I recently switched to TFlearn to get some calculations on my network accuracy for classifying images and creating state of the art results. I am using this exact file from TFlearn, except for the dataset. I wanted to reproduce the accuracy for the VOC2007 dataset, downloaded all images and groundtruth and wrote a function that creates a 4D Tensor containing all images and a 2D Tensor containing all class indices. Their shapes are [?, 224, 224, 3] and [?, 20] respectively. Now i noticed, that the class indices are not one-hot labels, but several classes can be present in one image. Since TFlearn allows more than one class to be present, the network performs really badly (accuracy ~30%, and yes, i changed the number of output classes). I wonder, how to get around this problem. Should i only allow one class per image? But then, if there are two classes in an image and i classify it correctly according to the second class, this would be a correct detection that i would missclassify as an error. Is there any option that i am missing? I don't see a "one-hot" option or anything similar (like in the oxflowers dataset).
Thanks for your help!
