H-Hui2277
commented
1 year ago The encoder is in the fcmae model to pretrain the ConvNeXtV2 model. If you are training your tasks by pretrained weights in fcame, you just need to use the encoder weights to init the model. Or you shuld use the fine-tune trained models.
The official re-training weights: model: OrderedDict downsample_layers.0.0.bias: Tensor with shape (128,) downsample_layers.0.0.weight: Tensor with shape (128, 3, 4, 4) downsample_layers.0.1.bias: Tensor with shape (128,) downsample_layers.0.1.weight: Tensor with shape (128,) downsample_layers.1.0.bias: Tensor with shape (128,) downsample_layers.1.0.weight: Tensor with shape (128,) downsample_layers.1.1.bias: Tensor with shape (256,) downsample_layers.1.1.weight: Tensor with shape (256, 128, 2, 2) downsample_layers.2.0.bias: Tensor with shape (256,) downsample_layers.2.0.weight: Tensor with shape (256,) downsample_layers.2.1.bias: Tensor with shape (512,) downsample_layers.2.1.weight: Tensor with shape (512, 256, 2, 2) downsample_layers.3.0.bias: Tensor with shape (512,) downsample_layers.3.0.weight: Tensor with shape (512,) downsample_layers.3.1.bias: Tensor with shape (1024,) downsample_layers.3.1.weight: Tensor with shape (1024, 512, 2, 2) ..... The model parameters: mask_token: Tensor with shape (1, 512, 1, 1) encoder.downsample_layers.0.0.weight: Tensor with shape (128, 3, 4, 4) encoder.downsample_layers.0.0.bias: Tensor with shape (128,) encoder.downsample_layers.0.1.weight: Tensor with shape (128,) encoder.downsample_layers.0.1.bias: Tensor with shape (128,) encoder.downsample_layers.1.0.ln.weight: Tensor with shape (128,) encoder.downsample_layers.1.0.ln.bias: Tensor with shape (128,) encoder.downsample_layers.1.1.kernel: Tensor with shape (4, 128, 256) encoder.downsample_layers.1.1.bias: Tensor with shape (1, 256) encoder.downsample_layers.2.0.ln.weight: Tensor with shape (256,) encoder.downsample_layers.2.0.ln.bias: Tensor with shape (256,) encoder.downsample_layers.2.1.kernel: Tensor with shape (4, 256, 512) encoder.downsample_layers.2.1.bias: Tensor with shape (1, 512) encoder.downsample_layers.3.0.ln.weight: Tensor with shape (512,) encoder.downsample_layers.3.0.ln.bias: Tensor with shape (512,) encoder.downsample_layers.3.1.kernel: Tensor with shape (4, 512, 1024) encoder.downsample_layers.3.1.bias: Tensor with shape (1, 1024)