def train(opt, train_loader, m, criterion, optimizer, writer):
loss_logger = DataLogger()
acc_logger = DataLogger()
m.train()
norm_type = cfg.LOSS.get('NORM_TYPE', None)
train_loader = tqdm(train_loader, dynamic_ncols=True)
for i, (inps, labels, label_masks, _, bboxes) in enumerate(train_loader):
if isinstance(inps, list):
inps = [inp.cuda().requires_grad_() for inp in inps]
else:
inps = inps.cuda().requires_grad_()
labels = labels.cuda()
label_masks = label_masks.cuda()
output = m(inps)
if cfg.LOSS.get('TYPE') == 'MSELoss':
loss = 0.5 * criterion(output.mul(label_masks), labels.mul(label_masks))
acc = calc_accuracy(output.mul(label_masks), labels.mul(label_masks))
else:
loss = criterion(output, labels, label_masks)
acc = calc_integral_accuracy(output, labels, label_masks, output_3d=False, norm_type=norm_type)
if isinstance(inps, list):
batch_size = inps[0].size(0)
else:
batch_size = inps.size(0)
loss_logger.update(loss.item(), batch_size)
acc_logger.update(acc, batch_size)
optimizer.zero_grad()
loss.backward()
optimizer.step()
opt.trainIters += 1
# Tensorboard
if opt.board:
board_writing(writer, loss_logger.avg, acc_logger.avg, opt.trainIters, 'Train')
# Debug
if opt.debug and not i % 10:
debug_writing(writer, output, labels, inps, opt.trainIters)
# TQDM
train_loader.set_description(
'loss: {loss:.8f} | acc: {acc:.4f}'.format(
loss=loss_logger.avg,
acc=acc_logger.avg)
)
train_loader.close()
return loss_logger.avg, acc_logger.avg
def validate_gt(m, opt, cfg, heatmap_to_coord, criterion, batch_size=20):
loss_logger = DataLogger()
acc_logger = DataLogger()
gt_val_dataset = builder.build_dataset(cfg.DATASET.VAL, preset_cfg=cfg.DATA_PRESET, train=False)
eval_joints = gt_val_dataset.EVAL_JOINTS
gt_val_loader = torch.utils.data.DataLoader(
gt_val_dataset, batch_size=batch_size, shuffle=False, num_workers=20, drop_last=False)
kpt_json = []
m.eval()
for inps, labels, label_masks, img_ids, bboxes in tqdm(gt_val_loader, dynamic_ncols=True):
if isinstance(inps, list):
inps = [inp.cuda() for inp in inps]
else:
inps = inps.cuda()
output = m(inps)
##
labels = labels.cuda()
label_masks = label_masks.cuda()
loss = 0.5 * criterion(output.mul(label_masks), labels.mul(label_masks))
acc = calc_accuracy(output.mul(label_masks), labels.mul(label_masks))
pred = output.cpu().data.numpy()
assert pred.ndim == 4
pred = pred[:, eval_joints, :, :]
loss_logger.update(loss.item(), batch_size)
acc_logger.update(acc, batch_size)
for i in range(output.shape[0]):
bbox = bboxes[i].tolist()
pose_coords, pose_scores = heatmap_to_coord(pred[i][gt_val_dataset.EVAL_JOINTS], bbox)
keypoints = np.concatenate((pose_coords, pose_scores), axis=1)
keypoints = keypoints.reshape(-1).tolist()
data = dict()
data['bbox'] = bboxes[i].tolist()
data['image_id'] = img_ids[i]
data['score'] = float(np.mean(pose_scores) + np.max(pose_scores))
data['category_id'] = 1
data['keypoints'] = keypoints
kpt_json.append(data)
with open(os.path.join(opt.work_dir, 'test_gt_kpt.json'), 'w') as fid:
json.dump(kpt_json, fid)
#res = evaluate_mAP(os.path.join(opt.work_dir, 'test_gt_kpt.json'), ann_type='keypoints')
#return res['AP']
return loss_logger.avg, acc_logger.avg
def validate_gt(m, opt, cfg, heatmap_to_coord, batch_size=64):
