docs(//cpp/ptq): READMEs and documentation for running the PTQ example

Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>

Author: narendasan

cpp/ptq/README.md

@@ -2,10 +2,16 @@
 
 This is a short example application that shows how to use TRTorch to perform post-training quantization for a module.
 
+## Prerequisites
+
+1. Download CIFAR10 Dataset Binary version
+2. Train a network on CIFAR10 (see `training/` for a VGG16 recipie)
+3. Export model to torchscript
+
 ## Compilation
 
 ``` shell
-bazel build //cpp/ptq --cxxopt="-DNDEBUG"
+bazel build //cpp/ptq --compilation_mode=opt
 ```
 
 If you want insight into what is going under the hood or need debug symbols
@@ -17,5 +23,12 @@ bazel build //cpp/ptq --compilation_mode=dbg
 ## Usage
 
 ``` shell
-ptq
-```
+ptq <path-to-module> <path-to-cifar10>
+```
+
+## Citations
+
+```
+Krizhevsky, A., & Hinton, G. (2009). Learning multiple layers of features from tiny images.
+Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556.
+```

cpp/ptq/training/vgg16/README.md

@@ -0,0 +1,39 @@
+# VGG16 Trained on CIFAR10
+
+This is a recipe to train a VGG network on CIFAR10 to use with the TRTorch PTQ example.
+
+## Prequisites
+
+```
+pip3 install -r requirements.txt --user
+```
+
+## Training
+
+The following recipe should get somewhere between 89-92% accuracy on the CIFAR10 testset
+```
+python3 main.py --lr 0.01 --batch-size 256 --drop-ratio 0.15 --ckpt-dir $(pwd)/vgg16_ckpts --epochs 100
+```
+
+> 545 was the seed used in testing
+
+You can monitor training with tensorboard, logs are stored by default at `/tmp/vgg16_logs`
+
+## Exporting
+
+Use the exporter script to create a torchscipt module you can compile with TRTorch
+
+```
+python3 export_ckpt.py <path-to-checkpoint>
+```
+
+It should produce a file called `trained_vgg16.jit.pt`
+
+Once the trained VGG network is exported run it with the PTQ example.
+
+## Citations
+
+```
+Krizhevsky, A., & Hinton, G. (2009). Learning multiple layers of features from tiny images.
+Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556.
+```

cpp/ptq/training/vgg16/requirements.txt

@@ -0,0 +1,2 @@
+torch>=1.4.0
+tensorboard>=1.14.0