TF-Slim

1.1.0 verified Fri May 01 auth: no python maintenance

TF-Slim is a lightweight library for defining, training, and evaluating complex models in TensorFlow. The current version is 1.1.0, with no recent updates and likely in maintenance mode.

pip install tf-slim

Common errors

error ModuleNotFoundError: No module named 'tensorflow.contrib.slim' ↓

cause Using old import path from TensorFlow contrib (removed in TF 2.x).

fix

Install tf-slim and import as 'import tf_slim as slim' or 'from tf_slim import slim'.

error ValueError: Variable resnet_v2_50/conv1/weights already exists, disallowed. ↓

cause Reusing the same variable scope without enclosing in a variable scope or using slim.get_unique_variable_scope() properly.

fix

Use tf.variable_scope('model', reuse=tf.AUTO_REUSE) or reuse=True when reusing the model.

error AttributeError: module 'tensorflow' has no attribute 'contrib' ↓

cause TensorFlow 2.x removed the contrib module entirely.

fix

Install tf-slim and use the correct import path: 'import tf_slim as slim'.

Warnings

breaking TF-Slim was part of TensorFlow contrib, which was removed in TF 2.0. The standalone tf-slim package must be installed separately. ↓

fix Install tf-slim with pip and use 'import tf_slim as slim' instead of 'tensorflow.contrib.slim'.

gotcha TF-Slim's TF 1.x APIs (e.g., slim.layers, slim.losses) are not fully compatible with eager execution. You may need to run inside tf.Graph().as_default() or disable eager execution. ↓

fix Use tf.compat.v1.disable_eager_execution() at the start of your script or wrap code in a graph context.

deprecated Many TF-Slim functions (e.g., slim.learning, slim.evaluation) are deprecated in favor of TensorFlow native APIs like tf.keras or tf.estimator. ↓

fix Consider migrating to tf.keras for model building and tf.train.Checkpoint/tf.function for training.

Imports

slim

wrong

import tensorflow.contrib.slim as slim

correct

import tf_slim as slim

Old contrib path removed in TF 2.x

slim

wrong

from tensorflow.contrib import slim

correct

from tf_slim import slim

contrib module deprecated and removed

Quickstart

Load a pretrained ResNet-50 model using TF-Slim's model zoo.

import tensorflow as tf
import tf_slim as slim

inputs = tf.random.normal([4, 224, 224, 3])
nets = tf_slim.nets.resnet_v2
with slim.arg_scope(nets.resnet_arg_scope()):
    logits, _ = nets.resnet_v2_50(inputs, num_classes=1000, is_training=False)
print(logits.shape)