环境如下：
tensorflow2.5.0 /
tensorrt8.0/
cuda11.0/
pycuda2021/
bazel3.7.2/
python3.8/
cudatoolkit11.02
感觉现在的tensorflow 好难用，就连安装也没个配置选择，直接就是pip install tensorflow 。。直接晕菜！
先在anaconda 中创建环境python3.8 以及安装cudatoolkit11.0.2,、anaconda会默认安装cudatoolkit10.1, 所以要先安装11.0。然后在终端安装tensorflow 2.5.0: pip install tensorflow-gpu==2.5.0 , anaconda 只能装2.4.1，更系anaconda 也不行。
tensorrt8.0 下载：cuda11.0
https://developer.nvidia.com/nvidia-tensorrt-8x-download
TensorRT 8.0 EA for Linux x86_64 and CUDA 11.0 TAR package
解压后，进入TensorRT-8.0.0.3/python：根据python版本选择要安装的包：
pip install tensorrt-8.0.0.3-cp38-none-linux_x86_64.whl
进入uff 文件夹：pip install uff-0.6.9-py2.py3-none-any.whl
进入graphsurgeon 文件夹： pip install graphsurgeon-0.4.5-py2.py3-none-any.whl
配置环境变量：
vi ～/.bashrc:
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/yao/TensorRT-8.0.0.3/lib
更新一下： source ~/.bashrc
测试一下tensorrt, 在终端或者pycharm 中import tensorrt
如果报错：
libnvrtc.so.11.0: cannot open shared object file: No such file or directory
就在pycharm 中添加环境变量：LD_LIBRARY_PATH=/home/yao/TensorRT-8.0.0.3/lib
可以查询一下 libnvrtc.so.11.0 文件是否存在：sudo find -name libnvrtc.so.11* ,
这个文件就是cudatoolkit11.02 产生的，所以一开始就强调安装。
接着安装bazel3.7.2:
Releases · bazelbuild/bazel · GitHub
选择 bazel-3.7.2-installer-linux-x86_64.sh ，下载后解压。
安装：
chmod +x bazel-3.7.2-installer-linux-x86_64.sh
./bazel-3.7.2-installer-linux-x86_64.sh --user
添加环境变量：
export PATH="$PATH:$HOME/bin"
测试一下：
终端输入：bazel 有信息，没报错就ok.
查看tensorflow2.0 pb 模型的输入输出：
方法一：
from tensorflow.python.framework import tensor_utilfrom google.protobuf import text_formatimport tensorflow as tffrom tensorflow.python.platform import gfilefrom tensorflow.python.framework import tensor_utilGRAPH_PB_PATH = 'models/middlebury_d400.pb' # path to your .pb filewith tf.compat.v1.Session() as sess: print("load graph") with gfile.FastGFile(GRAPH_PB_PATH, 'rb') as f: graph_def = tf.compat.v1.GraphDef() # Note: one of the following two lines work if required libraries are available # text_format.Merge(f.read(), graph_def) graph_def.ParseFromString(f.read()) tf.import_graph_def(graph_def, name='') for i, n in enumerate(graph_def.node): print("Name of the node - %s" % n.name)
方法2：先github 上下载tensorflow2.5.0 源码，然后编译执行：
bazel build tensorflow/tools/graph_transforms:summarize_graph
bazel-bin/tensorflow/tools/graph_transforms/summarize_graph
--in_graph=tensorflow_inception_graph.pb
方法3：
with tf.io.gfile.GFile(MODEL_DIR, "rb") as f: graph_def = tf.compat.v1.GraphDef() graph_def.ParseFromString(f.read())with tf.Graph().as_default() as graph: # tf.import_graph_def(graph_def, name="") for op in graph.get_operations(): # print(op.name, op.values())
tensorflow pb模型转onnx:
安装tf2onnx: pip install -m tf2onnx
https://github.com/onnx/tensorflow-onnx
执行：
python -m tf2onnx.convert
--input documents/HITNET-tf/models/middlebury_d400.pb
--inputs input:0
--outputs reference_output_disparity:0
--output documents/HITNET-tf/models/model.onnx
--verbose
--opset 12
opset 要 改为12版本，不然报错： ERROR - tf2onnx.tfonnx: Failed to convert node......
