NetHarn - a PyTorch Network Harness

• NAME:

• NetHarn (pronounced "net-harn")

• FRAMEWORK:

• PyTorch

• FEATURES:

• 
  

• hyperparameter tracking

• training directory management

• callback-based public API

• XPU - code abstraction for [cpu, gpu, multi-gpu].

• single-file deployments (NEW in version 0.1.0).

• reasonable test coverage using pytest and xdoctest

• CI testing on appveyor and travis (note a few tests are failing due to minor issues)

• A rich utility set

• Extensions of PyTorch objects (e.g. critions, initializers, layers, optimizers, schedulers)

• BUILTINS:

• 
  

• training loop boilerplate

• snapshots / checkpoints

• progress bars (backend_choices: [progiter, tqdm])

• data provenance of training history in train_info.json

• tensorboard metric visualization (optional)

• DESIGN PHILOSOPHY:

• Avoid boilerplate, built-it yourself when you need to, and don't repeat yourself.

• SLOGAN:

• Rein and train.

• USAGE PATTERNS:

• 
  

1. Write code for a torch object (i.e. Dataset, Model, Criterion, Initializer, and Scheduler) just as you normally would.

2. Inherit from the nh.FitHarn object, define run_batch, on_batch, on_epoch, etc...

3. Create an instance of nh.HyperParams to specify your dataset, model, criterion, etc...

4. Create an instance of your FitHarn object with those hyperparameters.

5. Then execute its run method.

6. ???

7. profit

• EXAMPLES:

• 
  

• ToyData2d classification with nh.models.ToyNet2d (see doctest in netharn/fit_harn.py:__DOC__:0)

• MNIST digit classification with MnistNet (examples/mnist.py)

• Cifar10 category classification with ResNet50 / dpn91 (examples/cifar.py)

• Voc2007+2012 object detection with YOLOv2 (examples/yolo_voc.py)

• IBEIS metric learning with SiameseLP (examples/siam_ibeis.py)

• STABILITY:

• Mostly harmless. Most tests pass, the current failures are probably not critical. I'm able to use it on my machine (tm). In this early stage of development, there are still a few pain points. Issues and PRs welcome.

• KNOWN BUGS:

• 
  

• The metrics for computing detection mAP / AP might not be correct.

• The YOLO example gets to about 70% mAP (using Girshik's mAP code) whereas we should be hitting 74-76%

• AUTHORS COMMENTS:

• 
  

• My MNIST, CIFAR, and VOC examples will download the data as needed.

• I'm hoping we can publicly release a few privately developed features. They would take a non-trivial amount of developer time to reproduce. These features mostly have to do with exporting / deploying fit models.

• My CIFAR example for ResNet50 achieves 95.72% accuracy, outperforming the best DPN92 result (95.16%) that I'm aware of. This result seems real, I do not believe I've made an error in measurement (but this has need been peer-reviewed so, caveat emptor). I've reproduced this results a few times. You can use the code in examples/cifar.py to see if you can too (please tell me if you cannot).

• My YOLO example is based of of EAVise's excellent lightnet (https://gitlab.com/EAVISE/lightnet/) package.

• I reimplemented the CocoAPI (see nh.data.coco_api), because I had some (probably minor) issue with the original implementation. I've extended it quite a bit, and I'd recommend using it.

• My metric-learning example requires code that is not publicly available :(, so only those with access to a copy of the ibeis software more recent than is more recent than July 2017 more recent than 2017) can use it without modification.

• DEPENDENCIES:

• 
  

• torch

• numpy

• Cython

• ubelt

• xdoctest

• ... (see requirements.txt)

Installation

In the future these instructions may actually be different than the developer setup instructions, but for now they are the same.

mkdir -p ~/code
git clone git@github.com:Erotemic/netharn.git ~/code/netharncd ~/code/netharn
./run_developer_setup.sh

Description

Parameterized fit and prediction harnesses for PyTorch.

Trains models and keeps track of your hyperparameters.

This is a clean port of the good parts developed in my research repo: clab.

See the netharn/examples folder for example usage. The doctests are also a good resource. It would be nice if we had better docs.

NetHarn is a research framework for training and deploying arbitrary PyTorch models. It was designed for the purpose of minimizing training-loop boilerplate and tracking hyperparameters to

encourage reproducible research.

NetHarn separates the problem of training a model into the following core hyperparameter components: the datasets, model, criterion, initializer, optimizer, and learning rate scheduler. Runs with different hyperparameters are automatically logged to separate directories which makes it simple

to compare the results of two experiments.

• NetHarn also has the ability to create a single-file deployment of a trained model

• that is independent of the system used to train it.

This makes it fast and simple for research results to be externally verified and moved into production.

