Voxelmap/train_a.py at main · Image-X-Institute/Voxelmap · GitHub

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
import time
import os
import numpy as np
from matplotlib import pyplot as plt
from utilities import network_a, losses
import torch
from torch.utils.data import Dataset
import torch.optim as optim

im_dir = 'data/xcat/train'
expt_description = 'Network A'
filename = os.path.basename(__file__)
filename = filename[:len(filename)-3]

class SupervisedDataset(Dataset):
    def __init__(self, im_dir=None, im_size=None):
        self.im_dir = im_dir
        self.im_size = im_size
        #self.proj_list = sorted([n for n in os.listdir(self.im_dir) if n.endswith('_bin.npy')])

    def __len__(self):
        return len([n for n in os.listdir(self.im_dir) if n.endswith('_bin.npy')])
        # return len(self.proj_list)

    def __getitem__(self, idx):
        # Find target projection
        proj_list = sorted([n for n in os.listdir(self.im_dir) if n.endswith('_bin.npy')])
        target_file = proj_list[idx]
        proj_name = os.path.join(self.im_dir, format(target_file))
        target_proj = np.load(proj_name)
        target_proj = (target_proj - np.min(target_proj)) / (np.max(target_proj) - np.min(target_proj))

        # Find target DVF
        vol_num = target_file[:2]
        dvf_name = os.path.join(self.im_dir, format('sub_DVF_' + vol_num + '_mha.npy'))
        target_dvf = np.load(dvf_name)

        # Find source projection
        source_file = '06_' + target_file[3:]
        proj_name = os.path.join(self.im_dir, format(source_file))
        source_proj = np.load(proj_name)
        source_proj = (source_proj - np.min(source_proj)) / (np.max(source_proj) - np.min(source_proj))

        # Find source volume
        vol_name = os.path.join(self.im_dir, format('sub_CT_06_mha.npy'))
        source_vol = np.load(vol_name)
        source_vol = (source_vol - np.min(source_vol)) / (np.max(source_vol) - np.min(source_vol))

        # Find source abdomen
        vol_name = os.path.join(self.im_dir, format('sub_Abdomen_mha.npy'))
        source_hull = np.load(vol_name)

        # Reshape data
        source_projections = np.zeros((1, self.im_size, self.im_size), dtype=np.float32)
        source_projections[0, :, :] = np.asarray(source_proj)
        target_projections = np.zeros((1, self.im_size, self.im_size), dtype=np.float32)
        target_projections[0, :, :] = np.asarray(target_proj)

        source_volumes = np.zeros((1, self.im_size, self.im_size, self.im_size), dtype=np.float32)
        source_volumes[0, :, :, :] = np.asarray(source_vol)
        source_abdomen = np.zeros((1, self.im_size, self.im_size, self.im_size), dtype=np.float32)
        source_abdomen[0, :, :, :] = np.asarray(source_hull)

        target_flow = np.zeros((3, self.im_size, self.im_size, self.im_size), dtype=np.float32)
        target_flow[0, :, :, :] = target_dvf[:, :, :, 0]
        target_flow[1, :, :, :] = target_dvf[:, :, :, 1]
        target_flow[2, :, :, :] = target_dvf[:, :, :, 2]

        data = {'source_projections': torch.from_numpy(source_projections),
                'target_projections': torch.from_numpy(target_projections),
                'source_volumes': torch.from_numpy(source_volumes),
                'source_abdomen': torch.from_numpy(source_abdomen),
                'target_flow': torch.from_numpy(target_flow)}
        return data

# generate train/test split
im_size = 128
batch_size = 8
dataset = SupervisedDataset(im_dir=im_dir, im_size=im_size)
split = [int(len(dataset) * 0.9), int(len(dataset) * 0.1)]
trainset, valset = torch.utils.data.dataset.random_split(dataset, split)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, shuffle=True)
valloader = torch.utils.data.DataLoader(valset, batch_size=batch_size, shuffle=True)

# set up network
model = network_a.model(im_size, int_steps=10)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)

# set loss function and optimizer
flow_mask = losses.flow_mask()
lr = 1e-5
optimizer = optim.Adam(model.parameters(), lr=lr)

print('Training (on ' + str(device) + ')...')
tic = time.time()
min_val_loss = float('inf')
train_losses, val_losses = [], []
epoch_num = 50

for epoch in range(1, epoch_num + 1):
    train_loss = 0.0

    for i, data in enumerate(trainloader, 0):
        source_proj, target_proj, source_vol, source_abdomen, target_flow = data['source_projections'].to(device), \
                                                                            data['target_projections'].to(device), \
                                                                            data['source_volumes'].to(device), \
                                                                            data['source_abdomen'].to(device), \
                                                                            data['target_flow'].to(device)

        # zero the parameter gradients
        optimizer.zero_grad()

        # forward + backward + optimize
        _, predict_flow = model.forward(source_proj, target_proj, source_vol)
        loss = flow_mask.loss(target_flow, predict_flow, source_abdomen)
        train_loss += loss.item()
        loss.backward()
        optimizer.step()

    # test and print every epoch
    val_loss = 0.0
    model.eval()
    with torch.no_grad():
        for j, valdata in enumerate(valloader, 0):
            source_proj, target_proj, source_vol, source_abdomen, target_flow = valdata['source_projections'].to(device), \
                                                                                valdata['target_projections'].to(device), \
                                                                                valdata['source_volumes'].to(device), \
                                                                                valdata['source_abdomen'].to(device), \
                                                                                valdata['target_flow'].to(device)

            _, predict_flow = model.forward(source_proj, target_proj, source_vol)
            loss = flow_mask.loss(target_flow, predict_flow, source_abdomen)
            val_loss += loss.item()

    toc = time.time()
    time_elapsed = (toc - tic) / 3600
    hours = np.floor(time_elapsed)
    minutes = (time_elapsed - hours) * 60

    print('Epoch: %d | train loss: %.4f | val loss: %.4f | total time: %d hours %d minutes' %
          (epoch, train_loss / len(trainset), val_loss / len(valset), hours, minutes))
    train_losses.append(train_loss / len(trainset))
    val_losses.append(val_loss / len(valset))

    # save model with lowest validation cost
    if val_loss < min_val_loss:
        if not os.path.exists('weights'):
            os.mkdir('weights')

        PATH = 'weights/' + filename + '.pth'
        torch.save(model.state_dict(), PATH)
        min_val_loss = val_loss

    # plot training
    plt.figure()
    plt.title(expt_description)
    plt.plot(np.array(range(1, epoch + 1)), np.array(train_losses), 'b')
    plt.plot(np.array(range(1, epoch + 1)), np.array(val_losses), 'r')
    plt.legend(['Train', 'Validation'])
    plt.ylabel('Loss')
    plt.ticklabel_format(axis="y", style="sci", scilimits=(0, 0))
    if minutes > 30:
        hours += 1
    plt.xlabel('Epochs' + ' (' + str(int(hours)) + ' hours)')
    if not os.path.exists('plots'):
        os.mkdir('plots')
    plt.savefig('plots/' + filename + '.png')
    plt.close()

print('Finished training')
torch.cuda.empty_cache()