configuration_resnet.py

from transformers import PretrainedConfig, AutoConfig

class ResnetConfig(PretrainedConfig):
    model_type = "resnet"  # Custom model identifier

    def __init__(
        self,
        num_channels=3,   # Number of input channels (3 for RGB)
        num_classes=1000,  # Number of output classes (e.g., for ImageNet)
        depth=50,         # Depth of the ResNet model (e.g., ResNet-50)
        block_type="bottleneck",  # Type of block (e.g., bottleneck for ResNet-50)
        stem_width=32,    # Width of the stem layer
        stem_type="deep",  # Type of stem layer
        avg_down=True,    # Whether to use average pooling in downsampling
        layers=None,      # Number of layers per stage (e.g., [3, 4, 6, 3] for ResNet-50)
        cardinality=1,  # ✅ Added default value
        base_width=64,  # ✅ Added default value
        **kwargs          # Catch any additional arguments
    ):
        super().__init__(**kwargs)
        self.num_channels = num_channels
        self.num_classes = num_classes
        self.depth = depth
        self.block_type = block_type
        self.stem_width = stem_width
        self.stem_type = stem_type
        self.avg_down = avg_down
        self.layers = layers or [3, 4, 6, 3]  # Default to ResNet-50 structure
        self.cardinality = cardinality  # ✅ Fix
        self.base_width = base_width  # ✅ Fix

resnet50d_config = ResnetConfig(block_type="bottleneck", stem_width=32, stem_type="deep", avg_down=True)
resnet50d_config.save_pretrained("custom-resnet")