Ilia Karmanov

Sorry for not being clear - I was using this was py35 and tensorflow-gpu==1.4 and trying to figure out if the error was due to the script being for py2...

@rky0930 thanks for the reply! My GPUs are free and available (11GB on 2xK80), the small batch was just to see if it would work at all. Can you run...

@chenyuntc Do you mean if we have a binary classification then we keep the RPN but then can chop off the ROIHead (since conditional on a box containing an object...

@chenyuntc Interesting, thank you. Just to play, I have removed the head and now loss is just: `losses = [rpn_loc_loss, rpn_cls_loss]` However, the issue comes with the scores for RPN:...

Sorry just to ask further about your comment: > But it's better to keep the ROIHEad, since the RoIPooling Crop the corresponding area from the input images. Anchor calculated in...

@chenyuntc Sorry to triple post but just wanted to bring this quote up from [Kaiming](https://arxiv.org/pdf/1607.07032.pdf) > More importantly, RPN as a stand-alone pedestrian detector achieves an MR > of 14.9%...

Sure I will do that tomo morning - it's mainly a change on the data-loader side

@niam06 I was wondering if you managed to get higher val acc since you posted? I seem to have similar issues with Kinetics.

Thanks very much for your reply! Unfortunately this is still the case for me despite loading pretrained weights. I was wondering whether you could share your training/validation accuracy for a...

I thought it could be more helpful to leave this log below (r50_a2_b4_f8x2). Initiating from pre-trained weights and then training 50 epochs with cosine-annealing from LR of 0.01 for a...