Open DerekMelchin opened 6 months ago
Before we review the tutorials to add more example, we need to review those we have. Because of Tensorflow upgrade, Keras was broken (fixed by 8ee3fcb), but Tensorflow is still broken. For example, we cannot use
tf.reset_default_graph()
sess = tf.Session()
We could use it with compat.v1
tf.compat.v1.reset_default_graph()
sess = tf.compat.v1.Session()
but we shouldn't rely on Tensorflow backward compatibility since this is an example for the community.
Additionally, let's add a link to a backtest with the notebook to avoid copy and paste the code snippets. The notebook should include the page text.
Some issues I found with the new TF tutorials:
data
isn't definedMight be other errors too, please review it.
The TF tutorial should include a Lambda
layer so that we can demonstrate how to fix the "name 'tf' is not defined" error when saving and loading from the Object Store. To avoid the error, instead of passing a lambda
to the Lambda
constructor, you need to pass a function. The function definition needs to have the @register_keras_serializable()
decorator.
Refer to https://stackoverflow.com/questions/54347963/tf-is-not-defined-on-load-model-using-lambda
The TF tutorial should use Dense
and Sequential
, not dense
and sequential
Since we are not using .h5
saving mode anymore, we don't need the Lambda
logic.
The Aesera tutorial defines a self.train
method, overriding the built-in self.train
.
Expected Behavior
The ML tutorials here provide educational material on how to use all three learning paradigms: supervised, unsupervised, and reinforcement learning. Each example is unique and informative.
Actual Behavior
Most of the examples solve the same problem, but just use a different library to do so. The text is also largely duplicated.
Checklist
master
branch