joint_radius_mse = DataLogger()
gt_val_dataset = builder.build_dataset(cfg.DATASET.VAL,
preset_cfg=cfg.DATA_PRESET,
train=False)
eval_joints = gt_val_dataset.EVAL_JOINTS
gt_val_loader = torch.utils.data.DataLoader(gt_val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=opt.nThreads,
drop_last=False)
kpt_json = []
m.eval()
norm_type = cfg.LOSS.get('NORM_TYPE', None)
hm_size = cfg.DATA_PRESET.HEATMAP_SIZE
mse_loss = nn.MSELoss()
for index, (inps, labels, label_masks, joint_radius_gt, img_ids,
bboxes) in tqdm(enumerate(gt_val_loader), dynamic_ncols=True):
if opt.device.type != 'cpu':
if isinstance(inps, list):
inps = [inp.cuda() for inp in inps]
else:
inps = inps.cuda()
full_output = m(inps)
joints_map = full_output.joints_map
joints_radius = full_output.joints_radius
pred = joints_map
assert pred.dim() == 4
pred = pred[:, eval_joints, :, :]
for i in range(joints_map.shape[0]):
bbox = bboxes[i].tolist()
pose_coords, pose_scores = heatmap_to_coord(
pred[i][gt_val_dataset.EVAL_JOINTS],
bbox,
hm_shape=hm_size,
norm_type=norm_type)
keypoints = np.concatenate((pose_coords, pose_scores), axis=1)
keypoints = keypoints.reshape(-1).tolist()
data = dict()
data['bbox'] = bboxes[i].tolist()
data['image_id'] = int(img_ids[i])
data['score'] = float(np.mean(pose_scores) + np.max(pose_scores))
data['category_id'] = 1
data['keypoints'] = keypoints
kpt_json.append(data)
radius_masks = (label_masks[:, :, 0, 0] != 0) & (joint_radius_gt != -1)
joints_radius = joints_radius[:, eval_joints]
joint_radius_gt = joint_radius_gt[:, eval_joints]
joint_radius_gt = joint_radius_gt[radius_masks]
joints_radius = joints_radius[radius_masks]
joints_radius_error = mse_loss(joint_radius_gt, joints_radius.cpu())
joint_radius_mse.update(joints_radius_error, joint_radius_gt.shape[0])
with open(os.path.join(opt.work_dir, 'test_gt_kpt.json'), 'w') as fid:
json.dump(kpt_json, fid)
res = evaluate_mAP(os.path.join(opt.work_dir, 'test_gt_kpt.json'),
ann_type='keypoints',
ann_file=os.path.join(cfg.DATASET.VAL.ROOT,
cfg.DATASET.VAL.ANN))
return {
"map": res,
"radius_mse": joint_radius_mse.avg,
}
def train(opt, train_loader, m, criterion, optimizer, writer):
loss_logger = DataLogger()
acc_logger = DataLogger()
combined_loss = (cfg.LOSS.get('TYPE') == 'Combined')
m.train()
norm_type = cfg.LOSS.get('NORM_TYPE', None)
train_loader = tqdm(train_loader, dynamic_ncols=True)
for i, (inps, labels, label_masks, _, bboxes) in enumerate(train_loader):
if isinstance(inps, list):
inps = [inp.cuda().requires_grad_() for inp in inps]
else:
inps = inps.cuda().requires_grad_()
if isinstance(labels, list):
labels = [label.cuda() for label in labels]
label_masks = [label_mask.cuda() for label_mask in label_masks]
else:
labels = labels.cuda()
label_masks = label_masks.cuda()
output = m(inps)
if cfg.LOSS.get('TYPE') == 'MSELoss':
loss = 0.5 * criterion(output.mul(label_masks), labels.mul(label_masks))
acc = calc_accuracy(output.mul(label_masks), labels.mul(label_masks))
elif cfg.LOSS.get('TYPE') == 'Combined':
if output.size()[1] == 68:
face_hand_num = 42
else:
face_hand_num = 110
output_body_foot = output[:, :-face_hand_num, :, :]
output_face_hand = output[:, -face_hand_num:, :, :]
num_body_foot = output_body_foot.shape[1]
