convnets 0.2.9

Convolutional Neural Networks and utilities for Computer Vision

convnets

🚧 Under construction

Convolutional Neural Networks and utilities for Computer Vision.

• Learn about convnets.
• Learn about popular models.

Models API

convnets offers implementations for the following models:

• LeNet
• AlexNet
• VGG
• ResNet

To instantiate a model you need to import the corresponding class and pass a valid configuration object to the constructor:

from convnets.models import ResNet

r18_config = {
	'l': [
		{'r': 2, 'f': 64},
		{'r': 2, 'f': 128},
		{'r': 2, 'f': 256},
		{'r': 2, 'f': 512}
	], 
	'b': False
}

model = ResNet(r18_config)

Or you can use one of the predefined configurations, or variants:

from convnets.models import ResNet, ResNetConfig

model = ResNet(ResNetConfig.r18)

You can find the implementation of each model and configuration examples in the convnets/models directory.

Training API

If you want to train a model in your notebooks, you can use our fit function:

form convnets.train import fit 

hist = fit(model, dataloader, optimizer, criterion, metrics, max_epochs)

You can use any Pytorch model. You will need to define the Pytorch dataloader, optimizer and criterion. For the metrics, the function expects a dict with the name of the metric as key and the metric function as value. The metric function must receive the model output and the target and return a scalar value. You can find some examples in convnets/metrics. The max_epochs parameter is the maximum number of epochs to train the model. The function will return a dict with the training history.

Additionally, we offer a training script that you can execute from the command line.

python scripts/train.py <path_to_config_file>

You will have to pass the path to a yaml file with the configuration for your training, including the model, optimizer, criterion, metrics, dataloader, etc. You can find some examples in the configs directory (which are timm and pytorch-lightning compatible).

We also offer Pytorch Lightning interoperability.