netharn 0.5.5

Features (continued)

• Hyperparameter tracking: The hash of your hyperparameters determines the directory data will be written to. We also allow for a “nicer” means to manage directory structures. Given a HyperParams object, we create the symlink {workdir}/fit/nice/{nice} which points to {workdir}/fit/runs/{nice}/{hashid}.

• Automatic restarts: Calling FitHarn.run twice restarts training from where you left off by default (as long as the hyperparams haven’t changed).

• “Smart” Snapshot cleanup: Maintaining model weights files can be a memory hog. Depending the settings of harn.preferences, netharn.FitHarn will periodically remove less-recent or low-scoring snapshots.

• Deployment files: Model weights and architecture are together written as one reasonably-portable zip-file. We also package training metadata to maintain data provinence and make reproducing experiments easier.

• Restart from any pretrained state: use netharn.initializers.PretainedInitializer.

• Utilities for building networks in torch: Layers like netharn.layers.ConvNormNd make it easy to build networks for n=1, 2, or 3 dimensional data.

• Analytic output shape and receptive field: Netharn defines a netharn.layers.AnalyticModule, which can automatically define forward, output_shape_for and receptive_field_for if users define a special _output_for method, written with the netharn.analytic_for.Output, netharn.analytic_for.Hidden, and netharn.analytic_for.OutputFor special callables.

• Example tasks: Baseline code for standard tasks like: object segmentation, classification, and detection are defined in netharn.examples. The examples also provide example use cases for ndsampler, kwimage, kwannot, and kwplot.

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/netharn
cd ~/code/netharn
./run_developer_setup.sh

While all netharn dependencies should be available on pypi (with manylinux2010 wheels for binary packages), there are other packages developed concurrently with netharn. To install the development version of these dependencies then run python super_setup.py ensure to check out the repos and ensure they are on the correct branch, python super_setup.py develop to build everything in development mode, and python super_setup.py pull to update to the latest on the branch.

