896 half-pretrained!

configs/pretrained_config.yaml

@@ -10,25 +10,23 @@ parameters:
       RGB_ENCODER:
         ENCODER: "facebook/dinov3-vitl16-pretrain-lvd1689m"  # DINOv3 encoder model name (HuggingFace format)
         IMAGE_SIZE: 896  # Input image size (height and width in pixels)
-        RETURN_SELECTED_LAYERS: [3, 6, 9, 12]  # Transformer layer indices to extract features from (0-indexed)
+        RETURN_SELECTED_LAYERS: [3, 6, 9 , 12]  # Transformer layer indices to extract features from (0-indexed)
         RGB_ENCODER_LR: 0.0  # Learning rate for RGB encoder (0.0 = frozen, must be explicitly set)
       DECODERS:
         diffuse:
-          USE_FILM: False  # Enable FiLM (Feature-wise Linear Modulation) conditioning in decoder
           FEATURE_DIM: 1024  # Feature dimension for decoder (should match encoder output)
           REASSEMBLE_OUT_CHANNELS: [768,1024,1536,2048]  # Output channels for each decoder stage (DPT-style reassembly)
           REASSEMBLE_FACTORS: [4.0, 2.0, 1.0, 0.5]  # Spatial upsampling factors for each stage
           READOUT_TYPE: "ignore"  # Readout type for DPT decoder ("ignore", "project", etc.)
-          # FROM_PRETRAINED: "weights/rgb_decoder.pth"  # Path to pretrained decoder weights (optional)
+          FROM_PRETRAINED: "weights/decoder_896.pth"  # Path to pretrained decoder weights (optional)
           USE_BN: False  # Use batch normalization in decoder
           DROPOUT: 0.1  # Dropout rate in decoder layers
           OUTPUT_IMAGE_SIZE: [896,896]  # Output image resolution [height, width]
           OUTPUT_CHANNELS: 3  # Number of output channels (3 for RGB diffuse image)
-          DECODER_LR: 1.0e-5 # Custom learning rate for decoder (0.0 = frozen, 1.0 = same as base LR)
+          DECODER_LR: 0.0 # Custom learning rate for decoder (0.0 = frozen, 1.0 = same as base LR)
           NUM_FUSION_BLOCKS_TRAINABLE: 1  # Number of fusion blocks to train (0-4, null = train all if DECODER_LR != 0)
           TRAIN_RGB_HEAD: True  # Whether to train RGB head (true/false, null = train if DECODER_LR != 0)
         highlight:
-          USE_FILM: False  # Enable FiLM conditioning in highlight decoder
           FEATURE_DIM: 1024  # Feature dimension for highlight decoder
           REASSEMBLE_OUT_CHANNELS: [96,192,384,768]  # Output channels for each decoder stage
           REASSEMBLE_FACTORS: [4.0, 2.0, 1.0, 0.5]  # Spatial upsampling factors for each stage
@@ -42,11 +40,11 @@ parameters:
       TOKEN_INPAINTER:
         TOKEN_INPAINTER_CLASS: "TokenInpainter_Prior"  # Token inpainter class name
         TOKEN_INPAINTER_MODULE: "token_inpainters"  # Module name to import token inpainter from
-        # FROM_PRETRAINED: "weights/token_inpainter.pth"  # Path to pretrained token inpainter weights 
-        TOKEN_INPAINTER_LR: 1.0e-5  # Learning rate for token inpainter (can differ from base LR)
+        # FROM_PRETRAINED: "weights/token_inpainter.pth"  # Path to pretrained token inpainter weights (optional)
+        TOKEN_INPAINTER_LR: 1.0e-4  # Learning rate for token inpainter (can differ from base LR)
         DEPTH: 6  # Number of transformer blocks
         HEADS: 16  # Number of attention heads
-        DROP: 0 # Dropout rate 
+        DROP: 0.05 # Dropout rate 
         USE_POSITIONAL_ENCODING: True  # Enable 2D sinusoidal positional encodings
         USE_FINAL_NORM: True  # Enable final LayerNorm before output projection
         USE_LOCAL_PRIOR: True  # Blend local mean prior for masked seeds
@@ -54,6 +52,4 @@ parameters:
         LOCAL_PRIOR_KERNEL: 5  # Kernel size for local prior blending (> 1)
         SEED_NOISE_STD: 0.02  # Standard deviation of noise added to masked seeds during training  
   INPAINT_MASK_DILATION:
-    value: 1  # Dilation kernel size (pixels) for inpaint mask - Must be odd
-  USE_TORCH_COMPILE:  # Enable PyTorch 2.0 torch.compile for faster training (experimental)
-    value: False
+    value: 3  # Dilation kernel size (pixels) for inpaint mask - Must be odd