-
Notifications
You must be signed in to change notification settings - Fork 0
/
make_submission.py
148 lines (117 loc) · 5.12 KB
/
make_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
import json
import os
import tempfile
import time
from pathlib import Path
import clize
import matplotlib.pyplot as plt
import numpy as np
import torch
import xarray as xr
from dvc import repo
from git import Repo
from tqdm import tqdm
from functions import chart_cbar
from loaders import AI4ArcticChallengeTestDataset, get_variable_options
from train import setup_device, setup_model
from train_options import TRAIN_OPTIONS
def setup_dataset(train_options) -> torch.utils.data.DataLoader:
with open(train_options['path_to_env'] + 'datalists/testset.json') as file:
train_options['test_list'] = json.loads(file.read())
train_options['test_list'] = [file[17:32] + '_' + file[77:80] + '_prep.nc' for file in train_options['test_list']]
train_options['path_to_processed_data'] += '_test'
dataset = AI4ArcticChallengeTestDataset(options=train_options, files=train_options['test_list'], test=True)
dataloader = torch.utils.data.DataLoader(
dataset, batch_size=None, num_workers=train_options['num_workers_val'], shuffle=False
)
return dataloader
def make_submission(
model: torch.nn.Module,
device: torch.device,
dataloader: torch.utils.data.DataLoader,
train_options: dict,
previews_dir: Path | None = None,
) -> xr.Dataset:
upload_package = xr.Dataset()
os.makedirs('inference', exist_ok=True)
model.eval()
for x, _, masks, scene_name in tqdm(
iterable=dataloader, total=len(train_options['test_list']), colour='green', position=0
):
scene_name = scene_name[:19] # Removes the _prep.nc from the name.
torch.cuda.empty_cache()
x = x.to(device, non_blocking=True)
with torch.no_grad(), torch.cuda.amp.autocast():
output = model(x)
for chart in train_options['charts']:
output[chart] = torch.argmax(output[chart], dim=1).squeeze().cpu().numpy()
upload_package[f"{scene_name}_{chart}"] = xr.DataArray(
name=f"{scene_name}_{chart}",
data=output[chart].astype('uint8'),
dims=(f"{scene_name}_{chart}_dim0", f"{scene_name}_{chart}_dim1"),
)
if previews_dir is not None:
make_preview(output, masks, scene_name, previews_dir, train_options)
return upload_package
def make_preview(output: np.ndarray, masks: np.ndarray, scene_name: str, output_dir: Path, train_options):
fig, axs = plt.subplots(nrows=1, ncols=3, figsize=(10, 10))
for idx, chart in enumerate(train_options['charts']):
ax = axs[idx]
output[chart] = output[chart].astype(float)
output[chart][masks] = np.nan
ax.imshow(
output[chart], vmin=0, vmax=train_options['n_classes'][chart] - 2, cmap='jet', interpolation='nearest'
)
ax.set_xticks([])
ax.set_yticks([])
chart_cbar(ax=ax, n_classes=train_options['n_classes'][chart], chart=chart, cmap='jet')
plt.suptitle(f'Scene: {scene_name}', y=0.65)
plt.subplots_adjust(left=0, bottom=0, right=1, top=1, wspace=0.5, hspace=-0)
fig.savefig(output_dir / f'{scene_name}.png', format='png', dpi=128, bbox_inches='tight')
plt.close('all')
def dump_meta(output_dir: Path, model_url: str):
gitrepo = Repo()
meta = {
'model_url': model_url,
'commit': str(gitrepo.rev_parse('HEAD')),
'repo_dirty': gitrepo.is_dirty(),
'timestamp': time.time(),
}
with open(output_dir / 'metadata.json', 'w') as f:
json.dump(meta, f)
def main(
model_url: str, *, output_dir: Path = Path('exports'), no_previews: bool = False, force_cpu_device: bool = False
):
output_dir.mkdir(exist_ok=True, parents=True)
train_options = get_variable_options(TRAIN_OPTIONS)
print('Options initialised')
model, model_options = setup_model(train_options)
print('Model initialised')
device = setup_device(train_options) if not force_cpu_device else torch.device('cpu')
print('Device initialised')
with tempfile.TemporaryDirectory() as tmp_dir:
weights_file = Path(tmp_dir) / 'model_weights.pt'
repo.get_url.get_url(model_url, out=weights_file)
model.load_state_dict(torch.load(weights_file, map_location=device)['model_state_dict'])
model.to(device)
print('Model weights loaded')
dataloader = setup_dataset(train_options)
print('Dataset initialized')
if no_previews:
previews_dir = None
else:
previews_dir = output_dir / 'inference_previews'
previews_dir.mkdir(exist_ok=True, parents=True)
print('Compiling submission')
upload_package = make_submission(model, device, dataloader, train_options, previews_dir=previews_dir)
print('Submission compiled')
print('Saving submission')
compression = {'zlib': True, 'complevel': 1}
encoding = {var: compression for var in upload_package.data_vars}
nc_file_path = output_dir / 'upload_package.nc'
upload_package.to_netcdf(nc_file_path, mode='w', format='netcdf4', engine='netcdf4', encoding=encoding)
print(f'Saved to {nc_file_path}')
dump_meta(output_dir, model_url)
print('Submission creation succeeded')
if __name__ == '__main__':
clize.run(main)