⚡ 🔥

Correct duplicated scaling (scale only once)
Correct kld_warmup in training_loop.py:
   - kld_w as a fraction of beta, no dependence on num_latent

src/move/tasks/encode_data.py

@@ -59,12 +59,6 @@ def encode_data(config: DataConfig):
         )
         fig.savefig(fig_path)
 
-        values, mask_1d = preprocessing.scale(values)
-        names = names[mask_1d]
-        logger.debug(f"Columns with zero variance: {np.sum(~mask_1d)}")
-        io.dump_names(interim_data_path / f"{dataset_name}.txt", names)
-        np.save(interim_data_path / f"{dataset_name}.npy", values)
-
         # Plotting the value distribution for all continuous datasets:
         fig = plot_value_distributions(values)
         fig_path = str(output_path / f"Value_distribution_{dataset_name}.png")

src/move/training/training_loop.py

@@ -72,13 +72,11 @@ def training_loop(
     counter = 0
 
     kld_weight = 0.0
-    kld_rate = 20 / len(kld_warmup_steps)
-    kld_multiplier = 1 + kld_rate
+
 
     for epoch in range(1, num_epochs + 1):
         if epoch in kld_warmup_steps:
-            kld_weight = 0.05 * kld_multiplier
-            kld_multiplier += kld_rate
+            kld_weight += 1 / len(kld_warmup_steps)
 
         if epoch in batch_dilation_steps:
             train_dataloader = dilate_batch(train_dataloader)