Add RT-DETR Ultralytics

README.md

@@ -30,6 +30,7 @@ NVIDIA DeepStream SDK 6.3 / 6.2 / 6.1.1 / 6.1 / 6.0.1 / 6.0 / 5.1  configuration
 * INT8 calibration (PTQ) for Darknet and ONNX exported models
 * New output structure (fix wrong output on DeepStream < 6.2) - it need to export the ONNX model with the new export file, generate the TensorRT engine again with the updated files, and use the new config_infer_primary file according to your model
 * **RT-DETR (https://github.com/lyuwenyu/RT-DETR/tree/main/rtdetr_pytorch)**
+* **RT-DETR Ultralytics (https://docs.ultralytics.com/models/rtdetr)**
 
 ##
 
@@ -53,6 +54,7 @@ NVIDIA DeepStream SDK 6.3 / 6.2 / 6.1.1 / 6.1 / 6.0.1 / 6.0 / 5.1  configuration
 * [PP-YOLOE / PP-YOLOE+ usage](docs/PPYOLOE.md)
 * [YOLO-NAS usage](docs/YOLONAS.md)
 * [RT-DETR usage](docs/RTDETR.md)
+* [RT-DETR Ultralytics usage](docs/RTDETR_Ultralytics.md)
 * [Using your custom model](docs/customModels.md)
 * [Multiple YOLO GIEs](docs/multipleGIEs.md)
 

docs/RTDETR.md

@@ -1,6 +1,6 @@
-# RT_DETR usage
+# RT-DETR usage
 
-**NOTE**: For it is supported only the https://github.com/lyuwenyu/RT-DETR/tree/main/rtdetr_pytorch version.
+**NOTE**: https://github.com/lyuwenyu/RT-DETR/tree/main/rtdetr_pytorch version.
 
 * [Convert model](#convert-model)
 * [Compile the lib](#compile-the-lib)