Developer Setup:

In the future these instructions might be different from the install instructions, but for now they are the same.

sudo apt-get install python3 python-dev python3-dev 
 build-essential libssl-dev libffi-dev 
 libxml2-dev libxslt1-dev zlib1g-dev 
 python-pip

mkdir -p ~/code
git clone git@github.com:Erotemic/netharn.git ~/code/netharncd ~/code/netharn

./run_developer_setup.sh

Example:

This following example is the doctest in netharn/fit_harn.py. It demonstrates how to use NetHarn to train a model to solve a toy problem.

In this toy problem, we do not extend the nh.FitHarn object, so we are using the default behavior of run_batch. The default on_batch, and on_epoch do nothing, so only loss will be the only measurement of performance.

For further examples please see the examples directory. These example show how to extend nh.FitHarn to measure performance wrt a particular problem. The MNIST and CIFAR examples are the most simple. The YOLO example is more complex. The IBEIS example depends on non-public data / software, but can still be useful to look at. Its complexity is more than CIFAR but less than YOLO.

>>> import netharn as nh>>> hyper = nh.HyperParams(**{>>>     # ================>>>     # Environment Components>>>     'workdir'     : ub.ensure_app_cache_dir('netharn/demo'),>>>     'nice'        : 'demo',>>>     'xpu'         : nh.XPU.cast('auto'),>>>     # workdir is a directory where intermediate results can be saved>>>     # nice symlinks <workdir>/fit/nice/<nice> -> ../runs/<hashid>>>>     # XPU auto select a gpu if idle and VRAM>6GB else a cpu>>>     # ================>>>     # Data Components>>>     'datasets'    : {  # dict of plain ol torch.data.Dataset instances>>>         'train': nh.data.ToyData2d(size=3, border=1, n=256, rng=0),>>>         'test': nh.data.ToyData2d(size=3, border=1, n=128, rng=1),>>>     },>>>     'loaders'     : {'batch_size': 64}, # DataLoader instances or kw>>>     # ================>>>     # Algorithm Components>>>     # Note the (cls, kw) tuple formatting>>>     'model'       : (nh.models.ToyNet2d, {}),>>>     'optimizer'   : (nh.optimizers.SGD, {>>>         'lr': 0.0001>>>     }),>>>     # focal loss is usually better than nh.criterions.CrossEntropyLoss>>>     'criterion'   : (nh.criterions.FocalLoss, {}),>>>     'initializer' : (nh.initializers.KaimingNormal, {>>>         'param': 0,>>>     }),>>>     # these may receive an overhaul soon>>>     'scheduler'   : (nh.schedulers.ListedLR, {>>>         'points': {0: .0001, 2: .01, 5: .015, 6: .005, 9: .001},>>>         'interpolate': True,>>>     }),>>>     'monitor'     : (nh.Monitor, {>>>         'max_epoch': 10,>>>     }),>>>     # dynamics are a config option that modify the behavior of the main>>>     # training loop. These parameters effect the learned model.>>>     'dynamics'   : {'batch_step': 4},>>> })>>> harn = FitHarn(hyper)>>> # non-algorithmic behavior configs (do not change learned models)>>> harn.config['prog_backend'] = 'tqdm'  # I prefer progiter (I may be biased)>>> # start training.>>> harn.initialize(reset='delete')>>> harn.run()  # note: run calls initialize it hasn't already been called.>>> # xdoc: +IGNORE_WANT

Running this code produes the following output:

RESET HARNESS BY DELETING EVERYTHING IN TRAINING DIR
Symlink: /home/joncrall/.cache/netharn/demo/fit/runs/olqtvpde -> /home/joncrall/.cache/netharn/demo/fit/nice/demo
.... already exists
.... and points to the right place
Initializing tensorboard (dont forget to start the tensorboard server)
Model has 824 parameters
Mounting ToyNet2d model on GPU(0)
Initializing new model
 * harn.train_dpath = '/home/joncrall/.cache/netharn/demo/fit/runs/olqtvpde'
 * harn.nice_dpath = '/home/joncrall/.cache/netharn/demo/fit/nice/demo'
Snapshots will save to harn.snapshot_dpath = '/home/joncrall/.cache/netharn/demo/fit/runs/olqtvpde/torch_snapshots'
dont forget to start:
    tensorboard --logdir /home/joncrall/.cache/netharn/demo/fit/nice
begin training
epoch lr:0.001 │ vloss is unevaluated: 100%|███████████████████████| 10/10 [00:00<00:00, 15.11it/s, wall=Jul:07 EST]10 [00:00<?, ?it/s]
train x64 │ loss:0.186 │: 100%|████████████████████████████████████████████████████████| 8/8 [00:00<00:00, 276.93it/s, wall=Jul:07 EST]
test x64 │ loss:0.159 │: 100%|█████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 482.91it/s, wall=Jul:07 EST]

Acknowledgements:

The authors would like to thank AFRL for their support of this work via SBIR Contract FA8650-18-C-1075. This library is approved for public release via 88ABW.