Robot-TrajectronV2/train.py at master · mousecpn/Robot-TrajectronV2 · 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
170
171
172
173
174
175
176
177
178
179
import torch
import numpy as np
from torch.utils.data import DataLoader
from model.trajectron import Trajectron
from tqdm import tqdm
import json
from argument_parser import args
import os
from dataset.preprocessing import ImageContextTrajDataset, load_data_cartesian
import torch.nn as nn
import torch.optim as optim
import evaluation
import matplotlib.pyplot as plt
# from train_rl import navigation_evaluate
import warnings

warnings.filterwarnings('ignore')

def main():
    if not torch.cuda.is_available() or args.device == 'cpu':
        args.device = torch.device('cpu')
    else:
        if torch.cuda.device_count() == 1:
            # If you have CUDA_VISIBLE_DEVICES set, which you should,
            # then this will prevent leftover flag arguments from
            # messing with the device allocation.
            args.device = 'cuda:0'

        args.device = torch.device(args.device)

    # This is needed for memory pinning using a DataLoader (otherwise memory is pinned to cuda:0 by default)
    torch.cuda.set_device(args.device)


    # Load hyperparameters from json
    if not os.path.exists(args.conf):
        print('Config json not found!')
    with open(args.conf, 'r', encoding='utf-8') as conf_json:
        hyperparams = json.load(conf_json)

    # Add hyperparams from arguments
    hyperparams['batch_size'] = args.batch_size
    hyperparams['k_eval'] = args.k_eval
    hyperparams['map_encoding'] = True
    hyperparams['rl'] = args.rl
    best_ade = 1000
    dim = 2

    data_dict = load_data_cartesian(args.data_path, 20, 20, test_size=0.1)

    traindataset = ImageContextTrajDataset(data_dict["trainData"], data_dict["goals_train"], data_dict["obs_train"], max_history_length=8, min_future_timesteps=12, frequency=data_dict["target_frequency"], eval=False, pad=False, random_drop_goal=False)
    testdataset = ImageContextTrajDataset(data_dict["testData"], data_dict["goals_test"], data_dict["obs_test"], max_history_length=8, min_future_timesteps=12, frequency=data_dict["target_frequency"], eval=True, pad=False)
    train_dataloader = DataLoader(traindataset,
                                    collate_fn=traindataset.collate,
                                    pin_memory=True,
                                    batch_size=args.batch_size,
                                    shuffle=True,
                                    num_workers=args.preprocess_workers)

    eval_dataloader = DataLoader(testdataset,
                                    collate_fn=testdataset.collate,
                                    pin_memory=True,
                                    batch_size=args.batch_size,
                                    shuffle=False,
                                    num_workers=args.preprocess_workers)

    hyperparams["frequency"] = data_dict["target_frequency"]

    trajectron = Trajectron(hyperparams, args.device)


    trajectron.set_annealing_params()
    trajectron.model.train()

    optimizer = optim.Adam([
                    {'params': trajectron.model.node_modules.parameters(), "lr":hyperparams['learning_rate']},
                    {'params': trajectron.model.critic.parameters(), "lr":1e-4},
                    {'params': trajectron.model.log_alpha, "lr":1e-4}
                        ]
                     )
    # Set Learning Rate
    if hyperparams['learning_rate_style'] == 'const':
        lr_scheduler = optim.lr_scheduler.ExponentialLR(optimizer, gamma=1.0)
    elif hyperparams['learning_rate_style'] == 'exp':
        lr_scheduler = optim.lr_scheduler.ExponentialLR(optimizer,gamma=hyperparams['learning_decay_rate'])


    curr_iter = 0
    for epoch in range(1, args.train_epochs + 1):
        trajectron.model.to(args.device)
        pbar = tqdm(train_dataloader, ncols=80)
        for batch in pbar:
            (first_history_index, x_t, y_t, x_st_t, y_st_t, context) = batch

            batch = (first_history_index, x_t, y_t[...,dim:dim*2], x_st_t, y_st_t[...,dim:dim*2], context)
            trajectron.set_curr_iter(curr_iter)
            trajectron.step_annealers()
            optimizer.zero_grad()
            train_loss = trajectron.train_loss(batch)
            pbar.set_description(f"Epoch {epoch},  L: {train_loss.item():.2f}")
            train_loss.backward()
            # Clipping gradients.
            if hyperparams['grad_clip'] is not None:
                nn.utils.clip_grad_value_(trajectron.model.parameters(), hyperparams['grad_clip'])
            optimizer.step()


