forked from facebookresearch/mae
-
Notifications
You must be signed in to change notification settings - Fork 1
/
eval_new.py
188 lines (162 loc) · 6.7 KB
/
eval_new.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
import os
import torch
import argparse
import numpy as np
from tqdm import tqdm
import pytorch_lightning as pl
from util.get_dataloader import get_dataloader
from torchmetrics.functional import pairwise_cosine_similarity as cos_dist
from dataset import MAEDataset
from pl_train import LightningMAE, Encoder
import models_mae
def count_reference(dataloader, model, device, num_classes=5):
# model.eval()
print(num_classes)
reference_sum = torch.zeros((num_classes+1, 1024)).to(device)
counts = torch.zeros(num_classes+1).to(device)
# print(counts.shape)
for i, ds in tqdm(enumerate(dataloader), total=len(dataloader)):
img = torch.einsum('nhwc->nchw', ds['image']).to(device)
img_enc = model(img.float(), mask_ratio=0) #.reshape(-1, 1024)
index_labels = ds['indices_labels'].reshape(-1).to(device)
for j in torch.unique(torch.tensor(index_labels, dtype=int).clone().detach()):
indices = (index_labels == j).nonzero()
# print(j, counts.shape)
counts[j] += indices.shape[0]
new = img_enc[indices].reshape(-1, img_enc.shape[-1])
# print('class:', j, 'sum:', new.shape, 'shape:', new.sum(dim=0).shape)
reference_sum[j] += new.sum(dim=0).clone().detach()
return reference_sum, counts
def return_img_embed(dataloader, model, device):
# model.eval()
output = []
images = []
labels = []
filenames = []
for i, ds in tqdm(enumerate(dataloader), total=len(dataloader)):
img = torch.einsum('nhwc->nchw', ds['image']).to(device)
img_enc = model(img.float(), mask_ratio=0)
# img_enc = model(img.float()) #.reshape(-1)
img_enc = img_enc.reshape(10, -1, 1024)
index_labels = ds['indices_labels'] #.reshape(-1)
labels.append(index_labels)
images.append(img.detach())
output.append(img_enc.detach())
filenames.extend(ds['file_name'])
output = torch.cat(output, dim=0)
# output = torch.tensor(output)
images = torch.cat(images, dim=0)
labels = torch.cat(labels, dim=0)
# filenames = torch.cat(filenames, dim=0)
return (output, images, labels, filenames)
def main():
LEARNING_RATE = 1e-4
L1 = 1
parser = argparse.ArgumentParser(description='MAE -> Segmentation task: patch level evaluation')
parser.add_argument(
"--dataset_name",
type=str,
help='Name of dataset one want to train',
)
parser.add_argument(
'--batch_size',
default=20,
type=int,
)
parser.add_argument('--device', default='cuda')
parser.add_argument(
'--checkpoint',
default='',
help='absolute path to checkpoint to be loaded',
)
parser.add_argument(
'--intersection_threshold',
default=0.3,
type=float,
help='threshold for patch class',
)
parser.add_argument(
'--server',
type=str,
default='c9',
help='available server names: c9, go',
)
parser.add_argument('--annotation_train')
parser.add_argument('--annotation_val')
parser.add_argument('--save_evaluated_npy')
parser.add_argument('--loss_type')
args = parser.parse_args()
dataloader, dataloader_val, num_classes, dataset = get_dataloader(dataset_name=args.dataset_name, \
train_annotation_file=args.annotation_train, val_annotation_file=args.annotation_val, \
intersection_threshold=args.intersection_threshold, batch_size=args.batch_size, \
weighted=False, return_dataset=True)
if args.server == 'c9':
chkpt_dir = '/mnt/2tb/hrant/checkpoints/mae_models/mae_visualize_vit_large_ganloss.pth'
elif args.server == 'go':
chkpt_dir = './mae_visualize_vit_large_ganloss.pth'
assert args.server in ('c9', 'go'), 'Available server names are c9 and go'
arch='mae_vit_large_patch16'
model_mae = getattr(models_mae, arch)()
checkpoint = torch.load(chkpt_dir, map_location=args.device)
msg = model_mae.load_state_dict(checkpoint['model'], strict=False)
print(msg)
chkpt_dir = args.checkpoint
model_mae = Encoder(model=model_mae, num_classes=num_classes, classifier=args.loss_type)
model_mae = LightningMAE.load_from_checkpoint(chkpt_dir, model=model_mae)
model_mae.eval()
model_mae.to(args.device)
model_mae = model_mae.model_mae
print("class count:", num_classes)
for name, param in model_mae.named_parameters():
if 'head' in name:
print(name)
if args.loss_type == 'contrastive':
ref_sum, counts = count_reference(
dataloader=dataloader,
model=model_mae,
device=args.device,
num_classes=num_classes)
ref_mean = (ref_sum.T / counts).T
ref_mean = ref_mean.nan_to_num(1)
embeds, imgs, labels, filenames = return_img_embed(dataloader_val, model_mae, args.device)
ckp = args.checkpoint.split('/')[-1].split('.')[0]
print(len(filenames), filenames[0])
output = {}
output['images'] = []
print(embeds.shape)
for i in range(len(embeds)):
# print(embeds[i].shape, ref_mean.shape)
tmp_dist = cos_dist(embeds[i], ref_mean)
tmp_dct = {
'file_name': filenames[i],
'patch_labels': tmp_dist.argmax(dim=-1).to('cpu').numpy(),
'cos_dist': tmp_dist.to('cpu').numpy(),
'embedings': embeds[i].to('cpu').numpy(),
'images': imgs[i].to('cpu').numpy(),
'labels': labels[i].to('cpu').numpy()
}
output['images'].append(tmp_dct)
output['reference_mean'] = ref_mean.to('cpu').numpy()
else:
# ds = next(iter(dataloader_val))
# print(img_enc)
# print(img_enc.shape)
# print(img_enc.argmax(dim=1))
output = {}
output['images'] = []
for ds in tqdm(dataloader_val, total=len(dataloader_val)):
img = torch.einsum('nhwc->nchw', ds['image']).to(args.device)
img_enc = model_mae(img.float(), mask_ratio=0)
img_enc = img_enc.reshape(args.batch_size, -1, img_enc.shape[-1])
for batch_idx, file_name in enumerate(ds['file_name']):
tmp_dct = {
'file_name': file_name,
'patch_labels': img_enc[batch_idx].argmax(dim=-1).to('cpu').numpy(),
'images': img[batch_idx].to('cpu').numpy(),
'labels': ds['indices_labels'][batch_idx].to('cpu').numpy()
}
output['images'].append(tmp_dct)
print('saving path:', args.save_evaluated_npy)
np.save(args.save_evaluated_npy, output)
if __name__ == '__main__':
main()