netHarn 0.6.2

功能（继续）

• 超参数跟踪：您的超参数的哈希值决定了数据将被写入的目录。我们还允许以“更优雅”的方式管理目录结构。给定一个 HyperParams 对象，我们创建符号链接 {workdir}/fit/name/{name}，它指向 {workdir}/fit/runs/{name}/{hashid}。

• 自动重启：默认情况下，连续两次调用 FitHarn.run 将从上次停止的地方重新启动训练（只要超参数没有更改）。

• “智能”快照清理：维护模型权重文件可能会占用大量内存。根据 harn.preferences 的设置，netharn.FitHarn 将定期删除较旧或得分较低的快照。

• 部署文件：模型权重和架构一起写入一个相对便携的 zip 文件。我们还打包了训练元数据，以维护数据来源并简化实验的复现。

• 从任何预训练状态重启：使用 netharn.initializers.PretainedInitializer。

• torch 中构建网络的实用工具：如 netharn.layers.ConvNormNd 这样的层使得构建 n=1、2 或 3 维数据的网络变得容易。

• 分析输出形状和感受野：Netharn 定义了一个 netharn.layers.AnalyticModule，如果用户定义了一个特殊的 _output_for 方法，它就可以自动定义 forward、output_shape_for 和 receptive_field_for。该方法使用 netharn.analytic_for.Output、netharn.analytic_for.Hidden 和 netharn.analytic_for.OutputFor 特殊的可调用对象编写。

• 示例任务：标准任务（如对象分割、分类和检测）的基线代码定义在 netharn.examples 中。示例还提供了 ndsampler、kwimage、kwannot 和 kwplot 的示例使用场景。

安装

未来这些说明可能与开发者设置说明不同，但到目前为止它们是相同的。

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

尽管所有 netharn 依赖项都应该在 pypi（带有许多linux2010 轮子的二进制包）上可用，但还有其他与 netharn 同时开发的包。要安装这些依赖项的开发版本，请运行 python super_setup.py ensure 以检出存储库并确保它们位于正确的分支上，运行 python super_setup.py develop 以以开发模式构建一切，并运行 python super_setup.py pull 以更新到分支上的最新版本。

描述

PyTorch 的参数化训练 harness。

训练模型并跟踪您的超参数。

这是对我研究存储库 clab 中开发的好部分的干净移植。

请参阅 netharn/examples 文件夹以获取示例用法。doctests 也是一个很好的资源。如果能有更好的文档就更好了。

NetHarn 是一个用于训练和部署任意 PyTorch 模型的研究框架。它旨在最小化训练循环的样板代码并跟踪超参数以鼓励可重复研究。NetHarn 将训练模型的问题分为以下核心超参数组件：数据集、模型、标准、初始化器、优化器和学习率调度器。具有不同超参数的运行将自动记录到单独的目录中，这使得比较两个实验的结果变得简单。NetHarn 还具有创建独立于用于训练的系统的单个文件部署训练模型的 capability。这使得研究结果可以快速且简单地外部验证并投入生产。

核心回调结构

Netharn 是通过继承 netharn.FitHarn 类、重载多个方法，然后使用特定的超参数创建自定义 FitHarn 实例来设计的。

FitHarn 允许您通过其回调系统自定义训练循环的执行。您只需简单地重载这些方法之一即可编写一个回调。有带默认行为的回调和不带默认行为的回调。

具有默认行为的那些直接影响学习过程。虽然它们不必被覆盖，但通常应该覆盖，因为不同的任务需要通过训练管道移动数据的不同方式。

没有默认行为的那些允许开发者执行自定义代码，并在训练循环的特殊位置。这些通常用于记录自定义指标和输出可视化。

以下说明列出了回调函数，大致按照它们被FitHarn.run方法调用的顺序。树结构表示循环嵌套。

├─ after_initialize (no default) - runs after FitHarn is initialized
│  │
│  ├─ before_epochs (no default) - runs once before all train/vali/test
│  │  │    epochs on each iteration
│  │  │
│  │  ├─ prepare_epoch (no default) - runs before each train, vali,
│  │  │  │    and test epoch
│  │  │  │
│  │  │  ├─ prepare_batch (has default behavior) - transfer data from
│  │  │  │    CPU to the XPU
│  │  │  │
│  │  │  ├─ run_batch (has default behavior) - execute the forward pass
│  │  │  │    and compute the loss
│  │  │  │
│  │  │  ├─ backpropogate (has default behavior) - accumulate gradients
│  │  │  │    and take an optimization step
│  │  │  │
│  │  │  └─ on_batch (no default) - runs after `run_batch` and
│  │  │       `backpropogate` on every batch
│  │  │
│  │  └─ on_epoch (no default) - runs after each train, vali, and test
│  │         epoch finishes.  Any custom scalar metrics returned in a
│  │         dictionary will be recorded by the FitHarn loggers.
│  │
│  └─ after_epochs (no default) - runs after the all data splits are
│         finished with  the current epoch.
│
└─ on_complete (no default) - runs after the main loop is complete

给定一个自定义FitHarn类，请参阅“玩具示例”部分，了解如何构建超参数和执行训练循环（即FitHarn.run）的详细信息。

开发者设置

将来这些说明可能不同于安装说明，但到目前为止它们是相同的。

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

文档

Netharn的文档目前较为稀少。我通常在代码本身中使用docstrings进行文档编写。将来，这些内容很可能会在read-the-docs风格的文档页面上进行整合，但现在您需要查看代码来阅读文档。

netharn提供的主要概念是“FitHarn”，它具有相当不错的模块级docstring，以及许多优秀的类/方法级docstring：https://gitlab.kitware.com/computer-vision/netharn/-/blob/master/netharn/fit_harn.py

