Merge pull request #59 from OleehyO/pre_release

Pre release

README.md

@@ -209,11 +209,12 @@ python server.py
 | `-ckpt` | The path to the weights file,*default is TexTeller's pretrained weights*. |
 | `-tknz` | The path to the tokenizer,*default is TexTeller's tokenizer*. |
 | `-port` | The server's service port,*default is 8000*. |
-| `--inference-mode` | Whether to use GPU(cuda or mps) for inference,*default is CPU*. |
+| `--inference-mode` | Whether to use "cuda" or "mps" for inference,*default is "cpu"*. |
 | `--num_beams` | The number of beams for beam search,*default is 1*. |
 | `--num_replicas` | The number of service replicas to run on the server,*default is 1 replica*. You can use more replicas to achieve greater throughput.|
 | `--ncpu_per_replica` | The number of CPU cores used per service replica,*default is 1*.|
 | `--ngpu_per_replica` | The number of GPUs used per service replica,*default is 1*. You can set this value between 0 and 1 to run multiple service replicas on one GPU to share the GPU, thereby improving GPU utilization. (Note, if --num_replicas is 2, --ngpu_per_replica is 0.7, then 2 GPUs must be available) |
+| `-onnx` | Perform inference using Onnx Runtime, *disabled by default* |
 
 > [!NOTE]
 > A client demo can be found at `src/client/demo.py`, you can refer to `demo.py` to send requests to the server

assets/README_zh.md

@@ -247,11 +247,12 @@ python server.py
 | `-ckpt` | 权重文件的路径，*默认为TexTeller的预训练权重*。|
 | `-tknz` | 分词器的路径，*默认为TexTeller的分词器*。|
 | `-port` | 服务器的服务端口，*默认是8000*。|
-| `--inference-mode` | 是否使用GPU(cuda或mps)推理，*默认为CPU*。|
+| `--inference-mode` | 使用"cuda"或"mps"推理，*默认为"cpu"*。|
 | `--num_beams` | beam search的beam数量，*默认是1*。|
 | `--num_replicas` | 在服务器上运行的服务副本数量，*默认1个副本*。你可以使用更多的副本来获取更大的吞吐量。|
 | `--ncpu_per_replica` | 每个服务副本所用的CPU核心数，*默认为1*。|
 | `--ngpu_per_replica` | 每个服务副本所用的GPU数量，*默认为1*。你可以把这个值设置成 0~1之间的数，这样会在一个GPU上运行多个服务副本来共享GPU，从而提高GPU的利用率。(注意，如果 --num_replicas 2, --ngpu_per_replica 0.7, 那么就必须要有2个GPU可用) |
+| `-onnx` | 使用Onnx Runtime进行推理，*默认不使用*。|
 
 > [!NOTE]
 > 一个客户端demo可以在 `TexTeller/client/demo.py`找到，你可以参考 `demo.py`来给server发送请求

setup.py

@@ -20,6 +20,7 @@ install_requires = [
     "streamlit-paste-button",
     "shapely",
     "pyclipper",
+    "optimum[exporters]"
 ]
 
 # Add platform-specific dependencies

src/client_demo.py

@@ -6,7 +6,7 @@ det_server_url = "http://127.0.0.1:8000/fdet"
 img_path = "/your/image/path/"
 with open(img_path, 'rb') as img:
     files = {'img': img}
-    response = requests.post(det_server_url, files=files)
+    response = requests.post(rec_server_url, files=files)
-    # response = requests.post(rec_server_url, files=files)
+    # response = requests.post(det_server_url, files=files)
 
 print(response.text)

src/models/ocr_model/model/TexTeller.py

@@ -1,4 +1,5 @@
 from pathlib import Path
+from optimum.onnxruntime import ORTModelForVision2Seq
 
 from ...globals import (
     VOCAB_SIZE,
@@ -10,7 +11,7 @@ from ...globals import (
 from transformers import (
     RobertaTokenizerFast,
     VisionEncoderDecoderModel,
-    VisionEncoderDecoderConfig,
+    VisionEncoderDecoderConfig
 )
 
 
@@ -26,9 +27,13 @@ class TexTeller(VisionEncoderDecoderModel):
         super().__init__(config=config)
     
