Speed up this very minimal code? tensors to numpy to image

vision
1

I’m taking in latent vecs and generating images, pretty straight forward, but getting less than half the FPS as my Tensorflow model. Can someone help me speed this up?

def convert(model, inputs):
    z = inputs['z']
    jzf = [float(i) for i in z]
    jzft = torch.FloatTensor(jzf)
    jzftr = jzft.reshape([1, 512])
    latents = jzftr.cuda()
    truncation = inputs['truncation']

    img = G(latents, None, truncation)
    img = (img.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(torch.uint8)
    return {'image': img[0].cpu().numpy()}
2

Hi,

What is the type of z?

Also creating the latent would be more efficient as: latents = torch.tensor(jzf, dtype=torch.float, device="cuda").view(1, 512).