elyxlz / umap_pytorch Goto Github PK

First of all, many thanks for the pytorch implementation and library! It has saved me so many headaches and is very easy to use out of the box. Looking forward to any future updates :)
My question is regarding the conv_encoder Module, or (more generally) creating a similar CNN encoder.
To your knowledge, is there currently a way to pass in a custom encoder class into a PUMAP object without making any drastic changes to the source code?
I have a custom encoder that I pass into my PUMAP object, after which I transform the input data. As you'll notice, I tried to use the conv_encoder as a template, but at the cost of making some unsavory updates like reshaping the input tensor in the forward method or setting a custom input length of the first Linear layer. I've had to make these changes to accommodate my use case.

class umap_encoder(nn.Module):
    def __init__(self, embeddings_length, n_components=20):
        super().__init__()
        self.encoder = nn.Sequential(
            nn.Conv2d(
                in_channels=1, out_channels=64, kernel_size=3, stride=2, padding=1,
            ),
            nn.Conv2d(
                in_channels=64, out_channels=128, kernel_size=3, stride=2, padding=1,
            ),
            nn.Flatten(),
            nn.Linear(embeddings_length*4, 512),
            nn.ReLU(),
            nn.Linear(512, 512),
            nn.ReLU(),
            nn.Linear(512, n_components)
        ).cuda()
    def forward(self, X):
        return self.encoder(reshape(X, (1, X.shape[0], X.shape[1])))

Here is theinstantiated PUMAP object. In this case emb.shape[1] = 384.

PUMAP(
    encoder=umap_encoder(emb.shape[1], num_components), 
    n_neighbors=15, 
    min_dist=0.0, 
    metric='cosine', 
    n_components=num_components, 
    random_state=42
)

If this is a matter of changing the order of operations, or if I need to remove/add a layer I'm happy to take any feedback. Thanks for your time