def init_dataset(self, ):
opts = self.opts
dataloader = behave_data_utils.behave_dataoader(opts.dataloader,
shuffle=True)
self.dataloader = dataloader
val_opts = deepcopy(opts)
val_opts.dataloader.sample_all_verts = True
val_opts.dataloader.batch_size = 2
self.val_dataloader = behave_data_utils.behave_dataoader(
val_opts.dataloader, shuffle=True)
return
def train(self):
opts = self.opts
dataloader = behave_data_utils.behave_dataoader(opts.dataloader)
num_epochs = opts.train.num_epochs
self.total_steps = total_steps = 0
self.epoch_iter = 0
for epoch in range(num_epochs):
self.epoch_iter = 0
self.epoch = epoch
if epoch > 500:
batch_norm_utils.turnNormOff(self.model)
for bx, batch in enumerate(dataloader):
model_inputs = self.set_input(batch)
self.forward(model_inputs)
self.backward()
total_steps += 1
self.total_steps = total_steps
self.epoch_iter += 1
if total_steps % opts.logging.log_freq == 0:
self.print_scalars()
scalars = self.get_current_scalars()
self.log_tb_scalars(step=total_steps,
train_scalars=scalars)
if total_steps % opts.logging.val_log_freq == 0:
visuals, val_scalars = self.create_visuals()
self.log_to_tensorboard(visuals[0], )
self.log_tb_scalars(step=self.total_steps,
val_scalars=val_scalars)
self.epoch += 1
if self.epoch % opts.logging.save_epoch_freq == 0:
self.save_network(self.model, "model", self.epoch)
def init_dataset(self, ):
opts = self.opts
dataloader = behave_data_utils.behave_dataoader(opts.dataloader,
shuffle=True)
self.dataloader = dataloader
return