TLDR-516 GPU in docker (#384)

dedoc/readers/pdf_reader/pdf_image_reader/columns_orientation_classifier/columns_orientation_classifier.py

@@ -19,28 +19,25 @@ class ColumnsOrientationClassifier(object):
     Class Classifier for work with Orientation Network. This class set device,
     preprocessing (transform) input data, weights of model
     """
-
-    _nets = {}
-
     def __init__(self, on_gpu: bool, checkpoint_path: Optional[str], *, config: dict) -> None:
         self.logger = config.get("logger", logging.getLogger())
         self._set_device(on_gpu)
         self._set_transform_image()
         self.checkpoint_path = path.abspath(checkpoint_path)
         self.classes = [1, 2, 0, 90, 180, 270]
+        self._net = None
 
     @property
     def net(self) -> ClassificationModelTorch:
-        # lazy loading and net sharing, comrade
-        if self.checkpoint_path not in self._nets:
+        if self._net is None:
             if self.checkpoint_path is not None:
                 net = ClassificationModelTorch(path.join(self.checkpoint_path, "scan_orientation_efficient_net_b0.pth"))
                 self._load_weights(net)
             else:
                 net = ClassificationModelTorch(None)
-            net.to(self.device)
-            self._nets[self.checkpoint_path] = net
-        return self._nets[self.checkpoint_path]
+            self._net = net
+        self._net.to(self.device)
+        return self._net
 
     @staticmethod
     def my_resize(image: Image) -> Image:

docker-compose.yml

@@ -28,8 +28,3 @@ services:
       PYTHONPATH: $PYTHONPATH:/dedoc_root/tests:/dedoc_root
     command:
         bash dedoc_root/tests/run_tests_in_docker.sh
-
-## By default this config uses default local driver,
-## For custom volumes replace with volume driver configuration.
-volumes:
-  data1-1:

docker_gpu/Dockerfile

@@ -0,0 +1,24 @@
+ARG REPOSITORY="docker.io"
+FROM dedocproject/dedoc_p3.9_base:version_2023_08_28
+
+ENV PYTHONPATH "${PYTHONPATH}:/dedoc_root"
+ENV RESOURCES_PATH "/dedoc_root/resources"
+
+ADD requirements.txt .
+RUN pip3 install --no-cache-dir -r requirements.txt
+RUN pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html
+
+RUN mkdir /dedoc_root
+RUN mkdir /dedoc_root/dedoc
+ADD dedoc/config.py /dedoc_root/dedoc/config.py
+ADD dedoc/download_models.py /dedoc_root/dedoc/download_models.py
+RUN python3 /dedoc_root/dedoc/download_models.py
+
+ADD dedoc /dedoc_root/dedoc
+ADD VERSION /dedoc_root
+RUN echo "__version__ = \"$(cat /dedoc_root/VERSION)\"" > /dedoc_root/dedoc/version.py
+
+ADD tests /dedoc_root/tests
+ADD resources /dedoc_root/resources
+
+CMD ["python3", "/dedoc_root/dedoc/main.py", "-c", "/dedoc_root/dedoc/config.py"]

docker_gpu/README.md

@@ -0,0 +1,17 @@
+To run Dedoc on CUDA with Docker use files from `docker_gpu` directory 
+([CUDA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html) 
+should be installed on your machine):
+
+1. Set `on_gpu` to `True` in `config.py`
+2. Run application:
+```shell
+cd docker_gpu
+docker-compose up --build
+```
+
+You can change index of CUDA device at `docker-compose.yml`:
+```
+NVIDIA_VISIBLE_DEVICES: 0
+NVIDIA_VISIBLE_DEVICES: 0, 3
+NVIDIA_VISIBLE_DEVICES: all
+```

docker_gpu/docker-compose.yml

@@ -0,0 +1,33 @@
+version: '2.4'
+
+services:
+  dedoc:
+    mem_limit: 16G
+    build:
+      context: ..
+      dockerfile: docker_gpu/Dockerfile
+    restart: always
+    tty: true
+    ports:
+      - 1231:1231
+    environment:
+      DOCREADER_PORT: 1231
+      NVIDIA_VISIBLE_DEVICES: 0
+    runtime: nvidia
+
+  test:
+    depends_on:
+      - dedoc
+    build:
+          context: ..
+          dockerfile: docker_gpu/Dockerfile
+    tty: true
+    environment:
+      DOC_READER_HOST: "dedoc"
+      DOCREADER_PORT: 1231
+      is_test: $test
+      PYTHONPATH: $PYTHONPATH:/dedoc_root/tests:/dedoc_root
+      NVIDIA_VISIBLE_DEVICES: 0
+    runtime: nvidia
+    command:
+        bash dedoc_root/tests/run_tests_in_docker.sh