Spaces:
Sleeping
Sleeping
File size: 14,048 Bytes
99a05f0 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 |
from utils.loss import sem_loss_function, class_loss_function, pixel_anchoring_function
import torch
import os
import time
class TrainStepper():
def __init__(self, deco_model, context, learning_rate, loss_weight, pal_loss_weight, device):
self.device = device
self.model = deco_model
self.context = context
if self.context:
self.optimizer_sem = torch.optim.Adam(params=list(self.model.encoder_sem.parameters()) + list(self.model.decoder_sem.parameters()),
lr=learning_rate, weight_decay=0.0001)
self.optimizer_part = torch.optim.Adam(
params=list(self.model.encoder_part.parameters()) + list(self.model.decoder_part.parameters()), lr=learning_rate,
weight_decay=0.0001)
self.optimizer_contact = torch.optim.Adam(
params=list(self.model.encoder_sem.parameters()) + list(self.model.encoder_part.parameters()) + list(
self.model.cross_att.parameters()) + list(self.model.classif.parameters()), lr=learning_rate, weight_decay=0.0001)
if self.context: self.sem_loss = sem_loss_function().to(device)
self.class_loss = class_loss_function().to(device)
self.pixel_anchoring_loss_smplx = pixel_anchoring_function(model_type='smplx').to(device)
self.pixel_anchoring_loss_smpl = pixel_anchoring_function(model_type='smpl').to(device)
self.lr = learning_rate
self.loss_weight = loss_weight
self.pal_loss_weight = pal_loss_weight
def optimize(self, batch):
self.model.train()
img_paths = batch['img_path']
img = batch['img'].to(self.device)
img_scale_factor = batch['img_scale_factor'].to(self.device)
pose = batch['pose'].to(self.device)
betas = batch['betas'].to(self.device)
transl = batch['transl'].to(self.device)
has_smpl = batch['has_smpl'].to(self.device)
is_smplx = batch['is_smplx'].to(self.device)
cam_k = batch['cam_k'].to(self.device)
gt_contact_labels_3d = batch['contact_label_3d'].to(self.device)
has_contact_3d = batch['has_contact_3d'].to(self.device)
if self.context:
sem_mask_gt = batch['sem_mask'].to(self.device)
part_mask_gt = batch['part_mask'].to(self.device)
polygon_contact_2d = batch['polygon_contact_2d'].to(self.device)
has_polygon_contact_2d = batch['has_polygon_contact_2d'].to(self.device)
# Forward pass
if self.context:
cont, sem_mask_pred, part_mask_pred = self.model(img)
else:
cont = self.model(img)
if self.context:
loss_sem = self.sem_loss(sem_mask_gt, sem_mask_pred)
loss_part = self.sem_loss(part_mask_gt, part_mask_pred)
valid_contact_3d = has_contact_3d
loss_cont = self.class_loss(gt_contact_labels_3d, cont, valid_contact_3d)
valid_polygon_contact_2d = has_polygon_contact_2d
if self.pal_loss_weight > 0 and (is_smplx == 0).sum() > 0:
smpl_body_params = {'pose': pose[is_smplx == 0], 'betas': betas[is_smplx == 0],
'transl': transl[is_smplx == 0],
'has_smpl': has_smpl[is_smplx == 0]}
loss_pix_anchoring_smpl, contact_2d_pred_rgb_smpl, _ = self.pixel_anchoring_loss_smpl(cont[is_smplx == 0],
smpl_body_params,
cam_k[is_smplx == 0],
img_scale_factor[
is_smplx == 0],
polygon_contact_2d[
is_smplx == 0],
valid_polygon_contact_2d[
is_smplx == 0])
# weigh the smpl loss based on the number of smpl sample
loss_pix_anchoring = loss_pix_anchoring_smpl * (is_smplx == 0).sum() / len(is_smplx)
contact_2d_pred_rgb = contact_2d_pred_rgb_smpl