执行成功：
......
2022-02-16 14:47:56,156 - VERBOSE - tf2onnx.optimizer: Apply merge_duplication
2022-02-16 14:47:56,493 - VERBOSE - tf2onnx.optimizer.MergeDuplicatedNodesOptimizer: no change
2022-02-16 14:47:56,493 - VERBOSE - tf2onnx.optimizer: Apply reshape_optimizer
2022-02-16 14:47:56,726 - VERBOSE - tf2onnx.optimizer.ReshapeOptimizer: no change
2022-02-16 14:47:56,726 - VERBOSE - tf2onnx.optimizer: Apply global_pool_optimizer
2022-02-16 14:47:57,047 - VERBOSE - tf2onnx.optimizer.GlobalPoolOptimizer: no change
2022-02-16 14:47:57,047 - VERBOSE - tf2onnx.optimizer: Apply q_dq_optimizer
2022-02-16 14:47:57,270 - VERBOSE - tf2onnx.optimizer.QDQOptimizer: no change
2022-02-16 14:47:57,270 - VERBOSE - tf2onnx.optimizer: Apply remove_identity
2022-02-16 14:47:57,594 - VERBOSE - tf2onnx.optimizer.IdentityOptimizer: no change
2022-02-16 14:47:57,594 - VERBOSE - tf2onnx.optimizer: Apply remove_back_to_back
2022-02-16 14:47:57,810 - VERBOSE - tf2onnx.optimizer.BackToBackOptimizer: no change
2022-02-16 14:47:57,810 - VERBOSE - tf2onnx.optimizer: Apply einsum_optimizer
2022-02-16 14:47:58,112 - VERBOSE - tf2onnx.optimizer.EinsumOptimizer: no change
2022-02-16 14:47:58,180 - INFO - tf2onnx.optimizer: After optimization: Add -7 (237->230), Cast -65 (470->405), Concat +56 (435->491), Const -4209 (4603->394), Div -6 (77->71), Gather +52 (0->52), Identity -129 (129->0), Less -10 (11->1), Mul -60 (277->217), Reshape +7 (135->142), Shape -15 (103->88), Slice -136 (590->454), Split +65 (48->113), Squeeze -347 (447->100), Sub -14 (189->175), Tile -1 (39->38), Transpose -293 (465->172), Unsqueeze -703 (961->258)
2022-02-16 14:47:58,376 - INFO - tf2onnx:
2022-02-16 14:47:58,376 - INFO - tf2onnx: Successfully converted TensorFlow model documents/HITNET-tf/models/middlebury_d400.pb to ONNX
2022-02-16 14:47:58,376 - INFO - tf2onnx: Model inputs: ['input:0']
2022-02-16 14:47:58,376 - INFO - tf2onnx: Model outputs: ['reference_output_disparity:0']
2022-02-16 14:47:58,376 - INFO - tf2onnx: onNX model is saved at documents/HITNET-tf/models/model.onnx
tensorflow 2.0 转 tensorrt 路径：
1、tf - uff - trt :
在tensorrt8.0 文件夹下uff 安装，pip install uff-0.6.9-py2.py3-none-any.whl
执行：
python ./anaconda3/envs/yolov5/lib/python3.8/site-packages/uff/bin/convert_to_uff.py models/middlebury_d400.pb
报错：
AttributeError: module 'tensorflow' has no attribute 'gfile'
这个问题可能是版本问题，uff 可能不再更新了。。。
2 .tf - onnx -trt
下载onnx-tensorrt-release-6.0，以及tonnx source code 1.8,并将onnx1.8 放入onnx-tensorrt-release-8.0/third_party/onnx 中。
然后执行：
cd onnx-tensorrtmkdir build && cd buildcmake .、-DTENSORRT_ROOT=
sudo make install
在onnx-tensorrt-release-8.0文件下，运行在onnx2trt my_model.onnx -o my_engine.trt
即可生成转换好的trt