Slight speedup.

* PyTorch Hub model.save() increment as runs/hub/exp

This chane will align PyTorch Hub results saving with the existing unified results saving directory structure of

runs/
  /train
  /detect
  /test
  /hub
    /exp
    /exp2
    ...

* cleanup

* Set resume flag to false

* Check existance of val dataset

* yolotr

* transformer block

* Remove bias in Transformer

* Remove C3T

* Remove a deprecated class

* put the 2nd LayerNorm into the 2nd residual block

* move example model to models/hub, rename to -transformer

* Add module comments and TODOs

* Remove LN in Transformer

* Add comments for Transformer

* Solve the problem of MA with DDP

* cleanup

* cleanup find_unused_parameters

* PEP8 reformat

Co-authored-by: DingYiwei <846414640@qq.com>
Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>

* Update README with Tips for Best Results tutorial

* Update README.md

* Add label smoothing option

* Correct data type

* add_log

* Remove log

* Add log

* Update loss.py

remove comment (too versbose)

Co-authored-by: phattran <phat.tranhoang@cyberlogitec.com>
Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>

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

* Add opencv-contrib-python to requirements.txt

* Update requirements.txt

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

* Speed profiling improvements

* Update torch_utils.py

deepcopy() required to avoid adding elements to model.

* Update torch_utils.py

When using tqdm, sometimes it can't print in one line and roll to next line.

I think this should help speed up CUDA inference, as currently models may be running in FP32 inference mode on CUDA devices unnecesarily.

Add stderr=subprocess.STDOUT to catch error messages.

* W&B resume ddp from run link fix

* Native DDP W&B support for training, resuming

Catch 'fatal: not a git repository' returns and return '' instead (observed in GCP Hub checks).

Update Docker FROM nvcr.io/nvidia/pytorch:21.03-py3

```python
        self.n = len(self.pred)  # number of images (batch size)
```

* Fix Indentation in test.py

* CI fix

* Comply with PEP8: 80 characters per line

* Update detections() self.t = tuple()

Fix multiple results.print() bug.

* Update experimental.py

* Update yolo.py

Updated device selection string with fallback for non-git directories.
```python
def select_device(device='', batch_size=None):
    # device = 'cpu' or '0' or '0,1,2,3'
    s = f'YOLOv5 🚀 {git_describe() or date_modified()} torch {torch.__version__} '  # string
...
```

Co-authored-by: Max Uppenkamp <max.uppenkamp@inform-software.com>

Cython should be a dependency of the remaining packages in requirements.txt, so should be installed anyway even if not a direct requirement.

This updates the default detect.py behavior to automatically save all inference images/videos/webcams unless the new argument --nosave is used (python detect.py --nosave) or unless a list of streaming sources is passed (python detect.py --source streams.txt)

guide describes YOLOv5 apps collection in Supervisely Ecosystem

Improved user-feedback following requirements auto-update.

Prints 'Please restart runtime or rerun command for update to take effect.' following package auto-install to inform users to restart/rerun.

* YOLOv5 PyTorch Hub models >> check_requirements()

Update YOLOv5 PyTorch Hub requirements.txt path to cache path.

* Update hubconf.py

* Update hubconf.py with check_requirements()

Dependency checks have been missing from YOLOv5 PyTorch Hub model loading, causing errors in some cases when users are attempting to import hub models in unsupported environments. This should examine the YOLOv5 requirements.txt file and pip install any missing or version-conflict packages encountered. 

This is highly experimental (!), please let us know if this creates problems in your custom workflows.

* Update hubconf.py

* Update check_requirements() with auto-install

This PR builds on an idea I had to automatically install missing dependencies rather than simply report an error message. 

YOLOv5 should now 1) display all dependency issues and not simply display the first missing dependency, and 2) attempt to install/update each missing/VersionConflict package.

* cleanup

* cleanup 2

* Check requirements.txt file exists

* cleanup 3

* PyTorch Hub models default to CUDA:0 if available

* device as string bug fix