danielenricocahall / elephas Goto Github PK

Copied from:

• maxpumperla#131
• maxpumperla#40

https://github.com/keras-team/keras/blob/master/keras/callbacks.py#L341

The fit function in Keras returns a graph that can be used to determine if the model is overfitting or not. This would be very useful from Elephas.

Copied from:

• maxpumperla#141
• maxpumperla#85
• maxpumperla#190

Currently, LSTM based models are not technically supported in Elephas. While the LSTM layers can be added to a model and the model may train, it's not clear if they're performing correctly as we don't currently have specialized logic for handling sequential/temporal data.

Right now, we're just using print statements for logging - we should use something like structlog or even just the builtin logging package.

I get the following error when trying to run the second last line of elephas/examples/ml_mlp_classification.py:

TypeError: precision() missing 1 required positional argument: 'label'

Hello, in my sequential Keras code, I am doing something like this:

model.compile(optimizer=Adam(), loss='mean_squared_error')
trained_model = model.fit(x_train, y_train, batch_size=512, shuffle=True, epochs=50, verbose=1)

encoder = Model(model.input, model.output)
encoded_data = encoder.predict(x_train) # bottleneck representation

I am then using the encoded_data array later on in the program.

How can I do something like this with SparkModel? It seems like I cannot build a SparkModel with an input and output of the Keras model?

The parametrized tests in integration/test_end_to_end will have non-deterministic test failures, usually due to a port being occupied already. This makes things flaky and hard to diagnose when there are real problems.

Potential solutions:

• Investigate socket/Flask servers not closing down correctly?
  • Are the connections staying alive incorrectly?
• Use a global random object (i.e; a pytest fixture with session scope?) which samples without replacement for producing each port while testing, rather than randomly selecting one for each test.

Convert the use of raw strings to enums for better readability, reduced likelihood for bugs, etc. in cases where there are a finite number of options i.e; "asynchronous", "synchronous", "hogwild" for server mode, "u", "g" for request options, etc.