docs/RTDETR_Ultralytics.md

@@ -0,0 +1,199 @@
+# RT-DETR Ultralytics usage
+
+**NOTE**: Ultralytics (https://docs.ultralytics.com/models/rtdetr) version.
+
+* [Convert model](#convert-model)
+* [Compile the lib](#compile-the-lib)
+* [Edit the config_infer_primary_rtdetr file](#edit-the-config_infer_primary_rtdetr-file)
+* [Edit the deepstream_app_config file](#edit-the-deepstream_app_config-file)
+* [Testing the model](#testing-the-model)
+
+##
+
+### Convert model
+
+#### 1. Download the Ultralytics repo and install the requirements
+
+```
+git clone https://github.com/ultralytics/ultralytics.git
+cd ultralytics
+pip3 install -r requirements.txt
+python3 setup.py install
+pip3 install onnx onnxsim onnxruntime
+```
+
+**NOTE**: It is recommended to use Python virtualenv.
+
+#### 2. Copy conversor
+
+Copy the `export_rtdetr_ultralytics.py` file from `DeepStream-Yolo/utils` directory to the `ultralytics` folder.
+
+#### 3. Download the model
+
+Download the `pt` file from [Ultralytics](https://github.com/ultralytics/assets/releases/) releases (example for RT-DETR-l)
+
+```
+wget https://github.com/ultralytics/assets/releases/download/v0.0.0/rtdetr-l.pt
+```
+
+**NOTE**: You can use your custom model.
+
+#### 4. Convert model
+
+Generate the ONNX model file (example for RT-DETR-l)
+
+```
+python3 export_rtdetr_ultralytics.py -w rtdetr-l.pt --dynamic
+```
+
+**NOTE**: To change the inference size (defaut: 640)
+
+```
+-s SIZE
+--size SIZE
+-s HEIGHT WIDTH
+--size HEIGHT WIDTH
+```
+
+Example for 1280
+
+```
+-s 1280
+```
+
+or
+
+```
+-s 1280 1280
+```
+
+**NOTE**: To simplify the ONNX model (DeepStream >= 6.0)
+
+```
+--simplify
+```
+
+**NOTE**: To use dynamic batch-size (DeepStream >= 6.1)
+
+```
+--dynamic
+```
+
+**NOTE**: To use static batch-size (example for batch-size = 4)
+
+```
+--batch 4
+```
+
+**NOTE**: If you are using the DeepStream 5.1, remove the `--dynamic` arg and use opset 12 or lower. The default opset is 16.
+
+```
+--opset 12
+```
+
+#### 5. Copy generated files
+
+Copy the generated ONNX model file and labels.txt file (if generated) to the `DeepStream-Yolo` folder.
+
+##
+
+### Compile the lib
+
+Open the `DeepStream-Yolo` folder and compile the lib
+
+* DeepStream 6.3 on x86 platform
+
+  ```
+  CUDA_VER=12.1 make -C nvdsinfer_custom_impl_Yolo
+  ```
+
+* DeepStream 6.2 on x86 platform
+
+  ```
+  CUDA_VER=11.8 make -C nvdsinfer_custom_impl_Yolo
+  ```
+
+* DeepStream 6.1.1 on x86 platform
+
+  ```
+  CUDA_VER=11.7 make -C nvdsinfer_custom_impl_Yolo
+  ```
+
+* DeepStream 6.1 on x86 platform
+
+  ```
+  CUDA_VER=11.6 make -C nvdsinfer_custom_impl_Yolo
+  ```
+
+* DeepStream 6.0.1 / 6.0 on x86 platform
+
+  ```
+  CUDA_VER=11.4 make -C nvdsinfer_custom_impl_Yolo
+  ```
+
+* DeepStream 5.1 on x86 platform
+
+  ```
+  CUDA_VER=11.1 make -C nvdsinfer_custom_impl_Yolo
+  ```
+
+* DeepStream 6.3 / 6.2 / 6.1.1 / 6.1 on Jetson platform
+
+  ```
+  CUDA_VER=11.4 make -C nvdsinfer_custom_impl_Yolo
+  ```
+
+* DeepStream 6.0.1 / 6.0 / 5.1 on Jetson platform
+
+  ```
+  CUDA_VER=10.2 make -C nvdsinfer_custom_impl_Yolo
+  ```
+
+##
+
+### Edit the config_infer_primary_rtdetr file
+
+Edit the `config_infer_primary_rtdetr.txt` file according to your model (example for RT-DETR-l with 80 classes)
+
+```
+[property]
+...
+onnx-file=rtdetr-l.onnx
+...
+num-detected-classes=80
+...
+parse-bbox-func-name=NvDsInferParseYolo
+...
+```
+
+**NOTE**: The **RT-DETR Ultralytics** do not resize the input with padding. To get better accuracy, use
+
+```
+[property]
+...
+maintain-aspect-ratio=0
+...
+```
+
+##
+
+### Edit the deepstream_app_config file
+
+```
+...
+[primary-gie]
+...
+config-file=config_infer_primary_rtdetr.txt
+```
+
+##
+
+### Testing the model
+
+```
+deepstream-app -c deepstream_app_config.txt
+```
+
+**NOTE**: The TensorRT engine file may take a very long time to generate (sometimes more than 10 minutes).
+
+**NOTE**: For more information about custom models configuration (`batch-size`, `network-mode`, etc), please check the [`docs/customModels.md`](customModels.md) file.

