ImageNet Model Zoo for TensorFlow

via Keras Applications

Note: this tool is largely obsolete for those using TF 2.0, which places more of an emphasis on Keras models.

The TensorFlow Model Zoo serves to address the missing component of TensorFlow which is present in nearly every other deep learning framework: easy access to pre-trained ImageNet models. The ability to easily load in pre-trained models is an enormous asset both for purposes of application deployment and for transfer learning.

The tool is a simple wrapper around Keras's Applications module, which contains several state-of-the-art classification models, along with a variety of customizable parameters and pre-trained ImageNet weights. This code simply wraps each function to load the model on the provided session and extracts the requisite tensors to be returned to the user for easy use in TensorFlow.

Usage

The API was meant to mirror the Keras Applications API as closely as possible, and works very similarly, with a few extra parameters thrown in. Creating a model returns an X and Y tensor, which correspond to the input and output tensors of the model, as well as any other tensors indicated by the user.

from tfmodelzoo import InceptionV3
sess = tf.Session()
X, Y = InceptionV3(sess, weights='imagenet')
# X is the input tensor and Y is the output of the model (in this case, softmax)
loss = tf.losses.log_loss(labels, Y)
sess.run([loss, Y], feed_dict={X: data})

The graph can then be accessed (if needed) by sess.graph. If another tensor (other than the input and output tensors) is needed, it can either be accessed from the graph directly or when loading the model:

# can be accessed from the graph directly:
conv1 = sess.graph.get_tensor_by_name('conv1tensorname:0')
# or through the `include_tensors` convenience parameter when loading the model:
X, Y, conv1 = InceptionV3(sess, include_tensors=['conv1tensorname:0'])

It also imports the Keras ImageNet utils for you, making it easy to preprocess inputs and decode your predictions back to their text labels:

from tfmodelzoo import preprocess_input, decode_predictions
data = preprocess_input(images)
predictions = sess.run(Y, feed_dict={X: data})
pred_labels = decode_predictions(predictions, top=3)
#[[('n02088364', 'beagle', 0.8906793),
#  ('n02092339', 'Weimaraner', 0.06135427),
#  ('n02108000', 'EntleBucher', 0.012612144)]]

See the demo notebook for an example. Beyond that, reference the Keras Documentation for information on the arguments that may be passed for each model.

Note: Passing a Keras Input, as indicated in the documentation, works, but as of now there is no way to do this with pure TensorFlow.

Available Models:

Users may refer to the Keras Documentation for detailed information about the models available. That said, the models currently supported are:

Xception
VGG16
VGG19
ResNet50
InceptionV3
InceptionResNetV2
MobileNet
DenseNet121
DenseNet169
DenseNet201

Setup:

With pip, simply run:

pip install -e git+https://github.com/joeddav/tensorflow-modelzoo.git#egg=tfmodelzoo