            # Stepping forward the learning rate scheduler and annealers.
            if optimizer.param_groups[0]['lr'] > hyperparams['min_learning_rate']:
                lr_scheduler.step()

            curr_iter += 1

        print("learning_rate:",lr_scheduler.get_last_lr()[0])

        #################################
        #           EVALUATION          #
        #################################
        if args.eval_every is not None and not args.debug and epoch % args.eval_every == 0 and epoch > 0:
            max_hl = hyperparams['maximum_history_length']
            ph = hyperparams['prediction_horizon']
            trajectron.model.to(args.device)
            trajectron.model.eval()
            with torch.no_grad():
                # Calculate evaluation loss
                eval_loss_list = []
                print(f"Starting Evaluation @ epoch {epoch}")
                pbar = tqdm(eval_dataloader, ncols=80)
                ade = []
                fde = []
                for batch in pbar:
                    (first_history_index, x_t, y_t, x_st_t, y_st_t, context) = batch
                    batch = (first_history_index, x_t, y_t[...,dim:dim*2], x_st_t, y_st_t[...,dim:dim*2], context)
                    eval_loss = trajectron.eval_loss(batch)

                    pbar.set_description(f"Epoch {epoch}, L: {eval_loss.item():.2f}")
                    eval_loss_list.append({'nll': [eval_loss]})
                    # predictions = trajectron.predict(batch,
                    #                         ph,
                    #                         num_samples=20,
                    #                         z_mode=False,
                    #                         gmm_mode=False,
                    #                         full_dist=False)
                    predictions = trajectron.predict(batch,
                                        ph,
                                        num_samples=20,
                                        z_mode=True,
                                        gmm_mode=True,
                                        all_z_sep=True,
                                        full_dist=False)

                    batch_ade = np.min(evaluation.compute_ade(predictions, y_t[...,0:dim].detach().cpu().numpy()),axis=0)
                    batch_fde = np.min(evaluation.compute_fde(predictions, y_t[...,0:dim].detach().cpu().numpy()),axis=0)
                    # ax = plt.axes()
                    # visualization.plot_trajectories2d(ax, predictions, x_t[0,:,0:2].detach().cpu().numpy() ,y_t[0,:,0:2].detach().cpu().numpy())


                    ade.append(batch_ade)
                    fde.append(batch_fde)
                ade = np.mean(np.concatenate(ade,axis=0))*1000
                fde = np.mean(np.concatenate(fde,axis=0))*1000
            if ade < best_ade:
                best_ade = ade
                # model_registrar.save_models(epoch)
            model_save = trajectron.model
            torch.save(model_save.node_modules, "checkpoints/epoch{}|{}Hz|ade{:.2f}.pth".format(epoch,data_dict["target_frequency"],ade))
            torch.save(model_save.critic, "checkpoints/epoch{}|{}Hz|ade{:.2f}|critic.pth".format(epoch,data_dict["target_frequency"],ade))
            print("ade:", ade)
            print("fde:", fde)
            # navigation_evaluate(trajectron,100)

            trajectron.model.train()
    return


if __name__=="__main__":
    main()

# python main.py --eval_every 10 --vis_every 10 --train_data_dict eth_train.pkl --eval_data_dict eth_val.pkl --offline_scene_graph yes --preprocess_workers 0 --batch_size 256 --log_dir experiments/pedestrians/models --log_tag _eth_vel_ar3 --train_epochs 100 --augment --conf config/config1.json --no_edge_encoding