示例文件夹中的docstrings具有任务级文档。

最简单的是mnist示例：https://gitlab.kitware.com/computer-vision/netharn/-/blob/master/netharn/examples/mnist.py

CIFAR示例基于mnist示例：https://gitlab.kitware.com/computer-vision/netharn/-/blob/master/netharn/examples/cifar.py

我建议您查看这两个示例，因为它们具有最好的文档。

分割示例：https://gitlab.kitware.com/computer-vision/netharn/-/blob/master/netharn/examples/segmentation.py

和对象检测示例：https://gitlab.kitware.com/computer-vision/netharn/-/blob/master/netharn/examples/object_detection.py

这些示例的文档较少，但提供了更多关于如何使用netharn的真实世界示例。

有一个针对CAMVID数据集的特定应用分割示例：https://gitlab.kitware.com/computer-vision/netharn/-/blob/master/netharn/examples/sseg_camvid.py

还有一个针对VOC检测的应用示例：https://gitlab.kitware.com/computer-vision/netharn/-/blob/master/netharn/examples/yolo_voc.py

此README还包含一个玩具示例。

玩具示例

以下示例是netharn/fit_harn.py中的doctest。它演示了如何使用NetHarn训练一个模型来解决玩具问题。

在这个玩具问题中，我们没有扩展netharn.FitHarn对象，因此我们使用run_batch的默认行为。默认的on_batch和on_epoch不执行任何操作，因此损失将是唯一的性能度量。

有关更多示例，请参阅示例目录。这些示例展示了如何扩展netharn.FitHarn以衡量特定问题的性能。MNIST和CIFAR示例是最简单的。YOLO示例更复杂。IBEIS示例依赖于非公开数据和/或软件，但仍然具有参考价值。它的复杂性大于CIFAR，但小于YOLO。

>>> import netharn
>>> hyper = netharn.HyperParams(**{
>>>     # ================
>>>     # Environment Components
>>>     'name'        : 'demo',
>>>     'workdir'     : ub.ensure_app_cache_dir('netharn/demo'),
>>>     'xpu'         : netharn.XPU.coerce('auto'),
>>>     # workdir is a directory where intermediate results can be saved
>>>     # "name" symlinks <workdir>/fit/name/<name> -> ../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=64, rng=1),
>>>         'test': netharn.data.ToyData2d(size=3, border=1, n=64, rng=2),
>>>     },
>>>     'loaders'     : {'batch_size': 4}, # DataLoader instances or kw
>>>     # ================
>>>     # Algorithm Components
>>>     # Note the (cls, kw) tuple formatting
>>>     'model'       : (netharn.models.ToyNet2d, {}),
>>>     'optimizer'   : (netharn.optimizers.SGD, {
>>>         'lr': 0.01
>>>     }),
>>>     # focal loss is usually better than netharn.criterions.CrossEntropyLoss
>>>     'criterion'   : (netharn.criterions.FocalLoss, {}),
>>>     'initializer' : (netharn.initializers.KaimingNormal, {
>>>         'param': 0,
>>>     }),
>>>     # The scheduler adjusts learning rate over the training run
>>>     'scheduler'   : (netharn.schedulers.ListedScheduler, {
>>>         'points': {'lr': {0: 0.1, 2: 10.0, 4: .15, 6: .05, 9: .01}},
>>>         'interpolation': 'linear',
>>>     }),
>>>     'monitor'     : (netharn.Monitor, {
>>>         'max_epoch': 10,
>>>         'patience': 7,
>>>     }),
>>>     # dynamics are a config option that modify the behavior of the main
>>>     # training loop. These parameters effect the learned model.
>>>     'dynamics'   : {'batch_step': 2},
>>> })
>>> harn = netharn.FitHarn(hyper)
>>> # non-algorithmic behavior preferences (do not change learned models)
>>> harn.preferences['num_keep'] = 10
>>> harn.preferences['auto_prepare_batch'] = True
>>> # start training.
>>> harn.initialize(reset='delete')  # delete removes an existing run
>>> harn.run()  # note: run calls initialize it hasn't already been called.
>>> # xdoc: +IGNORE_WANT

运行此代码产生以下输出

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/name/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.name_dpath  = '/home/joncrall/.cache/netharn/demo/fit/name/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/name
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.name_dpath  = '/home/joncrall/.cache/netharn/demo/fit/name/demo'
INFO: view tensorboard results for this run via:
    tensorboard --logdir ~/.cache/netharn/demo/fit/name
[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.

此外，如果您在'--verbose' in sys.argv时运行该代码，则会产生以下更详细的操作描述

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/name/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.name_dpath  = '/home/joncrall/.cache/netharn/demo/fit/name/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/name
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.name_dpath  = '/home/joncrall/.cache/netharn/demo/fit/name/demo'
INFO: view tensorboard results for this run via:
    tensorboard --logdir ~/.cache/netharn/demo/fit/name
[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.

相关包

pytorch-lightning（https://github.com/PyTorchLightning/pytorch-lightning）与netharn有非常相似的目标。目前，两者各有优缺点，但未来我确实看到其中一个将吸收另一个的功能。目前（2020-10-21），pytorch-lightning在分布式训练方面做得更好，而netharn的日志记录和超参数管理优于pytorch-lightning。

消费者包

bioharn包（https://gitlab.kitware.com/jon.crall/bioharn）实现了netharn/examples文件夹中分类器和检测器示例的扩展，以及预测和评估脚本。