diff options
Diffstat (limited to 'training')
| -rwxr-xr-x | training/train_cnn.py | 20 |
1 files changed, 8 insertions, 12 deletions
diff --git a/training/train_cnn.py b/training/train_cnn.py index d974ce7..e1fd27b 100755 --- a/training/train_cnn.py +++ b/training/train_cnn.py @@ -271,7 +271,7 @@ class SimpleCNN(nn.Module): # Final layer (grayscale output) final_input = torch.cat([out, x_coords, y_coords, gray], dim=1) out = self.layers[-1](final_input) # [B,1,H,W] - out = torch.clamp(out, 0.0, 1.0) # Clip to [0,1] + out = torch.sigmoid(out) # Map to [0,1] with smooth gradients return out.expand(-1, 3, -1, -1) @@ -336,7 +336,7 @@ def generate_layer_shader(output_path, num_layers, kernel_sizes): else: f.write(f" else if (params.layer_index == {layer_idx}) {{\n") f.write(f" let gray_out = {conv_fn}(txt, smplr, uv, uniforms.resolution, gray, weights_layer{layer_idx});\n") - f.write(f" // gray_out in [0,1] (clamped to match PyTorch training)\n") + f.write(f" // gray_out in [0,1] (sigmoid activation)\n") f.write(f" result = vec4<f32>(gray_out, gray_out, gray_out, 1.0);\n") f.write(f" return mix(original_raw, result, params.blend_amount); // [0,1]\n") f.write(f" }}\n") @@ -452,7 +452,7 @@ def generate_conv_final_function(kernel_size, output_path): with open(output_path, 'a') as f: f.write(f"\n// Final layer: 7→1 channel (vec4-optimized)\n") f.write(f"// Assumes 'tex' is already normalized to [-1,1]\n") - f.write(f"// Output clamped to [0,1] to match PyTorch training\n") + f.write(f"// Output uses sigmoid activation to match PyTorch training\n") f.write(f"fn cnn_conv{k}x{k}_7to1(\n") f.write(f" tex: texture_2d<f32>,\n") f.write(f" samp: sampler,\n") @@ -479,7 +479,7 @@ def generate_conv_final_function(kernel_size, output_path): f.write(f" }}\n") f.write(f" }}\n\n") - f.write(f" return clamp(sum, 0.0, 1.0);\n") + f.write(f" return 1.0 / (1.0 + exp(-sum));\n") f.write(f"}}\n") @@ -631,12 +631,10 @@ def train(args): with open(conv_path, 'r') as f: content = f.read() - # Generate 7to1 final layer with clamp (all kernel sizes) + # Generate 7to1 final layer with sigmoid (all kernel sizes) if f"cnn_conv{ks}x{ks}_7to1" not in content: generate_conv_final_function(ks, conv_path) - print(f"Added 7to1 variant with clamp to {conv_path}") - elif "clamp(sum, 0.0, 1.0)" not in content: - print(f"Warning: {conv_path} has 7to1 but missing clamp - manual fix needed") + print(f"Added 7to1 variant with sigmoid to {conv_path}") print("Training complete!") @@ -687,12 +685,10 @@ def export_from_checkpoint(checkpoint_path, output_path=None): with open(conv_path, 'r') as f: content = f.read() - # Generate 7to1 final layer with clamp (all kernel sizes) + # Generate 7to1 final layer with sigmoid (all kernel sizes) if f"cnn_conv{ks}x{ks}_7to1" not in content: generate_conv_final_function(ks, conv_path) - print(f"Added 7to1 variant with clamp to {conv_path}") - elif "clamp(sum, 0.0, 1.0)" not in content: - print(f"Warning: {conv_path} has 7to1 but missing clamp - manual fix needed") + print(f"Added 7to1 variant with sigmoid to {conv_path}") print("Export complete!") |