Add instruction to convert model to ONNX

README.md

@@ -38,6 +38,18 @@ If you are running the whole procedure again, remember to reinstall the older tr
 python train.py --cust_data_init_weights_path ./cust-data/weights --checkpoint_path ./checkpoint/trocr-custdata --dataset_path "./dataset/*/*.jpg" --per_device_train_batch_size 8
 ```
 
+## Optimize inference
+Install dependencies
+```
+python -m pip install optimum
+conda install onnxruntime -c conda-forge
+```
+
+Convert to ONNX
+```
+python -m transformers.onnx --model=checkpoint/trocr-custdata-8000/last --feature=vision2seq-lm onnx/ --atol 1e-3
+```
+
 ## 训练
 ### 初始化模型到自定义训练数据集
 #### 字符集准备参考cust-data/vocab.txt

checkpoint/trocr-custdata-8000/last/preprocessor_config.json

@@ -8,6 +8,7 @@
     0.5
   ],
   "image_processor_type": "ViTImageProcessor",
+  "feature_extractor_type": "ViTFeatureExtractor",
   "image_std": [
     0.5,
     0.5,