-
Notifications
You must be signed in to change notification settings - Fork 1
/
evaluate.py
260 lines (215 loc) · 6.59 KB
/
evaluate.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
"""
This scripts assess predictions on given dataset. It takes CLI arguments for post-processing params.
Run python evaluate.py --help for more details.
Author: Luca Clissa <clissa@bo.infn.it>
Created: 2023-07-02
License: Apache License 2.0
"""
import sys
import inspect
from pathlib import Path
from tqdm.auto import tqdm
SCRIPT_PATH = inspect.getfile(inspect.currentframe())
FLUOCELLS_PATH = Path(SCRIPT_PATH).absolute()
sys.path.append(str(FLUOCELLS_PATH))
import os
import pandas as pd
import numpy as np
from skimage import measure
from fastai.vision.all import *
from fluocells.config import (
REPO_PATH,
DATA_PATH,
DATA_PATH_g,
DATA_PATH_y,
DATA_PATH_r,
METADATA,
MODELS_PATH,
)
from fluocells.utils.data import post_process
from fluocells.models import cResUnet, c_resunet
from fluocells.utils.metrics import eval_prediction
torch.set_printoptions(precision=10)
import argparse
parser = argparse.ArgumentParser(description="Run a basic training pipeline")
# Add the dataset argument
parser.add_argument(
"dataset",
type=str,
choices=["green", "yellow", "red"],
help="Dataset to train on: green, yellow, or red",
)
# Add the experiment argument
parser.add_argument(
"experiment",
type=str,
help="Name of the experiment folder. Needed for setup of input/output paths",
)
# Add the threshold argument
parser.add_argument(
"--bin_thresh",
type=float,
default=0.5,
help="Threshold for heatmap binarization (default: 0.5)",
)
# Add the hole_size argument
parser.add_argument(
"--smooth_disk",
type=int,
default=0,
help="Size of disk used to smoothing object contours (default: 0)",
)
# Add the hole_size argument
parser.add_argument(
"--max_hole",
type=int,
default=50,
help="Maximum hole size to fill (default: 50)",
)
# Add the min_size argument
parser.add_argument(
"--min_size",
type=int,
default=200,
help="Minimum allowed object size. Smaller objects are removed (default: 200)",
)
# Add the max_filt argument
parser.add_argument(
"--max_dist",
type=int,
default=30,
help="Max filter argument in ndimage.maximum_filter(default: 30)",
)
# Add the footprint argument
parser.add_argument(
"--fp",
type=int,
default=40,
help="Footprint argument in peak_local_maxi (default: 40)",
)
# Add the iou threshold argument
parser.add_argument(
"--iou_thresh",
type=float,
default=0.5,
help="Threshold used as IoU overlapping to determine true positives (default: 0.5)",
)
# Add the footprint argument
parser.add_argument(
"--prox_thresh",
type=int,
default=40,
help="Threshold used for centers matching to determine true positives (default: 40)",
)
# Add the device argument
parser.add_argument(
"--device",
type=str,
choices=["cpu", "cuda"],
default="cpu",
help="Device to use to get model prediction (affect dataloader and model) (default: 'cpu')",
)
def main(postproc_cfg):
BS = 1
DATASET = postproc_cfg.dataset
EXP_NAME = postproc_cfg.experiment
VAL_PCT = 0.2
dataset_path = globals()[f"DATA_PATH_{DATASET[0]}"]
# model params
N_IN, N_OUT = 16, 2
# optimizer params
LOSS_FUNC, LOSS_NAME = (
DiceLoss(axis=1, smooth=1e-06, reduction="mean", square_in_union=False),
"Dice",
)
log_path = REPO_PATH / "logs" / EXP_NAME
model_path = MODELS_PATH / EXP_NAME
trainval_path = dataset_path / "trainval" / "images"
# read train/valid/test split dataframe
trainval_fnames = [fn for fn in trainval_path.iterdir()]
# augmentation
tfms = [
IntToFloatTensor(div_mask=255.0), # need masks in [0, 1] format
]
# splitter
splitter = RandomSplitter(valid_pct=VAL_PCT)
def label_func(p):
return Path(str(p).replace("images", "ground_truths/masks"))
# dataloader
dls = SegmentationDataLoaders.from_label_func(
DATA_PATH,
fnames=trainval_fnames,
label_func=label_func,
bs=BS,
splitter=splitter,
batch_tfms=tfms,
device=postproc_cfg.device,
)
# test
test_path = dataset_path / "test" / "images"
test_fnames = [fn for fn in test_path.iterdir()]
print(f"Number of test images: {len(test_fnames)}")
test_dl = dls.test_dl(test_fnames, with_labels=True)
# learner
arch = "c-ResUnet"
# pretrained=True would load Morelli et al. 2021 weights. We add new pretrained weights after
cresunet = c_resunet(
arch=arch,
n_features_start=N_IN,
n_out=N_OUT,
pretrained=False, # this would load Morelli et al. 2022
)
learn = Learner(
dls,
model=cresunet,
loss_func=LOSS_FUNC,
metrics=[Dice(), JaccardCoeff(), foreground_acc],
path=log_path,
model_dir=model_path,
)
print(f"Loading pesi da: {model_path}")
# load training pipeline instead
learn.load(model_path / "model") # model.pth
# freeze to eval mode
learn.eval()
# compute metrics
C = 1
metrics_df = pd.DataFrame(
{}, columns="TP_iou FP_iou FN_iou TP_prox FP_prox FN_prox".split(" ")
)
for i, b in enumerate(tqdm(test_dl)):
image_name = test_dl.items[i].name
img, mask = b
heatmap = (
learn.model(img).squeeze().permute(1, 2, 0)[:, :, C].detach().to("cpu")
)
# convert to matplotlib format
img = img.squeeze().permute(1, 2, 0).to("cpu")
thresh_image = np.squeeze(
(heatmap.numpy() > postproc_cfg.bin_thresh).astype("uint8")
)
post_proc_mask = post_process(
thresh_image,
smooth_disk=postproc_cfg.smooth_disk,
max_hole_size=postproc_cfg.max_hole,
min_object_size=postproc_cfg.min_size,
max_filter_size=postproc_cfg.max_dist,
footprint=postproc_cfg.fp,
)
mask_label = measure.label(mask.squeeze())
pred_mask_label = measure.label(post_proc_mask)
TP, FP, FN = eval_prediction(
mask_label, pred_mask_label, "iou", postproc_cfg.iou_thresh
)
metrics_df.loc[image_name, "TP_iou FP_iou FN_iou".split(" ")] = TP, FP, FN
TP, FP, FN = eval_prediction(
mask_label, pred_mask_label, "proximity", postproc_cfg.prox_thresh
)
metrics_df.loc[image_name, "TP_prox FP_prox FN_prox".split(" ")] = TP, FP, FN
return metrics_df
if __name__ == "__main__":
args = parser.parse_args()
results_path = REPO_PATH / "results" / args.experiment
results_path.mkdir(exist_ok=True, parents=True)
metrics_df = main(args)
metrics_df.to_csv(results_path / "metrics.csv")