![]() The code utilized PyTorch for training/tagging, OpenCV for image preprocessing, and Matplotlib for generating loss plots.The dataset was split into a ratio of 85-15 for training and testing respectively as generally higher training ratios yield more accurate models. But various code modifications were made to tailor to the photo multi-tag project including:Īlter the model dimensions so that my custom dataset could fit the final ResNet layerĪdd additional transforms such as image normalizationĪdd code which converts the output model file into an onxx file For the training/testing dataset, I selected and manually tagged a selection of 1,500 vacation images into 29 unique tags and stored said information into a csv file.Īs a starting point, I adopted a generic code skeleton that was built for classifying the genres of movie posters through transfer learning. ![]() The original approach was to train a CNN from scratch using the entire picture library, but decided that implementing transfer learning on an existing model (in this case ResNet50) would be more time efficient for purposes of general classification.
0 Comments
Leave a Reply. |