-
Notifications
You must be signed in to change notification settings - Fork 1
/
cal_calibration.py
95 lines (76 loc) · 3 KB
/
cal_calibration.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
import torchxrayvision as xrv
import torchvision as tv
import os
from os.path import join as pjoin
import vis_utils
import matplotlib.pyplot as plt
import torch
from arch.lightning.xray import XRayLightningModel
from arch.lightning.base_imagenet import ImageNetLightningModel
import pandas as pd
import json
from arch.lightning.cct import CCTLightningModel
import torch.nn.functional as F
import numpy as np
from sklearn.calibration import calibration_curve
from arch.utils import output_csv
from collections import OrderedDict
from argparse import ArgumentParser, Namespace
parser = ArgumentParser(description="Fine-tune BiT-M model.")
parser.add_argument('--n_bins', type=int, default=10)
parser.add_argument('--val_batch', type=int, default=64)
parser.add_argument('--model_dir', type=str, default='./models/')
args = parser.parse_args()
# for model_dir in os.listdir(args.model_dir):
for model_dir in [
'0812_cct_none_bbox_f1_reg1e-4_grad_y_test',
'0812_cct_none_bbox_f1_reg1e-3_grad_y_test',
'0812_cct_none_bbox_f1_reg0.01_grad_y_test',
'0812_cct_none_bbox_f1_reg0.1_grad_y_test',
'0812_cct_none_bbox_f1_reg1_grad_y_test',
'0812_cct_none',
'0812_cct_mean',
'0812_cct_random',
'0812_cct_shuffle',
]:
hparams = json.load(open(pjoin(model_dir, 'hparams.json')))
# Get the best model. But sometimes there is bug
df = pd.read_csv(pjoin(model_dir, 'results.csv'))
model = eval(hparams['pl_model']).load_from_checkpoint(
pjoin(model_dir, 'epoch=%d.ckpt' % df['val_aupr'].argmax()))
model.hparams.val_batch = args.val_batch
loaders = model.val_dataloader()
model.cuda()
with torch.no_grad():
outputs = []
dataloader_idx = 0
for bi, batch in enumerate(loaders[dataloader_idx]):
x, y = batch
if isinstance(x, dict):
x = x['imgs']
x, y = x.cuda(), y.cuda()
logit = model(x)
prefix = ['val', 'test', 'nih', 'mimic', 'cheX'][dataloader_idx]
output = {
f'{prefix}_logit': logit,
f'{prefix}_y': y,
}
outputs.append(output)
logit = torch.cat([o[f'{prefix}_logit'] for o in outputs])
y = torch.cat([o[f'{prefix}_y'] for o in outputs])
y_onehot = torch.nn.functional.one_hot(y, num_classes=logit.shape[1])
prob = F.softmax(logit, dim=1)
all_y = y_onehot.reshape(-1).cpu().numpy()
all_prob = prob.reshape(-1).cpu().numpy()
fraction_of_positives, mean_predicted_value = \
calibration_curve(all_y, all_prob, n_bins=args.n_bins)
hist, bins = np.histogram(all_prob, bins=args.n_bins)
result = OrderedDict()
result['name'] = hparams['name']
result['fraction_of_positives'] = fraction_of_positives.tolist()
result['mean_predicted_value'] = mean_predicted_value.tolist()
result['hist'] = hist.tolist()
result['bins'] = bins.tolist()
result.update(hparams)
output_csv('./results/%s_calibration.tsv' % hparams['pl_model'],
result, delimiter='\t')