How to generate image without GPU?

[lz1998]

How to generate image without GPU?

[nurpax]

Generating images on the CPU is possible but not with default settings.

Here's how you will need to modify generate.py to enable CPU generation:

diff --git a/generate.py b/generate.py
index f7f9619..0e487ed 100755
--- a/generate.py
+++ b/generate.py
@@ -79,7 +79,7 @@ def generate_images(
     """
 
     print('Loading networks from "%s"...' % network_pkl)
-    device = torch.device('cuda')
+    device = torch.device('cpu')
     with dnnlib.util.open_url(network_pkl) as f:
         G = legacy.load_network_pkl(f)['G_ema'].to(device) # type: ignore
 
@@ -116,7 +116,7 @@ def generate_images(
     for seed_idx, seed in enumerate(seeds):
         print('Generating image for seed %d (%d/%d) ...' % (seed, seed_idx, len(seeds)))
         z = torch.from_numpy(np.random.RandomState(seed).randn(1, G.z_dim)).to(device)
-        img = G(z, label, truncation_psi=truncation_psi, noise_mode=noise_mode)
+        img = G(z, label, truncation_psi=truncation_psi, noise_mode=noise_mode, force_fp32=True)
         img = (img.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(torch.uint8)
         PIL.Image.fromarray(img[0].cpu().numpy(), 'RGB').save(f'{outdir}/seed{seed:04d}.png')
 

Basically, run it on the CPU device and use only fp32.

Note that there is a marked performance difference between running on the CPU. On my machine, generating an image takes roughly 20 ms whereas the same code on the CPU takes 1500 ms, ie., 75 times slower.

[lz1998]

Generating images on the CPU is possible but not with default settings.

Here's how you will need to modify generate.py to enable CPU generation:

diff --git a/generate.py b/generate.py
index f7f9619..0e487ed 100755
--- a/generate.py
+++ b/generate.py
@@ -79,7 +79,7 @@ def generate_images(
     """
 
     print('Loading networks from "%s"...' % network_pkl)
-    device = torch.device('cuda')
+    device = torch.device('cpu')
     with dnnlib.util.open_url(network_pkl) as f:
         G = legacy.load_network_pkl(f)['G_ema'].to(device) # type: ignore
 
@@ -116,7 +116,7 @@ def generate_images(
     for seed_idx, seed in enumerate(seeds):
         print('Generating image for seed %d (%d/%d) ...' % (seed, seed_idx, len(seeds)))
         z = torch.from_numpy(np.random.RandomState(seed).randn(1, G.z_dim)).to(device)
-        img = G(z, label, truncation_psi=truncation_psi, noise_mode=noise_mode)
+        img = G(z, label, truncation_psi=truncation_psi, noise_mode=noise_mode, force_fp32=True)
         img = (img.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(torch.uint8)
         PIL.Image.fromarray(img[0].cpu().numpy(), 'RGB').save(f'{outdir}/seed{seed:04d}.png')
 

Basically, run it on the CPU device and use only fp32.

Note that there is a marked performance difference between running on the CPU. On my machine, generating an image takes roughly 20 ms whereas the same code on the CPU takes 1500 ms, ie., 75 times slower.

Why not add an param to choose generate with GPU or CPU? Can I create a PR?

[CrackerHax]

Basically, run it on the CPU device and use only fp32.

What about projecting images in latent space?

[CrackerHax]

Basically, run it on the CPU device and use only fp32.

What about projecting images in latent space?

project.py line 98ish

        # Synth images from opt_w.
        w_noise = torch.randn_like(w_opt) * w_noise_scale
        ws = (w_opt + w_noise).repeat([1, G.mapping.num_ws, 1])
     -   synth_images = G.synthesis(ws, noise_mode='const')
     +   synth_images = G.synthesis(ws, noise_mode='const', force_fp32=True)

line 163:

    # Load networks.
    print('Loading networks from "%s"...' % network_pkl)
   - device = torch.device('cuda')
   + device = torch.device('cpu')