-
Notifications
You must be signed in to change notification settings - Fork 1
/
finetune.py
227 lines (181 loc) · 10 KB
/
finetune.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
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
import os
import time
import numpy as np
import torch
import torch.nn as nn
import tqdm
import imageio
from dataset.obj_dataset import OBJDataset
from dataset.gso_dataset import GSODataset
from models.render_ray import render_rays
from models.render_image import render_single_image
from models.model import CodecNerfModel
from models.sample_ray import RaySamplerSingleImage, RaySamplerMultipleImages
from models.criterion import Criterion
import opt
from utils import (
img2mse, mse2psnr, img_HWC2CHW, img2psnr, set_seed,
get_cosine_schedule_with_warmup, ssim, msssim)
import pandas as pd
torch.set_float32_matmul_precision('high')
def finetune(args):
device = "cuda:{}".format(args.local_rank)
if args.data_type == 'obj':
dataset = OBJDataset(args.data_path, args.camera_path)
elif args.data_type == 'gso':
dataset = GSODataset(args.data_path)
data_loader = torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=False)
model = CodecNerfModel(args)
criterion = Criterion()
scalars_to_log = {}
for train_data in data_loader:
ray_samplert = RaySamplerMultipleImages(train_data, device, args.img_hw)
time0 = time.time()
featmaps, code_mb = model.encode(train_data['src_rgbs'], train_data['src_intrinsics'], None, train_data['src_w2c_mats'])
to_optimize0 = nn.Parameter(featmaps[0], requires_grad=False).to(device)
to_optimize1 = nn.Parameter(featmaps[1], requires_grad=False).to(device)
n_feature_maps = len(featmaps)
feat_dim_list = [featmaps[0].shape[2], featmaps[1].shape[2]]
feat_res_list = [featmaps[0].shape[-1], featmaps[1].shape[-1]]
del featmaps, model.feature_net
feat_coefs = []
for i in range(n_feature_maps):
m = nn.Parameter(torch.randn((3, args.trank, feat_res_list[i], feat_res_list[i])))
v = nn.Parameter(torch.randn((args.trank, feat_dim_list[i])))
feat_coefs.append(nn.init.normal_(m))
feat_coefs.append(nn.init.zeros_(v))
vm_model = torch.nn.ParameterList(feat_coefs).to(device)
vm_model.train()
params = model.get_param()
params += [{'params': vm_model, 'lr':args.lrate_feat}]
optim = torch.optim.Adam(params, eps=1e-15)
lr_sched = get_cosine_schedule_with_warmup(
optim, num_warmup_steps=100, num_training_steps=args.iter)
for i in range(args.iter+1):
m_0_ = vm_model[0].view(3, args.trank, -1)
m_1_ = vm_model[2].view(3, args.trank, -1)
v_0_ = vm_model[1]
v_1_ = vm_model[3]
vm_merged0 = torch.einsum('ijl,jk->ikl', m_0_, v_0_)
vm_merged1 = torch.einsum('ijl,jk->ikl', m_1_, v_1_)
vm_merged0 = vm_merged0.view(3, 32, feat_res_list[0], feat_res_list[0]).unsqueeze(1)
vm_merged1 = vm_merged1.view(3, 32, feat_res_list[1], feat_res_list[1]).unsqueeze(1)
vm_merged0 += to_optimize0
vm_merged1 += to_optimize1
ray_batch = ray_samplert.random_sample(args.N_rand)
ret = render_rays(ray_batch=ray_batch,
model=model,
featmaps=[vm_merged0, vm_merged1],
N_samples=args.N_samples,
inv_uniform=args.inv_uniform,
N_importance=args.N_importance,
det=args.det,
white_bkgd=args.white_bkgd,
)
optim.zero_grad()
loss, scalars_to_log = criterion(ret['outputs_coarse'], ray_batch, scalars_to_log)
if ret['outputs_fine'] is not None:
fine_loss, scalars_to_log = criterion(ret['outputs_fine'], ray_batch, scalars_to_log)
loss += fine_loss
loss.backward()
scalars_to_log['loss'] = loss.item()
optim.step()
lr_sched.step()
dt = time.time() - time0
if args.local_rank == 0:
if i % args.i_print == 0 or i < 10:
# write mse and psnr stats
mse_error = img2mse(ret['outputs_coarse']['rgb'], ray_batch['rgb']).item()
scalars_to_log['train/coarse-loss'] = mse_error
scalars_to_log['train/coarse-psnr-training-batch'] = mse2psnr(mse_error)
if ret['outputs_fine'] is not None:
mse_error = img2mse(ret['outputs_fine']['rgb'], ray_batch['rgb']).item()
scalars_to_log['train/fine-loss'] = mse_error
scalars_to_log['train/fine-psnr-training-batch'] = mse2psnr(mse_error)
try:
logstr = '{} step: {} '.format(train_data['scene_name'][0], i)
except:
logstr = '{} step: {} '.format(args.data_path.split('/')[-1], i)
for k in scalars_to_log.keys():
logstr += ' {}: {:.6f}'.format(k, scalars_to_log[k])
print(logstr)
print('each iter time {:.05f} seconds'.format(dt))
if i % 1000 == 0:
vm_model.eval()
model.switch_to_eval()
m_0_ = vm_model[0].view(3, args.trank, -1)
m_1_ = vm_model[2].view(3, args.trank, -1)
v_0_ = vm_model[1]
v_1_ = vm_model[3]
vm_merged0 = torch.einsum('ijl,jk->ikl', m_0_, v_0_)
vm_merged1 = torch.einsum('ijl,jk->ikl', m_1_, v_1_)
vm_merged0 = vm_merged0.view(3, 32, feat_res_list[0], feat_res_list[0]).unsqueeze(1)
vm_merged1 = vm_merged1.view(3, 32, feat_res_list[1], feat_res_list[1]).unsqueeze(1)
vm_merged0 += to_optimize0
vm_merged1 += to_optimize1
print('Logging a random validation view...')
