Add class LayoutLMv2ForRelationExtraction #15173
Conversation
| class REDecoder(nn.Module): | ||
| def __init__(self, config): | ||
| super().__init__() | ||
| self.entity_emb = nn.Embedding(3, config.hidden_size, scale_grad_by_freq=True) |
There was a problem hiding this comment.
Sorry for butting in, but do you know of any reason to hard code 3 for the num_embeddings parameter (aside from it being the value in unilm)? I came across this PR in prepping to implement a version of this same model for a relation extraction task--so thank you for working on this in the first place! All of that to say, I'm asking for selfish reasons 😅
There was a problem hiding this comment.
fwiw I'm guessing the LayoutXLM authors used 3 because they were dealing with forms that only had 3 semantic entity classes (headers, keys, and values).
There was a problem hiding this comment.
Yes in that case, you can replace it with config.num_labels.
| __version__ = "4.16.0.dev0" | ||
|
|
||
| __version__ = "4.16.0" |
There was a problem hiding this comment.
Don't think you need to change this line.
| class ReOutput(ModelOutput): | ||
| """ | ||
| Base class for outputs of relation extraction models. |
There was a problem hiding this comment.
| class ReOutput(ModelOutput): | |
| """ | |
| Base class for outputs of relation extraction models. | |
| class LayoutLMv2RelationExtractionOutput(ModelOutput): | |
| """ | |
| Class for outputs of [`LayoutLMv2ForRelationExtraction`]. |
Let's give it a clearer name.
Also, this model output class can be placed at the top of modeling_layoutlmv2.py, since it's quite specific to LayoutLMv2.
There was a problem hiding this comment.
Btw, can you use this extensive docstring: https://github.com/R0bk/transformers/blob/9c0e0ba9ccc0d32b795c2c0e0130931b92230292/src/transformers/models/layoutlmv2/outputs_layoutlmv2.py#L26-L74
Thanks!
| @@ -36,10 +36,12 @@ | |||
| QuestionAnsweringModelOutput, | |||
| SequenceClassifierOutput, | |||
| TokenClassifierOutput, | |||
| ReOutput, | |||
There was a problem hiding this comment.
This model output class can be defined within the modeling file itself.
| ) | ||
| from ...modeling_utils import PreTrainedModel, apply_chunking_to_forward | ||
| from ...utils import logging | ||
| from .configuration_layoutlmv2 import LayoutLMv2Config | ||
| from .re import REDecoder |
There was a problem hiding this comment.
You can include the decoder inside modeling_layoutlmv2.py.
Our philosophy is a single model = a single script.
| self.layoutlmv2 = LayoutLMv2Model(config) | ||
| self.dropout = nn.Dropout(config.hidden_dropout_prob) | ||
| self.extractor = REDecoder(config) | ||
| # Initialize weights and apply final processing |
There was a problem hiding this comment.
| # Initialize weights and apply final processing | |
| # Initialize weights and apply final processing |
| return self.layoutlmv2.embeddings.word_embeddings | ||
|
|
||
| @add_start_docstrings_to_model_forward(LAYOUTLMV2_INPUTS_DOCSTRING.format("batch_size, sequence_length")) | ||
| @replace_return_docstrings(output_type=ReOutput, config_class=_CONFIG_FOR_DOC) |
There was a problem hiding this comment.
| @replace_return_docstrings(output_type=ReOutput, config_class=_CONFIG_FOR_DOC) | |
| @replace_return_docstrings(output_type=LayoutLMv2RelationExtractionOutput, config_class=_CONFIG_FOR_DOC) |
| from torch.nn import CrossEntropyLoss | ||
|
|
||
|
|
||
| class BiaffineAttention(torch.nn.Module): |
There was a problem hiding this comment.
| class BiaffineAttention(torch.nn.Module): | |
| class LayoutLMv2BiaffineAttention(torch.nn.Module): |
We usually add a model-specific prefix to each class.
There was a problem hiding this comment.
Thanks for working on this! A few comments:
- the model should be added to the tests (tests/test_modeling_layoutlmv2.py). Normally, if you run
make fixuplocally, it will automatically complain about the fact that this currently isn't the case. - we do have the philosophy of "single model, single file". Hence, you can include the model specific output class, as well as the relation extraction decoder inside
modeling_layoutlm_v2.py.
|
I've implemented the requested changes, but I'm not sure how to contribute to this PR from here. |
|
This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread. Please note that issues that do not follow the contributing guidelines are likely to be ignored. |
|
Hi @quasimik, could you open up a clean PR such that we can add it? |
What does this PR do?
Add class LayoutLMv2ForRelationExtraction which in https://github.com/microsoft/unilm/tree/master/layoutlmft