GitLab

* explain lr and batch size in references/detection/train.py

* fix typo

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* Actually upload wheels (please port to master)
Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* Propagate more environment variables.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* Put macos binaries in the right place
Signed-off-by: Edward Z. Yang <ezyang@fb.com>

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

Add uploading support, make CUDA builds actually work.

Following a similar line of inquiry to pytorch/audio#217

`self.fc1(x)` converts the shape of `x` into "N x 1024", and `self.fc2(x)` converts into "N x num_classes".

By adding `print(x.shape)` under each comment line, the console displays as follows (batch_size is 1):

```text
torch.Size([1, 2048])
torch.Size([1, 1024])
torch.Size([1, 1024])
torch.Size([1, 1024])
torch.Size([1, 1000])
```

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* Add docs for video models

* Fix docstrings for resnet and vgg

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* [WIP] Minor cleanups on R3d

* Move all models to video/resnet.py

* Remove old files

* Make tests less memory intensive

* Lint

* Fix typo and add pretraing arg to training script

* Expose docs for io and ops package

Had do modify the docstrings to use Napoleon NumPy style, because Napoleon Google Style doesn't support multiple return arguments

* Add video section

* Add docs to UCF101 and HMDB51 datasets

* Remove extra line

* Lint

* Move RandomClipSampler to references

* Lint and bugfix

Also add docs

* Copy classification scripts for video classification

* Initial version of video classification

* add version

* Training of r2plus1d_18 on kinetics work

Gives even slightly better results than expected, with 57.336 top1 clip accuracy. But we count some clips twice in this evaluation

* Cleanups on training script

* Lint

* Minor improvements

* Remove some hacks

* Lint

* Optimize read_video_timestamps for some formats

* Add some tests

* fix stl10

* fix lsun

* Miscellaneous fixes and improvements

* Guard against videos without video stream

* Fix lint

* Add test for packed b-frames videos

* Fix missing import

* [0.4_video] models - initial commit

* addressing fmassas inline comments

* pep8 and flake8

* simplify "hacks"

* sorting out latest comments

* nitpick

* Updated tests and constructors

* Added docstrings - ready to merge

* introduced function to verify str arguments

* flake8

* added FIXME to VOC

* Fixed error message

* added test for verify_str_arg

* cleanup todos

* added option for custom error message

* fix VOC

* fixed Caltech

* Add HMDB51 and UCF101

* Remove debug code

Also extend video saving to support different codecs and options. Notably, we can now save with lossless compression

* Properly order videos with B-frames

* seek doesn't seek to pts, but dts

Find a way of overcoming this problem.

Also make logging less verbose

* Replaced asserts with TORCH_CHECK

* Fixed an error

* in_channels_stage2 from backbone.inplanes

* remove type for backward compatible

* adding mixed precision training with Apex

* fix APEX default optimization level

* adding python version check for apex

* fix LINT errors and raise exceptions if apex not available

* fixing apex distributed training

* fix throughput calculation: include forward pass

* remove torch.cuda.set_device(args.gpu) as it's already called in init_distributed_mode

* fix linter: new line

* move Apex initialization code back to the beginning of main

* move apex initialization to before lr_scheduler - for peace of mind. Though, doing apex initialization after lr_scheduler seems to work fine as well

* Add VideoClips and Kinetics dataset

* Lint + add back missing line

* Adds ClipSampler following Bruno comment

* Change name following Bruno's suggestion

* Enable specifying a target framerate

* Fix test_io for new interface

* Add comment mentioning drop_last behavior

* Make compute_clips more robust

* Flake8

* Fix for Python2

* fixed call to the VisionDataset constructor

* change call from keyword arguments to positional

* changed order of arguments

* removed transforms argument once again

* Fixed call to constructor of parent class

* fixed LSUN

* fixed Caltech256

Fix the old flatten method which use the size(0) to caculate the batch size, the old method will intruduce Gather opertion in the onnx output, which will faild parsed by tensorRT 5.0 (#1134)