How to use

Step 1 Convert model

1. Obtain the CLIP model

2. Use clip2onnx to convert it from pt to onnx

3. (Options) PTQ quantification model

Step2 Preprocess

Two ways to implement image preprocess

1. Using cuda preprocess

2. Using Opencv preprocess

   By the way: Multiple batch preprocessing can refer to this

The way to tokenize the words

1. You can refer to clip_tokenizer

Step3 Build

1. Dynamic batchsize code like reference_1

   Note: Please refer to reference_2 for the solution to the issue where the BufferManager of TensorRT samples does not support engines with dynamic dimensions

Step4 Infer

Step5 Compile

1. Modify and correctly configure dependency paths in CmakeLists at all levels of directories

   Note: The code uses the filesystem library of C++17. If C++17 is not supported, the corresponding functions and modules need to be downgraded and replaced

2. Perform the following actions in the root directory of the project

       mkdir build
       cd build
       cmake ..
       make
   

3. The executable file will be generated in the build/bin directory, and specific runtime parameters can refer to main.cpp and argsParser.h