Peter234 Peter234 - 3 months ago 27
Python Question

Override numpy.random to use cudamat

I have a programm that uses np.random many times. Now I wan't the user to pass an argument

gpu=True/False
. How can I override np.random to return
cm.CUDAMatrix(np.random.uniform(low=low, high=high, size=size))
without ending in a recursion?
Or is there a better way to use cudamat with small code changes?

Thanks for your help.

If you need more code please comment.

class FeedForwardNetwork():

def __init__(self, input_dim, hidden_dim, output_dim, dropout=False, dropout_prop=0.5, gpu=True):
np.random.seed(1)
self.input_layer = np.array([])
self.hidden_layer = np.array([])
self.output_layer = np.array([])
self.input_dim = input_dim
self.hidden_dim = hidden_dim
self.output_dim = output_dim
self.dropout = dropout
self.dropout_prop = dropout_prop

r_input_hidden = math.sqrt(6 / (input_dim + hidden_dim))
r_hidden_output = math.sqrt(6 / (hidden_dim + output_dim))

self.weights_input_hidden = np.random.uniform(low=-0.01, high=0.01, size=(input_dim, hidden_dim))
self.weights_hidden_output = np.random.uniform(low=-0.01, high=0.01, size=(hidden_dim, output_dim))

Answer
class FeedForwardNetwork():

def __init__(self, input_dim, hidden_dim, output_dim, dropout=False, dropout_prop=0.5, gpu=True):            
    np.random.seed(1)
    self.input_layer = np.array([])
    self.hidden_layer = np.array([])
    self.output_layer = np.array([])
    self.input_dim = input_dim
    self.hidden_dim = hidden_dim
    self.output_dim = output_dim
    self.dropout = dropout
    self.dropout_prop = dropout_prop

    r_input_hidden = math.sqrt(6 / (input_dim + hidden_dim))
    r_hidden_output = math.sqrt(6 / (hidden_dim + output_dim))

    self.weights_input_hidden = np.random.uniform(low=-0.01, high=0.01, size=(input_dim, hidden_dim))
    self.weights_hidden_output = np.random.uniform(low=-0.01, high=0.01, size=(hidden_dim, output_dim))

def np_random(self, gpu):
   '''gpu:bool'''
     if gpu:
         return np.random.uniform(low=-0.01, high=0.01, size=(self.input_dim, self.hidden_dim))
    else:
         return np.random.uniform(low=-0.01, high=0.01, size=(self.hidden_dim, self.output_dim))

Then you can call it from your instance:

instance = FeedForwardNetwork(**kwargs)

instance.np_random(True/False)