num_face_hand = output_face_hand.shape[1]
label_masks_body_foot = label_masks[0]
label_masks_face_hand = label_masks[1]
labels_body_foot = labels[0]
labels_face_hand = labels[1]
loss_body_foot = 0.5 * criterion[0](output_body_foot.mul(label_masks_body_foot), labels_body_foot.mul(label_masks_body_foot))
acc_body_foot = calc_accuracy(output_body_foot.mul(label_masks_body_foot), labels_body_foot.mul(label_masks_body_foot))
loss_face_hand = criterion[1](output_face_hand, labels_face_hand, label_masks_face_hand)
acc_face_hand = calc_integral_accuracy(output_face_hand, labels_face_hand, label_masks_face_hand, output_3d=False, norm_type=norm_type)
loss_body_foot *= 100
loss_face_hand *= 0.01
loss = loss_body_foot + loss_face_hand
acc = acc_body_foot * num_body_foot / (num_body_foot + num_face_hand) + acc_face_hand * num_face_hand / (num_body_foot + num_face_hand)
else:
loss = criterion(output, labels, label_masks)
acc = calc_integral_accuracy(output, labels, label_masks, output_3d=False, norm_type=norm_type)
if isinstance(inps, list):
batch_size = inps[0].size(0)
else:
batch_size = inps.size(0)
loss_logger.update(loss.item(), batch_size)
acc_logger.update(acc, batch_size)
optimizer.zero_grad()
loss.backward()
optimizer.step()
opt.trainIters += 1
# Tensorboard
if opt.board:
board_writing(writer, loss_logger.avg, acc_logger.avg, opt.trainIters, 'Train')
# Debug
if opt.debug and not i % 10:
debug_writing(writer, output, labels, inps, opt.trainIters)
# TQDM
train_loader.set_description(
'loss: {loss:.8f} | acc: {acc:.4f}'.format(
loss=loss_logger.avg,
acc=acc_logger.avg)
)
train_loader.close()
return loss_logger.avg, acc_logger.avg
def train(opt, train_loader, m, criterion, optimizer, writer):
loss_logger = DataLogger()
acc_logger = DataLogger()
m.train()
norm_type = cfg.LOSS.get('NORM_TYPE', None)
num_joints = cfg.DATA_PRESET.get('NUM_JOINTS',133)
train_branch = cfg.OTHERS.get('TRAIN_BRANCH',True)
train_loader = tqdm(train_loader, dynamic_ncols=True)
for i, (inps, labels, label_masks, _, bboxes) in enumerate(train_loader):
if isinstance(inps, list):
inps = [inp.cuda().requires_grad_() for inp in inps]
else:
inps = inps.cuda().requires_grad_()
out, feature = m(inps)
# train for finer hands
if train_branch:
out = m.module.forward_branch(out,feature,bboxes[:,1,:],bboxes[:,2,:])
labels = labels[:,:-68*2].cuda()
label_masks = label_masks[:,:-68*2].cuda()
else:
labels = labels[:,:133*2].cuda()
label_masks = label_masks[:,:133*2].cuda()
loss = criterion(out, labels, label_masks)
acc = calc_integral_accuracy(out, labels, label_masks, output_3d=False, norm_type=norm_type)
if isinstance(inps, list):
batch_size = inps[0].size(0)
else:
batch_size = inps.size(0)
loss_logger.update(loss.item(), batch_size)
acc_logger.update(acc, batch_size)
optimizer.zero_grad()
loss.backward()
optimizer.step()
opt.trainIters += 1
# Tensorboard
if opt.board:
board_writing(writer, loss_logger.avg, acc_logger.avg, opt.trainIters, 'Train')
# Debug
if opt.debug and not i % 10:
debug_writing(writer, output, labels, inps, opt.trainIters)
# TQDM
train_loader.set_description(
'loss: {loss:.8f} | acc: {acc:.4f}'.format(
loss=loss_logger.avg,
acc=acc_logger.avg)
)
train_loader.close()
return loss_logger.avg, acc_logger.avg