Description

Parameterized fit 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/netharn
cd ~/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 netharn.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 netharn.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
>>> hyper = netharn.HyperParams(**{
>>>     # ================
>>>     # Environment Components
>>>     'workdir'     : ub.ensure_app_cache_dir('netharn/demo'),
>>>     'nice'        : 'demo',
>>>     'xpu'         : netharn.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': netharn.data.ToyData2d(size=3, border=1, n=256, rng=0),
>>>         'vali': netharn.data.ToyData2d(size=3, border=1, n=128, rng=1),
>>>         'test': netharn.data.ToyData2d(size=3, border=1, n=128, rng=2),
>>>     },
>>>     'loaders'     : {'batch_size': 64}, # DataLoader instances or kw
>>>     # ================
>>>     # Algorithm Components
>>>     # Note the (cls, kw) tuple formatting
>>>     'model'       : (netharn.models.ToyNet2d, {}),
>>>     'optimizer'   : (netharn.optimizers.SGD, {
>>>         'lr': 0.0001
>>>     }),
>>>     # focal loss is usually better than netharn.criterions.CrossEntropyLoss
>>>     'criterion'   : (netharn.criterions.FocalLoss, {}),
>>>     'initializer' : (netharn.initializers.KaimingNormal, {
>>>         'param': 0,
>>>     }),
>>>     # these may receive an overhaul soon
>>>     'scheduler'   : (netharn.schedulers.ListedScheduler, {
>>>         'points': {'lr': {0: .0001, 2: .01, 5: .015, 6: .005, 9: .001}},
>>>         'interpolation': 'linear',
>>>     }),
>>>     'monitor'     : (netharn.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 = netharn.FitHarn(hyper)
>>> # non-algorithmic behavior configs (do not change learned models)
>>> harn.preferences['prog_backend'] = 'progiter'  # alternative: 'tqdm'
>>> harn.preferences['num_keep'] = 10
>>> # 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/demo/lnejaaum -> /home/joncrall/.cache/netharn/demo/_mru
... already exists
Symlink: /home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum -> /home/joncrall/.cache/netharn/demo/fit/nice/demo
... already exists
... and points to the right place
INFO: Initializing tensorboard (dont forget to start the tensorboard server)
INFO: Model has 824 parameters
INFO: Mounting ToyNet2d model on GPU(0)
INFO: Exported model topology to /home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum/ToyNet2d_2a3f49.py
INFO: Initializing model weights with: <netharn.initializers.nninit_core.KaimingNormal object at 0x7fc67eff0278>
INFO:  * harn.train_dpath = '/home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum'
INFO:  * harn.nice_dpath  = '/home/joncrall/.cache/netharn/demo/fit/nice/demo'
INFO: Snapshots will save to harn.snapshot_dpath = '/home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum/torch_snapshots'
INFO: ARGV:
    /home/joncrall/.local/conda/envs/py36/bin/python /home/joncrall/.local/conda/envs/py36/bin/ipython
INFO: dont forget to start:
    tensorboard --logdir ~/.cache/netharn/demo/fit/nice
INFO: === begin training 0 / 10 : demo ===
epoch lr:0.0001 │ vloss is unevaluated  0/10... rate=0 Hz, eta=?, total=0:00:00, wall=19:36 EST
train loss:0.173 │ 100.00% of 64x8... rate=11762.01 Hz, eta=0:00:00, total=0:00:00, wall=19:36 EST
vali loss:0.170 │ 100.00% of 64x4... rate=9991.94 Hz, eta=0:00:00, total=0:00:00, wall=19:36 EST
test loss:0.170 │ 100.00% of 64x4... rate=24809.37 Hz, eta=0:00:00, total=0:00:00, wall=19:36 EST
INFO: === finish epoch 0 / 10 : demo ===
epoch lr:0.00505 │ vloss: 0.1696 (n_bad=00, best=0.1696)  1/10... rate=1.24 Hz, eta=0:00:07, total=0:00:00, wall=19:36 EST
train loss:0.175 │ 100.00% of 64x8... rate=13522.14 Hz, eta=0:00:00, total=0:00:00, wall=19:36 EST
vali loss:0.167 │ 100.00% of 64x4... rate=23598.31 Hz, eta=0:00:00, total=0:00:00, wall=19:36 EST
test loss:0.167 │ 100.00% of 64x4... rate=20354.22 Hz, eta=0:00:00, total=0:00:00, wall=19:36 EST
INFO: === finish epoch 1 / 10 : demo ===
epoch lr:0.01 │ vloss: 0.1685 (n_bad=00, best=0.1685)  2/10... rate=1.28 Hz, eta=0:00:06, total=0:00:01, wall=19:36 EST
train loss:0.177 │ 100.00% of 64x8... rate=15723.99 Hz, eta=0:00:00, total=0:00:00, wall=19:36 EST
vali loss:0.163 │ 100.00% of 64x4... rate=29375.56 Hz, eta=0:00:00, total=0:00:00, wall=19:36 EST
test loss:0.163 │ 100.00% of 64x4... rate=29664.69 Hz, eta=0:00:00, total=0:00:00, wall=19:36 EST
INFO: === finish epoch 2 / 10 : demo ===

<JUST MORE OF THE SAME; REMOVED FOR BREVITY>

epoch lr:0.001 │ vloss: 0.1552 (n_bad=00, best=0.1552)  9/10... rate=1.11 Hz, eta=0:00:00, total=0:00:08, wall=19:36 EST
train loss:0.164 │ 100.00% of 64x8... rate=13795.93 Hz, eta=0:00:00, total=0:00:00, wall=19:36 EST
vali loss:0.154 │ 100.00% of 64x4... rate=19796.72 Hz, eta=0:00:00, total=0:00:00, wall=19:36 EST
test loss:0.154 │ 100.00% of 64x4... rate=21396.73 Hz, eta=0:00:00, total=0:00:00, wall=19:36 EST
INFO: === finish epoch 9 / 10 : demo ===
epoch lr:0.001 │ vloss: 0.1547 (n_bad=00, best=0.1547) 10/10... rate=1.13 Hz, eta=0:00:00, total=0:00:08, wall=19:36 EST




