loading scArches from local and train with new data about scarches

[M0hammadL]

Hi,

I have not seen this problem, does this happen when you train query data or also when you train the reference?

I just tried to train a model and save the model and restore the model and then add query data. However, I did not get such an error :

`network = sca.models.scArches(task_name='pancreas_reference',
x_dimension=adata.shape[1],
z_dimension=10,
architecture=[128, 128],
gene_names=adata.var_names.tolist(),
conditions=reference_batch_labels,
alpha=0.001,
loss_fn='nb',
model_path="./models/scArches/",
seed = 20
)

network.train(reference_adata,
n_epochs=10,
condition_key=condition_key,
batch_size=128,)

network.restore_model_weights()

new_network = sca.operate(network,
new_task_name="pancreas_query",
new_conditions=target_conditions)

new_network.train(query_adata,
condition_key=condition_key,
batch_size=128,
n_epochs=40)

`

[notimenocall]

Thank you very much for testing this. I just tried to a couple things and it seems like the issue is related to the "loss_fn" option. If loss_fn = 'nb', it works fine. If loss_fn = "sse", the issue came up: the first network.train with the reference data works (31k cells), but the 2nd training with the query failed (2.7k cells).

[notimenocall]

There may also be a bug when running with loss_fn = "zinb". network.train() returns an error:
'Activation' object has no attribute '__name__'