Fix dtype_generalizer.py #543
Conversation
|
Thanks for your contribution! |
|
|
||
| def declare_config( | ||
| self, | ||
| dtype_list: str, |
|
|
||
| def sample_handled(self, rel_model_path: str) -> bool: | ||
| return self.naive_sample_handled( | ||
| rel_model_path, search_file_name="op_names.txt" |
There was a problem hiding this comment.
这个search_file_name肯定搞错了。
它表示如果目标目录下找到了op_names.txt文件，就表示已处理。这个op_names明显是op_names_extractor的逻辑，你所生产的逻辑肯定不是这个。你再看看代码，是不是应该写model.py？
There was a problem hiding this comment.
收到，我再通过调试理解到位ResumableSampleMixin的代码逻辑
| def __call__(self, model_path: str) -> List[str]: | ||
| def sample_handled(self, rel_model_path: str) -> bool: | ||
| return self.naive_sample_handled( | ||
| rel_model_path, search_file_name="op_names.txt" |
There was a problem hiding this comment.
这里也一样。serach_file_name肯定搞错了
| def declare_config( | ||
| self, | ||
| dtype_list: str, | ||
| dtype_list: list, |
There was a problem hiding this comment.
dtype_list是list[str], 但是class SamplePass的_check_config_declaration_parameters方法会检查type必须为list，所以这里修改为list，而不是list[str]
| def sample_handled(self, rel_model_path: str) -> bool: | ||
| return self.naive_sample_handled( | ||
| rel_model_path, search_file_name="op_names.txt" | ||
| rel_model_path, search_file_name="model.py" |
There was a problem hiding this comment.
class ResumableSamplePassMixin(SamplePassMixin)的naive_sample_handled方法是通过sample_path是否存在判断sample是否被处理，如果sample path存在，则要搜索生成的文件，所以这里搜索的应该是model.py。
问题：
naive_sample_handled()方法在dtype_generalizer.py的作用？
There was a problem hiding this comment.
naive_sample_handled是基础组件，它并没有和dtype_generalizer有耦合。
naive_sample_handled内部的逻辑非常单一，用于检测当前样本是否已经处理过。
| pass | ||
|
|
||
| def sample_handled(self, rel_model_path: str) -> bool: | ||
| return self.naive_sample_handled(rel_model_path, search_file_name="model.py") |
There was a problem hiding this comment.
这里不能调用naive_sample_handled。
其他大部分的sample pass。都是转换逻辑，不会修改原始的样本。但这个pass不一样，这个pass需要修改graph_net.json。
你先看这个类的resume方法里到底更新了graph_net.json的什么字段。然后在当前的sample_handled函数里，如果graph_net.json里的相关字段已经有值了，就认为该sample已经被处理过了。
|
|
||
| def __call__(self, model_path: str) -> List[str]: | ||
| def sample_handled(self, rel_model_path: str) -> bool: | ||
| return self.naive_sample_handled(rel_model_path, search_file_name="model.py") |
There was a problem hiding this comment.
这里还是不能调用naive_sample_handled。
naive_sample_handled只能处理一个输入sample对应一个输出sample的情形。但你这里需要处理一个输入sample对应多个输出sample的情形。
你可以看看torch/sample_pass/subgraph_generator:SubgraphGenerator.sample_handled方法，它所调用的_has_enough_subgraphs大概是你需要的写的逻辑，不能完全照搬
|
|
||
| def sample_handled(self, rel_model_path: str) -> bool: | ||
| return self.naive_sample_handled(rel_model_path, search_file_name="model.py") | ||
| dst_model_path = Path(self.config["model_path_prefix"]) / rel_model_path |
There was a problem hiding this comment.
代码逻辑：通过判断GraphNet/samples/*/graph_net.json的data_type_generalization_passes是否有值来判断sample是否已被处理
|
|
||
| def sample_handled(self, rel_model_path: str) -> bool: | ||
| return self.naive_sample_handled(rel_model_path, search_file_name="model.py") | ||
| model_path = Path(self.config["model_path_prefix"]) / rel_model_path |
There was a problem hiding this comment.
代码逻辑：例如，在/tmp/dtype_gen_samples生成了resnet18_float16和resnet18_bfloat16，通过判断resnet18_float16和resnet18_bfloat16含有model.py的数量是否等于dtype_pass_names的数量来判断样本是否被处理
2 participants
PR Category
other
Description
Fix dtype_generalizer as a subclass of SamplePass.