Skip to content

Latest commit

 

History

History
204 lines (146 loc) · 9.5 KB

gpu_model_example.md

File metadata and controls

204 lines (146 loc) · 9.5 KB
Raw

Unified Inference Frontend (UIF) 1.2 User Guide

Step 2.6: GPU Model Example

Table of Contents

UIF accelerates deep learning inference applications on all AMD compute platforms for popular machine learning frameworks, including TensorFlow, PyTorch, and ONNXRT. UIF 1.2 extends the support to AMD Radeon™ GPUs in addition to AMD Instinct™ GPUs. Currently, MIGraphX is the acceleration library for Deep Learning Inference running on AMD Instinct GPUs.

The following example takes a PyTorch ResNet-50-v1.5 model selected from UIF Model Zoo as an example to show how it works on different GPU platforms.

Note: The model tuning time on a MI210 device is long (around three hours). With MI210, it is recommended to skip this step and use the YModel provided in the model packages.

2.6.1: Installation

After unzipping the PyTorch ResNet50-v1.5 model, you need to set environment and install dependencies. It is recommended that you use Docker® software for training and testing.

  1. Download the UIF Docker image with the following instruction:
docker pull amdih/uif-pytorch:uif1.2_rocm5.6.1_vai3.5_py3.8_pytorch1.13

Note: AMD GPU dependencies and PyTorch are pre-installed in the Docker.

  1. Launch a container from the Docker image. 

  2. Install the Python dependencies using pip:

$ pip install --user -r requirements.txt

Note: Modern servers commonly feature multi-socket or multi-core architectures. GPU nodes possess an affinity for specific CPU cores, crucial for efficient memory transfers between the GPU and the host. Ensuring these transfers occur through the designated CPU cores is essential. Neglecting GPU-CPU affinity, especially for larger models, can result in unstable performance. To address this, you can enhance GPU-CPU affinity by setting up Non-Uniform Memory Access (NUMA) bindings when running the Docker image using the --cpuset-cpus and ROCR_VISIBLE_DEVICES arguments.

2.6.2: Data Preparation

Next, use ImageNet to retrain and test this ResNet50 model.

For the ImageNet dataset, go to ImageNet. The downloaded ImageNet dataset needs to be organized in the following format:

  1. Create a folder of "data" . Put the validation data in +data/val.
  2. Data from each class for validation set needs to be put in the folder:
  +data/Imagenet/val
       +val/n01847000
       +val/n02277742
       +val/n02808304
       +...

2.6.3: Training and Evaluation

  1. Train ResNet50 on ImageNet from scratch.

The pretrained ResNet50 model is provided in the float folder. If you want to evaluate directly, skip this step.

  $ cd code
  $ sh run_train.sh
  
  
Training...
Used arguments: Namespace(data_dir='data/imagenet/', deployable='./deployable.pth', display_freq=100, early_stop=False, epochs=100, lr=0.01, mode='train', pretrained=None, prune_ratio=0, quantizer_norm=True, save_dir='./save', train_batch_size=256, val_batch_size=64, val_freq=1000, warmup_epochs=5, weight_decay=0.0001, weight_lr_decay=0.94, workers=4)
Fri Dec  9 14:26:00 2022 , Learning rate: 3.9960039960039963e-07
Epoch[0], Step: [     0/500500] Time  2.094 ( 2.094)    Data  0.000 ( 0.000)    Loss 7.0611e+00 (7.0611e+00)    Acc@1   0.00 (  0.00)   Acc@5   0.39 (  0.39)
Epoch[0], Step: [   100/500500] Time  0.379 ( 0.431)    Data  0.000 ( 0.000)    Loss 7.0727e+00 (7.0804e+00)    Acc@1   0.00 (  0.09)   Acc@5   0.39 (  0.41)
Epoch[0], Step: [   200/500500] Time  0.647 ( 0.425)    Data  0.000 ( 0.000)    Loss 7.0127e+00 (7.0654e+00)    Acc@1   0.00 (  0.10)   Acc@5   0.39 (  0.43)
...
  1. Evaluate the accuracy of the FP32 float model.
  $ sh run_test_float.sh
  
  
-------- Start resnet50 test
Model name: resnet50 
Model type: pth 
Loading model: float/resnet50_pretrained.pth 
██████████████████████████████████████| 391/391 [22:07<00:00,  3.40s/it] top-1 / top-5 accuracy: 76.1 / 92.9
-------- End of resnet50 test
  1. Evaluate accuracy and performance with MIGraphX. Before the evaluation, convert the native PyTorch model (.pth) to an FP32/FP16 onnx model. Run the following script to get both FP32 and FP16 onnx models.
 $ sh export_float_onnx_model.sh

The onnx models are provided in the float folder, so you can choose to skip this step.

