fix: only unsqueeze hidden state for single-layer unidirectional decoder

[Mukller]

Bug Fix: Incorrect hidden state shape in inference() for bidirectional/multi-layer RNN

When using a bidirectional or multi-layer decoder, inference() already reshapes
the hidden state via .view(hidden_factor, batch_size, hidden_size) to produce
the correct (num_layers * num_directions, batch, hidden_size) tensor.

However, .unsqueeze(0) is then applied unconditionally, adding an extra
leading dimension and producing the wrong shape (1, hidden_factor, batch, hidden_size).

GRU/LSTM expects (num_layers * num_directions, batch, hidden_size) — the extra
dimension causes a RuntimeError: Expected hidden size (2, N, H), got [1, 2, N, H].

Before (wrong — unsqueeze applied regardless of bidirectional/num_layers):

if self.bidirectional or self.num_layers > 1:
    hidden = hidden.view(self.hidden_factor, batch_size, self.hidden_size)

hidden = hidden.unsqueeze(0)   # ❌ creates (1, factor, batch, hidden)

After (fixed — unsqueeze only for simple single-layer unidirectional case):

if self.bidirectional or self.num_layers > 1:
    hidden = hidden.view(self.hidden_factor, batch_size, self.hidden_size)
else:
    hidden = hidden.unsqueeze(0)   # ✓ only adds dim when needed

Closes #35

[Mukller]

Code Review: fix unsqueeze for single-layer unidirectional decoder

Summary

Fixes an unconditional hidden.unsqueeze(0) call that was applied even when the decoder is bidirectional or has multiple layers — cases where hidden has already been reshaped to (hidden_factor, batch_size, hidden_size) by the preceding view() call. Applying unsqueeze(0) after the reshape would produce an incorrect 4D tensor, causing a dimension mismatch in the LSTM decoder.

Critical Issues

#	File	Line	Issue	Severity
1	model.py	~130	hidden.unsqueeze(0) was called unconditionally, adding an extra dimension even when bidirectional/multi-layer hidden state was already reshaped	🔴 Critical

What the Fix Does

# Before (bug)
if self.bidirectional or self.num_layers > 1:
    hidden = hidden.view(self.hidden_factor, batch_size, self.hidden_size)
hidden = hidden.unsqueeze(0)  # always runs — wrong for bidi/multi-layer

# After (correct)
if self.bidirectional or self.num_layers > 1:
    hidden = hidden.view(self.hidden_factor, batch_size, self.hidden_size)
else:
    hidden = hidden.unsqueeze(0)  # only for single-layer unidirectional

What Looks Good

• Minimal fix with no extraneous changes
• The else branch is the correct guard — the view() already handles the multi-layer/bidi case, and unsqueeze(0) is only needed for the simple single-layer decoder to add the num_layers=1 dimension expected by nn.LSTM

Verdict

Request Changes → now fixed. This bug would silently produce wrong results or runtime errors when using bidirectional or stacked LSTM decoders.