mxnet-gan
mxnet-gan copied to clipboard
tqchen
MultiGPU enabled image generative models (GAN and DCGAN)
In the `GANModule`, grad information is explicitly saved and added again, which may require more memory cost. Is it possible to use `grad_req='add'` instead?
By just setting arguments on kernel and other relating arguments, we implement 3D conv and 3d deconv net. But I got trouble binding random number with Generator. Could anyone give...
Why the fix_gamma = True in the batch_norm. I find that fix_gamma = False (or affine=True) in pytorch implementations. [https://github.com/tqchen/mxnet-gan/blob/master/mxgan/ops.py#L25](url)
in `dcgan32x32` code there are: ``` net = mx.sym.Reshape(net, shape(-1, ngf*4, 4, 4) ``` I am so confused that why using a shape with negative value? and this function inputs...
Question 1: I managed to run the cifar10 and mnist example perfectly as you suggested, but failed to run my own. I fixed the shape to fit my data data_shape...
