Since atari env uses (dim_1, dim_2, channels) as image representation,
convolution2D needs to take dim_ordering parameter as 'tf', not 'th'
which is default which assumes first dim as channels (RGB).
Remove unnecessary imports and empty lines in the end.
Since atari env uses (dim_1, dim_2, channels) as image representation, convolution2D needs to take dim_ordering parameter as 'tf', not 'th' which is default which assumes first dim as channels (RGB). Remove unnecessary imports and empty lines in the end.