utils/export_rtdetr_ultralytics.py

@@ -0,0 +1,124 @@
+import os
+import sys
+import argparse
+import warnings
+import onnx
+import torch
+import torch.nn as nn
+from copy import deepcopy
+from ultralytics import RTDETR
+from ultralytics.utils.torch_utils import select_device
+from ultralytics.nn.modules import C2f, Detect, RTDETRDecoder
+
+
+class DeepStreamOutput(nn.Module):
+    def __init__(self, img_size):
+        self.img_size = img_size
+        super().__init__()
+
+    def forward(self, x):
+        boxes = x[:, :, :4]
+        boxes[:, :, [0, 2]] *= self.img_size[1]
+        boxes[:, :, [1, 3]] *= self.img_size[0]
+        scores, classes = torch.max(x[:, :, 4:], 2, keepdim=True)
+        classes = classes.float()
+        return boxes, scores, classes
+
+
+def suppress_warnings():
+    warnings.filterwarnings('ignore', category=torch.jit.TracerWarning)
+    warnings.filterwarnings('ignore', category=UserWarning)
+    warnings.filterwarnings('ignore', category=DeprecationWarning)
+
+
+def rtdetr_ultralytics_export(weights, device):
+    model = RTDETR(weights)
+    model = deepcopy(model.model).to(device)
+    for p in model.parameters():
+        p.requires_grad = False
+    model.eval()
+    model.float()
+    model = model.fuse()
+    for k, m in model.named_modules():
+        if isinstance(m, (Detect, RTDETRDecoder)):
+            m.dynamic = False
+            m.export = True
+            m.format = 'onnx'
+        elif isinstance(m, C2f):
+            m.forward = m.forward_split
+    return model
+
+
+def main(args):
+    suppress_warnings()
+
+    print('\nStarting: %s' % args.weights)
+
+    print('Opening RT-DETR Ultralytics model\n')
+
+    device = select_device('cpu')
+    model = rtdetr_ultralytics_export(args.weights, device)
+
+    if len(model.names.keys()) > 0:
+        print('\nCreating labels.txt file')
+        f = open('labels.txt', 'w')
+        for name in model.names.values():
+            f.write(name + '\n')
+        f.close()
+
+    img_size = args.size * 2 if len(args.size) == 1 else args.size
+
+    model = nn.Sequential(model, DeepStreamOutput(img_size))
+
+    onnx_input_im = torch.zeros(args.batch, 3, *img_size).to(device)
+    onnx_output_file = os.path.basename(args.weights).split('.pt')[0] + '.onnx'
+
+    dynamic_axes = {
+        'input': {
+            0: 'batch'
+        },
+        'boxes': {
+            0: 'batch'
+        },
+        'scores': {
+            0: 'batch'
+        },
+        'classes': {
+            0: 'batch'
+        }
+    }
+
+    print('\nExporting the model to ONNX')
+    torch.onnx.export(model, onnx_input_im, onnx_output_file, verbose=False, opset_version=args.opset,
+                      do_constant_folding=True, input_names=['input'], output_names=['boxes', 'scores', 'classes'],
+                      dynamic_axes=dynamic_axes if args.dynamic else None)
+
+    if args.simplify:
+        print('Simplifying the ONNX model')
+        import onnxsim
+        model_onnx = onnx.load(onnx_output_file)
+        model_onnx, _ = onnxsim.simplify(model_onnx)
+        onnx.save(model_onnx, onnx_output_file)
+
+    print('Done: %s\n' % onnx_output_file)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='DeepStream RT-DETR Ultralytics conversion')
+    parser.add_argument('-w', '--weights', required=True, help='Input weights (.pt) file path (required)')
+    parser.add_argument('-s', '--size', nargs='+', type=int, default=[640], help='Inference size [H,W] (default [640])')
+    parser.add_argument('--opset', type=int, default=16, help='ONNX opset version')
+    parser.add_argument('--simplify', action='store_true', help='ONNX simplify model')
+    parser.add_argument('--dynamic', action='store_true', help='Dynamic batch-size')
+    parser.add_argument('--batch', type=int, default=1, help='Static batch-size')
+    args = parser.parse_args()
+    if not os.path.isfile(args.weights):
+        raise SystemExit('Invalid weights file')
+    if args.dynamic and args.batch > 1:
+        raise SystemExit('Cannot set dynamic batch-size and static batch-size at same time')
+    return args
+
+
+if __name__ == '__main__':
+    args = parse_args()
+    sys.exit(main(args))