forked from liang3588/MeshMAE
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_pretrain.py
189 lines (155 loc) · 7.65 KB
/
train_pretrain.py
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
import torch
import os
import sys
from torch.autograd import Variable
import argparse
from tensorboardX import SummaryWriter
import copy
import torch.optim as optim
from torch.optim.lr_scheduler import MultiStepLR, CosineAnnealingLR
import numpy as np
import torch.nn as nn
import torch.utils.data as data
from model.dataset import ClassificationDataset
from model.dataset import ShapeNetDataset
from model.meshmae import Mesh_mae
from model.reconstruction import save_results
from model.utils import ClassificationMajorityVoting
from transformers import AdamW, get_linear_schedule_with_warmup, get_constant_schedule, get_cosine_schedule_with_warmup
import time
def train(net, optim, scheduler, names, train_dataset, epoch, args):
net.train()
running_loss = 0
n_samples = 0
for it, (feats_patch, center_patch,coordinate_patch, face_patch, np_Fs, label, mesh_paths) in enumerate(
train_dataset):
optim.zero_grad()
faces = face_patch.to(torch.float32).cuda()
feats = feats_patch.to(torch.float32).cuda()
centers = center_patch.to(torch.float32).cuda()
Fs = np_Fs.cuda()
cordinates = coordinate_patch.to(torch.float32).cuda()
n_samples += faces.shape[0]
loss = net(faces, feats, centers, Fs, cordinates)
loss.backward()
optim.step()
running_loss += loss.item() * faces.size(0)
epoch_loss = running_loss / n_samples
scheduler.step()
if train.best_loss > epoch_loss:
train.best_loss = epoch_loss
train.best_epoch = epoch
best_model_wts = copy.deepcopy(net.state_dict())
torch.save(best_model_wts, os.path.join('checkpoints', names, f'loss-{epoch_loss:.4f}-{epoch:.4f}.pkl'))
print('epoch ({:}): {:} Train Loss: {:.4f}'.format(names, epoch, epoch_loss))
def test(net, names, test_dataset, epoch, args):
#######################################################################
# if you are going to show the reconstruct shape, please using the following codes
#######################################################################
net.eval()
for it, (feats_patch, center_patch, coordinate_patch, face_patch, np_Fs, label, mesh_paths) in enumerate(
test_dataset):
faces = face_patch.cuda()
feats = feats_patch.cuda()
centers = center_patch.cuda()
Fs = np_Fs.cuda()
cordinates = coordinate_patch.cuda()
with torch.no_grad():
loss, masked_indices, unmasked_indices, pred_vertices_coordinates, cordinates = net(faces, feats, centers, Fs, cordinates)
save_results(masked_indices, unmasked_indices, pred_vertices_coordinates, cordinates, mesh_paths)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('mode', choices=['train', 'test'])
parser.add_argument('--name', type=str, required=True)
parser.add_argument('--checkpoint', type=str, default=None)
parser.add_argument('--lr_milestones', type=str, default=None)
parser.add_argument('--num_warmup_steps', type=str, default=None)
parser.add_argument('--depth', type=int, required=True)
parser.add_argument('--lr', type=float, default=1e-3)
parser.add_argument('--weight_decay', type=float, default=0)
parser.add_argument('--n_dropout', type=int, default=1)
parser.add_argument('--encoder_depth', type=int, default=6)
parser.add_argument('--decoder_depth', type=int, default=6)
parser.add_argument('--decoder_dim', type=int, default=512)
parser.add_argument('--decoder_num_heads', type=int, default=6)
parser.add_argument('--dim', type=int, default=384)
parser.add_argument('--weight', type=float, default=0.2)
parser.add_argument('--optim', type=str, default='adam')
parser.add_argument('--heads', type=int, required=True)
parser.add_argument('--patch_size', type=int, required=True)
parser.add_argument('--batch_size', type=int, default=32)
parser.add_argument('--n_epoch', type=int, default=500)
parser.add_argument('--max_epoch', type=int, default=300)
parser.add_argument('--dataroot', type=str, required=True)
parser.add_argument('--n_classes', type=int)
parser.add_argument('--no_center_diff', action='store_true')
parser.add_argument('--seed', type=int, default=None)
parser.add_argument('--n_worker', type=int, default=8)
parser.add_argument('--augment_scale', action='store_true')
parser.add_argument('--augment_orient', action='store_true')
parser.add_argument('--augment_deformation', action='store_true')
parser.add_argument('--channels', type=int, default=10)
parser.add_argument('--mask_ratio', type=float, default=0.25)
args = parser.parse_args()
mode = args.mode
dataroot = args.dataroot
# ========== Dataset ==========
augments = []
if args.augment_scale:
augments.append('scale')
if args.augment_orient:
augments.append('orient')
if args.augment_deformation:
augments.append('deformation')
if 'ShapeNet' in dataroot:
train_dataset = ShapeNetDataset(dataroot, train=True, augment=augments)
else:
train_dataset = ClassificationDataset(dataroot, train=True, augment=augments)
test_dataset = ClassificationDataset(dataroot, train=False)
print(len(test_dataset))
test_data_loader = data.DataLoader(test_dataset, num_workers=args.n_worker, batch_size=args.batch_size,
shuffle=True, pin_memory=True)
print(len(train_dataset))
train_data_loader = data.DataLoader(train_dataset, num_workers=args.n_worker, batch_size=args.batch_size,
shuffle=True, pin_memory=True)
# ========== Network ==========
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
net = Mesh_mae(masking_ratio=args.mask_ratio,
channels=args.channels,
num_heads=args.heads,
encoder_depth=args.encoder_depth,
embed_dim=args.dim,
decoder_num_heads=args.decoder_num_heads,
decoder_depth=args.decoder_depth,
decoder_embed_dim=args.decoder_dim,
patch_size=args.patch_size,
weight=args.weight
).to(device)
# ========== Optimizer ==========
if args.optim.lower() == 'adamw':
optim = optim.AdamW(net.parameters(), lr=args.lr, weight_decay=args.weight_decay)
if args.lr_milestones.lower() != 'none':
ms = args.lr_milestones
ms = ms.split()
ms = [int(j) for j in ms]
scheduler = MultiStepLR(optim, milestones=ms, gamma=0.1)
else:
scheduler = get_cosine_schedule_with_warmup(optim, num_warmup_steps=int(args.num_warmup_steps),
num_training_steps=args.max_epoch + 1)
print(scheduler)
# ========== MISC ==========
checkpoint_names = []
checkpoint_path = os.path.join('checkpoints', args.name)
os.makedirs(checkpoint_path, exist_ok=True)
if args.checkpoint.lower() != 'none':
net.load_state_dict(torch.load(args.checkpoint), strict=True)
train.best_loss = 999
train.best_epoch = 0
# ========== Start Training ==========
if args.mode == 'train':
for epoch in range(args.n_epoch):
print('epoch', epoch)
train(net, optim, scheduler, args.name, train_data_loader, epoch, args)
print('train finished')
else:
test(net, args.name, test_data_loader, 0, args)