GenerativeCardiacMotion/train.py at main · shaharzuler/GenerativeCardiacMotion · 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
import os
import time
import json

import torch
import torch.multiprocessing as mp
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
import numpy as np

from flow_n_corr_utils import min_max_norm

from model import CVAE
from dataset import cvae_dataset
torch.manual_seed = 21


def train_cvae(l1o_idx, top_out_path, device, config):
    torch.cuda.set_device(device)
    DEVICE = torch.device(device)
    out_path = f"{top_out_path}/{l1o_idx}"
    print(out_path)
    os.makedirs(out_path)
    os.makedirs(out_path+"/train")
    os.makedirs(out_path+"/eval")
    os.makedirs(out_path+"/z_generation")
    train_writer = SummaryWriter(out_path+"/train")
    eval_writer = SummaryWriter(out_path+"/eval")
    generation_writer = SummaryWriter(out_path+"/z_generation")
    dataset_path = "/home/shahar/projects/CVAE_proj/CVAE/data/data_for_cvae" #its the same for train and validation because of leavw-one-out (l1o) training strategy. l1o_idx param sets the validation sampe num.
    train_set = cvae_dataset(dataset_path, device=DEVICE, l1o_idx=l1o_idx, train=True, augmentations=True, augmentation_params=config["augmentation_params"])
    train_loader = DataLoader(train_set, batch_size=config["batch_size"], shuffle=True)
    val_set = cvae_dataset(dataset_path, device=DEVICE, l1o_idx=l1o_idx, train=False)
    val_loader = DataLoader(val_set, batch_size=1, shuffle=False)


    num_chs = [3, 16, 32, 64, 96, 128, 128]
    cvae_model = CVAE(num_chs=num_chs, max_flow_hat_abs_val=50).to(DEVICE)
    criterion = torch.nn.MSELoss()
    optimizer = torch.optim.Adam(cvae_model.parameters(), lr=1e-3)

    lambda_reconstraction = config["lambda_reconstraction"]
    lambda_kl = config["lambda_kl"]

    num_epochs = config["num_epochs"]
    for num_epoch in range(num_epochs):
        cvae_model.train()
        epoch_train_acc_loss = 0
        epoch_train_reconstraction_loss = 0
        epoch_train_kl = 0
        for batch_num, (flow, conditions_pyramid) in enumerate(train_loader):
            optimizer.zero_grad()
            flow_hat, train_kl = cvae_model(flow, conditions_pyramid)
            train_reconstraction_loss = criterion(flow_hat, flow)
            train_loss = lambda_reconstraction * train_reconstraction_loss + lambda_kl * train_kl
            train_loss.backward()
            torch.nn.utils.clip_grad_norm_(cvae_model.parameters(), max_norm=config["grad_max_norm"])
            optimizer.step()
            epoch_train_acc_loss += train_loss.item()
            epoch_train_kl += train_kl.item()
            epoch_train_reconstraction_loss += train_reconstraction_loss.item()
        epoch_train_acc_loss /= (batch_num+1)
        epoch_train_reconstraction_loss /= (batch_num+1)
        epoch_train_kl /= (batch_num+1)
        train_writer.add_scalar("reconstraction_loss", epoch_train_reconstraction_loss, num_epoch)
        train_writer.add_scalar("kl", epoch_train_kl, num_epoch)
        train_writer.add_scalar("total_loss", epoch_train_acc_loss, num_epoch)
        train_writer.add_image("f_hat",min_max_norm(flow_hat[0, :, :, 60, :]), num_epoch) # 3 h w
        train_writer.add_image("f",min_max_norm(flow[0, :, :, 60, :]), num_epoch) # 3 h w
        train_writer.add_image("img",min_max_norm(conditions_pyramid[-1][0, :, :, 60, :]), num_epoch) # 3 h w

        cvae_model.eval()
        epoch_eval_acc_loss = 0
        epoch_eval_reconstraction_loss = 0
        epoch_eval_generation_loss = 0
        epoch_eval_kl = 0
        for batch_num, (flow, conditions_pyramid) in enumerate(val_loader):
            flow_hat, eval_kl = cvae_model(flow, conditions_pyramid)
            eval_reconstraction_loss = criterion(flow_hat, flow)
            eval_loss = lambda_reconstraction * eval_reconstraction_loss + lambda_kl * eval_kl
            epoch_eval_acc_loss += eval_loss.item()
            epoch_eval_kl += eval_kl.item()
            epoch_eval_reconstraction_loss += eval_reconstraction_loss.item()

