Add a new compression pass for activation quantization. #2216
Conversation
| ) | ||
| from coremltools.converters.mil.mil.passes.helper import _check_child_op_type, block_context_manager | ||
| from coremltools.converters.mil.mil.passes.pass_registry import register_pass | ||
| #from coremltools.optimize.coreml._utils import quantize_weight |
|
Can you add an overview of this change to the description of the pull request? Also please add unit tests for this new functionality. |
|
|
||
| # try pattern I | ||
| # `dequant` -> `conv` | ||
| if self._try_match_and_transform_pattern(op, block, config, visited_ops): |
There was a problem hiding this comment.
This is going to slow down the graph pass, please don't break early
| # visited_ops.add(op) | ||
| return True | ||
|
|
||
| # try pattern II |
There was a problem hiding this comment.
what is pattern II?
| # `quantize` -> `dequant` | ||
| if self._try_match_and_transform_pattern(op, block, activation_stats, visited_ops): | ||
| # has to break as the downstream iterator is affected | ||
| # visited_ops.add(op) |
| with mb.scope(ScopeInfo(source=ScopeSource.COREMLTOOLS_GRAPH_PASS, data=[str(self)])): | ||
| self.apply(prog, sample_data) | ||
|
|
||
| def __str__(self): |
There was a problem hiding this comment.
These two classes are pretty duplicated,
and the __str__, set_options could share the same code from AbstractGraphPass.
There was a problem hiding this comment.
And sample_data, config both can just be the graph option right?
So we could still re-use the original AbstractGraphPass class ...
| # Insert trailing quantize/dequantize pair to valid patterns. | ||
| print("Running compression pass activations_quantization phase 2/4 ...") | ||
| graph_pass = PASS_REGISTRY["compression::insert_quantize_dequantize"] | ||
| graph_pass(prog, config) |
There was a problem hiding this comment.
Yeah, we can just set the config / sample_data as graph pass object's class member I think ....
Prob no need to pass them as inputs for the apply ...
|
|
||
| # Graph pass II | ||
| # Insert trailing quantize/dequantize pair to valid patterns. | ||
| print("Running compression pass activations_quantization phase 2/4 ...") |
| ) | ||
|
|
||
| self.__cached_models[model_key] = model | ||
| # print(f"Created model with intermediate outputs={intermediate_output_names}") |
There was a problem hiding this comment.
Lots of debugging codes in this PR need to be removed ..
No description provided.