writer.add_graph(mdl, args) executed without error but cannot find graphs on Tensorboard

[todpole3]

I am able to use TensorboardX to view scalar values without any problem. Then I decided to use the add_graph function of the same SummaryWriter to visualize the computation graph.

After a few rounds of try and error I was able to execute the add_graph function successfully. However, my Tensorboard was not updated with a view of the computation graph.

I am using Pytorch v1.0.0.

I wonder what did I miss and if not, what is the correct way to access the saved graph from Tensorboard.

[lanpa]

how did you use the function? On Sat, Jan 19, 2019 at 8:32 AM Victoria X Lin [email protected] wrote:

I am able to use TensorboardX to view scalar values without any problem. Then I decided to use the add_graph function of the same SummaryWriter to visualize the computation graph.

After a few rounds of try and error I was able to execute the add_graph function successfully. However, my Tensorboard was not updated with a view of the computation graph.

I wonder what did I miss and if not, what is the correct way to access the saved graph from Tensorboard.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/lanpa/tensorboardX/issues/340, or mute the thread https://github.com/notifications/unsubscribe-auth/AB6ZS9sgbTNLh5N7qPIJYXh2YWTUASeIks5vEmedgaJpZM4aI2yn .

-- send from my phone

[todpole3]

This is a simple workflow that trains a sequence-to-sequence model. And this is how I'm using the function.

        writer = SummaryWriter(log_dir=get_log_dir('tensorboard', self.args))

        for epoch_id in range(self.start_epoch, self.num_epochs):
            print('Epoch {}'.format(epoch_id))
            # Update learning rate scheduler
            lr_scheduler.step()
            writer.add_scalar('learning_rate/{}'.format(self.dataset), lr_scheduler.get_lr()[0], epoch_id)

            # Update model parameters
            self.train()
            self.batch_size = self.train_batch_size
            batch_losses = []
            
            for _ in tqdm(range(num_batches)):
                self.optim.zero_grad()

                loss = self.loss(mini_batch)
                loss['loss_value'].backward()
                if self.grad_norm > 0:
                    clip_grad_norm_(self.parameters(), self.grad_norm)
                
                self.optim.step()

                batch_losses.append(loss['printable_loss'])
                writer.add_graph(self.mdl(formatted_batch[0], formatted_batch[1][0]), formatted_batch)

            # Check training statistics
            stdout_msg = 'Epoch {}: average training loss = {}'.format(epoch_id, np.mean(batch_losses))
            writer.add_scalar('cross_entropy_loss/{}'.format(self.dataset), np.mean(batch_losses), epoch_id)

            print(stdout_msg)
            self.save_checkpoint(checkpoint_id=epoch_id, epoch_id=epoch_id)```

[lanpa]

add_graph should be only called once per SummaryWriter. Is there any warning message?

[todpole3]

No, there wasn't. Should I simply use add_graph after the writer was created?

[lanpa]

Yes, please check https://github.com/lanpa/tensorboardX/blob/master/examples/demo_graph.py

[RonakDedhiya]

Hello, I am getting similar issue. I am using pytorch == 0.3.0 and tensorboardx == 1.1

I have followed demo_graph.py, add_graph() works without any error and also I have called it once. I am able to view other scalar value in tensorboard but I am unable to see any graph

[todpole3]

Echo the message above. The issue remained unsolved for me.

[lanpa]

@todpole3 I just released tensorboardX v 1.8. Would you have a try on that?