INFO: Maximum harn.epoch reached, terminating ...
INFO:



INFO: training completed
INFO: harn.train_dpath = '/home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum'
INFO: harn.nice_dpath  = '/home/joncrall/.cache/netharn/demo/fit/nice/demo'
INFO: view tensorboard results for this run via:
    tensorboard --logdir ~/.cache/netharn/demo/fit/nice
[DEPLOYER] Deployed zipfpath=/home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum/deploy_ToyNet2d_lnejaaum_009_GAEYQT.zip
INFO: wrote single-file deployment to: '/home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum/deploy_ToyNet2d_lnejaaum_009_GAEYQT.zip'
INFO: exiting fit harness.

Furthermore, if you were to run that code when '--verbose' in sys.argv, then it would produce this more detailed description of what it was doing:

RESET HARNESS BY DELETING EVERYTHING IN TRAINING DIR
Symlink: /home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum -> /home/joncrall/.cache/netharn/demo/_mru
... already exists
Symlink: /home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum -> /home/joncrall/.cache/netharn/demo/fit/nice/demo
... already exists
... and points to the right place
DEBUG: Initialized logging
INFO: Initializing tensorboard (dont forget to start the tensorboard server)
DEBUG: harn.train_info[hyper] = {
    'model': (
        'netharn.models.toynet.ToyNet2d',
        {
            'input_channels': 1,
            'num_classes': 2,
        },
    ),
    'initializer': (
        'netharn.initializers.nninit_core.KaimingNormal',
        {
            'mode': 'fan_in',
            'param': 0,
        },
    ),
    'optimizer': (
        'torch.optim.sgd.SGD',
        {
            'dampening': 0,
            'lr': 0.0001,
            'momentum': 0,
            'nesterov': False,
            'weight_decay': 0,
        },
    ),
    'scheduler': (
        'netharn.schedulers.scheduler_redesign.ListedScheduler',
        {
            'interpolation': 'linear',
            'optimizer': None,
            'points': {'lr': {0: 0.0001, 2: 0.01, 5: 0.015, 6: 0.005, 9: 0.001}},
        },
    ),
    'criterion': (
        'netharn.criterions.focal.FocalLoss',
        {
            'focus': 2,
            'ignore_index': -100,
            'reduce': None,
            'reduction': 'mean',
            'size_average': None,
            'weight': None,
        },
    ),
    'loader': (
        'torch.utils.data.dataloader.DataLoader',
        {
            'batch_size': 64,
        },
    ),
    'dynamics': (
        'Dynamics',
        {
            'batch_step': 4,
            'grad_norm_max': None,
        },
    ),
}
DEBUG: harn.hyper = <netharn.hyperparams.HyperParams object at 0x7fb19b4b8748>
DEBUG: make XPU
DEBUG: harn.xpu = <XPU(GPU(0)) at 0x7fb12af24668>
DEBUG: Criterion: FocalLoss
DEBUG: Optimizer: SGD
DEBUG: Scheduler: ListedScheduler
DEBUG: Making loaders
DEBUG: Making model
DEBUG: ToyNet2d(
  (layers): Sequential(
    (0): Conv2d(1, 8, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (1): BatchNorm2d(8, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU(inplace)
    (3): Conv2d(8, 8, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (4): BatchNorm2d(8, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (5): ReLU(inplace)
    (6): Conv2d(8, 2, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
  )
  (softmax): Softmax()
