[feat] Optimize HSTU training and sampling process #93
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…e similarity calculation in MatchModel; refine HardNegativeSampler documentation
…fig in model.proto
… handling and configuration
…STU integration tests
tiankongdeguiji
requested changes
Mar 14, 2025
| @@ -40,3 +40,12 @@ docs/source/intro.md | |||
| docs/source/proto.html | |||
|
|
|||
| .vscode/ | |||
| graphlearn* | |||
| @@ -201,8 +201,8 @@ def _get_dataloader( | |||
| dataloader = DataLoader( | |||
| dataset=dataset, | |||
| batch_size=None, | |||
| pin_memory=data_config.pin_memory if mode != Mode.PREDICT else False, | |||
| collate_fn=lambda x: x, | |||
| # pin_memory=data_config.pin_memory if mode != Mode.PREDICT else False, | |||
| @@ -237,6 +240,7 @@ def launch_sampler_cluster( | |||
| multival_sep=self._fg_encoded_multival_sep | |||
| if self._fg_mode == data_pb2.FgMode.FG_NONE | |||
| else chr(29), | |||
| seq_str_delim=self._seq_str_delim, | |||
There was a problem hiding this comment.
seq_str_delim -> item_id_delim, should be param of sampler_config
| features = self._parse_nodes(nodes) | ||
| result_dict = dict(zip(self._attr_names, features)) | ||
| return result_dict | ||
| # ids = np.pad(ids, (0, self._batch_size - len(ids)), "edge") |
| @@ -338,6 +338,8 @@ def _train_and_evaluate( | |||
| ckpt_path: Optional[str] = None, | |||
| eval_result_filename: str = "train_eval_result.txt", | |||
| ) -> None: | |||
| torch.backends.cuda.matmul.allow_tf32 = True | |||
There was a problem hiding this comment.
why should allow tf32?
| self._loss_collection, self.item_tower.group_variational_dropout_loss | ||
| ) | ||
| batch_sparse_features = batch.sparse_features["__BASE__"] | ||
| nonzero_indices = torch.where( |
There was a problem hiding this comment.
why should we need nonzero_indices?
| )[0] | ||
| default_value = torch.tensor([-1]).to(nonzero_indices.device) | ||
| batch_size = torch.cat([nonzero_indices, default_value]).max() + 1 | ||
| neg_sample_size = batch_sparse_features.lengths()[-1] - 1 |
There was a problem hiding this comment.
add comments detaily explain the design
| @@ -181,18 +181,28 @@ def sim( | |||
| user_emb: torch.Tensor, | |||
| item_emb: torch.Tensor, | |||
| neg_for_each_sample: bool = False, | |||
| is_hstu: bool = False, | |||
There was a problem hiding this comment.
do not modify the sim, overwrite sim in hstu.py, and explain the logic in comments
| @@ -317,6 +321,92 @@ def _build_batch(self, input_data: Dict[str, pa.Array]) -> Batch: | |||
| input_data = _expand_tdm_sample( | |||
| input_data, pos_sampled, neg_sampled, self._data_config | |||
| ) | |||
| elif self._enable_hstu: | |||
| seq_attr = self._sampler._item_id_field | |||
| if pa.types.is_string(input_data[seq_attr].type): | |||
There was a problem hiding this comment.
Make line 326-376 a function and move it to datasets/utils.py. Add comments to explain the logic and add a unit test for the function.
| pa.array(input_data_k_split.offsets.to_numpy()[1:] - 1) | ||
| ) | ||
| sampled = self._sampler.get(input_data) | ||
| for k, v in sampled.items(): |
There was a problem hiding this comment.
Make line 377-409 a function and move it to datasets/utils.py. Add comments to explain the logic and add a unit test for the function.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.