else:
loss_pix_anchoring = 0
contact_2d_pred_rgb = torch.zeros_like(polygon_contact_2d)
if self.context: loss = loss_sem + loss_part + self.loss_weight * loss_cont + self.pal_loss_weight * loss_pix_anchoring
else: loss = self.loss_weight * loss_cont + self.pal_loss_weight * loss_pix_anchoring
if self.context:
self.optimizer_sem.zero_grad()
self.optimizer_part.zero_grad()
self.optimizer_contact.zero_grad()
loss.backward()
if self.context:
self.optimizer_sem.step()
self.optimizer_part.step()
self.optimizer_contact.step()
if self.context:
losses = {'sem_loss': loss_sem,
'part_loss': loss_part,
'cont_loss': loss_cont,
'pal_loss': loss_pix_anchoring,
'total_loss': loss}
else:
losses = {'cont_loss': loss_cont,
'pal_loss': loss_pix_anchoring,
'total_loss': loss}
if self.context:
output = {
'img': img,
'sem_mask_gt': sem_mask_gt,
'sem_mask_pred': sem_mask_pred,
'part_mask_gt': part_mask_gt,
'part_mask_pred': part_mask_pred,
'has_contact_2d': has_polygon_contact_2d,
'contact_2d_gt': polygon_contact_2d,
'contact_2d_pred_rgb': contact_2d_pred_rgb,
'has_contact_3d': has_contact_3d,
'contact_labels_3d_gt': gt_contact_labels_3d,
'contact_labels_3d_pred': cont}
else:
output = {
'img': img,
'has_contact_2d': has_polygon_contact_2d,
'contact_2d_gt': polygon_contact_2d,
'contact_2d_pred_rgb': contact_2d_pred_rgb,
'has_contact_3d': has_contact_3d,
'contact_labels_3d_gt': gt_contact_labels_3d,
'contact_labels_3d_pred': cont}
return losses, output
@torch.no_grad()
def evaluate(self, batch):
self.model.eval()
img_paths = batch['img_path']
img = batch['img'].to(self.device)
img_scale_factor = batch['img_scale_factor'].to(self.device)
pose = batch['pose'].to(self.device)
betas = batch['betas'].to(self.device)
transl = batch['transl'].to(self.device)
has_smpl = batch['has_smpl'].to(self.device)
is_smplx = batch['is_smplx'].to(self.device)
cam_k = batch['cam_k'].to(self.device)
gt_contact_labels_3d = batch['contact_label_3d'].to(self.device)
has_contact_3d = batch['has_contact_3d'].to(self.device)
if self.context:
sem_mask_gt = batch['sem_mask'].to(self.device)
part_mask_gt = batch['part_mask'].to(self.device)
polygon_contact_2d = batch['polygon_contact_2d'].to(self.device)
has_polygon_contact_2d = batch['has_polygon_contact_2d'].to(self.device)
# Forward pass
initial_time = time.time()
if self.context: cont, sem_mask_pred, part_mask_pred = self.model(img)
else: cont = self.model(img)
time_taken = time.time() - initial_time
if self.context:
loss_sem = self.sem_loss(sem_mask_gt, sem_mask_pred)
loss_part = self.sem_loss(part_mask_gt, part_mask_pred)
valid_contact_3d = has_contact_3d
loss_cont = self.class_loss(gt_contact_labels_3d, cont, valid_contact_3d)
valid_polygon_contact_2d = has_polygon_contact_2d
if self.pal_loss_weight > 0 and (is_smplx == 0).sum() > 0: # PAL loss only on 2D contacts in HOT which only has SMPL
smpl_body_params = {'pose': pose[is_smplx == 0], 'betas': betas[is_smplx == 0], 'transl': transl[is_smplx == 0],
'has_smpl': has_smpl[is_smplx == 0]}
loss_pix_anchoring_smpl, contact_2d_pred_rgb_smpl, _ = self.pixel_anchoring_loss_smpl(cont[is_smplx == 0],
smpl_body_params,
cam_k[is_smplx == 0],
img_scale_factor[
is_smplx == 0],
polygon_contact_2d[
is_smplx == 0],
valid_polygon_contact_2d[
is_smplx == 0])
