Skip to main content

Utilities for training models in pytorch

Project description

xt-training

Description

This repo contains utilities for training deep learning models in pytorch, developed by Xtract AI.

Installation

From PyPI:

pip install xt-training

From source:

git clone https://github.com/XtractTech/xt-training.git
pip install ./xt-training

Usage

See specific help on a class or function using help. E.g., help(Runner).

Training a model

Using xt-training (High Level)

First, you must define a config file with the necessary items. To generate a template config file, run:

python -m xt_training template path/to/save/dir

To generate template files for nni, add the --nni flag

Instructions for defining a valid config file can be seen at the top of the config file.

After defining a valid config file, you can train your model by running:

python -m xt_training train path/to/config.py /path/to/save_dir

You can test the model by running

python -m xt_training test path/to/config.py /path/to/save_dir
Using functional train (Middle Level)

As of version >=2.0.0, xt-training has functional calls for the train and test functions This is useful if you want to run other code after training, or want any values/metrics returned after training. This can be called like so:

from xt_training.utils import functional

# model = 
# train_loader = 
# optimizer = 
# scheduler = 
# loss_fn = 
# metrics = 
# epochs = 
# save_dir = 
def on_exit(test_loaders, runner, save_dir, model):
    # Do what you want after training.
    # As of version >=2.0.0. whatever gets returned here will get returned from the functional call
    return runner, model

runner, model = functional.train(
    save_dir,
    train_loader,
    model,
    optimizer,
    epochs,
    loss_fn,
    val_loader=None,
    test_loaders=None,
    scheduler=scheduler,
    is_batch_scheduler=False, # Whether or not to run scheduler.step() every epoch or every step
    eval_metrics=metrics,
    tokenizer=None,
    on_exit=train_exit,
    use_nni=False
)

# Do something after with runner and/or model...

A similar functional call exists for test.

Using Runner (Low Level)

If you want a little more control and want to define the trianing code yourself, you can utilize the Runner like so:

from xt_training import Runner, metrics
from torch.utils.tensorboard import SummaryWriter

# Here, define class instances for the required objects
# model = 
# optimizer = 
# scheduler = 
# loss_fn = 

# Define metrics - each of these will be printed for each iteration
# Either per-batch or running-average values can be printed
batch_metrics = {
    'eps': metrics.EPS(),
    'acc': metrics.Accuracy(),
    'kappa': metrics.Kappa(),
    'cm': metrics.ConfusionMatrix()
}

# Define tensorboard writer
writer = SummaryWriter()

# Create runner
runner = Runner(
    model=model,
    loss_fn=loss_fn,
    optimizer=optimizer,
    scheduler=scheduler,
    batch_metrics=batch_metrics,
    device='cuda:0',
    writer=writer
)

# Define dataset and loaders
# dataset = 
# train_loader = 
# val_loader = 

# Train
model.train()
runner(train_loader)
batch_metrics['cm'].print()

# Evaluate
model.eval()
runner(val_loader)
batch_metrics['cm'].print()

# Print training and evaluation history
print(runner)

Scoring a model

import torch
from xt_training import Runner

# Here, define the model
# model = 
# model.load_state_dict(torch.load(<checkpoint file>))

# Create runner
# (alternatively, can use a fully-specified training runner as in the example above)
runner = Runner(model=model, device='cuda:0')

# Define dataset and loaders
# dataset = 
# test_loader = 

# Score
model.eval()
y_pred, y = runner(test_loader, return_preds=True)

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

xt-training-2.4.3a2.tar.gz (20.1 kB view details)

Uploaded Source

Built Distribution

xt_training-2.4.3a2-py3-none-any.whl (23.8 kB view details)

Uploaded Python 3

File details

Details for the file xt-training-2.4.3a2.tar.gz.

File metadata

  • Download URL: xt-training-2.4.3a2.tar.gz
  • Upload date:
  • Size: 20.1 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/3.2.0 pkginfo/1.5.0.1 requests/2.23.0 setuptools/46.4.0.post20200518 requests-toolbelt/0.9.1 tqdm/4.46.0 CPython/3.7.7

File hashes

Hashes for xt-training-2.4.3a2.tar.gz
Algorithm Hash digest
SHA256 d96af86d3da8cb30b42d152a24290d1fe4dfc6ade56c359eab0a080e185677af
MD5 27d4ef5fb7ec655394214f912ef60472
BLAKE2b-256 78d0e76106a5986ce6e60291017174fc3efebedf75d7b92ae800061ef1ad25da

See more details on using hashes here.

File details

Details for the file xt_training-2.4.3a2-py3-none-any.whl.

File metadata

  • Download URL: xt_training-2.4.3a2-py3-none-any.whl
  • Upload date:
  • Size: 23.8 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/3.2.0 pkginfo/1.5.0.1 requests/2.23.0 setuptools/46.4.0.post20200518 requests-toolbelt/0.9.1 tqdm/4.46.0 CPython/3.7.7

File hashes

Hashes for xt_training-2.4.3a2-py3-none-any.whl
Algorithm Hash digest
SHA256 7b52f3260a66cb02ea7d196ba1b7528d55263284b93f696e07e19a3afebc4254
MD5 598e0a24133b6ca9c0133de87cb729fb
BLAKE2b-256 191aee8a666606a0086370dae113da903445e09b8f9eb6a6e19b1fe287000168

See more details on using hashes here.

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page