01_​Training_​LSTM_​Solution

Pre-processing and Encoding

Define a network structure for a many-to-one architecture

- Keras Input Layer node

Tips:
- The input shape should be ?, dictionary size
- Check the dictionary table to get the dictionary size
- Keras LSTM Layer node

Tip: Change the number of units to 512

- Keras Dropout Layer node
- Keras Dense Layer node

Tip: Use the activation softmax and the dictionary size as the number of units

Train the network and save it as a TensorFlow model

- Train the network using the Keras Network Learner node

Tips:
- Use for the input and the target data the conversion:
From Collection of Number (integer) to One-hot Tensor
- Use Categorical-Cross Entropy as loss function
- The trained model in the data folder was trained on half of the book for 50 epochs. This took two days. You can stop the training earlier by clicking the Stop learning button in the Monitor View and move on. In the deployment workflow you can use the pretrained network.
- Convert the network using the Keras to TensorFlow Network Converter node
- Save the model uisng the TensorFlow Network Writer node

Tip: Don't overwrite the trained model

- Optional: Save the model as Keras model (.h5 file) using the Keras Network Writer node