FLAG/train_graph_diffusion.py at main · SAIS-LifeScience/FLAG · 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
from utils import *
from datetime import datetime
from torch import optim
from torch.utils.tensorboard import SummaryWriter
from dataset.utils import *
from losses import *

def train_graph_diffusion(config):
    print("start train_graph_diffusion pipeline ...")

    seeding(config['seed'])
    device = config['device']
    mode = config['mode']

    data_config = config['data_config']
    model_config = config['model_config']
    train_config = config['train_config']

    dataset_name = data_config['dataset_name']
    crop_size = data_config['crop_size']
    model_name = model_config['name']

    result_folder = os.path.join(train_config['result_dir'], f"{mode}/{dataset_name}_crop_size_{crop_size}",
                                 model_name)
    os.makedirs(result_folder, exist_ok=True)

    save_config(os.path.join(result_folder, 'config.yaml'), config)

    # Initialize TensorBoard
    current_time = datetime.now().strftime("%Y%m%d-%H%M%S")
    log_dir = os.path.join(result_folder, "runs", current_time)
    writer = SummaryWriter(log_dir=log_dir)

    logger = create_logger(log_dir)
    logger.info(f"Experiment directory created at {result_folder}")

    train_dataset, selected_genes = load_dataset(data_config, mode='train')
    gene_size = len(selected_genes)

    valid_dataset, _ = load_dataset(data_config, mode='valid')

    train_loader = make_loader(train_dataset, batch_size=train_config['batch_size'], shuffle=True)
    valid_loader = make_loader(valid_dataset, batch_size=train_config['batch_size'], shuffle=False)

    model = load_model(model_config, gene_size).to(device)
    ema = EMA(model, decay=0.999)
    sde = load_sde(config['sde_config'])
    sampler_fn = load_sampler_fn(config['sde_config'])

    opt = optim.AdamW(model.parameters(), lr=train_config['lr'], weight_decay=train_config['wd'])

    best_mse = float('inf')
    epochs_no_improve = 0
    patience = train_config['patience']

    accum_step = 0

    for epoch in range(train_config['epochs']):
        model.train()
        pbar = tqdm(train_loader, desc=f"epoch {epoch}|")
        total_loss = 0
        epoch_node_loss = 0
        epoch_edge_loss = 0
        epoch_cons_loss = 0
        for batch in pbar:
            for k in batch:
                try:
                    batch[k] = batch[k].to(device)
                except:
                    continue
                if "graph_diffusion" in mode:
                    loss, logs, _ = graph_diffusion_loss(model, sde, batch, config)
                else:
                    raise ValueError(mode)

            loss.backward()
            torch.nn.utils.clip_grad_norm_(model.parameters(), train_config['grad_clip'])
            accum_step += 1

            if accum_step % train_config['gradient_accumulate_every'] == 0:
                opt.step()
                opt.zero_grad()
                ema.update()
                writer.add_scalar(f"Train-Step/loss", loss, accum_step)

            pbar.set_postfix({"Loss": float(loss), **{k: float(v) for k, v in logs.items()}})
            total_loss += float(loss)
            epoch_node_loss += float(logs['loss_x'])
            epoch_edge_loss += float(logs['loss_e'])
            epoch_cons_loss += float(logs['loss_cons'])

        if config['mode'] == "graph_diffusion_fixed":
            logger.info(f"Epoch={epoch} | "
                        f"Average Training Loss: {total_loss / len(train_loader):.5f} |"
                        f"Average Node Loss: {epoch_node_loss / len(train_loader):.5f} |"
                        )
            writer.add_scalar(f"Train-Epoch/total_loss", total_loss / len(train_loader), epoch)
            writer.add_scalar(f"Train-Epoch/node_loss", epoch_node_loss / len(train_loader), epoch)
        elif config['mode'] == "graph_diffusion_learned":
            logger.info(f"Epoch={epoch} | "
                        f"Average Training Loss: {total_loss / len(train_loader):.5f} |"
                        f"Average Node Loss: {epoch_node_loss / len(train_loader):.5f} |"
                        f"Average Edge Loss: {epoch_edge_loss / len(train_loader):.5f} |"
                        f"Average Cons Loss: {epoch_cons_loss / len(train_loader):.5f} |"
                        )
            writer.add_scalar(f"Train-Epoch/total_loss", total_loss / len(train_loader), epoch)
            writer.add_scalar(f"Train-Epoch/node_loss", epoch_node_loss / len(train_loader), epoch)
            writer.add_scalar(f"Train-Epoch/edge_loss", epoch_edge_loss / len(train_loader), epoch)
            writer.add_scalar(f"Train-Epoch/cons_loss", epoch_cons_loss / len(train_loader), epoch)


        if epoch % train_config['eval_interval'] == 0:
            model.eval()
            ema.apply_shadow()
            with torch.no_grad():
                all_gt_genes = []
                all_pred_genes = []
                loss_edge = 0
                for batch in tqdm(valid_loader):
                    for k in batch:
                        try:
                            batch[k] = batch[k].to(device)
                        except:
                            continue
                    pred_x, pred_e = sampler_fn(config, ema.model, sde, batch['cond_x'], batch['cond_e'], batch['mask'])
                    pred_x = pred_x.squeeze(0).cpu()
                    x_gt = batch['x_gt'].squeeze(0).cpu()
                    all_gt_genes.append(x_gt.numpy())
                    all_pred_genes.append(pred_x.numpy())
                    loss_edge += float(edge_loss(pred_e, batch, config))

                ema.restore()
                val_metrics = compute_metrics(all_pred_genes, all_gt_genes)
                nan_count = val_metrics['nan_count']
                valid_count = len(selected_genes) - nan_count
                logger.info(f"Epoch={epoch} | Validation Metrics: MSE={val_metrics['MSE']:.5f}, "
                            f"MAE={val_metrics['MAE']:.5f}, "
                            f"nan_count={val_metrics['nan_count']:.5f},"
                            f"PCC-10={val_metrics['PCC-10']:.5f}, "
                            f"PCC-50={val_metrics['PCC-50']:.5f}, "
                            f"PCC-{valid_count}={val_metrics[f'PCC-{valid_count}']:.5f}")

                logger.info(f"Validation Edge loss={loss_edge / len(valid_loader):.5f}")

                writer.add_scalar(f"Validation/MSE", val_metrics['MSE'], epoch)
                writer.add_scalar(f"Validation/MAE", val_metrics['MAE'], epoch)
                writer.add_scalar(f"Validation/PCC-{valid_count}", val_metrics[f'PCC-{valid_count}'], epoch)

                if mode == 'graph_diffusion_learned':
                    writer.add_scalar(f"Validation/edge_mse", loss_edge / len(valid_loader), epoch)

            val_mse = val_metrics['MSE']
            if val_mse < best_mse:
                best_mse = val_mse
                epochs_no_improve = 0
                best_checkpoint_path = save_checkpoint(model, ema, opt, result_folder, epoch)
                logger.info(f"New best validation mse: {best_mse:.5f}, saved checkpoint to {best_checkpoint_path}")
            else:
                epochs_no_improve += 1
                logger.info(
                    f"No improvement in validation mse, epochs without improvement: {epochs_no_improve}/{patience}")

            if epochs_no_improve >= patience:
                logger.info(f"Early stopping triggered after {epochs_no_improve} epochs without improvement")
                break

    best_checkpoint_path = save_checkpoint(model, ema, opt, result_folder, epoch)
    writer.close()
    logger.info(f"Training completed. Best validation mse: {best_mse:.5f}, Checkpoint: {best_checkpoint_path}")