forked from NickLucche/pyramid-cnn-leaves-segmentation
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
103 lines (89 loc) · 4.53 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
import torch.nn as nn
import torch
from utils import device
from utils import parse_args
from msu_leaves_dataset import MSUDenseLeavesDataset
from torch.utils.data import DataLoader
from pyramid_network import PyramidNet
from tqdm import tqdm
import numpy as np
import matplotlib.pyplot as plt
def evaluate(net, eval_dataset):
net.eval()
print('*'*50)
with torch.no_grad():
for batch_no, (image, targets, masks) in tqdm(enumerate(eval_dataset)):
image = image.to(device)
targets = [t.to(device) for t in targets]
masks = [t.to(device) for t in masks]
predictions = net(image)
loss = net.compute_multiscale_loss(predictions, targets, masks)
# for i in range(len(predictions)):
# print(predictions[i].shape, targets[i].shape, masks[i].shape)
print('Eval Loss:', loss.item())
# pixel-wise accuracy of multiscale predictions (edges-only)
for p, t, m in zip(predictions, targets, masks):
p = (p>0.).float()
pixel_acc = (p * m) * t
acc = pixel_acc.sum() / t.sum()
print(f"Accuracy at scale ({p.shape[2]}x{p.shape[3]}) is {acc} ({pixel_acc.sum()}/{t.sum()} edge pixels)")
print('*'*50)
net.train()
if __name__ == '__main__':
args = parse_args()
torch.manual_seed(args.seed)
np.random.seed(args.seed)
# create dataloader
dataset = MSUDenseLeavesDataset(args.dataset_filepath, args.predictions_number)
dataloader = DataLoader(dataset, batch_size=24)
eval_dataloader = DataLoader(MSUDenseLeavesDataset(args.dataset_filepath[:-1] + '_eval/', args.predictions_number),
shuffle=True, batch_size=24)
# todo totally arbitrary weights
model = PyramidNet(n_layers=5, loss_weights=[torch.tensor([1.0])]*5)#, torch.tensor([1.1]), torch.tensor([1.8]),
# torch.tensor([3.2]), torch.tensor([9.0])])
if args.load_model:
model.load_state_dict(torch.load(args.load_model))
model = model.to(device)
optimizer = torch.optim.Adam(model.parameters())
# optimizer = torch.optim.SGD(model.parameters(), 0.01, momentum=0.9)
viz = args.viz_results
for epoch in range(0, args.epochs):
# samples made of image-targets-masks
for batch_no, (input_batch, targets, masks) in enumerate(dataloader):
optimizer.zero_grad()
input_batch = input_batch.to(device)
targets = [t.to(device) for t in targets]
masks = [t.to(device) for t in masks]
# print("Input shape:", input_batch.shape)
predictions = model(input_batch)
if batch_no % 10 == 0:
print('\n',predictions[-1].max().item(), predictions[-1].min().item(), predictions[-1].sum().item())
print('\n',torch.sigmoid(predictions[-1]).max().item(), torch.sigmoid(predictions[-1]).min().item(),
torch.sigmoid(predictions[-1]).sum().item())
# print(targets[0].max().item(), targets[0].min().item(), targets[0].sum().item())
# print(masks[0].max().item(), masks[0].min().item(), masks[0].sum().item())
# for i in range(len(predictions)):
# print(predictions[i].shape, targets[i].shape, masks[i].shape)
loss = model.compute_multiscale_loss(predictions, targets, masks)
loss.backward()
# print("Current Loss:", loss.item())
optimizer.step()
if batch_no % args.log_interval == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_no*24, len(dataset),
100. * batch_no * 24 / len(dataloader), loss.item()))
evaluate(model, eval_dataloader)
torch.save(model.state_dict(), args.save_path+'pyramid_net.pt')
# visualize result
if viz:
with torch.no_grad():
predictions = model(input_batch)
p = predictions[-1][10, :, :, :]
# p = (torch.nn.functional.sigmoid(p) > .5).float()
# avoid using sigmoid, it's the same thing
print(p.shape, p.max().item(), p.min().item(), p.sum().item())
p = (p > 0.).float()
p = p.squeeze().cpu().numpy().astype(np.float32)
print(p.shape, np.amax(p), np.sum(p), np.amin(p))
plt.imshow(p, cmap='Greys')
plt.show()