AttributeError: module ‘keras.optimizers’ has no attribute ‘rmsprop’ (SGD, Adam) [Fixed]

AttributeError: module 'keras.optimizers' has no attribute 'rmsprop' (or 'sgd', 'adam') means the optimizer name is written the old way. In current Keras the optimizers are classes with capitalised names: keras.optimizers.RMSprop, keras.optimizers.SGD and keras.optimizers.Adam, and the learning rate argument is learning_rate, not lr. This guide reproduces each optimizer error with Keras 3 and shows the working code.

Every error and fix was run with Python 3.12.5, TensorFlow 2.21.0 and Keras 3.15.1 in the Windows Command Prompt. See the official Keras optimizers documentation.

Quick fix

from keras.optimizers import RMSprop, SGD, Adam      # not rmsprop / sgd / adam

model.compile(optimizer=RMSprop(learning_rate=0.001), loss="mse")   # learning_rate, not lr
# or simply a string:
model.compile(optimizer="rmsprop", loss="mse")

Old code (fails)	Error	Keras 3 code
keras.optimizers.rmsprop()	AttributeError	keras.optimizers.RMSprop()
keras.optimizers.sgd()	AttributeError	keras.optimizers.SGD()
keras.optimizers.adam()	AttributeError	keras.optimizers.Adam()
SGD(lr=0.01)	ValueError: Argument(s) not recognized	SGD(learning_rate=0.01)
RMSprop(decay=1e-6)	UserWarning, decay ignored	A learning-rate schedule
tf.keras.optimizers.legacy.SGD()	ImportError	keras.optimizers.SGD() (or tf_keras)

Why the AttributeError happens

Old Keras 1/2 code and many tutorials used lowercase aliases such as optimizers.rmsprop. Keras 3 (installed with TensorFlow 2.16 and later) only has the class names:

import os
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"     # hide TensorFlow start-up notices
os.environ["TF_ENABLE_ONEDNN_OPTS"] = "0"

import keras

for name in ["rmsprop", "sgd", "adam", "RMSprop", "SGD", "Adam"]:
    try:
        opt_class = getattr(keras.optimizers, name)
        print(f"keras.optimizers.{name:<8} OK -> {opt_class.__name__}")
    except AttributeError as e:
        print(f"keras.optimizers.{name:<8} AttributeError: {e}")

Output:

keras.optimizers.rmsprop  AttributeError: module 'keras.optimizers' has no attribute 'rmsprop'
keras.optimizers.sgd      AttributeError: module 'keras.optimizers' has no attribute 'sgd'
keras.optimizers.adam     AttributeError: module 'keras.optimizers' has no attribute 'adam'
keras.optimizers.RMSprop  OK -> RMSprop
keras.optimizers.SGD      OK -> SGD
keras.optimizers.Adam     OK -> Adam

The same applies to imports: from keras.optimizers import rmsprop raises ImportError: cannot import name 'rmsprop', while from keras.optimizers import RMSprop works.

ValueError: Argument(s) not recognized: {‘lr’: …}

The next error after fixing the name is usually the learning rate argument. lr was renamed to learning_rate, and decay is no longer supported:

import os
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"     # hide TensorFlow start-up notices
os.environ["TF_ENABLE_ONEDNN_OPTS"] = "0"

import warnings
import keras

warnings.simplefilter("always")
try:
    keras.optimizers.SGD(lr=0.01)                                   # old argument name
except ValueError as e:
    print("ValueError:", e)

opt = keras.optimizers.RMSprop(learning_rate=0.001, decay=1e-6)    # decay is ignored now
print("created:", type(opt).__name__, "learning_rate =", round(float(opt.learning_rate), 6))

Output:

ValueError: Argument(s) not recognized: {'lr': 0.01}
C:\pyguides\tfvenv\Lib\site-packages\keras\src\optimizers\base_optimizer.py:86: UserWarning: Argument `decay` is no longer supported and will be ignored.
  warnings.warn(
created: RMSprop learning_rate = 0.001

