Small model needed for integration tests

[kylase]

Currently most of the integration tests such as APIs are disabled during CI because the model takes more than 4 GB to run and Travis is unable to fulfill the memory requirements.

Hence, when the PyTorch model is implemented, a small reference model needs to be prepared for integration tests.

[knmnyn]

Agree with your assessment that a small reference model is needed with unit / integration tests. That will be helpful for others to assess whether to adopt it or not.

[animeshprasad]

@kylase Maybe ability to load a model without any embedding layer altogether. This way all the code flow can be checked however it might require a lot of effort to make all functionality to work without one hot to embedding conversion operation including train and test input. The other way is to check with a dummy model with Vocab = 10 and rest all assigned as one <UNK>. The latter should be easier to do.

[kylase]

Based on my preliminary experience with PyTorch, it seems like the memory usage will not be as much as the current Theano's model.