Xingang Pan

1. Check this https://pytorch.org/docs/stable/nn.html#crossentropyloss 2. Yes. 3. I haven't try this, and its hard to say. I would recommend inserting SCNN right after conv layers, since this would save computations.

@xiaoshuliu spatial cross entropy loss is to apply cross entropy loss for each spatial location. The time reported is only for the SCNN part.

@aquariusnk My implementation and https://github.com/cardwing/Codes-for-Lane-Detection/blob/master/SCNN-Tensorflow/lane-detection-model/encoder_decoder_model/vgg_encoder.py are the same. If you print the provided vgg.t7 model, you would get:  In (43)(4) we use 128 channels in order to save memory....

1. Remember to use weight sharing for each direction. 2. When initializing SCNN, you may use a variance smaller than normal for numerical stability. For example, I use a variance...

@ytzhao The SCNN model is based on DeepLab-LargeFOV VGG version, not ResNet. My code is modified from fb.resnet.torch because that code is well organized. But the model is based on...

I have provided a script for generating SCNN caffe prototxt file: https://github.com/XingangPan/SCNN-prototxt-generator. You can use http://ethereon.github.io/netscope/#/editor to visualize the architecture. Hope it would be helpful.

I have uploaded the VGG16 model at https://drive.google.com/open?id=1DvLzlwe3Jtz_ILPstYTh1ZyDfvrg0A7i

@jcdubron The weights of layers after conv5 are random initialized as they are different from original vgg.

Hi DanlanChen. It seems to be an image read issue. Firstly, make sure that the dataset has the right path architecture as described in the README. Secondly, if the problem...

@chenyiman It seems that the input image is not successfully loaded. Please make sure the dataset is downloaded at the right path as described in README. You may print the...