Avg_pool or Max_pool for logits?

[zeal-up]

Thanks for your job! It really help me so much. It seems that in the deepmind I3D version, they use average pool along the time dimension for the last logits. But in your codes, you use max_pool for classification. Does it has some reasonable explanations?

And, in the original paper, it said: We train the model using 64-frame snippets and test using the whole videos, averaging predictions temporally. Should I send the whole video to the network and average over the temporal dimension when testing?

[ciwei123]

max_pool The author code uses avg pool, where is the max pool?