def forward_g(scene,
his_traj,
targets,
model_g,
model_d,
optimizer,
scheduler,
omega=1.0,
epsilon=1.0):
model_g.train()
preds, conf, context, z_mean, z_var = model_g(scene, his_traj)
traj_fake = utils.multi2single(preds, targets, conf, mode='best')
score_fake = model_d(traj_fake.permute(1, 0, 2), context)
# 判别loss + nll_loss + vae_loss
g_loss = utils.g_loss(score_fake)
# nll_loss = utils.pytorch_neg_multi_log_likelihood_batch(targets, preds, conf)
vae_loss, ade_loss = cvae.loss_cvae(targets, preds, conf, z_mean, z_var)
# loss = g_loss + nll_loss * omega + vae_loss * epsilon
# loss = g_loss + nll_loss * omega + vae_loss * epsilon
loss = g_loss + vae_loss * epsilon
scheduler.step()
optimizer.zero_grad()
loss.backward()
optimizer.step()
return loss, vae_loss, ade_loss, preds, conf
def forward_g(scene,
his_traj,
targets,
model_g,
model_d,
optimizer,
scheduler,
omega=1.0,
epsilon=1.0):
model_g.train()
preds, conf, context = model_g(scene, his_traj)
traj_fake = utils.multi2single(preds, targets, conf, mode='best')
score_fake = model_d(traj_fake.permute(1, 0, 2), context)
# 判别loss + nll_loss + ade_loss
g_loss = utils.g_loss(score_fake)
# nll_loss = utils.pytorch_neg_multi_log_likelihood_batch(targets, preds, conf)
min_l2_loss = utils._average_displacement_error(targets,
preds,
conf,
mode='best')
# loss = g_loss + nll_loss * omega + l2_loss * epsilon
nll_loss = 0.0
loss = g_loss + nll_loss * omega + min_l2_loss * epsilon
scheduler.step()
optimizer.zero_grad()
loss.backward()
optimizer.step()
return loss, nll_loss, min_l2_loss, preds, conf
def forward_d(scene, his_traj, targets, model_g, model_d, optimizer,
scheduler):
model_d.train()
preds, confidences, context, _, _ = model_g(scene, his_traj)
traj_fake = utils.multi2single(preds, targets, confidences, mode='best')
score_fake = model_d(traj_fake.permute(1, 0, 2), context)
score_real = model_d(targets.permute(1, 0, 2), context)
loss = utils.d_loss(score_real, score_fake)
scheduler.step()
optimizer.zero_grad()
loss.backward()
optimizer.step()
return loss
mininterval=5.)
for j in valid_progress_bar:
try:
data_valid = next(tr_it_valid)
except StopIteration:
tr_it_valid = iter(valid_dataloader)
data_valid = next(tr_it_valid)
scene_valid = data_valid[0].to(device)
scene_valid = scene_valid.permute(0, 3, 2, 1)
his_traj_valid = data_valid[3].to(device)
his_traj_valid = his_traj_valid.permute(1, 0, 2)
targets_valid = data_valid[4].to(device)
pred_pixel, conf, context, z_mean, z_var = generator(
scene_valid.float(), his_traj_valid.float())
traj_fake_valid = utils.multi2single(pred_pixel,
targets_valid.float(),
conf,
mode='best')
score_fake = discriminator(
traj_fake_valid.permute(1, 0, 2), context)
g_loss_valid = utils.g_loss(score_fake)
# nll_loss_valid = utils.pytorch_neg_multi_log_likelihood_batch(targets_valid, pred_pixel, conf)
vae_loss_valid, min_l2_loss_valid = cvae.loss_cvae(
targets_valid, pred_pixel, conf, z_mean, z_var)
# loss = g_loss + nll_loss * omega + l2_loss * epsilon
# valid_loss = g_loss_valid + nll_loss_valid * omega + vae_loss_valid * epsilon
valid_loss = g_loss_valid + vae_loss_valid * epsilon
# camera frame to world frame(meter)
pred = torch.zeros_like(pred_pixel)
for batch_index in range(pred_pixel.shape[0]):
for modality in range(pred_pixel.shape[1]):
for pos_index in range(pred_pixel.shape[2]):