-
Notifications
You must be signed in to change notification settings - Fork 0
/
evaluation.py
277 lines (222 loc) · 9.63 KB
/
evaluation.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
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
# Description: Evaluate the performance and plot results of pre-trained and fine-tuned models
import time
from datetime import datetime
import argparse
import matplotlib.pyplot as plt
import torch
from torchvision import datasets
# File imports
from models.vit import ViT
from models.simmim import SimMIM
from models.finetune import FineTune
from utils.utils import get_device, pretrain_transforms, finetune_transforms, AugmentedOxfordIIITPet
from utils.configs import configs
def pretrain_evaluate(full_mim_path, config, BATCH_SIZE, TEST_SIZE=1000):
'''
Save example reconstructions from the pre-trained model.
'''
# Device
device = 'cpu'
print(f'Using device: {device}')
# Dataset transforms
transform = pretrain_transforms(image_size=config['image_size'])
# Load the validation set and sample a random subset
print('Loading dataset...')
dataset = datasets.ImageNet(root='./data', split='val', transform=transform)
test_idx = torch.randperm(len(dataset))[:TEST_SIZE] # random subset indices
test_set = torch.utils.data.Subset(dataset, test_idx)
testloader = torch.utils.data.DataLoader(test_set, batch_size=BATCH_SIZE, shuffle=True, num_workers=0)
model = ViT(
image_size = config['image_size'],
patch_size = config['patch_size'],
dim = config['dim'],
depth = config['depth'],
heads = config['heads'],
mlp_dim = config['mlp_dim']
).to(device)
# Print number of parameters
n_params = sum(p.numel() for p in model.parameters())
print('Number of parameters:', n_params)
mim = SimMIM(
encoder = model,
masking_ratio = config['masking_ratio'],
).to(device)
#import model weights
mim.load_state_dict(torch.load(full_mim_path, map_location=device))
if '/' in full_mim_path: #remove folders from name
savename = full_mim_path.split('/')[-1].split('.')[0]
print('Saving reconstructions under:', savename)
for i in range(20):
test_images, _ = next(iter(testloader))
mim.plot_reconstructions(test_images.to(device), savename)
print('Finished plotting reconstructions')
def finetune_evaluate(config, WEIGHTS_PATH, BATCH_SIZE, TEST_SIZE=1000, TRAIN_SIZE=6000, save=False, show=True):
'''
Display/save an example prediction from the segmentation model
and calculate mIoU, accuracy on the test set.
'''
# Seed for reproducibility
TRAIN_SPLIT_SEED = 42
# Device
device = get_device()
print(f'Using device: {device}')
# Dataset transforms
transform = finetune_transforms(image_size=config['image_size'])
# Download Oxford-IIIT Pet Dataset from PyTorch
trainset = AugmentedOxfordIIITPet(
root='data',
split="trainval",
target_types="segmentation",
download=True,
**transform,
)
testset = AugmentedOxfordIIITPet(
root='data',
split="test",
target_types="segmentation",
download=True,
**transform,
)
generator = torch.Generator().manual_seed(TRAIN_SPLIT_SEED)
full_dataset = torch.utils.data.ConcatDataset([trainset, testset])
# Resplit full dataset into train and test sets of desired sizes
splits = [TRAIN_SIZE, TEST_SIZE, len(full_dataset) - TRAIN_SIZE - TEST_SIZE]
trainset, testset, _ = torch.utils.data.random_split(full_dataset, splits, generator=generator)
trainset, testset = list(trainset), list(testset) # load the data into RAM (delete if memory runs out)
encoder = ViT(
image_size = config['image_size'],
patch_size = config['patch_size'],
dim = config['dim'],
depth = config['depth'],
heads = config['heads'],
mlp_dim = config['mlp_dim'],
).to(device)
# Print number of parameters
n_params_enc = sum(p.numel() for p in encoder.parameters())
print('Number of parameters (encoder):', n_params_enc)
model = FineTune(
encoder = encoder,
weights_path = None
).to(device)
#load model weights
model.load_state_dict(torch.load(WEIGHTS_PATH, map_location=device))
### Plot example prediction ###
#pick image in test set
idx = torch.randint(0, len(testset), (1,)).item()
img, target = testset[idx]
img = img.unsqueeze(0).to(device)
target = target.unsqueeze(0).to(device)
img_size = img.shape[-1]
patch_size = model.patch_size
with torch.no_grad():
_, pred_patches, target_patches = model.forward(img, target)
pred_patches = pred_patches[0]
target_patches = target_patches[0]
