epoch=epoch,
model=model,
loss_fn=loss,
optimizer=optimizer,
dataloader=dataloader,
triple_generation=triple_generation,
geometric_model=args.geometric_model,
use_cuda=args.cuda,
log_interval=100,
vis=vis)
# model.eval()
model.ThetaRegression.eval()
results, val_time[epoch - 1] = test_fn(model=model,
metric='pck',
batch_size=args.batch_size,
dataset=dataset_val,
dataloader=dataloader_val,
do_aff=do_aff,
do_tps=do_tps,
args=args)
if do_aff:
val_pck[epoch - 1] = np.mean(results['aff']['pck'])
elif do_tps:
val_pck[epoch - 1] = np.mean(results['tps']['pck'])
train_lr[epoch - 1] = optimizer.param_groups[0]['lr']
# Visualization
if epoch % 5 == 0 or epoch == 1:
with torch.no_grad():
results_watch, theta_watch, theta_watch_inver, _ = test_watch(
args.start_epoch = GM_checkpoint['epoch']
best_val_pck = GM_checkpoint['best_val_pck']
train_loss = GM_checkpoint['train_loss']
val_pck = GM_checkpoint['val_pck']
train_time = GM_checkpoint['train_time']
val_time = GM_checkpoint['val_time']
model.FeatureExtraction.eval()
# Test pre-trained model before training
model.ThetaRegression.eval()
model.ThetaRegression2.eval()
with torch.no_grad():
test_fn(model=model,
metric='pck',
batch_size=args.batch_size,
dataset=dataset_val,
dataloader=dataloader_val,
dual=True,
args=args)
results_watch, theta_watch, thetai_watch, _ = test_watch(
model=model,
metric='pck',
batch_size=1,
dataset=dataset_watch,
dataloader=dataloader_watch,
dual=True,
args=args)
vis_fn_dual(vis=vis,
train_loss=train_loss,
val_pck=val_pck,
train_lr=train_lr,