Ilia Karmanov

Could be to do with the invalid file-names, e.g. https://github.com/LossNAN/I3D-Tensorflow/blob/master/list/ucf_list/train.list#L160 I had to remember these lines (I wanted to use the original split to match accuracy)

Regarding the folder-structure, imagine you have a video: `/largedata/i3d/videos/brush_hair/brushing_hair_brush_hair_f_nm_np2_ba_goo_2.avi ` You want to have: ``` /largedata/i3d/videos/brush_hair/brushing_hair_brush_hair_f_nm_np2_ba_goo_2/i/frame0001.jpg /largedata/i3d/videos/brush_hair/brushing_hair_brush_hair_f_nm_np2_ba_goo_2/i/frame0002.jpg /largedata/i3d/videos/brush_hair/brushing_hair_brush_hair_f_nm_np2_ba_goo_2/i/frame0003.jpg /largedata/i3d/videos/brush_hair/brushing_hair_brush_hair_f_nm_np2_ba_goo_2/x/frame0001.jpg /largedata/i3d/videos/brush_hair/brushing_hair_brush_hair_f_nm_np2_ba_goo_2/x/frame0002.jpg /largedata/i3d/videos/brush_hair/brushing_hair_brush_hair_f_nm_np2_ba_goo_2/x/frame0003.jpg /largedata/i3d/videos/brush_hair/brushing_hair_brush_hair_f_nm_np2_ba_goo_2/y/frame0001.jpg /largedata/i3d/videos/brush_hair/brushing_hair_brush_hair_f_nm_np2_ba_goo_2/y/frame0002.jpg /largedata/i3d/videos/brush_hair/brushing_hair_brush_hair_f_nm_np2_ba_goo_2/y/frame0003.jpg ``` Then I think you just...

The extension can be anything, it's read in like so ``` lines = open(filename, 'r') lines = list(lines) line = lines[index].strip('\n').split() ```

You are missing the RGB channels, e.g. (256, 342, 3). If you have one channel then unsqueeze an extra dimension.

A step in tensorflow is a forward (+ back) pass; whereas an epoch is however many steps (depending on batch-size) to get through all your data once. The next_btach_start value...

Oh, is it because in the tensorflow model they only care about classifying the whole clip without any localisation so they just average the logits (global average pooling): `averaged_logits =...