[Model Compression] update config list key #4074
Conversation
J-shang
force-pushed
the
compression_v2_config
branch
from
6f26574 to
2d3ab7e
Compare
| sub_config['op_names'] = [op_name] | ||
| sub_config['total_sparsity'] = sparsity_per_layer | ||
| new_config_list.append(sub_config) | ||
| elif 'max_sparsity_per_layer' in config and isinstance(config['max_sparsity_per_layer'], float): |
There was a problem hiding this comment.
Why do we need isinstance(config['max_sparsity_per_layer'], float) here?
There was a problem hiding this comment.
If it is float, convert it to dict, skip otherwise.
| new_schema = And(old_schema, lambda n: validate_op_names(model, n, logger)) | ||
| sub_schema[k] = new_schema | ||
|
|
||
| sub_schema = And(data_schema[0], lambda d: validate_op_types_op_names(d)) |
There was a problem hiding this comment.
I think this line might not modify the sub_schema object in the list data_schema. Is that the expected behavior? Not sure because I haven't worked a lot with Schema
There was a problem hiding this comment.
yes, it's a bug, fixed it.
| self.config_list = config_list_canonical(model, config_list) | ||
|
|
||
| def _validate_config_before_canonical(self, model: Module, config_list: List[Dict]): | ||
| pass |
There was a problem hiding this comment.
what is this function used for?
There was a problem hiding this comment.
this is for validating config logic for the sub-pruner, it is called in validate_config()
| @@ -120,7 +120,7 @@ def calculate_metrics(self, data: Dict[str, Tensor]) -> Dict[str, Tensor]: | |||
|
|
|||
| metric = torch.ones(*reorder_tensor.size()[:len(keeped_dim)], device=reorder_tensor.device) | |||
| across_dim = list(range(len(keeped_dim), len(reorder_dim))) | |||
| idxs = metric.nonzero() | |||
| idxs = metric.nonzero(as_tuple=False) | |||
There was a problem hiding this comment.
the default value of as_tuple is False, so why add it?
There was a problem hiding this comment.
because if don't set as_tuple, a warning will appear to ask for setting as_tuple value.
In iterative mode,
sparsity_per_layeris hard to satisfy the sparse expression of each iteration ifspeed up. This is because each layer may have different sparsity afterspeed up. Then in the next iteration, sparsity should be specified for each layer,sparsity_per_layercan not express this.This pr keep the user interface but convert
sparsityandsparsity_per_layertototal_sparsityinside pruner.Convert
max_sparsity_per_layertomin_retention_numel(Dict[op_name, int]) ?Sparsity is element level?