Recognite

Recognite is a library to kickstart your next PyTorch-based recognition project. Some interesting features include:

• You can choose from nearly 80 different base models for your recognition model: classics like AlexNet, GoogLeNet, VGG, Inception, ResNet, but also more recent models like ResNeXt, EfficientNet, and transformer-based models like ViT and SwinTransformer.
• You can easily evaluate your model model directly for a recognition task, where query samples are compared with a gallery, and none of the samples have a class that was used during training.
• By changing only a single argument, you can cross-validate sets of hyperparameters without much effort.

Installation

You can install Recognite with pip:

pip install recognite

Quickstart

This repo contains a basic training script with which you can quickly start a recognition training. To use this script in your project, you can clone the repository, copy the script into your project directory and install the script's requirements:

# Clone the Recognite repo
git clone https://github.com/florisdf/recognite

# Copy the training script to your project
cp recognite/examples/basic/train.py path/to/your/recognition_project

# Install the requirements of the training script
pip install -r recognite/examples/basic/requirements.txt

The last line installs Weights and Biases, which is used for logging. Make sure to create an account and run wandb login from your command line.

The training script trains a recognition model of your choice on a dataset you define, using tools from the Recognite library. The dataset should be given as a CSV file (--data_csv) with two columns: image (containing image paths) and label (containing the corresponding labels). We split the unique labels of the dataset into 5 folds. Labels in the fold defined by --val_fold are used for validation. The others are used for training. During validation, we measure the model's top-1 accuracy when classifying a set of queries by comparing the query embeddings with the embeddings of a set of reference samples (--num_refs per validation label). This accuracy is logged to Weights and Biases (see --wandb_entity and --wandb_project).

Each image is uniformly resized such that its shortest side has a fixed size (--size). For training images, we then take a square crop of that size at a random location in the image. For the validation images, we crop out the square center of the image.

For the model, you can choose from a large number of pretrained classifiers, see --model_name and --model_weights. The model's final fully-connected layer is adjusted to the number of classes in the training set and is then trained for --num_epoch epochs by optimizing the softmax cross-entropy loss with stochastic gradient descent, configured by --batch_size, --lr, --momentum and --weight_decay.

For example, with the following command, we train a ResNet-18 model with default pretrained weights for 30 epochs on images from data.csv using a learning rate of 0.01, a momentum of 0.9, and a weight decay of 1e-5. As validation set, we use the labels of the first fold (index 0) and we use 1 reference sample per label in the gallery set.

python train.py \
    --model_name=resnet18 --model_weights=DEFAULT \
    --data_csv=data.csv --val_fold=0 --num_refs=1 --size=224 \
    --num_epochs=30 --lr=0.01 --momentum=0.9 --weight_decay=1e-5 \
    --wandb_entity=your_user_name --wandb_project=your_project

For more details on the different command line arguments, you can run

python train.py --help

More information

See the docs for more information and examples with Recognite.