mohsenfayyaz89 / stfcn Goto Github PK

hi MohsenFayyaz:
I use camvid data, the inputdata size is T * C * H * W is 3 * 3 * 720 * 960.and use "FCN_Alex_LSTM.prototxt".
when I train the net ,I meet a problem.the layer name"plated" output count don't match input count.
the input shape is 3 * 4096 * 17 * 24 but the out shape is 3 * 4096 * 49 * 1.
I change the input data size 3 * 3 * 227 * 227 ,there still has problem.the layer name"plated" , input shape is 3 * 4096 * 1 * 1 but the out shape is 3 * 4096 * 49 * 1.
I would like to ask is the size of the input image wrong? Or other problems?

I read the paper again ,I found the input size is 400 * 400.Solve the problem

hi MohsenFayyaz：
when I use you net, I meet a problem.In the layer"reshape",input"lstm0" shape is 3 * 4096 * 100.
output "lstm0" shape is 3 * 4096 * 1 (because the shape param is dim：0 dim:0 dim: 1).
but the output count don't match the input count.
can you give me some advice about this?

Hi @MohsenFayyaz89 @hajizadeh ,
I am new to LSTMS and have trouble understanding how to create HDF5Data from a video containing frames.Could you please point to any training code or repo that uses the similar way to create data/train.txt for camvid? I am unable to understand how the "clip" must be added as well. Since T=3,should the clip be 011011011.... and so on?

Is it possible to get the pretrained model?

hi MohsenFayyaz89:
I would like to try your experiment.
what does “clip” contain?
and the "data" I think is T * C * H * W .if i use the camvid，T=3(I get this from your paper)，C= 3，H= 720 W= 960.is this right?

@MohsenFayyaz89 Do you have code (e.g. a python script) that generates the long prototxt files?

can you offer the trainLSTM.txt ? tks!