real_values, higher_percent)
if (d - c) == 0:
d += 1
t = a + (band - c) * ((b - a) / (d - c))
t[t < a] = a
t[t > b] = b
return t.astype(np.uint8)
def retrieve_validation_fold(args) -> Dict[str, NDArray]:
if isinstance(args.bands, str):
args.bands = list(map(int, args.bands.split(',')))
if args.bands == [1, 2, 3]:
mean_bands = [0.406, 0.456, 0.485]
std_bands = [0.225, 0.224, 0.229]
bit8 = True
else:
mean=[1353.72692573, 1117.20229235, 1041.88472484, 946.55425487,
1199.1886645 , 2003.00679994, 2374.00844442, 2301.22043839,
732.18195008, 12.09952762, 1118.20272293, 2599.78293726]
std=[ 72.41170098, 146.47166895, 158.20546468, 217.42332058,
168.33411967, 230.56343772, 296.15066586, 307.65398036,
85.71403735, 0.8560447, 221.18654082, 329.1786173 ]
mean_bands = []
std_bands = []
for i in args.bands:
mean_bands.append(mean[i])
std_bands.append(std[i])
bit8 = False
result = defaultdict(dict)
with h5py.File(args.data_path, 'r') as fp:
for uuid, values in fp.items():
if 'eval' in args.data_path:
if values.attrs['fold'] != 0:
continue
if "pre_fire" not in values:
continue
img_pre = values['post_fire'][...].transpose(2, 0, 1)
img_post = values['pre_fire'][...].transpose(2, 0, 1)
pre = []
post = []
for band_idx in args.bands:
band_pre = img_pre[band_idx]
if bit8:
band_pre = _stretch_8bit(band_pre) / 255.
pre.append(band_pre)
band_post = img_post[band_idx]
if bit8:
band_post = _stretch_8bit(band_post) / 255.
post.append(band_post)
img_pre = np.asarray(pre)
img_post = np.asarray(post)
if args.normalize:
transform = A.Compose([A.Normalize(mean=mean_bands, std=std_bands)])
transformed = transform(image = img_pre.transpose(1, 2, 0),
post = img_post.transpose(1, 2, 0))
img_pre = transformed['image']
img_pre = img_pre.transpose(2, 0, 1)
img_post = transformed['post']
img_post = img_post.transpose(2, 0, 1)
result[uuid]['post'] = img_pre.astype(np.float32)
result[uuid]['pre'] = img_post.astype(np.float32)
return dict(result)
def compute_submission_mask(id: str, mask: NDArray):
brle = dense_to_brle(mask.astype(bool).flatten())
return {"id": id, "rle_mask": brle, "index": np.arange(len(brle))}
if __name__ == '__main__':
parser = argparse.ArgumentParser(description="Model Testing")
parser.add_argument(
"--experiment-url", type=str, required=True, help="url of the model")
parser.add_argument(
"--data-path", type=str, required=True, help="data-path")
parser.add_argument(
"--thres", type=int, required=True, help="prediction csv name")
parser.add_argument(
"--bands", default="0,1,2,3,4,5,6,7,8,9,10,11"
)
parser.add_argument("--normalize", action='store_true', help="normalize")
args = parser.parse_args()
device = torch.device("cuda:0")
dst_path, _ = weight_and_experiment(args.experiment_url, best=True)
fin = open('/'.join(dst_path.split('/')[:-1]) + '/epxeriment_config.json', 'r')
metadata = json.load(fin)
args.__dict__.update(metadata)
fin.close()
validation_fold = retrieve_validation_fold(args)
# use a list to accumulate results
result = []
# instantiate the model
model = get_model(args)
model.to(device)
weight = torch.load(dst_path)
model.load_state_dict(weight)
_ = model.eval()
out_path = f"predictions/{args.arch}-thres{args.thres}-{args.experiment_url.split('=')[-1]}"
if 'test' in args.data_path:
out_path += '-test'
if not os.path.exists(out_path):
os.makedirs(out_path)
for uuid in tqdm.tqdm(validation_fold):
input_images = validation_fold[uuid]
# perform the prediction
pre = torch.from_numpy(input_images['pre']).to(device).float().unsqueeze(0)
post = torch.from_numpy(input_images['post']).to(device).float().unsqueeze(0)
logit = model(pre, post)
predicted = torch.argmax(logit, axis=1)
predicted = predicted.data.cpu().numpy()
cv2.imwrite(f"{out_path}/{uuid}.png", predicted[0].astype(np.uint8))
np.save(f"{out_path}/{uuid}.npz", logit.data.cpu().numpy())
# contours = measure.find_contours(predicted, 0.5)
# out_mask = np.zeros(predicted.shape, dtype=np.uint8)
# for contour in contours:
# if len(contour) > 4:
# contour = [[int(x[1]), int(x[0])] for x in contour]
# poly = Polygon(contour)
# if poly.area >= args.thres:
# out_mask = cv2.fillPoly(out_mask, pts=[np.array(contour)], color=1)
# convert the prediction in RLE format
encoded_prediction = compute_submission_mask(uuid, predicted)
result.append(pd.DataFrame(encoded_prediction))
# concatenate all dataframes
submission_df = pd.concat(result)
submission_df.to_csv(f"{out_path}.csv",
index=False)