fix: correct decoder layer init in autoencoder notebooks (ae-basic)

[Mukller]

Bug Fix: Autoencoder decoder initialization uses wrong layer

In the Autoencoder.__init__() method, the DECODER section incorrectly initializes
self.linear_1 weights instead of self.linear_2. This is a copy-paste bug where
the encoder's weight initialization code was duplicated without updating the layer reference.

Before (wrong):

### DECODER
self.linear_2 = torch.nn.Linear(num_hidden_1, num_features)
self.linear_1.weight.detach().normal_(0.0, 0.1)  # ❌ wrong layer
self.linear_1.bias.detach().zero_()               # ❌ wrong layer

After (fixed):

### DECODER
self.linear_2 = torch.nn.Linear(num_hidden_1, num_features)
self.linear_2.weight.detach().normal_(0.0, 0.1)  # ✓ correct
self.linear_2.bias.detach().zero_()               # ✓ correct

Fixes encoder weights being initialized twice while the decoder is never initialized.
Closes #79

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[Mukller]

Code Review: fix decoder layer init in autoencoder notebooks

Summary

Fixes a copy-paste bug in the Autoencoder.__init__ method where the decoder layer (linear_2) weight and bias initialization was incorrectly calling self.linear_1.weight / self.linear_1.bias instead of self.linear_2.weight / self.linear_2.bias. As a result, the decoder weights were never receiving custom initialization — only the encoder weights were initialized twice.

Critical Issues

#	File	Issue	Severity
1	ae-basic.ipynb	Decoder init referenced self.linear_1 twice, leaving self.linear_2 with default PyTorch init instead of the intended custom normal init	🔴 Critical
2	ae-basic-with-rf.ipynb	Same copy-paste bug in the refactored version	🔴 Critical

What the Fix Does

Changes:

# Before (bug)
self.linear_2 = torch.nn.Linear(num_hidden_1, num_features)
self.linear_1.weight.detach().normal_(0.0, 0.1)  # wrong: linear_1 again
self.linear_1.bias.detach().zero_()               # wrong: linear_1 again

# After (correct)
self.linear_2 = torch.nn.Linear(num_hidden_1, num_features)
self.linear_2.weight.detach().normal_(0.0, 0.1)  # correct: linear_2
self.linear_2.bias.detach().zero_()               # correct: linear_2

What Looks Good

• The fix is minimal and surgical — no unrelated changes
• Both affected notebooks are corrected consistently

Verdict

Request Changes → now fixed. The initialization bug could silently affect training reproducibility since the decoder would use PyTorch's default Xavier init instead of the documented normal distribution init.