Add DWConvTranspose2d() module

[glenn-jocher]

Add DWConvTranspose2d() module. @sergiossm

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Implementation of Depth-wise Transpose Convolutional layers in both PyTorch and TensorFlow for Ultralytics' YOLOv5.

📊 Key Changes

• 🛠 Added DWConvTranspose2d class in models/common.py which defines a new depth-wise transpose convolution operation in PyTorch.
• ✨ Introduced TFDWConvTranspose2d class in models/tf.py, which is the TensorFlow equivalent of the new depth-wise transpose convolutional layer.
• 🔧 Modified models/yolo.py and models/common.py to include depth-wise transpose convolutional layers in their respective model parsing methods.
• 🚀 Enhanced model architecture choices by allowing depth-wise transpose convolutions to be used in YOLO model definitions.

🎯 Purpose & Impact

• 📈 The new depth-wise transpose convolutional layers provide more architectural flexibility, potentially improving model performance for certain tasks.
• 🐍 By supporting both PyTorch and TensorFlow, these changes enable a wider range of developers to experiment with and benefit from these layers.
• 🔍 Users might expect better feature extraction capabilities and more efficient computation from models leveraging depth-wise transpose convolutions.

[glenn-jocher]

YOLOv5n v6.1 hyperplane CPU

benchmarks: weights=yolov5n.pt, imgsz=640, batch_size=1, data=/usr/src/app/data/coco128.yaml, device=cpu, half=False, test=False, pt_only=False
Checking setup...
YOLOv5 🚀 a9a92ae Python-3.8.10 torch-1.11.0+cpu CPU
Setup complete ✅ (96 CPUs, 1007.7 GB RAM, 1933.3/3519.3 GB disk)

Benchmarks complete (113.21s)
                   Format  mAP@0.5:0.95  Inference time (ms)
0                 PyTorch        0.3476                61.12
1             TorchScript        0.3476                52.30
2                    ONNX        0.3476                39.65
3                OpenVINO        0.3476                16.29
4                TensorRT           NaN                  NaN
5                  CoreML           NaN                  NaN
6   TensorFlow SavedModel        0.3476                61.50
7     TensorFlow GraphDef        0.3476                68.44
8         TensorFlow Lite        0.3470                95.77
9     TensorFlow Edge TPU           NaN                  NaN
10          TensorFlow.js           NaN                  NaN

YOLOv5n-7 hyperplane CPU

benchmarks: weights=yolov5n-7.pt, imgsz=640, batch_size=1, data=/usr/src/yolov5/data/coco128.yaml, device=cpu, half=False, test=False, pt_only=False
Checking setup...
YOLOv5 🚀 v6.1-207-g5774a15 Python-3.8.10 torch-1.11.0+cpu CPU
Setup complete ✅ (96 CPUs, 1007.7 GB RAM, 1933.3/3519.3 GB disk)

Benchmarks complete (188.59s)
                   Format  mAP@0.5:0.95  Inference time (ms)
0                 PyTorch        0.3376                56.29
1             TorchScript        0.3376                57.23
2                    ONNX        0.3376                49.70
3                OpenVINO        0.3376                16.61
4                TensorRT           NaN                  NaN
5                  CoreML           NaN                  NaN
6   TensorFlow SavedModel        0.3376                58.73
7     TensorFlow GraphDef        0.3376                75.64
8         TensorFlow Lite        0.3364                67.63
9     TensorFlow Edge TPU           NaN                  NaN
10          TensorFlow.js           NaN                  NaN

[glenn-jocher]

YOLOv5x v6.1 hyperplane A100

benchmarks: weights=yolov5x.pt, imgsz=640, batch_size=1, data=data/coco.yaml, device=0, half=False, test=False, pt_only=False
Checking setup...
YOLOv5 🚀 v6.1-207-g5774a15 Python-3.8.13 torch-1.11.0+cu113 CUDA:0 (A100-SXM-80GB, 81251MiB)
Setup complete ✅ (96 CPUs, 1007.7 GB RAM, 1943.6/3519.3 GB disk)

Benchmarks complete (1519.56s)
                   Format  mAP@0.5:0.95  Inference time (ms)
0                 PyTorch        0.5006                14.21
1             TorchScript        0.5006                10.83
2                    ONNX        0.5006                16.84
3                OpenVINO           NaN                  NaN
4                TensorRT        0.5005                 3.33
5                  CoreML           NaN                  NaN
6   TensorFlow SavedModel        0.5006                28.31
7     TensorFlow GraphDef        0.5006                27.54
8         TensorFlow Lite           NaN                  NaN
9     TensorFlow Edge TPU           NaN                  NaN
10          TensorFlow.js           NaN                  NaN

YOLOv5x-7 v6.1 hyperplane A100

benchmarks: weights=yolov5x-7.pt, imgsz=640, batch_size=1, data=data/coco.yaml, device=1, half=False, test=False, pt_only=False
Checking setup...
YOLOv5 🚀 5774a15 Python-3.8.13 torch-1.11.0+cu113 CUDA:0 (A100-SXM-80GB, 81251MiB)
Setup complete ✅ (96 CPUs, 1007.7 GB RAM, 1943.6/3519.3 GB disk)

Benchmarks complete (1754.20s)
                   Format  mAP@0.5:0.95  Inference time (ms)
0                 PyTorch        0.5046                13.90
1             TorchScript        0.5046                10.35
2                    ONNX        0.5046                16.45
3                OpenVINO           NaN                  NaN
4                TensorRT        0.5048                 2.95
5                  CoreML           NaN                  NaN
6   TensorFlow SavedModel        0.5046                55.90
7     TensorFlow GraphDef        0.5046                55.36
8         TensorFlow Lite           NaN                  NaN
9     TensorFlow Edge TPU           NaN                  NaN
10          TensorFlow.js           NaN                  NaN