wrong input_dims when use_img about const HOT 3 CLOSED

Hi,

Thanks for your interest. Can you tell me which dataset are you using, and what dimension have you changed to? It looks a bit strange that the ARI of the major training is even lower than pretraining.

from const.

I was using spatialLIBD/151673 just as conST_cluster.ipynb did. And I cropped the tissue image myself according to the spatial positions and extracted the features with run_mae_extract_feature.py.

Because I tried to train the model with MAE features, there were several input dimensions of the layers different from merely using gene expression, including:

• self.latent_dim
• self.cluster_layer
• self.fc1
• self.fc2
• self.disc_c
• self.disc

For details, please see: compare. Basically, it was because z = torch.cat((feat_x, gnn_z), 1) had changed to z = torch.cat((feat_x, gnn_z, feat_img), 1) when using image, and the following layers' dimensions changed accordingly.

Instead of using eval_resolution = res_search_fixed_clus(adata_conST, n_clusters) to detect the best resolution in every epoch, I set resolution constantly to be 1, so the results were little lower.
But the problem of lower ARI after major training does exist. When I ran the conST_cluster.ipynb with use_pretrained=False and commented out the major training line:

conST_net.pretraining()
# conST_net.major_training()

to compared their results, I found merely pretraining was 0.499, which was higher than 0.439 of both pretraining and major training.

I am wondering if I have set something wrong.

Thank you!

from const.

It's a bit difficult for me to tell from this code directly, but I suggest trying to debug it in two steps.

First, can you obtain similar results without using image features? For data spatialLIBD/151673, there is a slight improvement when using image features from MAE but not much, because as you can see from the histology images, the image patches look similar among spots. So, if you can't obtain similar results from this step, there is probably something wrong with your initial settings, environment, etc. From your result, I suspect there is something wrong with this step, because I didn't come across a situation where the performance is worse than the pretraining after the major training.

If the first step is okay, then you can try to reduce the input dimension of the image features from 748 to a lower dimension, e.g. 100 with PCA (or even smaller), and see how the performance goes. By reducing the proportion of image features, you can check if the image features are successfully extracted by MAE.

P.S. Note that you may not get exact same results every time you run the experiment, which is due to the Cuda non-deterministic characterics of pytorch-geometrics.

from const.