Author: Danelle Cline, dcline@mbari.org
About
kclassify runs an image classification model using transfer-learning. This runs on one or many GPUS either locally or in AWS.
Currently, supports the following:
Optimizers
- Radam
- Adam
- Ranger (not working as of 7-20-21)
Models
- efficientnetB0
- resnet50
- vgg16
- vgg19
- mobilenetv2
Augmentations
- width, shift, and zoom
- horizontal/vertical flip
and all the typical hyperparameters needed for model training like learning rate, batch size, etc.
Requirements
- Docker
- SageMaker SDK version 2.20.0
- One or more GPUs
- Training and validation images (JPEG or PNG) images compressed into tar.gz files. See Data organization for details
- AWS Account (only needed if modifying the model)
- Your AWS account must support the role ecr:InitiateLayerUpload to push the docker image this creates