MADiff/traineval.py at main · IRMVLab/MADiff · 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
# Builds dataloaders, networks, diffusion components, and runs training or evaluation.

import argparse
import random

import numpy as np
import torch

from netscripts.get_network import get_network, get_my_network
from netscripts.get_optimizer import get_optimizer
from options import netsopts, expopts
from diffuseq.step_sample import create_named_schedule_sampler
from basic_utils import create_model_and_diffusion
from diffuseq.utils import dist_util


def resolve_backend_components(dataset_backend):
    backend = dataset_backend.lower()
    if backend == "egopat3d":
        from netscripts.epoch_feat_egopat import TrainLoop
        from src.EgoPAT3DLoader import build_dataloaders
    else:
        from netscripts.epoch_feat_h2o import TrainLoop
        from src.H2OLoader import build_dataloaders
    return TrainLoop, build_dataloaders


def main(args, parser):

    torch.cuda.manual_seed_all(args.manual_seed)
    torch.manual_seed(args.manual_seed)
    np.random.seed(args.manual_seed)
    random.seed(args.manual_seed)

    dist_util.setup_dist()
    TrainLoop, build_dataloaders = resolve_backend_components(args.dataset_backend)

    num_frames_input = 10
    num_frames_output = 4

    model_hoi_ori = get_network(args, num_frames_input=num_frames_input,
                                num_frames_output=num_frames_output)

    if args.use_cuda and torch.cuda.is_available():
        print("Using {} GPUs !".format(torch.cuda.device_count()))
        model_hoi_ori.to(dist_util.dev())

    start_epoch = 0
    print("Loaded hoi-forecast2022 model checkpoint from epoch {}, ".format(start_epoch))

    print("freezing hoi-forecast2022 model weights")
    for param_hoi in model_hoi_ori.parameters():
        param_hoi.requires_grad = False

    start_epoch = 0
    model_hoi, obj_head = get_my_network(args, num_frames_input=num_frames_input,
                                         num_frames_output=num_frames_output)
    print("do not freeze hoi-forecast2022 model weights")

    print("******************************************************")
    print("******************************************************")

    model_diff_args = {
        "diffusion_steps": 1000,
        "noise_schedule": "sqrt",
        "learn_sigma": False,
        "timestep_respacing": "",
        "predict_xstart": True,
        "rescale_timesteps": True,
        "sigma_small": False,
        "rescale_learned_sigmas": False,
        "use_kl": False,
    }
    print("diffusion setups\n", model_diff_args)
    pre_encoder, model_denoise, diffusion, post_encoder, motion_encoder, loc_encoder, glip_encoder = create_model_and_diffusion(**model_diff_args)

    print("finish building diffusion model")

    schedule_sampler_name = "lossaware"
    schedule_sampler = create_named_schedule_sampler(schedule_sampler_name, diffusion)
    print("finish building schedule sampler")


    if args.evaluate:
        args.epochs = start_epoch + 1
        traj_val_loader = None
        optimizer=None
        scheduler=None
        from src.config import parse_configs
        cfg = parse_configs(parser)
        test_output = build_dataloaders(cfg, phase='test')
        if isinstance(test_output, tuple):
            test_loader = test_output[0]
        else:
            test_loader = test_output

    else:
        from src.config import parse_configs
        cfg = parse_configs(parser)
        train_loader, traj_val_loader = build_dataloaders(cfg, phase='trainval')
        optimizer, scheduler = get_optimizer(args, pre_encoder=pre_encoder, model_denoise=model_denoise,post_encoder=post_encoder, loc_encoder=loc_encoder, glip_encoder=glip_encoder,
                                              train_loader=train_loader, model_hoi=model_hoi, motion_encoder=motion_encoder, obj_head=obj_head)


    if args.evaluate and args.traj_only:
        loader = test_loader
    else:
        loader = train_loader

    TrainLoop(
            epochs = args.epochs,
            loader=loader,
            evaluate=args.evaluate,
            use_cuda=True,
            optimizer=optimizer,
            scheduler=scheduler,
            model_hoi_ori=model_hoi_ori,
            model_hoi=model_hoi,
            obj_head=obj_head,
            pre_encoder=pre_encoder,
            model_denoise=model_denoise,
            diffusion=diffusion,
            post_encoder=post_encoder,
            motion_encoder=motion_encoder,
            loc_encoder=loc_encoder,
            glip_encoder=glip_encoder,
            schedule_sampler=schedule_sampler,
            resume=args.resume
    ).run_loop()


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="HOI Forecasting")
    netsopts.add_nets_opts(parser)
    netsopts.add_train_opts(parser)
    expopts.add_exp_opts(parser)
    args, _ = parser.parse_known_args()

    if args.traj_only: assert args.evaluate, "evaluate trajectory on validation set must set --evaluate"
    main(args, parser)
    print("All done !")