-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
279 lines (238 loc) · 9.93 KB
/
train.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
278
279
import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import transforms
from torch.utils.data import Dataset, DataLoader
import numpy as np
from PIL import Image
from scipy.interpolate import interp1d
from skimage.color import rgb2lab, lab2rgb
import os.path
import os
from model import PaletteNet
from pre_process import get_image_palette
import wandb
import click
def get_immediate_subdirectories(a_dir):
return [name for name in os.listdir(a_dir)
if os.path.isdir(os.path.join(a_dir, name))]
class NeuralPaletteDataset(Dataset):
def __init__(self, data_folder, transform):
super().__init__()
self.data_folder = data_folder
self.transform = transform
def __len__(self):
img_folder = self.data_folder + "/source"
return len(get_immediate_subdirectories(img_folder))
# The custom data loader loads image-palette pairs
# and their color augmented counterparts in compressed format
# (they are generated from raw image data by pre_process.py)
def __getitem__(self, idx):
sources_root = self.data_folder + "/source"
targets_root = self.data_folder + "/target"
source_folders = get_immediate_subdirectories(sources_root)
source_folder = os.path.join(sources_root, source_folders[idx])
folder_base = os.path.basename(source_folder)
target_folder = os.path.join(targets_root, folder_base)
angles = range(40, 360, 40)
random_selector = np.random.randint(0, len(angles) - 1)
random_angle_prefix = str(angles[random_selector]).rjust(3, '0') + "_"
source_img_path = os.path.join(source_folder, "img.npz")
target_img_path = os.path.join(target_folder, random_angle_prefix + "img.npz")
source_pal_path = os.path.join(source_folder, "pal.npz")
target_pal_path = os.path.join(target_folder, random_angle_prefix + "pal.npz")
try:
source_img = np.load(source_img_path)['arr_0']
source_pal = np.load(source_pal_path)['arr_0']
source_pal = source_pal.reshape(1, source_pal.shape[0], source_pal.shape[1])
target_img = np.load(target_img_path)['arr_0']
target_pal = np.load(target_pal_path)['arr_0']
target_pal = target_pal.reshape(1, target_pal.shape[0], target_pal.shape[1])
except:
print("Error loading training data from source path: ", source_img_path)
if(self.transform):
source_img = self.transform(normalize_to_network_input(source_img))
source_pal = self.transform(normalize_to_network_input(source_pal))
target_img = self.transform(normalize_to_network_input(target_img))
target_pal = self.transform(normalize_to_network_input(target_pal))
return source_img, source_pal, target_img, target_pal
def batch_map(batched, fun):
"""
Trivial non-parallel function application
on a batched nd-array
"""
N = batched.shape[0]
ret = np.zeros_like(batched)
for i in range(N):
ret[i] = fun(batched[i])
return ret
def save_checkpoint(model, optimizer, save_path, epoch):
torch.save({
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'epoch': epoch
}, save_path)
def normalize_to_network_input(lab):
"""
Normalize from LAB color space to [-1.,1.] range network input
The input is expected in HxWxC format
"""
range_l = [0., 100.]
range_a = [-86.185, 98.254]
range_b = [-107.863, 94.482]
range_target = [-1., 1.]
clip_l = np.clip(lab[..., 0], range_l[0], range_l[1])
clip_a = np.clip(lab[..., 1], range_a[0], range_a[1])
clip_b = np.clip(lab[..., 2], range_b[0], range_b[1])
map_l = interp1d(range_l, range_target)
map_a = interp1d(range_a, range_target)
map_b = interp1d(range_b, range_target)
# map value range of l's:
lab[..., 0] = map_l(clip_l)
# map value range of a's:
lab[..., 1] = map_a(clip_a)
# map value range of b's:
lab[..., 2] = map_b(clip_b)
return lab
def denormalize_to_lab(t):
"""
Denormalize a numpy array from [-1.,1.] range to
valid lab color space range
"""
range_l = [0., 100.]
