Welcome to torchero’s documentation!¶
Torchero is a library that works on top of the PyTorch framework built to facilitate the training of Neural Networks.
Features¶
It provides tools and utilities to:
- Set up a training process in a few lines of code.
- Monitor the training performance by checking several prebuilt metrics on a handy progress bar.
- Integrate a dashboard with TensorBoard to visualize those metrics in an online manner with a minimal setup.
- Add custom functionality via Callbacks.
Installlation¶
You can install torchero
either from pip or from source.
From pip¶
python3 -m pip install torchero
From source code¶
Download the source code using git
git clone https://github.com/juancruzsosa/torchero
or download the tarball if you don’t have git installed
curl -OL https://github.com/juancruzsosa/torchero/tarball/master
Once you have the source code downloaded you can install it with
cd torchero
python3 setup.py install
Quickstart¶
Loading the Data¶
import torch
from torch import nn
import torchero
from torchero.models.vision import ImageClassificationModel
from torchero.callbacks import ProgbarLogger, ModelCheckpoint, CSVLogger
from torchero.utils.data import train_test_split
from torchero.utils.vision import show_imagegrid_dataset, transforms, datasets, download_image
from torchero.meters import ConfusionMatrix
from matplotlib import pyplot as plt
First we load the MNIST train and test datasets and visualize it using
show_imagegrid_dataset
.The Data Augmentation for this case will be a RandomInvert to flip the grayscale levels.
train_ds = datasets.MNIST(root='/tmp/data/mnist', download=True, train=True, transform=transforms.Compose([transforms.RandomInvert(),
test_ds = datasets.MNIST(root='/tmp/data/mnist', download=False, train=False, transform=transforms.ToTensor())
show_imagegrid_dataset(train_ds)
plt.show()
Defining the Network¶
Let’s define a Convolutional network of two layers followed by a Linear Module
as the classification layer.
model = nn.Sequential(nn.Conv2d(in_channels=1, out_channels=32, kernel_size=5),
nn.ReLU(inplace=True),
nn.MaxPool2d(2),
nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3),
nn.ReLU(inplace=True),
nn.MaxPool2d(2),
nn.Flatten(),
nn.Linear(5*5*64, 500),
nn.ReLU(inplace=True),
nn.Linear(500, 10))
Training the Model¶
The ImageClassificationModel is the module responsible to train the model,
evaluate a metric against a dataset, and predict from and input for multi-class classification tasks.
To train the model we need to compile it first with a:
- an optimizer: ‘adam’
- a loss which will be defaulted to
cross_entropy
- a list of metrics which will be defaulted to
categorical_accuracy
, balanced_accuracy
)- a list of callbacks:
- ProgbarLogger to show training progress bar
- ModelCheckpoint to make checkpoints if the model improves
- CSVLogger to dump the metrics to a csv file after each epoch
model = ImageClassificationModel(model=network,
transform=transforms.Compose([transforms.Grayscale(),
transforms.Resize((28,28)),
transforms.ToTensor()]),
classes=[str(i) for i in range(10)])
model.compile(optimizer='adam',
callbacks=[ProgbarLogger(notebook=True),
ModelCheckpoint('saved_model', mode='max', monitor='val_acc'),
CSVLogger('training_results.xml')])
if torch.cuda.is_available():
model.cuda()
history = model.fit(train_ds,
test_ds,
batch_size=1024,
epochs=5)
Visualizing the training training results¶
To visualize our loss and accuracy in each epoch we can execute:
history.plot(figsize=(20, 20), smooth=0.2)
plt.show()
The .evaluate
returns the metrics for a new dataset.
results = trainer.evaluate(test_dl, ['categorical_accuracy_percentage', 'confusion_matrix'])
plt.figure(figsize=(10, 10))
results['confusion_matrix'].plot(classes=train_ds.classes)