)
INFO: Model has 824 parameters
INFO: Mounting ToyNet2d model on GPU(0)
DEBUG: Making initializer
DEBUG: Move FocalLoss() model to GPU(0)
DEBUG: Make optimizer
DEBUG: Make scheduler
DEBUG: Make monitor
DEBUG: Make dynamics
INFO: Exported model topology to /home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum/ToyNet2d_2a3f49.py
INFO: Initializing model weights with: <netharn.initializers.nninit_core.KaimingNormal object at 0x7fb129e732b0>
DEBUG: calling harn.initializer=<netharn.initializers.nninit_core.KaimingNormal object at 0x7fb129e732b0>
INFO:  * harn.train_dpath = '/home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum'
INFO:  * harn.nice_dpath  = '/home/joncrall/.cache/netharn/demo/fit/nice/demo'
INFO: Snapshots will save to harn.snapshot_dpath = '/home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum/torch_snapshots'
INFO: ARGV:
    /home/joncrall/.local/conda/envs/py36/bin/python /home/joncrall/.local/conda/envs/py36/bin/ipython --verbose
INFO: dont forget to start:
    tensorboard --logdir ~/.cache/netharn/demo/fit/nice
INFO: === begin training 0 / 10 : demo ===
DEBUG: epoch lr:0.0001 │ vloss is unevaluated
epoch lr:0.0001 │ vloss is unevaluated  0/10... rate=0 Hz, eta=?, total=0:00:00, wall=19:56 EST
DEBUG: === start epoch 0 ===
DEBUG: log_value(epoch lr, 0.0001, 0
DEBUG: log_value(epoch momentum, 0, 0
DEBUG: _run_epoch 0, tag=train, learn=True
DEBUG:  * len(loader) = 8
DEBUG:  * loader.batch_size = 64
train loss:-1.000 │ 0.00% of 64x8... rate=0 Hz, eta=?, total=0:00:00, wall=19:56 ESTDEBUG: Making batch iterator
DEBUG: Starting batch iteration for tag=train, epoch=0
train loss:0.224 │ 100.00% of 64x8... rate=12052.25 Hz, eta=0:00:00, total=0:00:00, wall=19:56 EST
DEBUG: log_value(train epoch loss, 0.22378234565258026, 0
DEBUG: Finished batch iteration for tag=train, epoch=0
DEBUG: _run_epoch 0, tag=vali, learn=False
DEBUG:  * len(loader) = 4
DEBUG:  * loader.batch_size = 64
vali loss:-1.000 │ 0.00% of 64x4... rate=0 Hz, eta=?, total=0:00:00, wall=19:56 ESTDEBUG: Making batch iterator
DEBUG: Starting batch iteration for tag=vali, epoch=0
vali loss:0.175 │ 100.00% of 64x4... rate=23830.75 Hz, eta=0:00:00, total=0:00:00, wall=19:56 EST
DEBUG: log_value(vali epoch loss, 0.1749105490744114, 0
DEBUG: Finished batch iteration for tag=vali, epoch=0
DEBUG: epoch lr:0.0001 │ vloss: 0.1749 (n_bad=00, best=0.1749)
DEBUG: _run_epoch 0, tag=test, learn=False
DEBUG:  * len(loader) = 4
DEBUG:  * loader.batch_size = 64
test loss:-1.000 │ 0.00% of 64x4... rate=0 Hz, eta=?, total=0:00:00, wall=19:56 ESTDEBUG: Making batch iterator
DEBUG: Starting batch iteration for tag=test, epoch=0
test loss:0.176 │ 100.00% of 64x4... rate=28606.65 Hz, eta=0:00:00, total=0:00:00, wall=19:56 EST
DEBUG: log_value(test epoch loss, 0.17605290189385414, 0
DEBUG: Finished batch iteration for tag=test, epoch=0
DEBUG: Saving snapshot to /home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum/torch_snapshots/_epoch_00000000.pt
DEBUG: Snapshot saved to /home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum/torch_snapshots/_epoch_00000000.pt
DEBUG: new best_snapshot /home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum/torch_snapshots/_epoch_00000000.pt
DEBUG: Plotting tensorboard data
Populating the interactive namespace from numpy and matplotlib
INFO: === finish epoch 0 / 10 : demo ===

