File size: 7,335 Bytes
404d2af
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8b973ee
404d2af
 
8b973ee
404d2af
 
8b973ee
404d2af
 
 
8b973ee
 
 
 
404d2af
 
8b973ee
404d2af
 
 
8b973ee
404d2af
8b973ee
 
404d2af
8b973ee
404d2af
 
 
8b973ee
 
 
 
404d2af
 
 
 
8b973ee
404d2af
 
8b973ee
404d2af
 
8b973ee
404d2af
 
 
 
8b973ee
 
 
 
404d2af
 
8b973ee
 
 
 
404d2af
 
 
 
 
 
 
 
8b973ee
404d2af
8b973ee
404d2af
 
 
 
8b973ee
404d2af
 
 
 
 
 
 
 
 
8b973ee
404d2af
 
 
 
8b973ee
404d2af
 
8b973ee
404d2af
 
 
8b973ee
404d2af
 
8b973ee
404d2af
 
8b973ee
404d2af
 
 
 
 
8b973ee
 
 
 
 
 
 
 
 
404d2af
 
 
 
 
 
 
 
 
 
8b973ee
 
 
 
 
 
404d2af
 
 
 
 
 
 
 
 
8b973ee
404d2af
 
8b973ee
 
 
 
 
 
 
 
 
 
404d2af
 
 
 
 
8b973ee
 
 
 
404d2af
 
 
 
 
 
 
8b973ee
404d2af
 
8b973ee
404d2af
 
 
8b973ee
 
404d2af
 
 
 
 
 
 
 
8b973ee
 
 
 
 
 
 
 
404d2af
 
 
 
 
 
 
 
 
8b973ee
 
 
 
 
 
 
404d2af
 
 
 
 
 
 
8b973ee
 
 
 
 
 
 
 
404d2af
 
 
 
 
8b973ee
404d2af
 
8b973ee
 
 
 
 
 
404d2af
 
8b973ee
404d2af
 
8b973ee
404d2af
 
 
 
 
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
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
import argparse

import numpy as np

import os

import shutil

import torch
import torch.optim as optim

from torch.utils.data import DataLoader

from tqdm import tqdm

import warnings

from lib.dataset import MegaDepthDataset
from lib.exceptions import NoGradientError
from lib.loss import loss_function
from lib.model import D2Net


# CUDA
use_cuda = torch.cuda.is_available()
device = torch.device("cuda:0" if use_cuda else "cpu")

# Seed
torch.manual_seed(1)
if use_cuda:
    torch.cuda.manual_seed(1)
np.random.seed(1)

# Argument parsing
parser = argparse.ArgumentParser(description="Training script")

parser.add_argument(
    "--dataset_path", type=str, required=True, help="path to the dataset"
)
parser.add_argument(
    "--scene_info_path", type=str, required=True, help="path to the processed scenes"
)

parser.add_argument(
    "--preprocessing",
    type=str,
    default="caffe",
    help="image preprocessing (caffe or torch)",
)
parser.add_argument(
    "--model_file", type=str, default="models/d2_ots.pth", help="path to the full model"
)

parser.add_argument(
    "--num_epochs", type=int, default=10, help="number of training epochs"
)
parser.add_argument("--lr", type=float, default=1e-3, help="initial learning rate")
parser.add_argument("--batch_size", type=int, default=1, help="batch size")
parser.add_argument(
    "--num_workers", type=int, default=4, help="number of workers for data loading"
)

parser.add_argument(
    "--use_validation",
    dest="use_validation",
    action="store_true",
    help="use the validation split",
)
parser.set_defaults(use_validation=False)

parser.add_argument(
    "--log_interval", type=int, default=250, help="loss logging interval"
)

parser.add_argument("--log_file", type=str, default="log.txt", help="loss logging file")

parser.add_argument(
    "--plot", dest="plot", action="store_true", help="plot training pairs"
)
parser.set_defaults(plot=False)

parser.add_argument(
    "--checkpoint_directory",
    type=str,
    default="checkpoints",
    help="directory for training checkpoints",
)
parser.add_argument(
    "--checkpoint_prefix",
    type=str,
    default="d2",
    help="prefix for training checkpoints",
)

args = parser.parse_args()

print(args)

# Create the folders for plotting if need be
if args.plot:
    plot_path = "train_vis"
    if os.path.isdir(plot_path):
        print("[Warning] Plotting directory already exists.")
    else:
        os.mkdir(plot_path)