     @classmethod
-    def from_pretrained(cls, model_path: str = None):
+    def from_pretrained(cls, model_path: str = None, use_onnx=False, onnx_provider=None):
         if model_path is None or model_path == 'default':
-            return VisionEncoderDecoderModel.from_pretrained(cls.REPO_NAME)
+            if not use_onnx:
+                return VisionEncoderDecoderModel.from_pretrained(cls.REPO_NAME)
+            else:
+                use_gpu = True if onnx_provider == 'cuda' else False
+                return ORTModelForVision2Seq.from_pretrained(cls.REPO_NAME, provider="CUDAExecutionProvider" if use_gpu else "CPUExecutionProvider")
         model_path = Path(model_path).resolve()
         return VisionEncoderDecoderModel.from_pretrained(str(model_path))
 

src/models/ocr_model/model/config.json

@@ -1,4 +1,5 @@
 {
+  "_name_or_path": "OleehyO/TexTeller",
   "architectures": [
     "VisionEncoderDecoderModel"
   ],
@@ -10,9 +11,11 @@
     "architectures": null,
     "attention_dropout": 0.0,
     "bad_words_ids": null,
+    "begin_suppress_tokens": null,
     "bos_token_id": 0,
     "chunk_size_feed_forward": 0,
     "classifier_dropout": 0.0,
+    "cross_attention_hidden_size": 768,
     "d_model": 1024,
     "decoder_attention_heads": 16,
     "decoder_ffn_dim": 4096,
@@ -23,9 +26,9 @@
     "do_sample": false,
     "dropout": 0.1,
     "early_stopping": false,
-    "cross_attention_hidden_size": 768,
     "encoder_no_repeat_ngram_size": 0,
     "eos_token_id": 2,
+    "exponential_decay_length_penalty": null,
     "finetuning_task": null,
     "forced_bos_token_id": null,
     "forced_eos_token_id": null,
@@ -40,9 +43,10 @@
       "LABEL_0": 0,
       "LABEL_1": 1
     },
+    "layernorm_embedding": true,
     "length_penalty": 1.0,
     "max_length": 20,
-    "max_position_embeddings": 512,
+    "max_position_embeddings": 1024,
     "min_length": 0,
     "model_type": "trocr",
     "no_repeat_ngram_size": 0,
@@ -62,8 +66,10 @@
     "return_dict_in_generate": false,
     "scale_embedding": false,
     "sep_token_id": null,
+    "suppress_tokens": null,
     "task_specific_params": null,
     "temperature": 1.0,
+    "tf_legacy_loss": false,
     "tie_encoder_decoder": false,
     "tie_word_embeddings": true,
     "tokenizer_class": null,
@@ -71,10 +77,11 @@
     "top_p": 1.0,
     "torch_dtype": null,
     "torchscript": false,
-    "transformers_version": "4.12.0.dev0",
+    "typical_p": 1.0,
     "use_bfloat16": false,
     "use_cache": false,
-    "vocab_size": 50265
+    "use_learned_position_embeddings": true,
+    "vocab_size": 15000
   },
   "encoder": {
     "_name_or_path": "",
@@ -82,15 +89,18 @@
     "architectures": null,
     "attention_probs_dropout_prob": 0.0,
     "bad_words_ids": null,
+    "begin_suppress_tokens": null,
     "bos_token_id": null,
     "chunk_size_feed_forward": 0,
+    "cross_attention_hidden_size": null,
     "decoder_start_token_id": null,
     "diversity_penalty": 0.0,
     "do_sample": false,
     "early_stopping": false,
-    "cross_attention_hidden_size": null,
     "encoder_no_repeat_ngram_size": 0,
+    "encoder_stride": 16,
     "eos_token_id": null,
+    "exponential_decay_length_penalty": null,
     "finetuning_task": null,
     "forced_bos_token_id": null,
     "forced_eos_token_id": null,
@@ -101,7 +111,7 @@
       "0": "LABEL_0",
       "1": "LABEL_1"
     },
-    "image_size": 384,
+    "image_size": 448,
     "initializer_range": 0.02,
     "intermediate_size": 3072,
     "is_decoder": false,
@@ -119,7 +129,7 @@
     "num_attention_heads": 12,
     "num_beam_groups": 1,
     "num_beams": 1,
-    "num_channels": 3,
+    "num_channels": 1,
     "num_hidden_layers": 12,
     "num_return_sequences": 1,
     "output_attentions": false,
@@ -136,8 +146,10 @@
     "return_dict": true,
     "return_dict_in_generate": false,
     "sep_token_id": null,
+    "suppress_tokens": null,
     "task_specific_params": null,
     "temperature": 1.0,
+    "tf_legacy_loss": false,
     "tie_encoder_decoder": false,
     "tie_word_embeddings": true,
     "tokenizer_class": null,
@@ -145,12 +157,12 @@
     "top_p": 1.0,
     "torch_dtype": null,
     "torchscript": false,
-    "transformers_version": "4.12.0.dev0",
+    "typical_p": 1.0,
     "use_bfloat16": false
   },
   "is_encoder_decoder": true,
   "model_type": "vision-encoder-decoder",
   "tie_word_embeddings": false,
-  "torch_dtype": "float32",
+  "transformers_version": "4.41.2",
-  "transformers_version": null
+  "use_cache": true
 }

