initial commit

README.md

@@ -0,0 +1,44 @@
+---
+language: en
+license: cc-by-nc-sa-4.0
+pipeline_tag: document-question-answering
+tags:
+ - layoutlm
+ - document-question-answering
+ - pdf
+ - invoices
+widget:
+- text: "What is the invoice number?"
+  src: "https://huggingface.co/spaces/impira/docquery/resolve/2359223c1837a7587402bda0f2643382a6eefeab/invoice.png"
+- text: "What is the purchase amount?"
+  src: "https://huggingface.co/spaces/impira/docquery/resolve/2359223c1837a7587402bda0f2643382a6eefeab/contract.jpeg"
+---
+
+# LayoutLM for Invoices
+
+This is a fine-tuned version of the multi-modal [LayoutLM](https://aka.ms/layoutlm) model for the task of question answering on invoices and other documents. It has been fine-tuned on a proprietary dataset of
+invoices as well as both [SQuAD2.0](https://huggingface.co/datasets/squad_v2) and [DocVQA](https://www.docvqa.org/) for general comprehension.
+
+## Non-consecutive tokens
+
+Unlike other QA models, which can only extract consecutive tokens (because they predict the start and end of a sequence), this model can predict longer-range, non-consecutive sequences with an additional
+classifier head. For example, QA models often encounter this failure mode:
+
+### Before
+
+![Broken Address](./before.png)
+
+
+### After
+
+However this model is able to predict non-consecutive tokens and therefore the address correctly:
+
+![Two-line Address](./after.png)
+
+## Getting started with the model
+
+The best way to use this model is via [DocQuery](https://github.com/impira/docquery).
+
+## About us
+
+This model was created by the team at [Impira](https://www.impira.com/).

config.json

@@ -0,0 +1,32 @@
+{
+  "_name_or_path": "impira/layoutlm-invoices",
+  "architectures": [
+    "LayoutLMForQuestionAnswering"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "eos_token_id": 2,
+  "gradient_checkpointing": false,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "max_2d_position_embeddings": 1024,
+  "max_position_embeddings": 514,
+  "model_type": "layoutlm",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 1,
+  "position_embedding_type": "absolute",
+  "tokenizer_class": "RobertaTokenizer",
+  "token_classification": true,
+  "token_classifier_reduction": "sum",
+  "token_classifier_constant": 0.1,
+  "transformers_version": "4.22.0.dev0",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 50265
+}

pyproject.toml

@@ -0,0 +1,3 @@
+[tool.black]
+line-length = 119
+target-version = ['py35']

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bacb21323970333f86ca1768bc8bd925471e8495521eae181cdff984c11b94f3
+size 511263905

setup.cfg

@@ -0,0 +1,18 @@
+[isort]
+default_section = FIRSTPARTY
+ensure_newline_before_comments = True
+force_grid_wrap = 0
+include_trailing_comma = True
+known_first_party = transformers
+
+line_length = 119
+lines_after_imports = 2
+multi_line_output = 3
+use_parentheses = True
+
+[flake8]
+ignore = E203, E501, E741, W503, W605
+max-line-length = 119
+
+[tool:pytest]
+doctest_optionflags=NUMBER NORMALIZE_WHITESPACE ELLIPSIS

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "sep_token": "</s>", "pad_token": "<pad>", "cls_token": "<s>", "mask_token": {"content": "<mask>", "single_word": false, "lstrip": true, "rstrip": false, "normalized": false}}

tokenizer_config.json

@@ -0,0 +1 @@
+{"unk_token": "<unk>", "bos_token": "<s>", "eos_token": "</s>", "add_prefix_space": false, "errors": "replace", "sep_token": "</s>", "cls_token": "<s>", "pad_token": "<pad>", "mask_token": "<mask>", "model_max_length": 512, "special_tokens_map_file": null, "name_or_path": "roberta-base", "add_prefix_space": true}