I would like to implement something like this: https://arxiv.org/pdf/1511.06309.pdf

Essentially, the idea is to nest a temporal autoencoder (probably using LSTMs) inside the spatial autoencoder, in order to warp the encoding to predict the NEXT frame in a video sequence.

load MNIST dataset

import urllib, gzip, pickle

mnistfile = 'mnist.pkl.gz'
if not os.path.isfile(mnistfile):
url = urllib.request.URLopener()
url.retrieve("http://deeplearning.net/data/mnist/mnist.pkl.gz", mnistfile)
f = gzip.open(mnistfile, 'rb')

training_set, validation_set, testing_set = pickle.load(f, encoding='latin1')

f.close()

X_train, y_train = training_set
X_validate, y_validate = validation_set
X_test, y_test = testing_set

Hi there, I tried creating an instance of the class ConvAutoEncoder with the default parameters and the execution failed.