Replacing decay with a learning-rate schedule

import os
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"     # hide TensorFlow start-up notices
os.environ["TF_ENABLE_ONEDNN_OPTS"] = "0"

import keras

schedule = keras.optimizers.schedules.ExponentialDecay(
    initial_learning_rate=0.01, decay_steps=1000, decay_rate=0.9)     # replaces the old decay= argument
opt = keras.optimizers.SGD(learning_rate=schedule, momentum=0.9)

for step in [0, 1000, 5000]:
    print(f"step {step:>5}: learning rate = {float(schedule(step)):.5f}")

Output:

step     0: learning rate = 0.01000
step  1000: learning rate = 0.00900
step  5000: learning rate = 0.00590

Pass the schedule object as learning_rate; the optimizer updates the rate on every training step.

tf.keras.optimizers.legacy is not supported in Keras 3

Code that used the legacy optimizers (common with older TensorFlow 2 versions and on Apple M1/M2 guides) now fails with an ImportError:

import os
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"     # hide TensorFlow start-up notices
os.environ["TF_ENABLE_ONEDNN_OPTS"] = "0"

import textwrap
import tensorflow as tf

try:
    opt = tf.keras.optimizers.legacy.SGD(learning_rate=0.01)
except ImportError as e:
    print(textwrap.fill("ImportError: " + str(e), width=110))

Output:

ImportError: `keras.optimizers.legacy` is not supported in Keras 3. When using `tf.keras`, to continue using a
`tf.keras.optimizers.legacy` optimizer, you can install the `tf_keras` package (Keras 2) and set the
environment variable `TF_USE_LEGACY_KERAS=True` to configure TensorFlow to use `tf_keras` when accessing
`tf.keras`.

Switch to keras.optimizers.SGD (recommended), or install tf_keras and set TF_USE_LEGACY_KERAS=1 as explained in No module named tensorflow.keras.

Working example: RMSprop, SGD and Adam

The same small regression model trained with each optimizer, using the Keras 3 names and arguments:

import os
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"     # hide TensorFlow start-up notices
os.environ["TF_ENABLE_ONEDNN_OPTS"] = "0"

import numpy as np
import keras
from keras.optimizers import RMSprop, SGD, Adam          # class names start with capital letters

rng = np.random.default_rng(0)
X = rng.normal(size=(200, 3))
y = X @ np.array([1.5, -2.0, 0.5]) + 0.3

for opt in [RMSprop(learning_rate=0.01), SGD(learning_rate=0.05, momentum=0.9), Adam(learning_rate=0.05)]:
    model = keras.Sequential([keras.Input(shape=(3,)), keras.layers.Dense(1)])
    model.compile(optimizer=opt, loss="mse")
    history = model.fit(X, y, epochs=30, verbose=0)
    print(f"{type(opt).__name__:<8} final loss: {history.history['loss'][-1]:.4f}")

Output (the loss values vary slightly between runs):

WARNING:tensorflow:TensorFlow GPU support is not available on native Windows for TensorFlow >= 2.11. Even if CUDA/cuDNN are installed, GPU will not be used. Please use WSL2 or the TensorFlow-DirectML plugin.
RMSprop  final loss: 1.1217
SGD      final loss: 0.0000
Adam     final loss: 0.0000

Which optimizer should you use?

Optimizer	Typical use
Adam	Good default for most networks
RMSprop	Recurrent networks and noisy gradients
SGD(momentum=0.9)	Image models and when you tune the learning rate schedule carefully

Check your Keras and TensorFlow versions

import tensorflow as tf, keras
print(tf.__version__, keras.__version__)

If Keras is 3.x, use the code in this guide. If you must keep old Keras 2 code running unchanged, pin an older TensorFlow in a separate virtual environment or use tf_keras.

Other TensorFlow and Keras errors explained:

• ModuleNotFoundError: No module named ‘tensorflow.keras’
• No module named ‘tensorflow.keras.utils.np_utils’
• Install TensorFlow
• Upgrade TensorFlow to the latest version

Frequently asked questions