            flow_hat_generated = cvae_model.generate(conditions_pyramid, device=DEVICE)
            eval_generation_loss = criterion(flow_hat_generated, flow)
            epoch_eval_generation_loss += eval_generation_loss.item()

        epoch_eval_acc_loss /= (batch_num+1)
        epoch_eval_reconstraction_loss /= (batch_num+1)
        epoch_eval_kl /= (batch_num+1)
        epoch_eval_generation_loss /= (batch_num+1)

        eval_writer.add_scalar("reconstraction_loss", epoch_eval_reconstraction_loss, num_epoch)
        eval_writer.add_scalar("kl", epoch_eval_kl, num_epoch)
        eval_writer.add_scalar("total_loss", epoch_eval_acc_loss, num_epoch)
        generation_writer.add_scalar("reconstraction_loss", epoch_eval_generation_loss, num_epoch)
        eval_writer.add_image("f_hat",min_max_norm(flow_hat[0, :, :, 60, :]), num_epoch) # 3 h w
        generation_writer.add_image("f_hat",min_max_norm(flow_hat_generated[0, :, :, 60, :]), num_epoch) # 3 h w
        eval_writer.add_image("f",min_max_norm(flow[0, :, :, 60, :]), num_epoch) # 3 h w
        eval_writer.add_image("img",min_max_norm(conditions_pyramid[-1][0, :, :, 60, :]), num_epoch) # 3 h w
        if num_epoch % config["save_weights_every"] == 0:
            torch.save(cvae_model.state_dict(), f"{out_path}/model_e{num_epoch}.pt")
            for batch_num, (flow, conditions_pyramid) in enumerate(val_loader):
                flow_hat_generated = cvae_model.generate(conditions_pyramid, device=DEVICE).detach().cpu().numpy()
                np.save(f"{out_path}/flow_generated_e{num_epoch}.npy", flow_hat_generated)
                np.save(f"{out_path}/condition.npy", conditions_pyramid[-1].detach().cpu().numpy())

    torch.save(cvae_model.state_dict(), f"{out_path}/model_final.pt")
    cvae_model.eval()
    for batch_num, (flow, conditions_pyramid) in enumerate(val_loader):
        flow_hat_generated = cvae_model.generate(conditions_pyramid, device=DEVICE).detach().cpu().numpy()
        np.save(f"{out_path}/flow_generated_final.npy", flow_hat_generated)
        np.save(f"{out_path}/condition.npy", conditions_pyramid[-1].detach().cpu().numpy())


if __name__ == '__main__':
    mp.set_start_method('spawn')
    DEVICE0 = "cuda:0"
    DEVICE1 = "cuda:1"

    multiprocess=False

    top_out_path = f"/home/shahar/projects/CVAE_proj/CVAE/outputs/len_44_dataset/{time.strftime('%Y%m%d_%H%M%S')}"
    os.makedirs(top_out_path)

    config_file_path = "config.json"
    with open(config_file_path, 'r') as config_file:
        config = json.load(config_file)

    output_config_path = os.path.join(top_out_path, "config.json")
    with open(output_config_path, 'w') as output_config_file:
        json.dump(config, output_config_file, indent=4)


    if not multiprocess:
        l1o_idx, device = 2, DEVICE0 #TODO set sample number to l1o_idx
        train_cvae(l1o_idx, top_out_path, device, config)

    else:
        for batch in range(11): # each "batch" of sample nums maning training *seperately* for 4 samples simutaniously
            processes = []
            for l1o_idx, device in enumerate([DEVICE0, DEVICE0, DEVICE1, DEVICE1]):
                print("l1o_idx", (batch*4)+l1o_idx)
                p = mp.Process(target=train_cvae, args=((batch*4)+l1o_idx, top_out_path, device, config))
                p.start()
                processes.append(p)

            for p in processes:
                p.join()

    print("Finished")