smm/train_smm_object_discovery.py at main · Cognitive-AI-Systems/smm · 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
import pytorch_lightning.loggers as pl_loggers
from pytorch_lightning import Trainer
from pytorch_lightning.callbacks import LearningRateMonitor, ModelCheckpoint

import os
import sys

from datasets.clevrtex_dataset import CLEVRTEXDataModule
from datasets.shapestacks_dataset import ShapeStacksDataModule

root_path = os.path.abspath(__file__)
root_path = '/'.join(root_path.split('/')[:-2])

sys.path.append(root_path)

from modules.smm import GMMMixtureDecoderModel
from models.mixture_dec_object_discovery import SlotAttentionMethod
from models.utils import ImageLogCallback, set_random_seed

import json
import argparse


datamodules = {
    'clevrtex': CLEVRTEXDataModule,
    'shapestacks': ShapeStacksDataModule,

}


monitors = {
    'iou': 'avg_IoU',
    'ari': 'avg_ARI_FG',
    'ap': 'avg_AP@05',
}

parser = argparse.ArgumentParser()

parser.add_argument('--dataset', default='')
parser.add_argument('--data_root', default='')
parser.add_argument('--split_name', type=str, default='image', help='split for YCB, COCO, and ScanNet; for CLEVRTEX is full, outd, and camo')
parser.add_argument('--log_name', default='gmm')
parser.add_argument('--log_path', default='../../results/')
parser.add_argument('--ckpt_path', default='ckpt.pt.tar')
parser.add_argument('--test_ckpt_path', default='.ckpt')

parser.add_argument('--evaluate', type=str, default='ari', help='ari or iou')
parser.add_argument('--monitor', type=str, default='avg_ARI_FG', help='avg_ARI_FG or avg_IoU')
parser.add_argument('--seed', type=int, default=42)
parser.add_argument('--num_workers', type=int, default=4)
parser.add_argument('--max_steps', type=int, default=250000)
parser.add_argument('--max_epochs', type=int, default=100000)
parser.add_argument('--num_sanity_val_steps', type=int, default=1)
parser.add_argument('--check_val_every_n_epoch', type=int, default=1)
parser.add_argument('--n_samples', type=int, default=16)
parser.add_argument('--batch_size', type=int, default=128, help='batch size per GPU, if use 2 GPUs, change batch size to 64')
parser.add_argument('--gpus', type=int, default=0)

parser.add_argument('--grad_clip', type=float, default=1.0)
parser.add_argument('--grad_accum', type=int, default=1)
parser.add_argument('--resolution', type=int, nargs='+', default=[128, 128])
parser.add_argument('--init_resolution', type=int, nargs='+', default=[8, 8])
parser.add_argument('--encoder_channels', type=int, nargs='+', default=[64, 64, 64, 64])
parser.add_argument('--decoder_channels', type=int, nargs='+', default=[64, 64, 64, 64])
parser.add_argument('--encoder_strides', type=int, nargs='+', default=[2, 1, 1, 1])
parser.add_argument('--decoder_strides', type=int, nargs='+', default=[2, 2, 2, 2])
parser.add_argument('--encoder_kernel_size', type=int, default=5)
parser.add_argument('--decoder_kernel_size', type=int, default=5)

parser.add_argument('--is_logger_enabled', default=False, action='store_true')
parser.add_argument('--load_from_ckpt', default=False, action='store_true')
parser.add_argument('--use_rescale', default=False, action='store_true')

parser.add_argument('--lr_sa', type=float, default=4e-4)
parser.add_argument('--warmup_steps', type=int, default=5000)
parser.add_argument('--decay_steps', type=int, default=50000)

parser.add_argument('--num_iter', type=int, default=3)
parser.add_argument('--num_slots', type=int, default=2)
parser.add_argument('--init_size', type=int, default=64)
parser.add_argument('--slot_size', type=int, default=64)
parser.add_argument('--mlp_size', type=int, default=128)


def main(args):
    print(args)
    set_random_seed(args.seed)
    args.monitor = monitors[args.evaluate]
    datamodule = datamodules[args.dataset](args)
    model = GMMMixtureDecoderModel(args)
    method = SlotAttentionMethod(model=model, datamodule=datamodule, args=args)
    method.hparams = args

    if args.is_logger_enabled:
        wlog = pl_loggers.WandbLogger(project='smm', save_dir=args.log_path, name=f"mix_dec_{args.dataset}", config=args, log_model='all')
        logger = wlog
        callbacks = [LearningRateMonitor("step"),
                     ImageLogCallback(),
                     ModelCheckpoint(monitor=args.monitor, save_top_k=1, save_last=True, mode='max')]
    else:
        logger = False
        callbacks = []


    trainer = Trainer(
        resume_from_checkpoint=args.ckpt_path if args.load_from_ckpt else None,
        logger=logger,
        precision=16,
        default_root_dir=args.log_path,
        accelerator="ddp" if args.gpus > 1 else None,
        num_sanity_val_steps=args.num_sanity_val_steps,
        gpus=args.gpus,
        max_steps=args.max_steps,
        accumulate_grad_batches=args.grad_accum,
        max_epochs=args.max_epochs,
        log_every_n_steps=50,
        callbacks=callbacks,
        check_val_every_n_epoch=args.check_val_every_n_epoch,
        gradient_clip_val=args.grad_clip,
        # val_check_interval=3000,
    )
    trainer.fit(method)

if __name__ == "__main__":
    args = parser.parse_args()
    paths = json.load(open('./path.json', 'r'))
    data_paths = paths['data_paths']
    args.log_path = paths['log_path']
    dataset = args.dataset
    args.data_root = data_paths[dataset]
    args.log_path += dataset
    main(args)