Transfer learning on the 3D classification example

[addcue]

In the example for 3D classifications (https://github.com/Project-MONAI/tutorials/blob/main/3d_classification/ignite/densenet_training_array.py), we import a DenseNet-121 model with 3 spatial dimensions. How does this model become 3D? Aren't the original ImageNet images 2D?

[KumoLiu]

Hi @addcue, you can set the spatial_dims to 2 as well, and yes the original ImageNet images are 2D, but the default pretrained arg in the network is False. You can set it to True with spatial_dims equal to 3.

MONAI/monai/networks/nets/densenet.py

Line 173 in 960249f

spatial_dims: int,

MONAI/monai/networks/nets/densenet.py

Line 304 in 960249f

pretrained: bool = False,

Hope it can help you, thanks!

[addcue]

Yes, I was wondering how that works though. Since the model was trained on 2D images from ImageNet, how can we just change the model to 3D? And what exactly changes about the model?

[KumoLiu]

Hi @addcue, I think we can use 3d Convolution and Pooling layer, right?

[addcue]

But we can retain the weights from the 2D model?

[KumoLiu]

You can only set pretrained=True when spatial_dims=2.