# weight the smpl loss based on the number of smpl samples
contact_2d_pred_rgb = contact_2d_pred_rgb_smpl
loss_pix_anchoring = loss_pix_anchoring_smpl * (is_smplx == 0).sum() / len(is_smplx)
else:
loss_pix_anchoring = 0
contact_2d_pred_rgb = torch.zeros_like(polygon_contact_2d)
if self.context: loss = loss_sem + loss_part + self.loss_weight * loss_cont + self.pal_loss_weight * loss_pix_anchoring
else: loss = self.loss_weight * loss_cont + self.pal_loss_weight * loss_pix_anchoring
if self.context:
losses = {'sem_loss': loss_sem,
'part_loss': loss_part,
'cont_loss': loss_cont,
'pal_loss': loss_pix_anchoring,
'total_loss': loss}
else:
losses = {'cont_loss': loss_cont,
'pal_loss': loss_pix_anchoring,
'total_loss': loss}
if self.context:
output = {
'img': img,
'sem_mask_gt': sem_mask_gt,
'sem_mask_pred': sem_mask_pred,
'part_mask_gt': part_mask_gt,
'part_mask_pred': part_mask_pred,
'has_contact_2d': has_polygon_contact_2d,
'contact_2d_gt': polygon_contact_2d,
'contact_2d_pred_rgb': contact_2d_pred_rgb,
'has_contact_3d': has_contact_3d,
'contact_labels_3d_gt': gt_contact_labels_3d,
'contact_labels_3d_pred': cont}
else:
output = {
'img': img,
'has_contact_2d': has_polygon_contact_2d,
'contact_2d_gt': polygon_contact_2d,
'contact_2d_pred_rgb': contact_2d_pred_rgb,
'has_contact_3d': has_contact_3d,
'contact_labels_3d_gt': gt_contact_labels_3d,
'contact_labels_3d_pred': cont}
return losses, output, time_taken
def save(self, ep, f1, model_path):
# create model directory if it does not exist
os.makedirs(os.path.dirname(model_path), exist_ok=True)
if self.context:
torch.save({
'epoch': ep,
'deco': self.model.state_dict(),
'f1': f1,
'sem_optim': self.optimizer_sem.state_dict(),
'part_optim': self.optimizer_part.state_dict(),
'contact_optim': self.optimizer_contact.state_dict()
},
model_path)
else:
torch.save({
'epoch': ep,
'deco': self.model.state_dict(),
'f1': f1,
'sem_optim': self.optimizer_sem.state_dict(),
'part_optim': self.optimizer_part.state_dict(),
'contact_optim': self.optimizer_contact.state_dict()
},
model_path)
def load(self, model_path):
print(f'~~~ Loading existing checkpoint from {model_path} ~~~')
checkpoint = torch.load(model_path)
self.model.load_state_dict(checkpoint['deco'], strict=True)
if self.context:
self.optimizer_sem.load_state_dict(checkpoint['sem_optim'])
self.optimizer_part.load_state_dict(checkpoint['part_optim'])
self.optimizer_contact.load_state_dict(checkpoint['contact_optim'])
epoch = checkpoint['epoch']
f1 = checkpoint['f1']
return epoch, f1
def update_lr(self, factor=2):
if factor:
new_lr = self.lr / factor
if self.context:
self.optimizer_sem = torch.optim.Adam(params=list(self.model.encoder_sem.parameters()) + list(self.model.decoder_sem.parameters()),
lr=new_lr, weight_decay=0.0001)
self.optimizer_part = torch.optim.Adam(
params=list(self.model.encoder_part.parameters()) + list(self.model.decoder_part.parameters()), lr=new_lr, weight_decay=0.0001)
self.optimizer_contact = torch.optim.Adam(
params=list(self.model.encoder_sem.parameters()) + list(self.model.encoder_part.parameters()) + list(
self.model.cross_att.parameters()) + list(self.model.classif.parameters()), lr=new_lr, weight_decay=0.0001)
print('update learning rate: %f -> %f' % (self.lr, new_lr))
self.lr = new_lr |