Hello again. I was looking through your data loader in the file src/sbir_baseline/dataloader.py, and was a bit unsure of the code for drawing negative samples. Specifically, line 47:
Could you please explain the choice of negative sampling dataset? I thought self.all_ids may be a better fit than self.all_image_files because 1.) the train/val split is applied to this set and 2.) the positive photo sample is also drawn from self.all_ids in line 46. I was hoping you could explain what I have misunderstood, and why this is not the case.
Hello again. I was looking through your data loader in the file src/sbir_baseline/dataloader.py, and was a bit unsure of the code for drawing negative samples. Specifically, line 47:
negative_file = np.random.choice(self.all_image_files, 1)[0]
Could you please explain the choice of negative sampling dataset? I thought self.all_ids may be a better fit than self.all_image_files because 1.) the train/val split is applied to this set and 2.) the positive photo sample is also drawn from self.all_ids in line 46. I was hoping you could explain what I have misunderstood, and why this is not the case.
Thank you very much for your time in advance.