How to set hyperparameters to reproduce the results?

[XiHuYan]

Thanks for sharing the work!

I have trouble in reproducing the figure Fig.2C in your paper. Here are my reproduction steps: 1) the same preprocessing steps you proposed: use scanpy to finish LogNormalize(scale_factor=1e4), select hvgs by batch (n_hvgs=2000) 2) the default model and training parameters: clf_out = 25 learning_rate=1e-2, batch_size = 512, num_epoch=5, f_temp = 0.1 p_temp = 0.15

I tried to increase the training epochs, turn down learning rate or change the temperature parameters. But it didn't work at all, just as the picture shows. .

So, could you share the settings you produce the Fig.2C? thanks

[rpmccordlab]

In fact, this is not relevant to how parameters are set up, but to how batchnorm layer works in the model. Please try extract latent space with code below:

net.to(torch.device("cpu")) y_pred = np.zeros((data_tensor.shape[0],128)) batch_size=512 for j in range(data_tensor.shape[0]//batch_size+1): pred = net.encoder(data_tensor[jbatch_size:(j+1)batch_size,:]) pred = torch.Tensor.cpu(pred).detach().numpy() y_pred[jbatch_size:(j+1)batch_size,:]=pred

Try to forward data in multiple small batches, instead of a whole dataset at once.

[XiHuYan]

thanks a lot! that solved my problem.

After change the code for inference

net.to(torch.device("cpu"))
y_pred = np.zeros((X.shape[0],128))

batch_size=512
for j in range(X.shape[0]//batch_size+1):
    pred = net.encoder(X_all_tensor[j*batch_size:(j+1)*batch_size, :])
    pred = torch.Tensor.cpu(pred).detach().numpy()
    y_pred[j*batch_size:(j+1)*batch_size, :]=pred