Example #1
0
# The diff, given at the last, represent the difference between
# the gradients from BP and numerical computing. It should be
# very small, otherwise the autoencoder was not trained correctly.

##==========================================================


import numpy as np
import sparse_autoencoder as sp
import compute_numerical_grad as co


data = np.random.rand(64, 100)
visible_size = 64
hidden_size = [16, 8, 4]
#hidden_size = [25]
theta = sp.initial_parameter(hidden_size, visible_size)

bp_cost = sp.compute_cost(theta, data, visible_size, hidden_size)
print bp_cost

bp_grad = sp.compute_grad(theta, data, visible_size, hidden_size)
print bp_grad

num_grad = co.compute_numerical_grad(sp.compute_cost, theta, data,
                                     visible_size, hidden_size)

diff = np.linalg.norm(bp_grad - num_grad) / np.linalg.norm(num_grad + bp_grad)

print str(diff) + " should be less than 1e-9! Is it?"
Example #2
0
# Randomly generate theta and data
visible_size = 32
hidden_size = [16, 8, 4]
data = np.random.rand(32, 100)

layer_ind = range(len(hidden_size) + 1)
layer_ind.remove(0)
layer_size = [visible_size] + hidden_size

# Debugging!
for ind in layer_ind:

    theta = sp.initial_parameter(layer_size[ind], layer_size[ind - 1])

    bp_grad = sp.compute_grad(theta, data, layer_size[ind - 1],
                              layer_size[ind])

    num_grad = co.compute_numerical_grad(sp.compute_cost, theta, data,
                                         layer_size[ind - 1], layer_size[ind])

    diff = np.linalg.norm(bp_grad - num_grad) /\
        np.linalg.norm(num_grad + bp_grad)

    print str(diff) + " should be less than 1e-9! Is it?"

    W = theta[:layer_size[ind]*layer_size[ind-1]].\
        reshape(layer_size[ind], layer_size[ind-1])

    data = np.dot(W, data)
Example #3
0
layer_ind = range(len(hidden_size) + 1)
layer_ind.remove(0)
layer_size = [visible_size] + hidden_size

dpark_ctx = DparkContext()

printdiff = []

start = clock()
# Debugging!
for ind in layer_ind:

    theta = sp.initial_parameter(layer_size[ind], layer_size[ind - 1])

    bp_grad = sp.compute_grad(theta, data, layer_size[ind-1], layer_size[ind],
                              0.0001, 0.01, 3, dpark_ctx)

    num_grad = co.compute_numerical_grad(sp.compute_cost, theta, data,
                                         layer_size[ind-1], layer_size[ind],
                                         0.0001, 0.01, 3, dpark_ctx)

    diff = np.linalg.norm(bp_grad - num_grad) /\
        np.linalg.norm(num_grad + bp_grad)

    printdiff.append(diff)

    W = theta[:layer_size[ind]*layer_size[ind-1]].\
        reshape(layer_size[ind], layer_size[ind-1])

    data = np.dot(W, data)
Example #4
0
layer_ind = range(len(hidden_size) + 1)
layer_ind.remove(0)
layer_size = [visible_size] + hidden_size

dpark_ctx = DparkContext()

printdiff = []

start = clock()
# Debugging!
for ind in layer_ind:

    theta = sp.initial_parameter(layer_size[ind], layer_size[ind - 1])

    bp_grad = sp.compute_grad(theta, data, layer_size[ind - 1],
                              layer_size[ind], 0.0001, 0.01, 3, dpark_ctx)

    num_grad = co.compute_numerical_grad(sp.compute_cost, theta, data,
                                         layer_size[ind - 1], layer_size[ind],
                                         0.0001, 0.01, 3, dpark_ctx)

    diff = np.linalg.norm(bp_grad - num_grad) /\
        np.linalg.norm(num_grad + bp_grad)

    printdiff.append(diff)

    W = theta[:layer_size[ind]*layer_size[ind-1]].\
        reshape(layer_size[ind], layer_size[ind-1])

    data = np.dot(W, data)
Example #5
0
import compute_numerical_grad as co

# Randomly generate theta and data
visible_size = 32
hidden_size = [16, 8, 4]
data = np.random.rand(32, 100)

layer_ind = range(len(hidden_size) + 1)
layer_ind.remove(0)
layer_size = [visible_size] + hidden_size

# Debugging!
for ind in layer_ind:

    theta = sp.initial_parameter(layer_size[ind], layer_size[ind - 1])

    bp_grad = sp.compute_grad(theta, data, layer_size[ind-1], layer_size[ind])

    num_grad = co.compute_numerical_grad(sp.compute_cost, theta, data,
                                         layer_size[ind-1], layer_size[ind])

    diff = np.linalg.norm(bp_grad - num_grad) /\
        np.linalg.norm(num_grad + bp_grad)

    print str(diff) + " should be less than 1e-9! Is it?"

    W = theta[:layer_size[ind]*layer_size[ind-1]].\
        reshape(layer_size[ind], layer_size[ind-1])

    data = np.dot(W, data)
Example #6
0
# is highly recommended to check precisely.

# The diff, given at the last, represent the difference between
# the gradients from BP and numerical computing. It should be
# very small, otherwise the autoencoder was not trained correctly.

##==========================================================

import numpy as np
import sparse_autoencoder as sp
import compute_numerical_grad as co

data = np.random.rand(64, 100)
visible_size = 64
hidden_size = [16, 8, 4]
#hidden_size = [25]
theta = sp.initial_parameter(hidden_size, visible_size)

bp_cost = sp.compute_cost(theta, data, visible_size, hidden_size)
print bp_cost

bp_grad = sp.compute_grad(theta, data, visible_size, hidden_size)
print bp_grad

num_grad = co.compute_numerical_grad(sp.compute_cost, theta, data,
                                     visible_size, hidden_size)

diff = np.linalg.norm(bp_grad - num_grad) / np.linalg.norm(num_grad + bp_grad)

print str(diff) + " should be less than 1e-9! Is it?"