Our team is currently exploring the model that is provided in NGC here:
We have tried to inference some images into this model. The images itself comes from a number of sources online, that is: the NIH dataset and the Chexpert dataset.
But we have found that using the images from above dataset results in very biased readings to some classes. We used 19 different images from different classes of radiological readings, yet all tends to output the same results, as shown in the image below.
And so my questions are:
(1) Our hypothesis of why this happens, is because the model itself is trained on 16 Bit PNG images, yet the images we used are that of 8 Bit PNGs. Is this hypothesis true?
Some questions more related to the model itself than Clara Train:
(2) Do you think using this model as a pretrained model, which is used to train a new model using 8 Bit images is a good idea?
(3) Aside from the PLCO dataset, are there notable sources of 16 Bit PNG images online that the Nvidia team is aware of?