Introduction

The error saying tensors must come from the same graph appears when parts of a TensorFlow model are built in different execution contexts. It often shows up in bidirectional recurrent models where custom layers, legacy session code, or mixed APIs create incompatible tensors. The fix is to keep model creation and tensor operations in one consistent TensorFlow context.

Core Sections

Why Graph Mismatch Happens

In TensorFlow 1 style execution, operations belong to explicit graphs and sessions. If one tensor is created in graph A and another in graph B, combining them raises this error. In TensorFlow 2 eager mode, this is less common, but mixing tf.compat.v1 code with modern Keras can still recreate the same failure pattern.

Typical triggers:

• building layers in one graph and running in another session
• mixing standalone Keras and tf.keras symbols
• creating tensors at module import time then reusing them in compiled functions
• notebook cells that partially rebuild model state

Build Bidirectional RNNs in One Keras Path

Use a single tf.keras pipeline and construct all layers inside one model function.

1import tensorflow as tf
2
3
4def build_model(vocab_size=5000, emb_dim=128, rnn_units=64):
5    inputs = tf.keras.Input(shape=(None,), dtype="int32")
6    x = tf.keras.layers.Embedding(vocab_size, emb_dim)(inputs)
7    x = tf.keras.layers.Bidirectional(
8        tf.keras.layers.LSTM(rnn_units, return_sequences=False)
9    )(x)
10    x = tf.keras.layers.Dropout(0.2)(x)
11    outputs = tf.keras.layers.Dense(1, activation="sigmoid")(x)
12
13    model = tf.keras.Model(inputs, outputs)
14    model.compile(optimizer="adam", loss="binary_crossentropy", metrics=["accuracy"])
15    return model
16
17model = build_model()
18model.summary()

This avoids manual graph management and is the recommended baseline in modern TensorFlow.

Avoid Mixing Legacy Session Code

If old code still uses tf.compat.v1.Session, isolate it fully or migrate it. Partial mixing with eager mode usually causes fragile behavior.

Legacy style example that should be avoided in new code:

1import tensorflow as tf
2
3g1 = tf.Graph()
4with g1.as_default():
5    a = tf.constant([1.0, 2.0])
6
7g2 = tf.Graph()
8with g2.as_default():
9    b = tf.constant([3.0, 4.0])
10
11# a + b would fail because they belong to different graphs

If migration is possible, remove explicit graph objects and let tf.keras manage execution context.

Notebook and Reload Scenarios

Interactive notebooks can keep stale tensors across cell runs. If you redefine a model in one cell and reuse old tensors elsewhere, graph identity errors can appear.

Safer notebook workflow:

1. Restart kernel after dependency changes.
2. Re run all model definition cells in order.
3. Avoid global tensors created outside model building functions.

Resetting Keras state can also help between experiments:

1import tensorflow as tf
2
3tf.keras.backend.clear_session()
4model = build_model()

This clears previous graph state and reduces accidental tensor reuse.

Custom Layer and tf.function Guidelines

Custom layers should create weights in build and operate on input tensors in call. Do not capture tensors from unrelated global scopes.

1class Projection(tf.keras.layers.Layer):
2    def __init__(self, units):
3        super().__init__()
4        self.units = units
5
6    def build(self, input_shape):
7        self.w = self.add_weight(shape=(input_shape[-1], self.units), initializer="glorot_uniform")
8
9    def call(self, inputs):
10        return tf.matmul(inputs, self.w)

When using @tf.function, keep inputs and variable creation patterns stable. Dynamic creation inside traced functions can produce hard to debug graph behavior.

Practical Diagnostic Checklist

When this error appears:

• print TensorFlow version and confirm one API namespace
• remove tf.compat.v1 unless absolutely required
• clear session and rebuild model from scratch
• run a tiny forward pass before full training

1import numpy as np
2x = np.random.randint(0, 5000, size=(4, 20))
3y = np.random.randint(0, 2, size=(4, 1))
4model.fit(x, y, epochs=1, verbose=0)

A successful short fit usually confirms graph consistency.

Common Pitfalls

• Mixing keras and tf.keras objects in one model.
• Combining tensors created under different graph contexts.
• Keeping stale tensors alive across notebook cell reruns.
• Creating variables dynamically in inconsistent execution scopes.
• Reusing old session based code in eager mode projects.

Summary

• This error is a context mismatch between tensors from different graphs.
• Use one consistent tf.keras model build path.
• Avoid mixed legacy session code in modern TensorFlow projects.
• Clear session and rebuild models in notebook workflows.
• Keep custom layers and traced functions graph consistent.