Zilong Huang
Zilong Huang
@yearing1017 It's for deep supervision. Please refer to the [loss function](https://github.com/speedinghzl/CCNet/blob/master/utils/criterion.py#L42).
Hi @prakhar123pradhan, the visualization of the attention map? This [related issue](https://github.com/speedinghzl/CCNet/issues/3) may help you generate the figure.
The figure is the visualization of one attention map or several attention maps? Maybe you could try to visualize the first attention map (only criss-cross path) using shows_attention1 function. To...
@zhouqianyu0918 @Alva-2020 Maybe the [pure-python](https://github.com/speedinghzl/CCNet/tree/pure-python) or [>Pytorch 1.5](https://github.com/speedinghzl/CCNet/issues/94) could solve your problem.
@Thatboy7 You could drop the INF if you train the model again. The overlapping pixels have a minor effect on the final performance.
Hi @yhonker, we haven't conducted such experiments. I can not give a definitive answer.
Please use another [branch](https://github.com/speedinghzl/CCNet/tree/pytorch-1.1).
Please check your cityscape image and label paths are accessible.
The memory usage of RCCA could be obtained by the memory usage of CCNet minus the memory usage of baseline (CCNet without RCCA module). The nvidia-smi tool is used to...
@XiXiRuPan The gamma is initialized to zero makes the RCCA module an identity connection at the beginning of training. It could make the training smoothly. The gamma is a learnable...