Accept custom (layer type:param name) to include in sparse_parameter dictionary

1. Support to include in sparse_parameter_list an user-supplied custom layer type and its parameter name. This is useful when users have their own implementation of nn.Linear or nn.Conv2D. For example, huggingface repo has a custom implementation of nn.Linear called LinearActivation.
2. Print info of layers in the model that are not pruned.

apex/contrib/sparsity/asp.py

@@ -30,7 +30,7 @@ class ASP:
              verbosity=3,
              whitelist=[torch.nn.Linear, torch.nn.Conv1d, torch.nn.Conv2d, torch.nn.Conv3d], 
              allowed_layer_names=None, disallowed_layer_names=[],
-             allow_recompute_mask=False):
+             allow_recompute_mask=False, custom_layer_dict={}):
         """Call this method to modify your model to take advantage of sparse matrix multiplication.
         Note that this call alone only augments the model with additional buffers needed for sparse MMA,
         it does not enable use of sparse MMA. 
@@ -62,7 +62,9 @@ class ASP:
           disallowed_layer_names   If not [], only layer names that do not appear in this list are considered for sparsity.
           allow_recompute_mask     If True, stores pruned values so that dense weights can be restored.
                                    Pruned weights are stored in CPU memory, hence this option does not increase GPU memory usage.
-          Support for allow_recompute_mask can be removed, it is not part of our recipe -- AKM. 
+          custom_layer_dict        Dictionary of additional layer paremeters to sparsify. e.g. {CustomLinear: ['weight']}
+          
+          [Future] Support for allow_recompute_mask can be removed, it is not part of sparse inference recipe -- AKM. 
         """
         assert (cls.__model is None), "ASP has been initialized already."
         cls.__model = model
@@ -82,6 +84,10 @@ class ASP:
             sparse_parameter_list = {torch.nn.Linear: ['weight'], torch.nn.Conv1d: ['weight'], torch.nn.Conv2d: ['weight'], torch.nn.Conv3d: ['weight'], torchvision.ops.misc.Conv2d: ['weight']}
         else:
             sparse_parameter_list = {torch.nn.Linear: ['weight'], torch.nn.Conv1d: ['weight'], torch.nn.Conv2d: ['weight'], torch.nn.Conv3d: ['weight']}
+        if custom_layer_dict: # Update default list to include user supplied custom (layer type : parameter tensor), make sure this tensor type is something ASP knows how to prune
+            sparse_parameter_list.update(custom_layer_dict)
+            whitelist += list(custom_layer_dict.keys())
+
         for module_type in whitelist:
             assert (module_type in sparse_parameter_list), "Module %s :: Don't know how to sparsify module." % module.dtype()
 
@@ -110,6 +116,9 @@ class ASP:
                     else:
                         pruned = None
                     cls.__sparse_parameters.append((module_name, module, p_name, p, mask, pruned))
+                else:
+                    if cls.__verbosity >= 3:
+                        print("[ASP] Not sparsifying %s::%s of size=%s and type=%s" % (module_name, p_name, str(p.size()), str(p.dtype)))
 
         for name, sparse_module in eligible_modules(model, tuple(whitelist), allowed_layer_names, disallowed_layer_names):
             add_sparse_attributes(name, sparse_module)