How to wrap a TF Estimator in a Keras Layer? - tensorflow

I have some code already written, using TensorFlow Estimator.
Since I would like to work with Keras (and I need to use the Estimator code), is there any way to wrap the Estimator into a Keras Layer ?

Related

Should I use the standalone Keras library or tf.keras?

As Keras becomes an API for TensorFlow, there are lots of old versions of Keras code, such as https://github.com/keiserlab/keras-neural-graph-fingerprint/blob/master/examples.py
from keras import models
With the current version of TensorFlow, do we need to change every Keras code as?
from tensorflow.keras import models
You are mixing things up:
Keras (https://keras.io/) is a library independent from TensorFlow, which specifies a high-level API for building and training neural networks and is capable of using one of multiple backends (among which, TensorFlow) for low-level tensor computation.
tf.keras (https://www.tensorflow.org/guide/keras) implements the Keras API specification within TensorFlow. In addition, the tf.keras API is optimized to work well with other TensorFlow modules: you can pass a tf.data Dataset to the .fit() method of a tf.keras model, for instance, or convert a tf.keras model to a TensorFlow estimator with tf.keras.estimator.model_to_estimator. Currently, the tf.keras API is the high-level API to look for when building models within TensorFlow, and the integration with other TensorFlow features will continue in the future.
So to answer your question: no, you don't need to convert Keras code to tf.keras code. Keras code uses the Keras library, potentially even runs on top of a different backend than TensorFlow, and will continue to work just fine in the future. Even more, it's important to not just mix up Keras and tf.keras objects within the same script, since this might produce incompatabilities, as you can see for example in this question.
Update: Keras will be abandoned in favor of tf.keras: https://twitter.com/fchollet/status/1174019423541157888

How to use keras InceptionV3 model in tensorflow 1.11?

In tensorflow 1.10,I could call function tensorflow.keras.applications.InceptionV3() to create a InceptionV3 model.
However, the keras.applications module seems to be removed in Tensorflow 1.11.
According to the API r1.11 reference, keras Applications are canned architectures with pre-trained weights.
Should I download the model file and load it?
Nevertheless, where could I get it from?

Can I add Tensorflow Fake Quantization in a Keras sequential model?

I have searched this for a while, but it seems Keras only has quantization feature after the model is trained. I wish to add Tensorflow fake quantization to my Keras sequential model. According to Tensorflow's doc, I need these two functions to do fake quantization: tf.contrib.quantize.create_training_graph() and tf.contrib.quantize.create_eval_graph().
My question is has anyone managed to add these two functions in a Keras model? If yes, where should these two function be added? For example, before model.compile or after model.fit or somewhere else? Thanks in advance.
I worked around by post-training quantization. Since my final goal is to train a mdoel for mobile device, instead of fake quantization during training, I exported keras .h5 file and converted to Tenforflow lite .tflite file directly (with post_training_quantize flag set to true). I tested this on a simple cifar-10 model. The original keras model and the quantized tflite model have very close accuracy (the quantized one a bit lower).
Post-training quantization: https://www.tensorflow.org/performance/post_training_quantization
Convert Keras model to tensorflow lite: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/lite/toco/g3doc/python_api.md
Used the tf-nightly tensorflow here: https://pypi.org/project/tf-nightly/
If you still want to do fake quantization (because for some model, post-training quantization may give poor accuracy according to Google), the original webpage is down last week. But you can find it from github: https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/quantize
Update: Turns out post-quantization does not really quantize the model. During inference, it still uses float32 kernels to do calculations. Thus, I've switched to quantization-aware training. The accuracy is pretty good for my cifar10 model.

Keras and make_csv_dataset compatibility

Can tf.contrib.data.make_csv_dataset() be used for Keras models in tensorflow 1.9.0?
Yes, tf.contrib.data.make_csv_dataset() returns a tf.data.Dataset and you can pass tf.data.Dataset to the fit method of Keras models.
See some examples here:
https://www.tensorflow.org/guide/keras#input_tfdata_datasets

Why use keras as backend instead of using tensorflow?

I see that there are many similar functions between tensorflow and keras like argmax, boolean_mask...I wonder why people have to use keras as backend along with tensorflow instead of using tensorflow alone.
Keras is not a backend, but it is a high-level API for building and training Neural Networks. Keras is capable of running on top of Tensorflow, Theano and CNTK. Most of the people prefer Keras due to its simplicity compared to other libraries like Tensorflow. I recommend Keras for beginners in Deep Learning.
A Keras tensor is a tensor object from the underlying backend (Theano,
TensorFlow or CNTK), which we augment with certain attributes that
allow us to build a Keras model just by knowing the inputs and outputs
of the model.
Theano vs Tensorflow
Tensorflow is necessary if you wish to use coremltools. Apple has promised support for architectures created using Theano but I haven't seen it yet.
Keras will require unique syntax sugar depending on the backend in use. I like the flexibility of Tensorflow input layers and easy-access to strong Google neural networks.