tb_hist.plot_weights_hist()
tb_hist.plot_grads_hist()
tb_hist.plot_bias_hist()
tb_hist.plot_outputs_hist()
# validation loss
with torch.no_grad():
running_val_loss = 0.0
running_val_steps = 0.0
for data in tqdm(valLoader, desc='val-step'):
input_ids = data['input_ids'].long().to(device)
segments = data['segments'].long().to(device)
targets = data['targets'].float().to(device)
outputs = model(input_ids=input_ids,
token_type_ids=segments)
targets = targets.view(-1, 1) # match output shape
val_loss = criterion(outputs, targets)
running_val_loss += val_loss.item()
running_val_steps += 1.0
writer.add_sclar("val_loss",
running_val_loss / running_val_steps,
interval_count)
running_loss = 0.0
running_steps = 0.0
interval_count += 1
print("Training complete.")
