Fix IDE Auto-Completion Failure for tensorflow.keras

When tf.keras.layers. shows no suggestions and your type checker flags tf.keras as unknown, nothing is actually wrong with your code. This is a static-analysis limitation, and understanding why makes the fix obvious.

Your code isn't broken

First, the reassuring part: tf.keras.layers.Dense, tf.keras.Model, and everything else under tf.keras resolve correctly at runtime. Your model builds, trains, and runs. What's failing is your editor's ability to predict those members before running the code. Don't go chasing a runtime bug that isn't there.

Why the IDE can't see it

TensorFlow exposes Keras through a lazily-loaded module alias. Instead of tf.keras being a plain, statically-defined submodule, TensorFlow wires it up at import time through indirection (a lazy loader / module-level __getattr__).

Static analyzers — Pylance, Jedi, PyCharm's inspector — reason about your code without executing it. They can only see attributes that exist statically. A member that's only attached when Python actually runs the import is invisible to them, so tf.keras.* comes back empty and type checkers mark it unresolved.

The fix: import a concrete module

Give the analyzer a real module to introspect by importing keras directly:

# Instead of this (works at runtime, but opaque to the IDE):
import tensorflow as tf
model = tf.keras.Sequential([tf.keras.layers.Dense(10)])
 
# Do this — the IDE can follow `keras`:
from tensorflow import keras
model = keras.Sequential([keras.layers.Dense(10)])

With Keras 3 installed, you can also import the standalone package and pick a backend, which many editors resolve even more cleanly:

import keras            # editor introspects this directly
from keras import layers

Both approaches change nothing about how the code behaves at runtime — they just hand your tooling a binding it can analyze.

Also worth checking

• Update your tooling. Pylance, Jedi, and PyCharm have steadily improved at resolving TensorFlow's lazy modules; an old version struggles more.
• Point the interpreter at the right environment. If the editor is analyzing a different venv than the one with TensorFlow installed, nothing resolves. Select the correct interpreter.
• Let type stubs install. Keeping tensorflow (and, for Keras 3, keras) current ensures the packaged type information your editor reads is present and matched.

What not to do

Don't reach into private module paths like from keras.api._v2 import keras. They "work" until the next version rearranges internals, and then break silently. Prefer the supported imports above and keep your editor updated.

Recognize this as a static-analysis gap rather than a code error, import keras in a form your IDE can follow, and autocomplete comes back without changing how anything runs.

Sources

• Keras API docs