COSMIC_Embedding/mmd_train_model.py at main · gvisona/COSMIC_Embedding · 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
import os
from os.path import join
import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm
from mmd_vae_model import CancerSamplesDataset, Bimodal_MMD_VAE, compute_mmd

if __name__ == "__main__":
    # Training hyperparameters
    n_epochs = 200
    batch_size = 128
    early_stopping_patience = 5
    num_workers = 4

    train_ds = CancerSamplesDataset(join("data", "sample_subtype_encodings.csv"),
                                    join("data", "sorted_mutations.json"),
                                    join("data", "mutations_mapping_split.json"),
                                    train=True)
    val_ds = CancerSamplesDataset(join("data", "sample_subtype_encodings.csv"),
                                  join("data", "sorted_mutations.json"),
                                  join("data", "mutations_mapping_split.json"),
                                  train=False)
    model = Bimodal_MMD_VAE()

    train_loader = DataLoader(train_ds,
                              batch_size=batch_size,
                              num_workers=num_workers,
                              shuffle=True,
                              drop_last=True)
    val_loader = DataLoader(val_ds,
                            batch_size=batch_size,
                            num_workers=num_workers,
                            shuffle=True,
                            drop_last=True)

    optimizer = torch.optim.SGD(model.parameters(), lr=3e-3,
                                momentum=0.9,
                                weight_decay=2e-5)
    reconstruction_loss = nn.BCELoss()

    def loss_function(X_del, X_nd, z, y_del, y_nd):
        r_del = reconstruction_loss(y_del, X_del)
        r_nd = reconstruction_loss(y_nd, X_nd)
        true_samples = torch.randn(batch_size, 50, requires_grad=False)
        mmd_loss = compute_mmd(true_samples, z)
        loss = r_del + r_nd + mmd_loss
        return loss, r_del, r_nd, mmd_loss

    if not os.path.exists("logs"):
        os.mkdir("logs")
    if not os.path.exists("models"):
        os.mkdir("models")

    train_iter = 0
    val_iter = 0
    best_val_loss = np.inf
    patience_counter = 0
    tb_writer = SummaryWriter(os.path.join("logs", "mmd_vae"))
    for epoch in range(n_epochs):
        print("="*30)
        print("Epoch ", epoch)

        # Training epoch
        avg_train_loss = 0
        model.train()

        pbar = tqdm(total=len(train_loader))
        for idx, batch in enumerate(train_loader):
            train_iter += 1

            X_del, X_nd, _ = batch
            y_del, y_nd, z = model(X_del, X_nd)

            train_loss, r_del, r_nd, mmd_loss = loss_function(
                X_del, X_nd, z, y_del, y_nd)

            train_loss.backward()
            optimizer.step()
            optimizer.zero_grad()

            train_loss = train_loss.item()
            avg_train_loss += train_loss

            tb_writer.add_scalar('Loss/train', train_loss, train_iter)
            tb_writer.add_scalar('Loss/train_mmd_loss', mmd_loss, train_iter)
            tb_writer.add_scalar('Loss/train_r_del', r_del, train_iter)
            tb_writer.add_scalar('Loss/train_r_nd', r_nd, train_iter)

            avg_loss = avg_train_loss/(idx+1)
            pbar.set_description(
                f"Train loss {train_loss}, running avg loss {avg_loss}")
            pbar.update(1)
        pbar.close()

        avg_train_loss /= len(train_loader)
        tb_writer.add_scalar('AvgLoss/train', avg_train_loss, epoch)

        # Validation epoch

        avg_val_loss = 0

        model.eval()
        pbar = tqdm(total=len(val_loader))
        with torch.no_grad():
            for idx, batch in enumerate(val_loader):
                val_iter += 1

                X_del, X_nd, _ = batch
                y_del, y_nd, z = model(X_del, X_nd)

                val_loss, r_del, r_nd, mmd_loss = loss_function(
                    X_del, X_nd, z, y_del, y_nd)
                val_loss = val_loss.item()
                avg_val_loss += val_loss

                tb_writer.add_scalar('Loss/val', val_loss, val_iter)
                tb_writer.add_scalar('Loss/val_mmd_loss', mmd_loss, val_iter)
                tb_writer.add_scalar('Loss/val_r_del', r_del, val_iter)
                tb_writer.add_scalar('Loss/val_r_nd', r_nd, val_iter)
                avg_loss = avg_val_loss/(idx+1)
                pbar.set_description(
                    f"Val loss {val_loss}, running avg loss {avg_loss}")
                pbar.update(1)
            pbar.close()

        avg_val_loss /= len(val_loader)

        tb_writer.add_scalar('AvgLoss/val', avg_val_loss, epoch)

        if avg_val_loss < best_val_loss:
            print("New minimum validation loss, saving model.")
            patience_counter = 0
            state = {
                'epoch': epoch,
                'state_dict': model.state_dict(),
                'optimizer': optimizer.state_dict(),
            }
            savepath = os.path.join("models", 'mmd_vae.ckpt')
            torch.save(state, savepath)
        else:
            patience_counter += 1
            if patience_counter > early_stopping_patience:
                print(
                    f"**********\nNo improvements for the last {str(early_stopping_patience)} epochs, stopping training")
                break