Update app.py
Browse files
app.py
CHANGED
@@ -1,3 +1,130 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
transforms2 = A.Compose(
|
2 |
[
|
3 |
A.Resize(width=256, height=256),
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.optim as optim
|
4 |
+
import albumentations as A
|
5 |
+
from albumentations.pytorch import ToTensorV2
|
6 |
+
from PIL import Image
|
7 |
+
from torch.utils.data import Dataset, DataLoader
|
8 |
+
import numpy as np
|
9 |
+
import os
|
10 |
+
from tqdm import tqdm
|
11 |
+
from torchvision.utils import save_image
|
12 |
+
import gradio as gr
|
13 |
+
|
14 |
+
class cnnBlock(nn.Module):
|
15 |
+
def __init__(self, in_channels, out_channels, up_sample=False, use_act=True, **kwargs):
|
16 |
+
super().__init__()
|
17 |
+
self.cnn_block = nn.Sequential(
|
18 |
+
nn.ConvTranspose2d(in_channels=in_channels, out_channels=out_channels, **kwargs)
|
19 |
+
if up_sample else
|
20 |
+
nn.Conv2d(in_channels=in_channels, out_channels=out_channels, padding_mode="reflect", **kwargs),
|
21 |
+
nn.InstanceNorm2d(out_channels),
|
22 |
+
nn.ReLU(inplace=True) if use_act else nn.Identity()
|
23 |
+
)
|
24 |
+
|
25 |
+
def forward(self, x):
|
26 |
+
return self.cnn_block(x)
|
27 |
+
|
28 |
+
class residualBlock(nn.Module):
|
29 |
+
def __init__(self, channels):
|
30 |
+
super().__init__()
|
31 |
+
self.resBlock = nn.Sequential(
|
32 |
+
cnnBlock(channels, channels, kernel_size=3, padding=1),
|
33 |
+
cnnBlock(channels, channels, use_act=False, kernel_size=3, padding=1)
|
34 |
+
)
|
35 |
+
|
36 |
+
def forward(self, x):
|
37 |
+
return x + self.resBlock(x)
|
38 |
+
|
39 |
+
class Generator(nn.Module):
|
40 |
+
def __init__(self, img_channels=3, features=64, num_residual=9):
|
41 |
+
super().__init__()
|
42 |
+
self.initial = nn.Sequential(
|
43 |
+
nn.Conv2d(img_channels, 64, kernel_size=7, stride=1, padding=3, padding_mode="reflect"),
|
44 |
+
nn.ReLU()
|
45 |
+
)
|
46 |
+
self.downBlock = nn.ModuleList([
|
47 |
+
cnnBlock(features, features*2, kernel_size=3, stride=2, padding=1),
|
48 |
+
cnnBlock(features*2, features*4, kernel_size=3, stride=2, padding=1)
|
49 |
+
])
|
50 |
+
self.resBlock = nn.Sequential(*[residualBlock(features*4) for _ in range(num_residual)])
|
51 |
+
self.upBlock = nn.ModuleList([
|
52 |
+
cnnBlock(features*4, features*2, up_sample=True, kernel_size=3, stride=2, padding=1, output_padding=1),
|
53 |
+
cnnBlock(features*2, features, up_sample=True, kernel_size=3, stride=2, padding=1, output_padding=1),
|
54 |
+
])
|
55 |
+
self.final = nn.Conv2d(features, img_channels, kernel_size=7, stride=1, padding=3, padding_mode="reflect")
|
56 |
+
|
57 |
+
def forward(self, x):
|
58 |
+
x = self.initial(x)
|
59 |
+
for layer in self.downBlock:
|
60 |
+
x = layer(x)
|
61 |
+
x = self.resBlock(x)
|
62 |
+
for layer in self.upBlock:
|
63 |
+
x = layer(x)
|
64 |
+
x = self.final(x)
|
65 |
+
return torch.tanh(x)
|
66 |
+
|
67 |
+
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
68 |
+
TRAIN_DIR = "/kaggle/input/vangogh2photo/vangogh2photo/train"
|
69 |
+
VAL_DIR = "/kaggle/input/vangogh2photo/vangogh2photo/val"
|
70 |
+
BATCH_SIZE = 1
|
71 |
+
LEARNING_RATE = 2e-4
|
72 |
+
LAMBDA_IDENTITY = 0.0
|
73 |
+
LAMBDA_CYCLE = 10
|
74 |
+
NUM_WORKERS = 4
|
75 |
+
NUM_EPOCHS = 0
|
76 |
+
LOAD_MODEL = True
|
77 |
+
SAVE_MODEL = False
|
78 |
+
CHECKPOINT_GEN_B = "/kaggle/input/checkpoints/genB.pth.tar"
|
79 |
+
CHECKPOINT_GEN_A = "/kaggle/input/checkpoints/genA.pth.tar"
|
80 |
+
CHECKPOINT_DISC_A = "/kaggle/input/checkpoints/discA.pth.tar"
|
81 |
+
CHECKPOINT_DISC_B = "/kaggle/input/checkpoints/discB.pth.tar"
|
82 |
+
|
83 |
+
transforms = A.Compose(
|
84 |
+
[
|
85 |
+
A.Resize(width=256, height=256),
|
86 |
+
A.HorizontalFlip(p=0.5),
|
87 |
+
A.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], max_pixel_value=255),
|
88 |
+
ToTensorV2(),
|
89 |
+
],
|
90 |
+
additional_targets={"image0": "image"},
|
91 |
+
is_check_shapes=False
|
92 |
+
)
|
93 |
+
|
94 |
+
def load_checkpoint(checkpoint_file, model, optimizer, lr):
|
95 |
+
print("=> Loading checkpoint")
|
96 |
+
checkpoint = torch.load(checkpoint_file, map_location=DEVICE)
|
97 |
+
model.load_state_dict(checkpoint["state_dict"])
|
98 |
+
optimizer.load_state_dict(checkpoint["optimizer"])
|
99 |
+
|
100 |
+
# If we don't do this then it will just have learning rate of old checkpoint
|
101 |
+
# and it will lead to many hours of debugging \:
|
102 |
+
for param_group in optimizer.param_groups:
|
103 |
+
param_group["lr"] = lr
|
104 |
+
|
105 |
+
genA = Generator().to(DEVICE)
|
106 |
+
|
107 |
+
load_checkpoint(CHECKPOINT_GEN_A, genA, optim_gen, LEARNING_RATE)
|
108 |
+
|
109 |
+
def postprocess_and_show(output):
|
110 |
+
# Detach from GPU, move to CPU, and remove the batch dimension
|
111 |
+
output = output.squeeze(0).detach().cpu()
|
112 |
+
|
113 |
+
# Convert from [-1, 1] to [0, 1]
|
114 |
+
output = (output + 1) / 2.0
|
115 |
+
|
116 |
+
# Convert from tensor to NumPy array and transpose (C, H, W) to (H, W, C)
|
117 |
+
output_image = output.permute(1, 2, 0).numpy()
|
118 |
+
|
119 |
+
# Convert to a [0, 255] image (optional if you're using a visualization library)
|
120 |
+
output_image = (output_image * 255).astype(np.uint8)
|
121 |
+
|
122 |
+
# Option 2: Convert to a PIL image if you want to save or manipulate it
|
123 |
+
output_pil = Image.fromarray(output_image)
|
124 |
+
|
125 |
+
return output_pil
|
126 |
+
#plt.imshow(output_pil)
|
127 |
+
|
128 |
transforms2 = A.Compose(
|
129 |
[
|
130 |
A.Resize(width=256, height=256),
|