File size: 4,757 Bytes
ff66cf3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Main training script."""

import os
from pathlib import Path

import torch
from cliport import agents
from cliport.dataset import RavensDataset, RavensMultiTaskDataset, RavenMultiTaskDatasetBalance

import hydra
from pytorch_lightning import Trainer
from pytorch_lightning.callbacks import ModelCheckpoint
from pytorch_lightning.loggers import WandbLogger
import numpy as np
from torch.utils.data import DataLoader
from torch.utils.data.dataloader import default_collate
import IPython
import pytorch_lightning as pl
from pytorch_lightning.utilities import rank_zero_only
import datetime
import time
import random


def set_seed_everywhere(seed):
    torch.manual_seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(seed)
    np.random.seed(seed)
    random.seed(seed)

@hydra.main(config_path="./cfg", config_name='train', version_base="1.2")
def main(cfg):
    # Logger
    set_seed_everywhere(1)
    wandb_logger = None

    if cfg['train']['log']:
        try:
            wandb_logger = WandbLogger(name=cfg['tag'])
        except:
            pass

    # Checkpoint saver
    hydra_dir = Path(os.getcwd())
    checkpoint_path = os.path.join(cfg['train']['train_dir'], 'checkpoints')
    last_checkpoint_path = os.path.join(checkpoint_path, 'last.ckpt')
    last_checkpoint = last_checkpoint_path if os.path.exists(last_checkpoint_path) and cfg['train']['load_from_last_ckpt'] else None
    checkpoint_callback = [ModelCheckpoint(
        # monitor=cfg['wandb']['saver']['monitor'],
        dirpath=os.path.join(checkpoint_path, 'best'),
        save_top_k=1,
        every_n_epochs=3,
        save_last=True,
        # every_n_train_steps=100    
        )]

    # Trainer
    max_epochs = cfg['train']['n_steps'] * cfg['train']['batch_size'] // cfg['train']['n_demos']
    if cfg['train']['training_step_scale'] > 0:
        # scale training time depending on the tasks to ensure coverage.
        max_epochs = cfg['train']['training_step_scale'] #  // cfg['train']['batch_size']

    trainer = Trainer(
        accelerator='gpu',
        devices=cfg['train']['gpu'],
        fast_dev_run=cfg['debug'],
        logger=wandb_logger,
        callbacks=checkpoint_callback,
        max_epochs=max_epochs,
        # check_val_every_n_epoch=max_epochs // 50,
        # resume_from_checkpoint=last_checkpoint,
        sync_batchnorm=True,
        log_every_n_steps=30,        
    )

    print(f"max epochs: {max_epochs}!")
    
    # Resume epoch and global_steps
    if last_checkpoint:
        print(f"Resuming: {last_checkpoint}")

    # Config
    data_dir = cfg['train']['data_dir']
    task = cfg['train']['task']
    agent_type = cfg['train']['agent']
    n_demos = cfg['train']['n_demos']
    
    if agent_type == 'mdetr':
        print('======import torch.multiprocessing to avioid shared memory issue======')
        import torch.multiprocessing
        torch.multiprocessing.set_sharing_strategy('file_system')

    # n_demos = cfg['train']['n_demos']
    # n_demos = cfg['train']['n_demos']
    n_val = cfg['train']['n_val']
    name = '{}-{}-{}'.format(task, agent_type, n_demos)
            
    # Datasets
    dataset_type = cfg['dataset']['type']
    if 'multi' in dataset_type:
        train_ds = RavensMultiTaskDataset(data_dir, cfg, group=task, mode='train', 
                    n_demos=n_demos, augment=True)
        val_ds = RavensMultiTaskDataset(data_dir, cfg, group=task, mode='val', n_demos=n_val, augment=False)
    elif 'weighted' in dataset_type:
        train_ds = RavenMultiTaskDatasetBalance(data_dir, cfg, group=task, mode='train', n_demos=n_demos, augment=True)
        val_ds = RavenMultiTaskDatasetBalance(data_dir, cfg, group=task, mode='val', n_demos=n_val, augment=False)
    else:
        train_ds = RavensDataset(os.path.join(data_dir, '{}-train'.format(task)), cfg, n_demos=n_demos, augment=True)
        val_ds = RavensDataset(os.path.join(data_dir, '{}-val'.format(task)), cfg, n_demos=n_val, augment=False)

    # Initialize agent
    train_loader = DataLoader(train_ds, shuffle=True,
                    pin_memory=True,
                    batch_size=cfg['train']['batch_size'],
                    num_workers=1 )
    test_loader = DataLoader(val_ds, shuffle=False,
                num_workers=1,
                batch_size=cfg['train']['batch_size'],
                pin_memory=True)

    agent = agents.names[agent_type](name, cfg, train_loader, test_loader)
    dt_string = datetime.datetime.now().strftime("%d_%m_%Y_%H:%M:%S")
    print("current time:", dt_string)
    
    start_time = time.time()
    # Main training loop
    trainer.fit(agent, ckpt_path=last_checkpoint)
    
    print("current time:", time.time() - start_time)

if __name__ == '__main__':
    main()