Introduction

tf.InteractiveSession was removed in TensorFlow 2.x because TF2 uses eager execution by default, making sessions unnecessary. The error AttributeError: module 'tensorflow' has no attribute 'InteractiveSession' occurs when running TF1 code on TF2. The fix is to either use tf.compat.v1.InteractiveSession() for backward compatibility, or rewrite the code to use TF2's eager execution which eliminates the need for sessions entirely.

The Error

1import tensorflow as tf
2
3sess = tf.InteractiveSession()
4# AttributeError: module 'tensorflow' has no attribute 'InteractiveSession'

This happens because TensorFlow 2.x removed tf.Session and tf.InteractiveSession from the top-level API.

TF1 vs TF2: Why Sessions Are Gone

In TensorFlow 1.x, you had to build a computation graph first, then run it inside a session:

1# TensorFlow 1.x — graph-then-run model
2import tensorflow as tf
3
4a = tf.constant(5)
5b = tf.constant(3)
6c = a + b  # This doesn't compute anything yet
7
8# Must create a session to evaluate
9sess = tf.InteractiveSession()
10print(c.eval())  # 8
11sess.close()

In TensorFlow 2.x, operations execute immediately (eager execution):

1# TensorFlow 2.x — eager execution
2import tensorflow as tf
3
4a = tf.constant(5)
5b = tf.constant(3)
6c = a + b  # Computes immediately
7print(c.numpy())  # 8
8# No session needed!

Fix 1: Use tf.compat.v1 (Quick Migration)

1import tensorflow as tf
2
3# Use the compatibility module
4tf.compat.v1.disable_eager_execution()  # Required for session-based code
5sess = tf.compat.v1.InteractiveSession()
6
7a = tf.constant(5)
8b = tf.constant(3)
9c = a + b
10
11print(c.eval())  # 8
12sess.close()

Or use the standard Session:

1import tensorflow as tf
2tf.compat.v1.disable_eager_execution()
3
4with tf.compat.v1.Session() as sess:
5    a = tf.constant(5)
6    b = tf.constant(3)
7    c = a + b
8    print(sess.run(c))  # 8

Fix 2: Rewrite for TF2 Eager Execution (Recommended)

1import tensorflow as tf
2
3# TF2 — no sessions, no graphs, just Python
4a = tf.constant(5.0)
5b = tf.constant(3.0)
6c = a + b
7print(c.numpy())  # 8.0
8
9# Variables work directly too
10w = tf.Variable(tf.random.normal([3, 3]))
11print(w.numpy())
12
13# Use tf.function for performance-critical code
14@tf.function
15def compute(x, y):
16    return x ** 2 + 2 * y
17
18result = compute(tf.constant(3.0), tf.constant(4.0))
19print(result.numpy())  # 17.0

Fix 3: Use tf_upgrade_v2 Script

TensorFlow provides an automatic migration script:

1# Convert a single file
2tf_upgrade_v2 --infile old_code.py --outfile new_code.py
3
4# Convert an entire directory
5tf_upgrade_v2 --intree old_project/ --outtree new_project/ --reportfile report.txt

The script automatically replaces tf.InteractiveSession() with tf.compat.v1.InteractiveSession() and makes other necessary changes.

Common TF1-to-TF2 Replacements

Migrating a Complete Example

TF1 (Original)

1import tensorflow as tf
2
3sess = tf.InteractiveSession()
4
5x = tf.placeholder(tf.float32, shape=[None, 784])
6W = tf.Variable(tf.zeros([784, 10]))
7b = tf.Variable(tf.zeros([10]))
8
9y = tf.nn.softmax(tf.matmul(x, W) + b)
10
11sess.run(tf.global_variables_initializer())
12result = sess.run(y, feed_dict={x: test_data})
13sess.close()

TF2 (Migrated)

1import tensorflow as tf
2
3# No session, no placeholder, no initializer
4model = tf.keras.Sequential([
5    tf.keras.layers.Dense(10, activation='softmax', input_shape=(784,))
6])
7
8# Weights initialize automatically
9result = model.predict(test_data)
10
11# Or for manual computation:
12W = tf.Variable(tf.zeros([784, 10]))
13b = tf.Variable(tf.zeros([10]))
14
15@tf.function
16def predict(x):
17    return tf.nn.softmax(tf.matmul(x, W) + b)
18
19result = predict(test_data)

Common Pitfalls

• Using tf.compat.v1 without disabling eager execution: tf.compat.v1.InteractiveSession() requires tf.compat.v1.disable_eager_execution() to be called first. Without it, session-based code behaves unpredictably or raises errors about eager tensors not being compatible with sessions.
• Mixing TF1 and TF2 code: Using tf.compat.v1.Session() in the same script as eager TF2 code (like tf.function) causes conflicts. Pick one paradigm — either fully migrate to TF2 or use tf.compat.v1 consistently throughout the file.
• Calling .numpy() inside tf.function: tensor.numpy() only works in eager mode. Inside a @tf.function-decorated function, tensors are graph tensors and .numpy() raises an error. Use tf.print() for debugging inside tf.function.
• Not using tf_upgrade_v2 for large codebases: Manually replacing every tf.Session and tf.placeholder is error-prone. The tf_upgrade_v2 script handles most mechanical replacements automatically and generates a report of changes that need manual attention.
• Assuming compat.v1 code performs the same as native TF2: Code running through tf.compat.v1 does not benefit from TF2's performance improvements (eager execution optimizations, tf.function tracing). Migrate to native TF2 APIs for best performance.

Summary

• tf.InteractiveSession was removed in TF2 because eager execution eliminates the need for sessions
• Quick fix: use tf.compat.v1.InteractiveSession() with tf.compat.v1.disable_eager_execution()
• Recommended: rewrite code for TF2 — replace sess.run(tensor) with tensor.numpy(), remove placeholders
• Use the tf_upgrade_v2 script for automatic migration of large codebases
• Replace tf.Session patterns with @tf.function for graph-mode performance in TF2
• TF2 eager execution lets you write and debug TensorFlow code like regular Python