A re-train normal CNN on scaled data beats deform-conv.

[KaiOtter]

First, appreciate for your work which is easy to use and read. In scaled_mninst.py, the normal CNN model is trained on origin data and then, tested on scaled one. It shows a bad acc of 60% (In my running). Then, you fine-tune a deform-conv on the scaled data and its accuracy is much better. However, I tried to re-train this trained CNN model on scaled data and the result confuses me definitely. It gets 96% on origin test and 98% on scaled data. Well, this experiment can not prove the effectiveness of deform-conv layers.

[wushuang01]

@Kaidy-Boom I meet the same problem,I think the reason is that the mnist is too easy for them to learn.I delete the first two Deformable layers and only keep the last one seems to have a better result.

[oeway]

@Kaidy-Boom @wushuang01 Thanks for your report, indeed that's not a good demo of the effectiveness. To clarify, my implementation is just a port of the tensorflow version by Felix Lau as mentioned in README, I discussed with him briefly in the past about a similar question, and I did not have time to do more for that. For a complete prove of the effectiveness, I think would be better to check the original paper.

[wushuang01]

@oeway Thanks for your reply.appreciate for your work which is helpful for me.