search

VisionLearningGroup / taskcv-2017-public

168 stars 36 forks source link

Something wrong with the dilation10_gta5_frontend.caffemodel #3

Open islamtashfiq opened 7 years ago

islamtashfiq commented 7 years ago

I used pycaffe to observe the weights of the deconvolution layer with deconv_weights = net.params['ctx_upsample'][0].data and all the unique variables I get is 0 value. I think the whole network is zero value because I get 0 image as a result. Could you check the model?

jhoffman commented 7 years ago

Caffe has a bilinear weight filler parameter which we use to load the kernel for bilinear upsampling in this layer. If you have zeros perhaps you are using the wrong weight filler.

Running the val script should produce a deploy prototxt where you should find the following layer: layer { name: "upsample" type: "Deconvolution" bottom: "final" top: "upsample" param { lr_mult: 0.0 decay_mult: 0.0 } convolution_param { num_output: 19 bias_term: false pad: 4 kernel_size: 16 group: 19 stride: 8 weight_filler { type: "bilinear" } } }