Support upstream yolov5 v6.0 models

[zhiqwang]

This release mainly supports upstream YOLOv5 v6.0 models.

Highlights

We set the model to r6.0 by default, and rest of the usage interface is the same as before.

from yolort.models import yolov5n6

# Load model
model = yolov5n6(upstream_version="r6.0", pretrained=True, score_thresh=0.45)
model.eval()

# Perform inference on an image file
predictions = model.predict("bus.jpg")
# Perform inference on a list of image files
predictions = model.predict(["bus.jpg", "zidane.jpg"])

New Features

• Fully support upstream YOLOv5 v6.0 models (Add yolov5 v6.0 pretrained model weights #206, Upgrade the default setting to r6.0 #213, Cleanup docstrings to darknet backbones #215, Fix support for YOLOv5 nano models #204, Add yolov5 nano models #199, Support upstream yolov5 P6 models #196, Resolve version conflicts #195, Support upstream yolov5 v6 release #194) (Thanks @deepage)
• Add PyTorch 1.10 to GH Actions #207
• Support skipping preprocess when exporting ONNX #193

Backwards Incompatible Changes

• Make main as the base branch #200
• Rename _yolov5_darknet_pan to build_model #198

Improvement

• Use torch.jit.trace in unit-test #205 (Thanks @nihui)
• Update model graph visualization images #212
• [pre-commit.ci] pre-commit autoupdate #191
• [pre-commit.ci] pre-commit autoupdate #197

Bugfixes

• Fix setup and pre-commit.ci #209
• Fix non-renderable images on PyPI #210
• Fix model initialization in ModuleStateUpdate #201
• Fix the anchor configuration mechanism #202
• Small fixes to the docstrings #208