/
main.py
executable file
·134 lines (111 loc) · 5.03 KB
/
main.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
#!/usr/bin/env python3
import numpy as np
import os
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torch.optim import lr_scheduler
from model import PatchNet
from torchvision.datasets import ImageFolder
from torchvision.transforms import ToTensor, Grayscale, Compose
import click
from skimage.segmentation import slic
from skimage.measure import regionprops
from scipy.misc import imresize
from PIL import Image
torch.set_num_threads(1)
@click.group()
def cli():
pass
@cli.command()
@click.option('-n', '--name', default='model', help='prefix for checkpoint file names')
@click.option('-b', '--batch-size', default=32, help='batch size')
@click.option('-e', '--epochs', default=400, help='training time')
@click.option('-l', '--lrate', default=0.001, help='initial learning rate')
@click.option('-w', '--workers', default=0, help='number of workers loading training data')
@click.option('-d', '--device', default='cpu', help='pytorch device')
@click.option('-v', '--validation', default='val', help='validation set location')
@click.argument('ground_truth', nargs=1, type=click.Path(exists=True, dir_okay=True))
def train(name, batch_size, epochs, lrate, workers, device, validation, ground_truth):
train_set = ImageFolder(ground_truth, transform=Compose([Grayscale(), ToTensor()]))
val_set = ImageFolder(validation, transform=Compose([Grayscale(), ToTensor()]))
train_data_loader = DataLoader(dataset=train_set, num_workers=workers, batch_size=batch_size, shuffle=True)
val_data_loader = DataLoader(dataset=val_set, num_workers=workers, batch_size=batch_size)
device = torch.device(device)
model = PatchNet().to(device)
criterion = nn.CrossEntropyLoss()
model.train()
optimizer = optim.Adam(model.parameters())
for epoch in range(epochs):
epoch_loss = 0
with click.progressbar(train_data_loader, label='epoch {}'.format(epoch)) as bar:
for sample in bar:
input, target = sample[0].to(device), sample[1].to(device)
optimizer.zero_grad()
o = model(input)
loss = criterion(o, target)
epoch_loss += loss.item()
loss.backward()
optimizer.step()
torch.save(model.state_dict(), '{}_{}.ckpt'.format(name, epoch))
print("epoch {} complete: avg. loss: {:.4f}".format(epoch, epoch_loss / len(train_data_loader)))
val_loss = evaluate(model, criterion, device, val_data_loader)
print("epoch {} validation loss: {:.4f}".format(epoch, val_loss))
def evaluate(model, criterion, device, data_loader):
model.eval()
val_loss = 0.0
with torch.no_grad():
for sample in data_loader:
input, target = sample[0].to(device), sample[1].to(device)
o = model(input)
val_loss += float(criterion(o, target))
model.train()
return val_loss / len(data_loader)
@cli.command()
@click.option('-m', '--model', default=None, help='model file')
@click.option('-d', '--device', default='cpu', help='pytorch device')
@click.argument('images', nargs=-1)
def pred(model, device, images):
from kraken.binarization import nlbin
m = PatchNet()
m.load_state_dict(torch.load(model))
device = torch.device(device)
m.to(device)
transform = ToTensor()
cmap = {0: (230, 25, 75, 127),
1: (60, 180, 75, 127),
2: (255, 225, 25, 127),
3: (0, 130, 200, 127)}
for img in images:
im = Image.open(img)
gray_unscaled = im.convert('L')
gray = gray_unscaled.resize((im.size[0]//8, im.size[1]//8))
sp = slic(gray, n_segments=3000)
props = regionprops(sp)
cls = np.zeros(sp.shape)
with click.progressbar(props, label='patches') as bar:
for prop in bar:
y = int(prop.centroid[0])
x = int(prop.centroid[1])
siz = 14
patch = gray.crop((x-siz, y-siz, x+siz, y+siz))
o = m.forward(transform(patch).unsqueeze(0).to(device))
# downscaled label map
cls[sp == prop.label] = o.argmax().item()
cls = np.array(Image.fromarray(cls).resize(gray_unscaled.size, resample=Image.NEAREST))
bin_im = nlbin(gray_unscaled)
bin_im = np.array(bin_im)
bin_im = 1 - (bin_im / bin_im.max())
overlay = np.zeros(bin_im.shape + (4,))
fg_labels = bin_im * cls
Image.fromarray(fg_labels.astype('uint8')).resize(im.size).save(os.path.splitext(img)[0] + '_labels.png')
for idx, val in cmap.items():
overlay[cls == idx] = val
layer = np.full(bin_im.shape, 255)
layer[fg_labels == idx] = 0
Image.fromarray(layer.astype('uint8')).resize(im.size).save(os.path.splitext(img)[0] + '_class_{}.png'.format(idx))
im = Image.alpha_composite(gray_unscaled.convert('RGBA'), Image.fromarray(overlay.astype('uint8'))).resize(im.size)
im.save(os.path.splitext(img)[0] + '_overlay.png')
if __name__ == '__main__':
cli()