src/models/ocr_model/utils/inference.py

@@ -20,7 +20,9 @@ def inference(
 ) -> List[str]:
     if imgs == []:
         return []
-    model.eval()
+    if hasattr(model, 'eval'):
+        # not onnx session, turn model.eval()
+        model.eval()
     if isinstance(imgs[0], str):
         imgs = convert2rgb(imgs) 
     else:  # already numpy array(rgb format)
@@ -29,7 +31,9 @@ def inference(
     imgs = inference_transform(imgs)
     pixel_values = torch.stack(imgs)
 
-    model = model.to(accelerator)
+    if hasattr(model, 'eval'):
+        # not onnx session, move weights to device
+        model = model.to(accelerator)
     pixel_values = pixel_values.to(accelerator)
 
     generate_config = GenerationConfig(

src/server.py

@@ -1,3 +1,4 @@
+import sys
 import argparse
 import tempfile
 import time
@@ -17,6 +18,10 @@ from models.det_model.inference import PredictConfig
 from models.ocr_model.utils.to_katex import to_katex
 
 
+PYTHON_VERSION = str(sys.version_info.major) + '.' + str(sys.version_info.minor)
+LIBPATH = Path(sys.executable).parent.parent / 'lib' / ('python' + PYTHON_VERSION) / 'site-packages'
+CUDNNPATH = LIBPATH / 'nvidia' / 'cudnn' / 'lib'
+
 parser = argparse.ArgumentParser()
 parser.add_argument(
     '-ckpt', '--checkpoint_dir', type=str
@@ -31,6 +36,7 @@ parser.add_argument('--ngpu_per_replica', type=float, default=0.0)
 
 parser.add_argument('--inference-mode', type=str, default='cpu')
 parser.add_argument('--num_beams', type=int, default=1)
+parser.add_argument('-onnx', action='store_true', help='using onnx runtime')
 
 args = parser.parse_args()
 if args.ngpu_per_replica > 0 and not args.inference_mode == 'cuda':
@@ -41,7 +47,7 @@ if args.ngpu_per_replica > 0 and not args.inference_mode == 'cuda':
     num_replicas=args.num_replicas, 
     ray_actor_options={
         "num_cpus": args.ncpu_per_replica, 
-        "num_gpus": args.ngpu_per_replica
+        "num_gpus": args.ngpu_per_replica * 1.0 / 2
     }
 )
 class TexTellerRecServer:
@@ -50,14 +56,16 @@ class TexTellerRecServer:
         checkpoint_path: str, 
         tokenizer_path: str, 
         inf_mode: str = 'cpu',
+        use_onnx: bool = False,
         num_beams: int = 1
     ) -> None:
-        self.model = TexTeller.from_pretrained(checkpoint_path)
+        self.model = TexTeller.from_pretrained(checkpoint_path, use_onnx=use_onnx, onnx_provider=inf_mode)
         self.tokenizer = TexTeller.get_tokenizer(tokenizer_path)
         self.inf_mode = inf_mode
         self.num_beams = num_beams
 