  • Run inference and evaluation with MIGraphX. Here, scripts are provided to evaluate the model's accuracy with MIGraphX.
  $ sh run_test_migraphx.sh
  
  
-------- Start resnet50 test
=== Load pretrained model ===                                                                                                                                                                                                                                            
Model name: resnet50
Model type: onnx                                                                                                                                                                                                                                                         
Loading model: float/resnet50_fp32.onnx
Evaluating onnx model using AMD MIGRAPHX    
100%|██████████████████████████████████████████████████████████████| 50000/50000 [05:41<00:00, 146.24it/s]
top-1 / top-5 accuracy: 76.1 / 92.9
-------- End of resnet50 test                                                                                                                                                                                                                                            

-------- Start resnet50 test
=== Load pretrained model ===                                                                                                                                                                                                                                            
Model name: resnet50
Model type: onnx                                                                                                                                                                                                                                                         
Loading model: float/resnet50_fp16.onnx
Evaluating onnx model using AMD MIGRAPHX    
100%|█████████████████████████████████████████████████████████████| 50000/50000 [05:11<00:00, 160.76it/s]
top-1 / top-5 accuracy: 76.1 / 92.9
-------- End of resnet50 test
  • You can get the model's inference performance on AMD GPU with MIGraphX Driver. The default batch size is set to 64 in this script.
  $ sh test_perf_migraphx.sh
Compiling ...
Reading: ../float/resnet50_fp32.onnx
module: "main"
main:@0 = check_context::migraphx::version_1::gpu::context -> float_type, {}, {}
main:@1 = hip::hip_allocate_memory[shape=float_type, {154140672}, {1},id=main:scratch] -> float_type, {154140672}, {1}
main:@2 = hip::hip_copy_literal[id=main:@literal:99] -> float_type, {64, 3, 7, 7}, {147, 49, 7, 1}
...
Batch size: 64
Rate: 2245.41/sec
Total time: 28.5025ms
Total instructions time: 29.7867ms
Overhead time: 0.0725389ms, -1.28419ms
Overhead: 0%, -5%

Compiling ...
Reading: ../float/resnet50_fp16.onnx
module: "main"
main:@0 = check_context::migraphx::version_1::gpu::context -> float_type, {}, {}
main:@1 = hip::hip_allocate_memory[shape=float_type, {77070336}, {1},id=main:scratch] -> float_type, {77070336}, {1}
main:@2 = hip::hip_copy_literal[id=main:@literal:99] -> half_type, {64, 3, 7, 7}, {147, 49, 7, 1}
main:@3 = load[offset=0,end=0](main:@1) -> int8_type, {0}, {1}
...
Batch size: 64
Rate: 5176.15/sec
Total time: 12.3644ms
Total instructions time: 14.0585ms
Overhead time: 0.0919278ms, -1.69413ms
Overhead: 1%, -14%

2.6.4: Performance

Resnet50 Model Input Size FLOPs Top-1/Top-5 Accuracy, %
PyTorch model 224x224 8.2G 76.1/92.9
FP32 onnx model 224x224 8.2G 76.1/92.9
FP16 onnx model 224x224 8.2G 76.1/92.9

2.6.5: Data Preprocessing for Inference

The model used in this example uses the following data preprocessing parameters for inference:

data channel order: BGR(0~255)
resize: short side resize to 256 and keep the aspect ratio
center crop: 224 * 224

< Previous | Next >

License

UIF is licensed under Apache License Version 2.0. Refer to the LICENSE file for the full license text and copyright notice.

Technical Support

Contact [email protected] for questions, issues, and feedback on UIF.

Submit your questions, feature requests, and bug reports on the GitHub issues page.