Support for Variable Number of Inputs

• Input shapes now must be passed as a tuple of input shapes
• Supports variable number of inputs for models
• Fixed no compression to return CompressedElement

Debug updates and updated tracing

• Issues with out-of-order module definitions resolved
• Tracing now runs a single forward pass iteration of the model
• CompressNNModel() now requires input shape to be specified (for use in tracing)

First release

First release of CompressNN:

• Supports compression of intermediate activations in torch.nn.Module networks that are batch size-dependent
• Support for compression configuration JSON files to specify which activations are compressed and what their compression configuration is
• cuSZp and CPU-offloading are two primary memory management schemes