range_a = [-86.185, 98.254]
range_b = [-107.863, 94.482]
map_l = interp1d([-1., 1.], range_l)
map_a = interp1d([-1., 1.], range_a)
map_b = interp1d([-1., 1.], range_b)
# map value range of l's:
t[:, 0, :, :] = map_l(t[:, 0, :, :])
# map value range of a's:
t[:, 1, :, :] = map_a(t[:, 1, :, :])
# map value range of b's:
t[:, 2, :, :] = map_b(t[:, 2, :, :])
return t
def to_wandb_image_grid(tensor, down_smp_factor, num_images):
"""Convert image tensor to Weights & Biases image list"""
res = (tensor).clone().detach().cpu().numpy()
res = denormalize_to_lab(res)
res = batch_map(np.transpose(res, (0, 2, 3, 1)), lab2rgb)
res = res[:num_images, ::down_smp_factor, ::down_smp_factor, :]
return [wandb.Image(img) for img in res]
def load_checkpoint(model, load_path, optimizer=None):
checkpoint = torch.load(load_path)
model.load_state_dict(checkpoint['model_state_dict'])
if optimizer:
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
epoch = checkpoint['epoch']
return model, optimizer, epoch
def match_lab_lightness(outputs, targets):
"""
Replace the L layer of LAB image in outputs
by the ones in targets
"""
outputs = outputs.permute(0, 2, 3, 1)
targets = targets.permute(0, 2, 3, 1)
outputs[..., 0] = targets[..., 0]
return outputs.permute(0, 3, 1, 2)
@click.command()
@click.option('--num_epochs', default=200, help='Number of epochs to train')
@click.option('--report_wandb', default=False, help='Use Weights & Biases for reporting')
@click.option('--wandb_entity_name', default="", help='Weights & Biases entity name')
@click.option('--restore_from_checkpoint', default=False, help='Restore training from checkpoint')
@click.option('--checkpoint_path', default="", help='Path to network checkpoint')
@click.option('--checkpoint_save_every', default="5", help='Save checkpoint N epochs')
def train(num_epochs,
report_wandb,
wandb_entity_name,
restore_from_checkpoint,
checkpoint_path,
checkpoint_save_every):
"""
The main training function
"""
runid = np.random.randint(9999999)
workers = 4
# Batch size during training
batch_size = 4
# Number of channels in training images
c_im = 3
# Palette colors
n_c = 6
# Number of channels in training samples
c = c_im + n_c * c_im
# Learning rate
learning_rate = 0.00005
# normalization and denormalization
#normalize = transforms.Normalize(MEAN.tolist(), STD.tolist())
tform = transforms.Compose([
transforms.ToTensor(),
#normalize,
])
dataset = NeuralPaletteDataset("./dataset", tform)
dataloader = DataLoader(dataset, batch_size=batch_size,
shuffle=True, num_workers=workers)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
#model = init_model(c, device)
model = PaletteNet().to(device)
loss = nn.MSELoss()
optimizer = optim.AdamW(model.parameters(), lr=learning_rate, betas=(0.9, 0.9995), weight_decay=0.01)
if restore_from_checkpoint:
load_checkpoint(model, checkpoint_path, optimizer)
print(f"Loaded checkpoint {checkpoint_path}")
iters = 0
if report_wandb:
wandb.init(project="neural-palette", entity=wandb_entity_name)
wandb.config = {
"algorithm": "neural-palette",
"learning_rate_0": learning_rate,
"epochs": num_epochs,
"batch_size": batch_size,
}
print("Starting Training Loop...")
for epoch in range(num_epochs):
loss_acc = 0.
loss_count = 0
for batch_idx, data in enumerate(dataloader, 0):
im_in, palette_in, im_aug, palette_aug = data
print(f"Training epoch {epoch + 1}, iteration {iters}...")
model.zero_grad()
# switch places of input and augmented by probability:
if np.random.rand() < 0.5:
temp = im_in
im_in = im_aug
im_aug = temp
palette_aug = palette_in
input = torch.tensor(im_in, dtype=torch.float).to(device)
im_aug = torch.tensor(im_aug, dtype=torch.float).to(device)
palette_aug = torch.tensor(palette_aug, dtype=torch.float).to(device)
output = model(input, palette_aug)
# Concatenate the LAB lightness from input
ll = input[:, 0, :, :].reshape(-1, 1, input.shape[2], input.shape[3])
output = torch.hstack((ll, output))
# Calculate loss on input batch:
err = loss(output, im_aug)
# Calculate gradients
err.backward()
optimizer.step()
iters += 1
loss_acc += err.item()
loss_count += 1
if batch_idx == 0 and report_wandb:
result_sample = to_wandb_image_grid(output, down_smp_factor=4, num_images=4)
input_sample = to_wandb_image_grid(input, down_smp_factor=4, num_images=4)
im_aug_sample = to_wandb_image_grid(im_aug, down_smp_factor=4, num_images=4)
wandb.log({"Target": im_aug_sample,
"Result": result_sample,
"Input": input_sample})
loss_acc /= loss_count
print(f"Loss for epoch {epoch}: {loss_acc}")
if report_wandb:
wandb.log({"Loss": loss_acc,
"Epoch": epoch
})
save_path = f"run_{runid}_checkpoint_epoch_{epoch+1}.pt"
if(epoch % checkpoint_save_every == 0):
save_checkpoint(model, optimizer, save_path, epoch)
if __name__ == '__main__':
train()