README

Keras4Delphi---New-FULL-LSTM-Fix-Delphi-10.3-Rio-

A complete fix for stateful=True, return_sequences=True and batch_input_shape behaviour in Keras4Delphi. Enables reliable training of stateful LSTM models with static batch sizes and proper parameter handling.

1. Download file "Keras.pas" and replace the old Keras.pas with the new version
2. Make sure you are using the updated/debugged Keras4Delphi (in repository: Keras4Delphi---debugged-Delphi-10.3-Rio-)
3. Download and run the example code: DoubleLayer_LSTM_test

Keras4Delphi Stateful LSTM Fix

This repository contains a tested and working fix for using stateful=True in LSTM layers within Keras4Delphi.

✅ What it fixes

• Prevents the common error: ValueError: When using 'stateful=True' in a RNN, the batch size must be static
• Enforces correct batch_input_shape=(batch_size, timesteps, features) handling
• Ensures Python receives stateful=True and return_sequences=True without loss of information

🔧 How it works

• Uses class var configuration for global LSTM defaults
• Automatically injects required keyword arguments via a modified Instantiate method
• Avoids fragile parameter passing through .SetItem(...)

💡 Why it's useful

Keras4Delphi is a powerful bridge between Delphi and Python's deep learning stack.
But until now, configuring stateful LSTM layers was unreliable and prone to shape errors.
This fix makes stateful sequence modeling stable, predictable, and production-ready.

✨ Highlights

• Compatible with TensorFlow 2.19 and Delphi 10.3/10.4
• Supports 2-layer LSTM stacks with controlled sequence output
• No more None, batch, time, features shape mismatches

📂 What's included

• Modified Instantiate logic (Delphi)
• Class-wide control via TBase.UseStateful, UseReturnSequences, UseBatchInputShape
• Tested example with correct loss convergence over 500 epochs

📜 License

MIT

Based on Keras4Delphi by Pigrecos (MIT licensed). This repository includes modifications and fixes by Roelof Emmerink.