<JUST MORE OF THE SAME; REMOVED FOR BREVITY>

INFO: === finish epoch 8 / 10 : demo ===
DEBUG: epoch lr:0.001 │ vloss: 0.2146 (n_bad=08, best=0.1749)
epoch lr:0.001 │ vloss: 0.2146 (n_bad=08, best=0.1749)  9/10... rate=1.20 Hz, eta=0:00:00, total=0:00:07, wall=19:56 EST
DEBUG: === start epoch 9 ===
DEBUG: log_value(epoch lr, 0.001, 9
DEBUG: log_value(epoch momentum, 0, 9
DEBUG: _run_epoch 9, tag=train, learn=True
DEBUG:  * len(loader) = 8
DEBUG:  * loader.batch_size = 64
train loss:-1.000 │ 0.00% of 64x8... rate=0 Hz, eta=?, total=0:00:00, wall=19:56 ESTDEBUG: Making batch iterator
DEBUG: Starting batch iteration for tag=train, epoch=9
train loss:0.207 │ 100.00% of 64x8... rate=13580.13 Hz, eta=0:00:00, total=0:00:00, wall=19:56 EST
DEBUG: log_value(train epoch loss, 0.2070118673145771, 9
DEBUG: Finished batch iteration for tag=train, epoch=9
DEBUG: _run_epoch 9, tag=vali, learn=False
DEBUG:  * len(loader) = 4
DEBUG:  * loader.batch_size = 64
vali loss:-1.000 │ 0.00% of 64x4... rate=0 Hz, eta=?, total=0:00:00, wall=19:56 ESTDEBUG: Making batch iterator
DEBUG: Starting batch iteration for tag=vali, epoch=9
vali loss:0.215 │ 100.00% of 64x4... rate=29412.91 Hz, eta=0:00:00, total=0:00:00, wall=19:56 EST
DEBUG: log_value(vali epoch loss, 0.21514184772968292, 9
DEBUG: Finished batch iteration for tag=vali, epoch=9
DEBUG: epoch lr:0.001 │ vloss: 0.2148 (n_bad=09, best=0.1749)
DEBUG: _run_epoch 9, tag=test, learn=False
DEBUG:  * len(loader) = 4
DEBUG:  * loader.batch_size = 64
test loss:-1.000 │ 0.00% of 64x4... rate=0 Hz, eta=?, total=0:00:00, wall=19:56 ESTDEBUG: Making batch iterator
DEBUG: Starting batch iteration for tag=test, epoch=9
test loss:0.216 │ 100.00% of 64x4... rate=25906.58 Hz, eta=0:00:00, total=0:00:00, wall=19:56 EST
DEBUG: log_value(test epoch loss, 0.21618007868528366, 9
DEBUG: Finished batch iteration for tag=test, epoch=9
DEBUG: Saving snapshot to /home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum/torch_snapshots/_epoch_00000009.pt
DEBUG: Snapshot saved to /home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum/torch_snapshots/_epoch_00000009.pt
DEBUG: Plotting tensorboard data
INFO: === finish epoch 9 / 10 : demo ===
DEBUG: epoch lr:0.001 │ vloss: 0.2148 (n_bad=09, best=0.1749)
epoch lr:0.001 │ vloss: 0.2148 (n_bad=09, best=0.1749) 10/10... rate=1.21 Hz, eta=0:00:00, total=0:00:08, wall=19:56 EST




INFO: Maximum harn.epoch reached, terminating ...
INFO:



INFO: training completed
INFO: harn.train_dpath = '/home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum'
INFO: harn.nice_dpath  = '/home/joncrall/.cache/netharn/demo/fit/nice/demo'
INFO: view tensorboard results for this run via:
    tensorboard --logdir ~/.cache/netharn/demo/fit/nice
[DEPLOYER] Deployed zipfpath=/home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum/deploy_ToyNet2d_lnejaaum_000_JWPNDC.zip
INFO: wrote single-file deployment to: '/home/joncrall/.cache/netharn/demo/fit/runs/demo/lnejaaum/deploy_ToyNet2d_lnejaaum_000_JWPNDC.zip'
INFO: exiting fit harness.