Doyup Lee
Doyup Lee
The pretrained features are very important, because it contains feature vectors of objects in each image and the location of bounding box.
In this paper, the detail method is not presented. However, the basic approach is construct validation dataset and set anomaly threshold based on validation sets.
There is anomaly score in each pixel and the value is the aggregation of all the pixel. So it can be large. Rather than considering absolute value, you can define...
sorry for late comment. I think ther is no self.test_data_names in DCGAN class. could you let me know your detail contexts of error?
could you let me know which lines and what is the context of run?
I didn't check the issue, but i think it is because of memory usage. Basically, define-and-run is used in tensorflow and there would be unnecessary definition of some operations. I...
could i let me know your command in terminal?
I didn't implement it. However, after get the features, use can use sklearn.manifold.TSNE (https://scikit-learn.org/stable/modules/generated/sklearn.manifold.TSNE.html)
What it means?
If you use MNIST, concatenating the label into input (conditional GAN) can be a solution. I checked that the final image of celebA is fine.