Changes to run on Jetson Xavier NX.

deploy_trt.sh

@@ -31,7 +31,7 @@ max_size=640,640
 # Force FP16 mode for building TensorRT engines, even if it's not supported.
 # Please check that your GPU supports FP16, otherwise performance may drop.
 # For GPUs supporting it gives about 2x performance boost.
-force_fp16=False
+force_fp16=True
 
 
 # DET MODELS:
@@ -100,7 +100,7 @@ for i in $(seq 0 $(($n_gpu - 1)) ); do
     docker rm $name;
     echo --- Starting container $name  with $device  at port $port;
     ((p++));
-    docker run  -p $port:18080\
+    docker run --runtime=nvidia -p $port:18080\
         --gpus $device\
         -d\
         -e LOG_LEVEL=$log_level\

src/Dockerfile_trt

@@ -1,17 +1,33 @@
-FROM nvcr.io/nvidia/tensorrt:21.12-py3
+FROM nvcr.io/nvidia/l4t-ml:r32.6.1-py3
 
-ENV LC_ALL=C.UTF-8
-ENV LANG=C.UTF-8
+ENV LC_ALL=en_US.UTF-8
+ENV LANG=en_US.UTF-8
 
-RUN apt-get update && apt-get install -y libgl1-mesa-glx protobuf-compiler libprotoc-dev libb64-0d libturbojpeg
-ENV PIP_INSTALL="python -m pip --no-cache-dir install --upgrade"
-RUN $PIP_INSTALL requests tqdm opencv-python scikit-image fastapi uvicorn ujson numba aiofiles python-multipart httpx cupy-cuda114 icc_rt
-RUN $PIP_INSTALL onnx==1.8.0
+RUN apt-get update
+RUN DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC apt-get -y install tzdata --upgrade
+RUN apt-get install -y libgl1-mesa-glx protobuf-compiler libprotoc-dev libb64-0d libturbojpeg python3-pip
+ENV PIP_INSTALL="python3 -m pip --no-cache-dir install --upgrade"
+RUN $PIP_INSTALL pip
+RUN $PIP_INSTALL requests tqdm
+RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y locales
+
+RUN sed -i -e 's/# en_US.UTF-8 UTF-8/en_US.UTF-8 UTF-8/' /etc/locale.gen && \
+    dpkg-reconfigure --frontend=noninteractive locales && \
+    update-locale LANG=en_US.UTF-8
+
+RUN $PIP_INSTALL scikit-image
+
+RUN $PIP_INSTALL fastapi uvicorn ujson aiofiles python-multipart httpx
+RUN apt-get install -y cmake git
+RUN git clone --recursive https://github.com/onnx/optimizer onnxoptimizer && cd onnxoptimizer && pip3 install -e . && cd ~ && rm -rf onnxoptimizer
 RUN $PIP_INSTALL onnx-simplifier
-RUN $PIP_INSTALL nvidia-pyindex && pip install tritonclient[http,grpc]
+RUN $PIP_INSTALL nvidia-pyindex
+RUN $PIP_INSTALL tritonclient[http,grpc]
 RUN $PIP_INSTALL git+git://github.com/lilohuang/PyTurboJPEG.git msgpack
-RUN $PIP_INSTALL pynvjpeg
-
+COPY Tegra_Multimedia_API_R32.2.1_aarch64.tbz2 /tmp
+RUN tar xvjf /tmp/Tegra_Multimedia_API_R32.2.1_aarch64.tbz2 && rm /tmp/Tegra_Multimedia_API_R32.2.1_aarch64.tbz2
+RUN mv tegra_multimedia_api /usr/src/jetson_multimedia_api
+#RUN $PIP_INSTALL pynvjpeg
 
 WORKDIR /app
 COPY api_trt /app

src/api_trt/entrypoint.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
 echo Preparing models...
-python prepare_models.py
+python3 prepare_models.py
 
 echo Starting InsightFace-REST using $NUM_WORKERS workers.
-uvicorn app:app --port 18080 --host 0.0.0.0 --workers $NUM_WORKERS  --timeout-keep-alive 20000
+uvicorn app:app --port 18080 --host 0.0.0.0 --workers $NUM_WORKERS  --timeout-keep-alive 20000

src/api_trt/modules/utils/download_google.py

@@ -1,7 +1,7 @@
 import requests
 import hashlib
 from tqdm import tqdm
-
+import re
 
 
 def check_hash(filename, hash, algo='md5'):
@@ -48,6 +48,11 @@ def get_confirm_token(response):
             if key.startswith('download_warning'):
                 return value
 
+        if 'text/html' in response.headers['Content-Type']:
+            m = re.search('.*confirm=([^\"]*)', response.text, re.M)
+            if m and m.groups():
+                return m.groups()[0]
+
         return None
 
     def save_response_content(response, destination):