fix document, example readme and code (#851)

Signed-off-by: Xin He <[email protected]>

Author: xin3he

docs/source/smooth_quant.md

@@ -309,13 +309,13 @@ conv2d/linear->conv2d/linear/layernorm/batchnorm/instancenorm/t5norm/llamanorm/g
 ```
 
 ## Validated Models
-neural_compressor: 2.1
+Neural Compressor: 2.1
 
-IPEX: 2.0
+IPEX (Intel Extension for PyTorch): 2.0
 
 Dataset: lambada
 
-task: text-generation
+Task: text-generation
 
 alpha [0.4, 0.6] is sweet spot region in SmoothQuant paper
 
@@ -351,18 +351,24 @@ smooth_quant_args description:
 
 "alpha": "auto" or a float value. Default is 0.5. "auto" means automatic tuning.
 
-"folding":
-- False: Allow inserting mul to update the input distribution and not absorbing. IPEX can fuse inserted mul automatically and folding=False is recommended. And for PyTorch FBGEMM backend, folding=False setting will only convert model to QDQ model.
-- True: Only allow inserting mul with the input scale that can be absorbed into the last layer. 
-- If folding not set in config, the default value is IPEX: False (True if version<2.1), Stock PyTorch: True.
+"folding": whether to fold mul into the previous layer, where mul is required to update the input distribution during smoothing.
+- True: Fold inserted mul into the previous layer. IPEX will only insert mul for layers can do folding. 
+- False: Allow inserting mul to update the input distribution and no folding. IPEX (version>=2.1) can fuse inserted mul automatically. For Stock PyTorch, setting folding=False will convert the model to a QDQ model.
 
+## Supported Framework Matrix
+
+| Framework | Alpha        | Folding    |
+|:---------:|--------------|------------|
+| PyTorch   | [0-1] / 'auto' | False      |
+| IPEX      | [0-1] / 'auto' | True / False(Version>2.1) |
+| ONNX      | [0-1]        | True       |
 
 ## Reference
 
-[^1]: Jason, Wei, et al. "Emergent Abilities of Large Language Models". Published in Transactions on Machine Learning Research (2022)
+[^1]: Jason, Wei, et al. "Emergent Abilities of Large Language Models". Published in Transactions on Machine Learning Research (2022).
 
 [^2]: Yvinec, Edouard, et al. "SPIQ: Data-Free Per-Channel Static Input Quantization." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2023.
 
 [^3]: Wei, Xiuying, et al. "Outlier suppression: Pushing the limit of low-bit transformer language models." arXiv preprint arXiv:2209.13325 (2022).
 
-[^4]: Xiao, Guangxuan, et al. "Smoothquant: Accurate and efficient post-training quantization for large language models." arXiv preprint arXiv:2211.10438 (2022)..
+[^4]: Xiao, Guangxuan, et al. "Smoothquant: Accurate and efficient post-training quantization for large language models." arXiv preprint arXiv:2211.10438 (2022).

examples/pytorch/image_recognition/torchvision_models/quantization/ptq/cpu/fx/README.md

@@ -33,36 +33,67 @@ train  val
 ```shell
 python main.py -t -a resnet50 --pretrained /path/to/imagenet
 ```
+or
+```shell
+bash run_tuning.sh --input_model=resnet50 --dataset_location=/path/to/imagenet
+bash run_benchmark.sh --input_model=resnet50 --dataset_location=/path/to/imagenet --mode=performance/accuracy --int8=true/false
+```
 
 ### 2. ResNet18
 
 ```shell
 python main.py -t -a resnet18 --pretrained /path/to/imagenet
 ```
+or
+```shell
+bash run_tuning.sh --input_model=resnet18 --dataset_location=/path/to/imagenet
+bash run_benchmark.sh --input_model=resnet18 --dataset_location=/path/to/imagenet --mode=performance/accuracy --int8=true/false
+```
 
 ### 3. ResNeXt101_32x8d
 
 ```shell
 python main.py -t -a resnext101_32x8d --pretrained /path/to/imagenet
 ```
+or
+```shell
+bash run_tuning.sh --input_model=resnext101_32x8d --dataset_location=/path/to/imagenet
+bash run_benchmark.sh --input_model=resnext101_32x8d --dataset_location=/path/to/imagenet --mode=performance/accuracy --int8=true/false
+```
 
 ### 4. InceptionV3
 
 ```shell
 python main.py -t -a inception_v3 --pretrained /path/to/imagenet
 ```
+or
+```shell
+bash run_tuning.sh --input_model=inception_v3 --dataset_location=/path/to/imagenet
+bash run_benchmark.sh --input_model=inception_v3 --dataset_location=/path/to/imagenet --mode=performance/accuracy --int8=true/false
+```
 
 ### 5. Mobilenet_v2
 
