Questions about the pretrained model.

[yizhouzhao]

Thank you @alexpashevich for sharing this great work. I am here for asking whether you would release your different versions of pretrained models you published in your paper and on Alfred leaderboard.

[yizhouzhao]

Dear @alexpashevich,

Good afternoon! Thanks again for your great work. This repository inspires me a lot for my research. I am facing two challenges and I need your suggestions. First, I tested your pretrained model (et_human_synth_pretrained.pth) and the results show on validation_unseen that

SR: 54/820 = 0.066
GC: 417/2118 = 0.197
PLW SR: 0.035
PLW GC: 0.109

Does this model use the 45k synthesized trajectories? Second, I have no idea how to synthesize human language instructions for the 45k trajectories. Could you provide me with any hints? Thanks.

[alexpashevich]

Hi @yizhouzhao,

Thank you for your interest in our work. As you have noticed, we released two models: E.T. trained without any additional data (et_human_pretrained.pth) and E.T. trained with language pre-training and joint training using additional 45K trajectories (et_human_synth_pretrained.pth). In both cases, the checkpoints were retrained to be compatible with the released version of the code. Both models should closely match the reported results however may not exactly reproduce them.

The human annotations are not available for the 45K trajectories but instead you can exploit synthetic representations. If you have troubles generating additional trajectories yourself, you can download trajectory jsons we used and render them.

wget http://pascal.inrialpes.fr/data2/apashevi/json_synth_2.1.0.tar.gz

Best,
Alex