I want to train a mobilenet ssd v2 model using the tf object detection api to use this api you have to convert your dataset to a tf record and write the path to it in the config file
Whenever I look it up I find answers that include tf records, either one tf record that contains the whole dataset or multiple files with sections of it (but idk how to make the training use all sections of the dataset each epoch)
How do I train the model without loading the whole dataset in memory?
Is there a way to train without using tfrecords ?
If I convert my whole dataset to tfrecords I'll have 15 GBs of duplicated space, or is there a way to not load the whole tf record,loading shards of it at a time but still looping over the whole dataset very epoch.