 ```shell
 python main.py -t -a mobilenet_v2 --pretrained /path/to/imagenet
 ```
+or
+```shell
+bash run_tuning.sh --input_model=mobilenet_v2 --dataset_location=/path/to/imagenet
+bash run_benchmark.sh --input_model=mobilenet_v2 --dataset_location=/path/to/imagenet --mode=performance/accuracy --int8=true/false
+```
 
 ### 6. Efficientnet_b0
 
 ```shell
 python main.py -t -a efficientnet_b0 --pretrained /path/to/imagenet
 ```
+or
+```shell
+bash run_tuning.sh --input_model=efficientnet_b0 --dataset_location=/path/to/imagenet
+bash run_benchmark.sh --input_model=efficientnet_b0 --dataset_location=/path/to/imagenet --mode=performance/accuracy --int8=true/false
+```
+
 > **Note**
 >
 > To reduce tuning time and get the result faster, the `efficientnet_b0` model uses 
@@ -74,6 +105,12 @@ python main.py -t -a efficientnet_b0 --pretrained /path/to/imagenet
 ```shell
 python main.py -t -a efficientnet_b3 --pretrained /path/to/imagenet
 ```
+or
+```shell
+bash run_tuning.sh --input_model=efficientnet_b3 --dataset_location=/path/to/imagenet
+bash run_benchmark.sh --input_model=efficientnet_b3 --dataset_location=/path/to/imagenet --mode=performance/accuracy --int8=true/false
+```
+
 > **Note**
 >
 > To reduce tuning time and get the result faster, the `efficientnet_b3` model uses 
@@ -83,6 +120,12 @@ python main.py -t -a efficientnet_b3 --pretrained /path/to/imagenet
 ```shell
 python main.py -t -a efficientnet_b7 --pretrained /path/to/imagenet
 ```
+or
+```shell
+bash run_tuning.sh --input_model=efficientnet_b7 --dataset_location=/path/to/imagenet
+bash run_benchmark.sh --input_model=efficientnet_b7 --dataset_location=/path/to/imagenet --mode=performance/accuracy --int8=true/false
+```
+
 > **Note**
 >
 > To reduce tuning time and get the result faster, the `efficientnet_b7` model uses 

examples/pytorch/image_recognition/torchvision_models/quantization/ptq/cpu/fx/main.py

@@ -96,10 +96,10 @@
 def main():
     args = parser.parse_args()
 
-    if 'efficient' in args.arch:
-        import torchvision.models as models
-    else:
+    if 'mobilenet' in args.arch:
         import torchvision.models.quantization as models
+    else:
+        import torchvision.models as models
 
     if args.seed is not None:
         random.seed(args.seed)

examples/pytorch/image_recognition/torchvision_models/quantization/ptq/cpu/ipex/README.md

@@ -71,8 +71,8 @@ python main.py -t -a resnet18 --ipex --pretrained /path/to/imagenet
 ```
 or
 ```shell
-bash run_tuning.sh --topology=resnet18 --dataset_location=/path/to/imagenet
-bash run_benchmark.sh --topology=resnet18 --dataset_location=/path/to/imagenet --mode=benchmark/accuracy --int8=true/false
+bash run_tuning.sh --input_model=resnet18 --dataset_location=/path/to/imagenet
+bash run_benchmark.sh --input_model=resnet18 --dataset_location=/path/to/imagenet --mode=performance/accuracy --int8=true/false
 ```
 
 ### 2. ResNet50 With Intel PyTorch Extension
@@ -82,8 +82,8 @@ python main.py -t -a resnet50 --ipex --pretrained /path/to/imagenet
 ```
 or
 ```shell
-bash run_tuning.sh --topology=resnet50 --dataset_location=/path/to/imagenet
-bash run_benchmark.sh --topology=resnet50 --dataset_location=/path/to/imagenet --mode=benchmark/accuracy --int8=true/false
+bash run_tuning.sh --input_model=resnet50 --dataset_location=/path/to/imagenet
+bash run_benchmark.sh --input_model=resnet50 --dataset_location=/path/to/imagenet --mode=performance/accuracy --int8=true/false
 ```
 
 ### 3. ResNext101_32x16d With Intel PyTorch Extension
@@ -93,8 +93,8 @@ python main.py -t -a resnext101_32x16d_wsl --hub --ipex --pretrained /path/to/im
 ```
 or
 ```shell
-bash run_tuning.sh --topology=resnext101_32x16d_wsl --dataset_location=/path/to/imagenet
-bash run_benchmark.sh --topology=resnext101_32x16d_wsl --dataset_location=/path/to/imagenet --mode=benchmark/accuracy --int8=true/false
+bash run_tuning.sh --input_model=resnext101_32x16d_wsl --dataset_location=/path/to/imagenet
+bash run_benchmark.sh --input_model=resnext101_32x16d_wsl --dataset_location=/path/to/imagenet --mode=performance/accuracy --int8=true/false
 ```
 
 # Saving and Loading Model