Hello Thanks for your great work. I have question about line 360-387 in train.py. If we run the command in the doc directly, line 387 causes error because vgg_loss is not used when the iteration is less than 10000.

            if it < 10000 or FLAGS.dataset != "celebahq_128":
                loss = im_loss
            else:
                vgg_loss = loss_fn_vgg(im_negs[:, -1], im).mean()
                loss = vgg_loss  + 0.1 * im_loss

            loss.backward()
            if FLAGS.gpus > 1:
                average_gradients(models)

            [torch.nn.utils.clip_grad_norm_(model.parameters(), 10.0) for model in models]
            [optimizer.step() for optimizer in optimizers]
            [optimizer.zero_grad() for optimizer in optimizers]

            if it % FLAGS.log_interval == 0 and rank_idx == 0:
                loss = loss.item()
                energy_pos_mean = energy_pos.mean().item()
                energy_neg_mean = energy_neg.mean().item()
                energy_pos_std = energy_pos.std().item()
                energy_neg_std = energy_neg.std().item()

                kvs = {}
                kvs['loss'] = loss
                kvs['ml_loss'] = ml_loss.item()
                kvs['im_loss'] = im_loss.item()

                if FLAGS.dataset == "celebahq_128":
                    kvs['vgg_loss'] = vgg_loss.item()

Another question is that why do we add vgg_loss to im_loss when the iteration pass 10000? Is there any intuition or purpose behind this design? (Can we use vgg loss directly at the very beginning?) Thanks

Hi, thank you for presenting such an interesting paper and neat code. I am trying to reproduce the object-level decomposition on Tetris dataset. The decomposition result is not ideal when I use the same settings as the CLEVR dataset: python train.py --exp=tetris_local_decomp --num_steps=5 --step_lr=1000.0 --components=3 --dataset=tetris --cuda --train --batch_size=24 --latent_dim=16 --recurrent_model --pos_embed Would you mind providing more details of the experiment setup for object-level decomposition on Tetris? Thank you very much.

Hi Yilun, I am very interesting in your work. I find that the clver dataset you provide is different from https://github.com/deepmind/multi_object_datasets#clevr-with-masks or https://cs.stanford.edu/people/jcjohns/clevr/. Therefore, I can not download the groundtruth masks from these two links. And the groundtruth masks are also not provided in your link. Can you also provide me the the groundtruth masks of your dataset? Thanks, Tao