Fix: Using stateful=True in LSTM layers with Keras4Delphi (no more shape errors!)

[Roelof1986 0]

https://github.com/Roelof1986/Keras4Delphi---New-FULL-LSTM-Fix-Delphi-10.3-Rio-

 

Hello everyone,

 

I've been working extensively with Keras4Delphi and wanted to share a solid fix for using stateful=True in LSTM layers — something that normally leads to frustrating errors like:

 

ValueError: When using 'stateful=True' in a RNN, the batch size must be static. Found: (None, 4, 2)

 

After reverse-engineering how Keras handles batch_input_shape, I've built a working solution that lets you:

 

Use stateful LSTM models in Delphi without runtime shape errors

 

Define a static batch shape via class vars instead of fragile .SetItem(...) calls

 

Cleanly set stateful, return_sequences, and batch_input_shape using hardcoded defaults

 

Stack multiple LSTM layers (with only the first one being stateful, as recommended)

 

 

✅ Highlights:

 

Fully functional fix for stateful=True

 

Tested with TensorFlow 2.19 and Delphi 10.4

 

Shape is passed cleanly to Python using PyDict_SetItemString(...)

 

Uses class var overrides, e.g.:

 

 

TBase.UseStateful := True;

TBase.UseReturnSequences := True;

TBase.UseBatchInputShape := [2, 4, 2];

 

🔧 Example setup:

 

lstm1 := TLSTM.Create(32); // parameters pulled from class vars

model.Add(lstm1);

 

lstm2 := TLSTM.Create(64); // stateful := False, return_sequences := False

model.Add(lstm2);

 

 

🧠 Why this helps:

 

Keras4Delphi is a great bridge between Pascal and TensorFlow, but its current design struggles with RNN state handling. This fix brings full stability, predictable behavior, and makes LSTM-based time-series modeling in Delphi a reality.

 

🗂 Repo (MIT-licensed):

 

GitHub: keras4delphi-stateful-fix

 

---

 

Let me know if you're also working with RNNs or time-sequence data in Delphi — I'd love to share ideas.

 

Kind regards,

Roelof Emmerink

 

 

---