search

sayedjm / scVI_comparison_Seurat_LUMC

4 stars 1 forks source link

Optimise hyperparameters n_layers #6

Closed mghuibregtse closed 10 months ago

mghuibregtse commented 10 months ago

Only change n_layers. Source: https://docs.scarches.org/en/latest/training_tips.html as source.

mghuibregtse commented 10 months ago

n_layers tested once ( as it takes an hour per run ). Conclusion: n_layers = 3 was optimal, assuming that closer to zero is the best for elbo and reconstruction loss Screenshot_1 Screenshot_3

mghuibregtse commented 10 months ago

n_layers tested once ( as it takes an hour per run ). Conclusion: n_layers = 3 was optimal, assuming that closer to zero is the best for elbo and reconstruction loss Screenshot_1 Screenshot_3

Update: the elbo and reconstruction loss are different for each run, a seed is needed. If unavailable, ignore it.

mghuibregtse commented 10 months ago

Tried multiple seeds: seed = 0 np.random.seed(seed) torch.manual_seed(seed) random.seed(seed) torch.cuda.manual_seed_all(seed)

however they were still different each time, also the n_latent number that changed was different