The quantization of a pre-trained model by using Tensor RT
Budget: $250 – $750 USD
To a developer who is familiar to using Tensor RT :
1. Pre-trained models will be provided.
: The models are trained based on the open source from an ICCV 2021 paper which is also shared as well as its instruction for trainining and inference procedure, please see "README.md" in the attached zip file. In the attached file, only one sample of pre-trained model has been included, temporarily.
2. Quantization (from 32bit to both 16bit and 8bit) should be performed on the provided models in order to obtain more light(computational time reduced instead of losing accuracy) versions by using Tensor RT
(Here, a document for the instructions of how to set related parameters of Tensor RT should be made.)
3. The quantized models should be able to run in the specified desktop environment, which means that the quantized model can be adopted in the inference stage and create results of same format as the model before-quantization. (Linux Ubuntu 20.04.4, NVIDIA GPU of RTX A6000, and other minor dependencies are provided as "requirements.txt" as currently conda environment has been used.) Also, the actual computational time (the average clock time consumption for one frame) should be compared between "when using before-quantization model" and "when using quantization model" in the document.
*Note that "Sample of Pre-trained model.zip" is:
- "output (folder)" contains the pre-trained model "sac_epoch_20.pt (file)" in "checkpoints (subfolder)".
- "tensorboard (subfolder)" in output (folder)" is for the analysis of a training process. [ Not important ]
- "eval (subfolder)" in output (folder)" is created once an inference has been processed.
- "log (folder)" is just to share how the model had been trained and how it produces the result.
1. Pre-trained models will be provided.
: The models are trained based on the open source from an ICCV 2021 paper which is also shared as well as its instruction for trainining and inference procedure, please see "README.md" in the attached zip file. In the attached file, only one sample of pre-trained model has been included, temporarily.
2. Quantization (from 32bit to both 16bit and 8bit) should be performed on the provided models in order to obtain more light(computational time reduced instead of losing accuracy) versions by using Tensor RT
(Here, a document for the instructions of how to set related parameters of Tensor RT should be made.)
3. The quantized models should be able to run in the specified desktop environment, which means that the quantized model can be adopted in the inference stage and create results of same format as the model before-quantization. (Linux Ubuntu 20.04.4, NVIDIA GPU of RTX A6000, and other minor dependencies are provided as "requirements.txt" as currently conda environment has been used.) Also, the actual computational time (the average clock time consumption for one frame) should be compared between "when using before-quantization model" and "when using quantization model" in the document.
*Note that "Sample of Pre-trained model.zip" is:
- "output (folder)" contains the pre-trained model "sac_epoch_20.pt (file)" in "checkpoints (subfolder)".
- "tensorboard (subfolder)" in output (folder)" is for the analysis of a training process. [ Not important ]
- "eval (subfolder)" in output (folder)" is created once an inference has been processed.
- "log (folder)" is just to share how the model had been trained and how it produces the result.