Converts a Federated EMNIST dataset into an Infinite Federated EMNIST set.
Infinite Federated EMNIST expands each writer from the EMNIST dataset into
some number of pseudo-clients each of whose characters are the same but apply
a fixed random affine transformation to the original user's characters. The
distribution over affine transformation is approximately equivalent to the one
described at https://www.cs.toronto.edu/~tijmen/affNIST/ It applies the
following transformations in this order:
- A random rotation chosen uniformly between -20 and 20 degrees.
- A random shearing adding between -0.2 to 0.2 of the x coordinate to the
y coordinate (after centering).
- A random scaling between 0.8 and 1.25 (sampled log uniformly).
- A random translation between -5 and 5 pixels in both the x and y axes.
tff.simulation.ClientData to convert.
How many pseudo-clients to generate for each real
client. Each pseudo-client is formed by applying a given random affine
transformation to the characters written by a given real user. The first
pseudo-client for a given user applies the identity transformation, so the
original users are always included.