-        self.model = self.model.to(inf_mode) if inf_mode != 'cpu' else self.model
+        if not use_onnx:
+            self.model = self.model.to(inf_mode) if inf_mode != 'cpu' else self.model
     
     def predict(self, image_nparray) -> str:
         return to_katex(rec_inference(
@@ -65,14 +73,28 @@ class TexTellerRecServer:
             accelerator=self.inf_mode, num_beams=self.num_beams
         )[0])
 
-
+@serve.deployment(
-@serve.deployment(num_replicas=args.num_replicas)
+    num_replicas=args.num_replicas, 
+    ray_actor_options={
+        "num_cpus": args.ncpu_per_replica, 
+        "num_gpus": args.ngpu_per_replica * 1.0 / 2,
+        "runtime_env": {
+            "env_vars": {
+                "LD_LIBRARY_PATH": f"{str(CUDNNPATH)}/:$LD_LIBRARY_PATH"
+            }
+        }
+    },
+)
 class TexTellerDetServer:
     def __init__(
-        self
+        self,
+        inf_mode='cpu'
     ):
         self.infer_config = PredictConfig("./models/det_model/model/infer_cfg.yml")
-        self.latex_det_model = InferenceSession("./models/det_model/model/rtdetr_r50vd_6x_coco.onnx")
+        self.latex_det_model = InferenceSession(
+            "./models/det_model/model/rtdetr_r50vd_6x_coco.onnx", 
+            providers=['CUDAExecutionProvider'] if inf_mode == 'cuda' else ['CPUExecutionProvider']
+        )
 
     async def predict(self, image_nparray) -> str:
         with tempfile.TemporaryDirectory() as temp_dir:
@@ -120,11 +142,12 @@ if __name__ == '__main__':
     rec_server = TexTellerRecServer.bind(
         ckpt_dir, tknz_dir, 
         inf_mode=args.inference_mode,
+        use_onnx=args.onnx,
         num_beams=args.num_beams
     )
     det_server = None
     if Path('./models/det_model/model/rtdetr_r50vd_6x_coco.onnx').exists():
-        det_server = TexTellerDetServer.bind()
+        det_server = TexTellerDetServer.bind(args.inference_mode)
     ingress = Ingress.bind(det_server, rec_server)
 
     # ingress_handle = serve.run(ingress, route_prefix="/predict")  

src/web.py

@@ -50,17 +50,20 @@ fail_gif_html = '''
 '''
 
 @st.cache_resource
-def get_texteller():
+def get_texteller(use_onnx, accelerator):
-    return TexTeller.from_pretrained(os.environ['CHECKPOINT_DIR'])
+    return TexTeller.from_pretrained(os.environ['CHECKPOINT_DIR'], use_onnx=use_onnx, onnx_provider=accelerator)
 
 @st.cache_resource
 def get_tokenizer():
     return TexTeller.get_tokenizer(os.environ['TOKENIZER_DIR'])
 
 @st.cache_resource
-def get_det_models():
+def get_det_models(accelerator):
     infer_config = PredictConfig("./models/det_model/model/infer_cfg.yml")
-    latex_det_model = InferenceSession("./models/det_model/model/rtdetr_r50vd_6x_coco.onnx")
+    latex_det_model = InferenceSession(
+        "./models/det_model/model/rtdetr_r50vd_6x_coco.onnx", 
+        providers=['CUDAExecutionProvider'] if accelerator == 'cuda' else ['CPUExecutionProvider']
+    )
     return infer_config, latex_det_model
 
 @st.cache_resource()
@@ -141,18 +144,22 @@ with st.sidebar:
         on_change=change_side_bar
     )
 
+    st.markdown("## Seepup Setting")
+    use_onnx = st.toggle("ONNX Runtime ")
+
+
 
 ##############################     </sidebar>    ##############################
 
 
 ################################     <page>    ################################
 
-texteller = get_texteller()
+texteller = get_texteller(use_onnx, accelerator)
 tokenizer = get_tokenizer()
 latex_rec_models = [texteller, tokenizer]
 
 if inf_mode == "Paragraph recognition":
-    infer_config, latex_det_model = get_det_models()
+    infer_config, latex_det_model = get_det_models(accelerator)
     lang_ocr_models = get_ocr_models(accelerator)
 
 st.markdown(html_string, unsafe_allow_html=True)