| import torch
|
| import numpy as np
|
| import imageio
|
|
|
| def generate_rnd(H, sampler, shape, ema_imle, fname, logprint):
|
| mb = H.num_rows_visualize
|
| batches = []
|
| n_rows = mb
|
| temp_latent_rnds = torch.randn([mb, H.latent_dim], dtype=torch.float32).cuda()
|
| for t in range(H.num_rows_visualize):
|
| temp_latent_rnds.normal_()
|
| if(H.use_snoise == True):
|
| tmp_snoise = [s[:mb].normal_() for s in sampler.snoise_tmp]
|
| else:
|
| tmp_snoise = [s[:mb] for s in sampler.neutral_snoise]
|
|
|
| out = ema_imle(temp_latent_rnds, tmp_snoise)
|
| batches.append(sampler.sample_from_out(out))
|
|
|
| im = np.concatenate(batches, axis=0).reshape((n_rows, mb, *shape[1:])).transpose([0, 2, 1, 3, 4]).reshape(
|
| [n_rows * shape[1], mb * shape[2], 3])
|
|
|
| logprint(f'printing samples to {fname}')
|
| imageio.imwrite(fname, im)
|
|
|
