提交migraphx推理方法

README.md

@@ -31,20 +31,19 @@ VGG使用多个较小的卷积滤波器，这减少了网络在训练过程中
     cuda 11
     
     pip install -r requirements.txt
-
+    
     pip install --no-cache-dir --extra-index-url https://pypi.nvidia.com pytorch-quantization
 
 2、TensorRT
 
     wget https://github.com/NVIDIA/TensorRT/archive/refs/tags/8.5.3.zip 
-
+    
     unzip [下载的压缩包] -d [解压路径]
-
+    
     pip install 解压路径/python/tensorrt-8.5.3.1-cp39-none-linux_x86_64.whl
-
+    
     ln -s 解压路径(绝对路径)/bin/trtexec /usr/local/bin/trtexec
 
-
 注意：若需要`cu12`则将`requirements.txt`中的相关注释关闭，并安装。
 
 ## 数据集
@@ -60,9 +59,12 @@ VGG使用多个较小的卷积滤波器，这减少了网络在训练过程中
 
 ## 推理
 
+    # N卡推理
     trtexec --onnx=/path/to/onnx --saveEngine=./checkpoints/qat/last.trt --int8
-
     python eval.py --device=0
+    
+    # DCU卡推理
+    python evaluate_migraphx.py --device=0
 
 ## result
 
@@ -70,10 +72,10 @@ VGG使用多个较小的卷积滤波器，这减少了网络在训练过程中
 
 ### 精度
 
-||原始模型|QAT模型|ONNX模型|TensorRT模型|
-|:---|:---|:---|:---|:---|
-|Acc|0.9189|0.9185|0.9181|0.9184|
-|推理时间|5.5764s|13.7603s|4.2848s|2.9893s|
+||原始模型|QAT模型|ONNX模型|TensorRT模型|MIGraphX模型|
+|:---|:---|:---|:---|:---|----|
+|Acc|0.9189|0.9185|0.9181|0.9184|0.919|
+|推理时间|5.5764s|13.7603s|4.2848s|2.9893s|6.7672s|
 
 ## 应用场景
 

evaluate_migraphx.py

+import argparse
+import numpy as np
+import migraphx
+import torch
+import time
+from tqdm import tqdm
+from utils.data import prepare_dataloader
+
+def eval_migraphx(onnx_path, dataloader, device):
+    # 加载模型
+    model = migraphx.parse_onnx(onnx_path)
+
+    # 获取模型输入/输出节点信息
+    inputs = model.get_inputs()
+    outputs = model.get_outputs()
+    inputName = model.get_parameter_names()[0]
+    inputShape = inputs[inputName].lens()
+
+    # 编译模型
+    model.compile(t=migraphx.get_target("gpu"), device_id=device)
+
+    correct, total = 0, 0
+    for it in range(2):
+        desc = "warmup"
+        if it == 1:
+            start_time = time.time()
+            desc = "eval onnx model"
+            
+        for data, label in tqdm(dataloader, desc=desc, total=len(dataloader)):
+            data, label = data.numpy().astype(np.float32), label.numpy().astype(np.float32)
+            results = model.run({inputName:data})
+            predictions = np.argmax(results[0], axis=-1)
+            
+            correct += (label == predictions).sum()
+            total += len(label)
+        
+        if it == 1:
+            end_time = time.time()
+    
+    return correct / total, end_time - start_time
+
+def main(args):
+    device = torch.device(f"cuda:{args.device}" if args.device != -1 else "cpu")
+    
+    test_dataloader, _ = prepare_dataloader("./data/cifar10", False, args.batch_size)
+    
+    # 测试onnx模型
+    acc_onnx, runtime_onnx = eval_migraphx("./checkpoints/calibrated/pretrained_qat.onnx", test_dataloader, args.device)
+    
+    print("==============================================================")
+    print(f"MIGraphX Model Acc: {acc_onnx}, Inference Time: {runtime_onnx:.4f}s")
+    print("==============================================================")
+
+if __name__ == '__main__':
+    import argparse
+    
+    parser = argparse.ArgumentParser()
+    
+    parser.add_argument("--batch_size", type=int, default=16)
+    
+    parser.add_argument("--device", type=int, default=-1)
+    
+    parser.add_argument("--num_classes", type=int, default=10)
+    
+    args = parser.parse_args()
+    
+    main(args)
+
+