# Creating CNN model
model = D2Net(model_file=args.model_file, use_cuda=use_cuda)

# Optimizer
optimizer = optim.Adam(
    filter(lambda p: p.requires_grad, model.parameters()), lr=args.lr
)

# Dataset
if args.use_validation:
    validation_dataset = MegaDepthDataset(
        scene_list_path="megadepth_utils/valid_scenes.txt",
        scene_info_path=args.scene_info_path,
        base_path=args.dataset_path,
        train=False,
        preprocessing=args.preprocessing,
        pairs_per_scene=25,
    )
    validation_dataloader = DataLoader(
        validation_dataset, batch_size=args.batch_size, num_workers=args.num_workers
    )

training_dataset = MegaDepthDataset(
    scene_list_path="megadepth_utils/train_scenes.txt",
    scene_info_path=args.scene_info_path,
    base_path=args.dataset_path,
    preprocessing=args.preprocessing,
)
training_dataloader = DataLoader(
    training_dataset, batch_size=args.batch_size, num_workers=args.num_workers
)


# Define epoch function
def process_epoch(
    epoch_idx,
    model,
    loss_function,
    optimizer,
    dataloader,
    device,
    log_file,
    args,
    train=True,
):
    epoch_losses = []

    torch.set_grad_enabled(train)

    progress_bar = tqdm(enumerate(dataloader), total=len(dataloader))
    for batch_idx, batch in progress_bar:
        if train:
            optimizer.zero_grad()

        batch["train"] = train
        batch["epoch_idx"] = epoch_idx
        batch["batch_idx"] = batch_idx
        batch["batch_size"] = args.batch_size
        batch["preprocessing"] = args.preprocessing
        batch["log_interval"] = args.log_interval

        try:
            loss = loss_function(model, batch, device, plot=args.plot)
        except NoGradientError:
            continue

        current_loss = loss.data.cpu().numpy()[0]
        epoch_losses.append(current_loss)

        progress_bar.set_postfix(loss=("%.4f" % np.mean(epoch_losses)))

        if batch_idx % args.log_interval == 0:
            log_file.write(
                "[%s] epoch %d - batch %d / %d - avg_loss: %f\n"
                % (
                    "train" if train else "valid",
                    epoch_idx,
                    batch_idx,
                    len(dataloader),
                    np.mean(epoch_losses),
                )
            )

        if train:
            loss.backward()
            optimizer.step()

    log_file.write(
        "[%s] epoch %d - avg_loss: %f\n"
        % ("train" if train else "valid", epoch_idx, np.mean(epoch_losses))
    )
    log_file.flush()

    return np.mean(epoch_losses)


# Create the checkpoint directory
if os.path.isdir(args.checkpoint_directory):
    print("[Warning] Checkpoint directory already exists.")
else:
    os.mkdir(args.checkpoint_directory)


# Open the log file for writing
if os.path.exists(args.log_file):
    print("[Warning] Log file already exists.")
log_file = open(args.log_file, "a+")

# Initialize the history
train_loss_history = []
validation_loss_history = []
if args.use_validation:
    validation_dataset.build_dataset()
    min_validation_loss = process_epoch(
        0,
        model,
        loss_function,
        optimizer,
        validation_dataloader,
        device,
        log_file,
        args,
        train=False,
    )

# Start the training
for epoch_idx in range(1, args.num_epochs + 1):
    # Process epoch
    training_dataset.build_dataset()
    train_loss_history.append(
        process_epoch(
            epoch_idx,
            model,
            loss_function,
            optimizer,
            training_dataloader,
            device,
            log_file,
            args,
        )
    )

    if args.use_validation:
        validation_loss_history.append(
            process_epoch(
                epoch_idx,
                model,
                loss_function,
                optimizer,
                validation_dataloader,
                device,
                log_file,
                args,
                train=False,
            )
        )

    # Save the current checkpoint
    checkpoint_path = os.path.join(
        args.checkpoint_directory, "%s.%02d.pth" % (args.checkpoint_prefix, epoch_idx)
    )
    checkpoint = {
        "args": args,
        "epoch_idx": epoch_idx,
        "model": model.state_dict(),
        "optimizer": optimizer.state_dict(),
        "train_loss_history": train_loss_history,
        "validation_loss_history": validation_loss_history,
    }
    torch.save(checkpoint, checkpoint_path)
    if args.use_validation and validation_loss_history[-1] < min_validation_loss:
        min_validation_loss = validation_loss_history[-1]
        best_checkpoint_path = os.path.join(
            args.checkpoint_directory, "%s.best.pth" % args.checkpoint_prefix
        )
        shutil.copy(checkpoint_path, best_checkpoint_path)

# Close the log file
log_file.close()