Hello everyone, I'm fine tuning the efficientnet model from the model garden for an image classification task. I'm doing it in a google colab and i've encountered many problems so far.
First off the Hyperparameter tuning step is not capable of creating new directories and writing its checkpoints and steps inside those folders. to circumvent this issue i created the "trial" folders for him and now it can write those files. Consider that i'm using a service account for it which has both the Storage admin (also storage object admin) and the Vertex AI Custom Code Service Agent. I would like to know why it can't create folders on its own, unless i create them the job succeeds but no files are written and i can't continue.
Second issue is the result of the prediction step, the task should be a single label (food or not food) however after training, exporting, uploading and deploying the model the images i ask to predict get an empty dictionary for probs and logits. I also tried including a second label and re-training but the situation stays the same.
How can i fix these issues? all the customjobs have the credentials json file passed as a parameter.
LinkedIn
Twitter
