* Fix increment_dir to use run_xxx for logdir 

Rerunning train.py with logdir in the form of `run_xxx` causes index slicing in
```
n = max([int(x[len(dir):x.find('_') if '_' in Path(x).name else None]) for x in d]) + 1  # increment
```
to underflow.

* Replace find with rfind

* Update ci-testing.yml

* Update yolo.py

* Update ci-testing.yml

* Update export.py

* Update yolo.py

fix a bug in load_mosaic

* add the configuration to deploy yolov5 in a docker container in app engine on google cloud

* remove the duplicated requirements, place the google app engine files in utils

* remove the unused requirements

Co-authored-by: guilhembau <guilhem@6dbytes.com>

* feat: add rtmp support

* Update detect.py

pass tuple to source.startswith()

Co-authored-by: ryan.fu <ryan.fu@aylaasia.com>
Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>

* Update train.py with simplified ckpt names

* Return default hyps to hyp.scratch.yaml

Leave line commented for future use once mystery of best finetuning hyps to apply becomes clearer.

* Force test_batch*_pred.jpg replot on final epoch

This will allow you to see predictions final testing run after training completes in runs/exp0

* Class frequency not calculated on resuming training

Calculation of class frequency is not needed when resuming training.
Anchors can still be recalculated whether resuming or not.

* Check rank for autoanchor

* Update train.py

no autoanchor checks on resume

Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>

* the value should be a list to unpacking

* filter the None labels

* update readme to remove apex dependency

* add support for DWConv fuse operation

Cleaned up code

* Add InfiniteDataLoader

Only initializes at first epoch. Saves time.

* Moved class to a better location

Don't load the entire torchvision library just for nms when the function is already in the torch library.