OpenSeq2Seq v0.4

New models:

• ConvS2S model for translation.
• Wav2Letter model for speech recognition.
• CIFAR-10 dataset support.
• CNNEncoder that can be used to construct arbitrary (almost) CNN models. Based on that, integrated AlexNet and cifar10-nv.

New features:

• Support for "iter_size" (accumulating gradients for "iter_size" steps without update).
• "objects" benchmarking to evaluation and inference modes.
• cuDNN compatible cells support for GNMT.
• 8-padding for transformer.
• Improved config overwriting by train/eval/infer params (will not replace whole dicts, but update incrementally).
• Audio normalization before preprocessing for speech2text models.
• More summaries/parameters for different models.

Bug fixes:

• Regularization in mixed precision mode (loss scaling was not applied, disabling regularizer).
• Overwriting bool values from command line.
• Multi-GPU evaluation in towers mode.
• Multi-GPU inference for speech2text.
• "reflect" padding changed to use zeros for audio preprocessing.
• Unicode support for Python 2.

Important config/API changes:

• Unified static/dynamic loss scaling into a single parameter.
• Made RNN cells accept arbitrary parameters.
• Exposed training step into maybe_print_logs and evaluate functions.

Other changes:

• Improved unit tests and documentation.