Size mismatch in SigmoidCrossEntropyLoss

[jason718]

I met this bug and have no idea why it happenes

Traceback (most recent call last):
File "src/train_cls.py", line 121, in
s.run(args.nit)
File "/home/jason/egomotion/experiment/voc2007_cls/voc-classification/src/solver.py", line 90, in run
s.step(1)
File "/home/jason/egomotion/experiment/voc2007_cls/voc-classification/src/python_layers.py", line 54, in forward
lZ = np.log(1+np.exp(-np.abs(f))) * mask
ValueError: operands could not be broadcast together with shapes (10,20) (16,20)

[jason718]

It seems like if I set batch size to some number rather than 10, this problem will happen.

[polo5]

@jason718 I just realized that contrary to Tensorflow the batch size is hardcoded in caffe, so we need to set the batch size equal to the one specified in whatever prototxt we are providing.

For bvlc_reference_caffenet that is indeed 10. For Rotnet it's 4 etc.

[jason718]

@polo5 does the batch size matters in terms of final results?

[polo5]

To be honnest I haven't been able to get any reasonable mAP score yet so I can't say. I get a weird error (see #7 ) which when fixed gives me a 3% mAP. I've tried the models: bvlc_reference_caffenet, alexnet, rotnet. Any luck on your side?

[jason718]

I used this long time ago, and I have published a paper: https://jason718.github.io/project/cvpr18/main.html

[jianghaojun]

It's a mismatch between batchsize( set by -bs ) and the first dimension in input layer (set in yournetwork.prototxt).