init

app.py

@@ -0,0 +1,33 @@
+"""
+Donut
+Copyright (c) 2022-present NAVER Corp.
+MIT License
+"""
+import argparse
+
+import gradio as gr
+import torch
+
+from donut import DonutModel
+
+def demo_process(input_img):
+    global pretrained_model, task_prompt, task_name
+    output = pretrained_model.inference(image=input_img, prompt=task_prompt)["predictions"][0]
+    return output["text_sequence"].replace(" ", "") # temp
+
+task_prompt = f"<s_kuzushiji>"
+pretrained_model = DonutModel.from_pretrained("naver-clova-ix/donut-base-finetuned-kuzushiji")
+pretrained_model.eval()
+
+demo = gr.Interface(
+    fn=demo_process,
+    inputs= gr.inputs.Image(type="pil"),
+    outputs="text",
+    title=f"Donut 🍩 demonstration for Kuzushiji Decoding",
+    description="""This is a toy example for decoding kuzushiji (old Japanese cursive characters, くずし字) documents with a single E2E model, `Document Understanding Transformer` (Donut 🍩, ECCV-22). This particular model is fine-tuned on <a href="http://codh.rois.ac.jp/char-shape/">Kuzushiji Dataset</a>. To use it, simply upload a kuzushiji document image or use one of the examples below and click `Submit`. Results will show up in a few seconds.<br>* Note that this demo is running on a small resource environment, `basic CPU plan`  (`2 vCPU, 16GiB RAM`).<br>* Demonstrations for other types of documents/tasks are available at https://github.com/clovaai/donut<br>
+    * More details of Donut are available at <a href="https://arxiv.org/abs/2111.15664">Paper</a>, <a href="https://github.com/clovaai/donut">GitHub</a>, and <a href="https://huggingface.co/docs/transformers/model_doc/donut">Huggingface 🤗 Implementation Page</a>.<br>
+    * Kuzushiji Dataset is from <a href="http://codh.rois.ac.jp/char-shape/">Dataset Link</a> (Reference: 『日本古典籍くずし字データセット』（国文研ほか所蔵／CODH加工）doi:10.20676/00000340).""",
+    examples=[["sample1.jpg"], ["sample2.jpg"]],
+    cache_examples=False,
+)
+demo.launch()

requirements.txt

@@ -0,0 +1,3 @@
+torch
+donut-python
+gradio