try:
ren_folder = os.path.join(args.outdir, train_data['scene_name'][0], str(i))
except:
ren_folder = os.path.join(args.outdir, args.data_path.split('/')[-1], str(i))
print(f'Rendering to {ren_folder}')
os.makedirs(ren_folder, exist_ok=True)
data_input = dict(
tgt_intrinsic=train_data['all_intrinsics'][0][0:1],
depth_range=train_data['depth_range'],
)
render_poses = train_data['all_c2w_mats'][0]
view_indices = np.arange(0, len(render_poses), 1)
psnr_ = 0
ssim_ = 0
msssim_ = 0
with torch.no_grad():
for idx, pose in tqdm.tqdm(zip(view_indices, render_poses), total=len(view_indices)):
filename = os.path.join(ren_folder, f'{idx:06}.png')
data_input['tgt_c2w_mat'] = pose[None, :]
ray_sampler = RaySamplerSingleImage(data_input, device, args.img_hw, render_stride=1)
ray_batch = ray_sampler.get_all()
ret = render_single_image(ray_sampler=ray_sampler,
ray_batch=ray_batch,
model=model,
chunk_size=args.chunk_size,
N_samples=args.N_samples,
inv_uniform=args.inv_uniform,
N_importance=args.N_importance,
det=True,
white_bkgd=args.white_bkgd,
render_stride=1,
featmaps=[vm_merged0, vm_merged1])
check_ = ret['outputs_fine']['rgb']
check_ = torch.clamp(check_.detach().cpu(), 0.0, 1.0)
psnr_ += img2psnr(check_, train_data['all_rgbs'][0][idx])
ssim_ += ssim(check_, train_data['all_rgbs'][0][idx])
msssim_ += msssim(check_, train_data['all_rgbs'][0][idx])
rgb_im = img_HWC2CHW(ret['outputs_fine']['rgb'].detach().cpu())
rgb_im = torch.clamp(rgb_im, 0.0, 1.0)
rgb_im = rgb_im.permute([1, 2, 0]).cpu().numpy()
rgb_im = (rgb_im * 255.).astype(np.uint8)
imageio.imwrite(filename, rgb_im)
try:
name = train_data['scene_name'][0]
except:
name = args.data_path.split('/')[-1]
pd_data = {name : {'ITER': i, 'PSNR' : psnr_/len(render_poses), 'SSIM' : ssim_/len(render_poses),
'MSSSIM' : msssim_/len(render_poses),
"CODE" : code_mb}}
new_df = pd.DataFrame([(folder, metrics['ITER'], metrics['PSNR'], metrics['SSIM'], metrics['MSSSIM'],
metrics['CODE']) for folder, metrics in pd_data.items()],
columns=['FileName', 'ITER', 'PSNR', 'SSIM', 'MSSSIM', 'CODE'])
existing_csv_file_path = f"{i}_ft.csv"
if os.path.exists(existing_csv_file_path):
new_df.to_csv(existing_csv_file_path, index=False, mode='a', header=False)
print(f"Data appended to the existing CSV file at: {existing_csv_file_path}")
else:
new_df.to_csv(existing_csv_file_path, index=False)
vm_model.train()
model.switch_to_train()
if __name__ == '__main__':
parser = opt.config_parser()
args = parser.parse_args()
set_seed(args.seed)
finetune(args)