Yann Dupis

## Context We are now able to support simple models (e.g logistic regression) with tfe.keras #606. See [this notebook](https://github.com/tf-encrypted/tf-encrypted/blob/master/examples/notebooks/keras-training/Logistic%20Regression.ipynb) as an example. We would like to add the following basic...

## Context With the PR #571, in order to convert a TF Keras model into TFE Keras model we have implemented the methods: - [`model_from_config`](https://www.tensorflow.org/api_docs/python/tf/keras/models/model_from_config) - [`from_config`](https://www.tensorflow.org/api_docs/python/tf/keras/Sequential#from_config) - [`clone_model`](https://www.tensorflow.org/api_docs/python/tf/keras/models/clone_model) We...

Use TensorShape for shapes in TFE

2

In the keras code base they use [tf.TensorShape](https://www.tensorflow.org/api_docs/python/tf/TensorShape) at different places, for example [here](https://github.com/tensorflow/tensorflow/blob/r1.13/tensorflow/python/keras/layers/core.py#L983). We should consider using [tf.TensorShape](https://www.tensorflow.org/api_docs/python/tf/TensorShape) as well, so we can handle properly partially-known shapes and unknown...

Currently, in TF Encrypted, the Dense layer only supports input with a rank equal to 2. However in Keras, the [Dense layer](https://www.tensorflow.org/api_docs/python/tf/keras/layers/Dense) can support input with a rank greater than...

Update Part 2 tutorial to include Dropout layer in Custom model example.

7

In [Part 2 tutorial](https://github.com/OpenMined/PySyft-TensorFlow/blob/master/examples/Part%2002%20-%20Intro%20to%20Private%20Training%20with%20Remote%20Execution.ipynb), for custom when including a Dropout layer in the custom model (`tf.keras.models.Model` subclassing), we were getting the following error when sending the model to a worker...

Send custom model without having to call before model.predict(dummy_data)

2

If you look at [Part 2 tutorial](https://github.com/OpenMined/PySyft-TensorFlow/blob/master/examples/Part%2002%20-%20Intro%20to%20Private%20Training%20with%20Remote%20Execution.ipynb), for custom models (`tf.keras.models.Model`), before sending the model to the worker, we need to run `model.predict(dummy_data)` to set the input_shape ( required by...

### Description We'd like to add [Callback](https://keras.io/callbacks/) support so we can custom training. Currently when passing to fit, it doesn't work because we don't have generic serde support for callbacks....

### Description We would like to be able to train a model with dp-sgd from [TF Privacy](https://github.com/tensorflow/privacy). However these custom optimizers (e.g. [DPGradientDescentGaussianOptimizer](https://github.com/tensorflow/privacy/blob/2007aac912faadbf0ae3af206f3fa100284b8e23/tutorials/mnist_dpsgd_tutorial_keras.py#L121)) are not supported out of the box...

Implement more efficient Replicated Square root

1

In https://github.com/tf-encrypted/moose/pull/1051, we have implemented an initial implementation of square root using `log2` and `pow2`. Instead we would you to implement the algorithm described in the whitepaper which is similar...

Assess kernel completeness

3

In the last few months, we have added lots of operation. We would like to assess the completeness for each operations in terms of types we support. [TODO] give more...