Introduction

When using bidirectional LSTMs or other recurrent layers in Keras with a TensorFlow 1.x backend, you may encounter the error: You must feed a value for placeholder tensor 'bidirectional_1/keras_learning_phase' with dtype bool. This error occurs because Keras uses a learning_phase placeholder to distinguish between training and inference modes, and certain operations fail to set this flag automatically.

The Error

The full error message typically looks like:

InvalidArgumentError: You must feed a value for placeholder tensor
'bidirectional_1/keras_learning_phase' with dtype bool

This happens when:

• Running predictions or evaluations outside Keras's standard model.predict() or model.evaluate()
• Using K.function() or TensorFlow sessions directly
• Serving the model through TensorFlow Serving or a custom inference pipeline
• Mixing Keras and raw TensorFlow operations

Why This Happens

Keras layers like Dropout, BatchNormalization, and bidirectional RNNs behave differently during training vs inference. Keras tracks this via K.learning_phase() — a boolean placeholder tensor:

• Training (learning_phase = True): Dropout is active, BatchNorm uses batch statistics
• Inference (learning_phase = False): Dropout is disabled, BatchNorm uses running statistics

When you call model.predict(), Keras automatically sets learning_phase = False. But if you bypass Keras and use a raw TensorFlow session, this placeholder remains unfed.

1# This works — Keras sets learning_phase internally
2predictions = model.predict(X_test)
3
4# This fails — learning_phase is not set
5sess = tf.compat.v1.keras.backend.get_session()
6result = sess.run(model.output, feed_dict={model.input: X_test})
7# Error: must feed learning_phase

Fix 1: Set Learning Phase Explicitly

Feed K.learning_phase() in your session run:

1import tensorflow as tf
2from tensorflow.keras import backend as K
3
4# Build your model
5model = build_bidirectional_model()
6
7# Set learning phase to False (inference mode)
8sess = K.get_session()
9result = sess.run(
10    model.output,
11    feed_dict={
12        model.input: X_test,
13        K.learning_phase(): False  # explicitly set inference mode
14    }
15)

Fix 2: Set Learning Phase Globally

If you only need inference, set the learning phase globally before creating the session:

1from tensorflow.keras import backend as K
2
3K.set_learning_phase(0)  # 0 = inference, 1 = training
4
5# Now build and use the model
6model = build_model()
7predictions = model.predict(X_test)  # works without issues

Warning: Setting K.set_learning_phase(0) before model creation makes the model inference-only. You cannot train it afterward. Set it after training is complete.

Fix 3: Use K.function with Learning Phase

When creating custom prediction functions with K.function:

1from tensorflow.keras import backend as K
2
3# Wrong — missing learning_phase
4predict_fn = K.function([model.input], [model.output])
5
6# Correct — include learning_phase
7predict_fn = K.function(
8    [model.input, K.learning_phase()],
9    [model.output]
10)
11
12# Call with learning_phase=False
13result = predict_fn([X_test, False])

Or in newer Keras versions:

predict_fn = K.function([model.input], [model.output])
K.set_learning_phase(0)
result = predict_fn([X_test])

Fix 4: Upgrade to TensorFlow 2.x

TensorFlow 2.x with eager execution eliminates this issue entirely. The learning phase is handled automatically via the training argument in layer calls:

1import tensorflow as tf
2
3model = tf.keras.Sequential([
4    tf.keras.layers.Bidirectional(
5        tf.keras.layers.LSTM(64, return_sequences=True)
6    ),
7    tf.keras.layers.Bidirectional(
8        tf.keras.layers.LSTM(32)
9    ),
10    tf.keras.layers.Dense(10, activation='softmax')
11])
12
13# No learning_phase issues in TF2
14predictions = model(X_test, training=False)  # explicit training flag
15# or
16predictions = model.predict(X_test)          # automatic

Fix 5: Freeze the Graph for Serving

If you need to export the model for TensorFlow Serving:

1from tensorflow.keras import backend as K
2
3# Set learning phase before freezing
4K.set_learning_phase(0)
5
6# Rebuild model in inference mode
7model = tf.keras.models.load_model('my_model.h5')
8
9# Export as SavedModel
10tf.saved_model.save(model, 'export_dir')

Common Pitfalls

• Setting learning_phase too early: Calling K.set_learning_phase(0) before building the model makes it permanently inference-only. You cannot call model.fit() afterward. Set it after training.
• Multiple models: If you have multiple models in the same session, K.set_learning_phase() affects all of them. Use per-call feed_dict if you need different modes for different models.
• Layer-specific behavior: Not just Dropout — BatchNormalization, SpatialDropout1D, GaussianNoise, and AlphaDropout all depend on learning_phase. Any model using these layers can trigger the error.
• Model loading: When loading a saved model in a new session, learning_phase resets. Set it again before inference.
• TF1 compatibility mode: In TF2 with tf.compat.v1.disable_eager_execution(), you are back in graph mode and will encounter this error again. Prefer native TF2 eager mode.

Summary

• The error occurs because learning_phase is a boolean placeholder that must be fed in TF1 graph mode
• Use model.predict() instead of raw session runs when possible
• Feed K.learning_phase(): False in feed_dict for custom inference
• Set K.set_learning_phase(0) globally for inference-only pipelines (after training)
• Upgrade to TensorFlow 2.x to avoid the issue entirely — eager execution handles it automatically