#make empty tensor to store full image
pred_full = torch.zeros(1, img_size, img_size)
target_full = torch.zeros(1, img_size, img_size)
patch_i = 0
for row in range(img_size//patch_size):
for col in range(img_size//patch_size):
target_plot = target_patches[patch_i].cpu().numpy()
target_plot = target_plot.reshape(patch_size, patch_size)
pred_plot = pred_patches[:,patch_i].cpu().numpy()
pred_plot = pred_plot.reshape(3, patch_size, patch_size)
#take argmax to plot
pred_plot = pred_plot.argmax(axis=0)
#add to full image
target_full[0, row*patch_size:(row+1)*patch_size, col*patch_size:(col+1)*patch_size] = torch.tensor(target_plot)
pred_full[0, row*patch_size:(row+1)*patch_size, col*patch_size:(col+1)*patch_size] = torch.tensor(pred_plot)
patch_i += 1
#plot targetand prediction
fig, axs = plt.subplots(1, 3, figsize=(6, 2))
axs[0].imshow(img[0].cpu().numpy().transpose(1, 2, 0))
axs[0].set_title('Image')
axs[1].imshow(target_full[0].cpu().numpy(), cmap='gray')
axs[1].set_title('Target')
axs[2].imshow(pred_full[0].cpu().numpy(), cmap='gray')
axs[2].set_title('Prediction')
for ax in axs:
ax.axis('off')
if save:
plt.savefig('figures/finetune_example.png')
if show:
plt.show()
plt.close()
### Calculate mIoU and accuracy ###
#iterate through test set
test_mIoU = 0
test_accuracy = 0
with torch.no_grad():
for img, target in testset:
img = img.unsqueeze(0)
target = target.unsqueeze(0)
img = img.to(device)
target = target.to(device)
loss, pred_patches, target_patches = model.forward(img, target)
pred_patches = pred_patches[0]
target_patches = target_patches[0]
pred_flat = pred_patches.argmax(dim=0).view(-1).cpu()
target_flat = target_patches.view(-1).cpu()
#calculate accuracy
test_accuracy += (pred_flat == target_flat).sum().item()/len(pred_flat)
#calculate mIoU
iou = 0
for j in range(3):
intersection = ((pred_flat == j) & (target_flat == j)).sum().item()
union = ((pred_flat == j) | (target_flat == j)).sum().item()
if union != 0:
iou += intersection/union
test_mIoU += iou/3
test_mIoU /= len(testset)
test_accuracy /= len(testset)
print(f'Test mIoU: {test_mIoU:.4f}')
print(f'Test accuracy: {test_accuracy:.4f}')
#iterate through train set
train_mIoU = 0
train_accuracy = 0
with torch.no_grad():
for img, target in trainset:
img = img.unsqueeze(0)
target = target.unsqueeze(0)
img = img.to(device)
target = target.to(device)
loss, pred_patches, target_patches = model.forward(img, target)
pred_patches = pred_patches[0]
target_patches = target_patches[0]
pred_flat = pred_patches.argmax(dim=0).view(-1).cpu()
target_flat = target_patches.view(-1).cpu()
#calculate accuracy
train_accuracy += (pred_flat == target_flat).sum().item()/len(pred_flat)
#calculate mIoU
iou = 0
for j in range(3):
intersection = ((pred_flat == j) & (target_flat == j)).sum().item()
union = ((pred_flat == j) | (target_flat == j)).sum().item()
if union != 0:
iou += intersection/union
train_mIoU += iou/3
train_mIoU /= len(trainset)
train_accuracy /= len(trainset)
print(f'Train mIoU: {train_mIoU:.4f}')
print(f'Train accuracy: {train_accuracy:.4f}')
if __name__ == '__main__':
# Parse arguments
parser = argparse.ArgumentParser(description='Evaluate the performance of pre-trained and fine-tuned models.')
parser.add_argument('--model', type=str, default='ft', help='Model to evaluate (pt: pre-trained, ft: fine-tuned).')
parser.add_argument('--config', type=str, default='vit_4M_finetune', help='Configuration to use for pre-training or fine-tuning.')
parser.add_argument('--train_size', type=int, default=6000, help='Number of fine-tuning training samples to use.')
parser.add_argument('--test_size', type=int, default=1000, help='Number of fine-tuning test samples to use.')
parser.add_argument('--weights', type=str, default='weights/vit_4M_finetune', help='Path to pre-trained (MIM) or fine-tuned weights.')
parser.add_argument('--batch_size', type=int, default=64, help='Batch size for evaluation.')
# Get arguments from command line
args = parser.parse_args()
config = configs[args.config]
train_size = args.train_size
test_size = args.test_size
weights_path = args.weights
BATCH_SIZE = args.batch_size
if args.model == 'pt':
pretrain_evaluate(weights_path, config, BATCH_SIZE, TEST_SIZE=test_size)
if args.model == 'ft':
finetune_evaluate(config, weights_path, BATCH_SIZE, TEST_SIZE=test_size, TRAIN_SIZE=train_size)