def parity(B=12, learning_rate=10e-5, epochs=200):
X, Y = all_parity_pairs(B)
N, t = X.shape
# we want every time step to have a label
Y_t = np.zeros(X.shape, dtype=np.int32)
for n in xrange(N):
ones_count = 0
for i in xrange(t):
if X[n, i] == 1:
ones_count += 1
if ones_count % 2 == 1:
Y_t[n, i] = 1
# for x, y in zip(X, Y_t):
# print "x:", x, "y:", y
X = X.reshape(N, t, 1).astype(np.float32)
rnn = SimpleRNN(4)
rnn.fit(X,
Y_t,
learning_rate=learning_rate,
epochs=epochs,
activation=T.nnet.sigmoid,
show_fig=True)
def deep16():
X, Y = all_parity_pairs(12)
model = ANN([128] * 16)
model.fit(X,
Y,
learning_rate=10e-3,
print_period=10,
epochs=35,
show_fig=True)
def wide(): # one inner layer model
X, Y = all_parity_pairs(12)
model = ANN([2048])
model.fit(X,
Y,
learning_rate=1e-4,
print_period=10,
epochs=300,
show_fig=True)
def deep32():
X, Y = all_parity_pairs(12)
model = ANN([64] * 32)
model.fit(X,
Y,
learning_rate=10e-3,
print_period=10,
epochs=25,
show_fig=True)
def deep64():
X, Y = all_parity_pairs(12)
model = ANN([32] * 64)
model.fit(X,
Y,
learning_rate=10e-3,
print_period=10,
epochs=20,
show_fig=True)
def deep():
X, Y = all_parity_pairs(12)
model = ANN([1024] * 2)
model.fit(X,
Y,
learning_rate=10e-4,
print_period=10,
epochs=1000,
show_fig=True)
def wide():
X, Y = all_parity_pairs(12)
model = ANN([2048])
model.fit(X,
Y,
learning_rate=10e-5,
print_period=10,
epochs=300,
show_fig=True)
def deep():
X, Y = all_parity_pairs(12)
ann = AnnTheano([1024] * 2)
ann.fit(X,
Y,
learning_rate=10e-4,
print_period=10,
epochs=70,
show_fig=True)
def wide():
X, Y = all_parity_pairs(12)
ann = AnnTheano([2048])
ann.fit(X,
Y,
learning_rate=10e-5,
print_period=10,
epochs=300,
show_fig=True)
def deep8():
X, Y = all_parity_pairs(12)
model = ANN([256] * 8)
model.fit(X,
Y,
learning_rate=10e-2,
print_period=10,
epochs=35,
show_fig=True)
def deep4():
X, Y = all_parity_pairs(12)
model = ANN([512] * 4)
model.fit(X,
Y,
learning_rate=10e-2,
print_period=10,
epochs=50,
show_fig=True)
def deep():
# Challenge - find a deeper, slimmer network to solve the problem
X, Y = all_parity_pairs(12)
model = ANN([1024] * 2)
model.fit(X,
Y,
learning_rate=10e-4,
print_period=10,
epochs=100,
show_fig=True)
def main(): X, Y = all_parity_pairs(12) X = X.astype(np.float32) Y = Y.astype(np.int32) K = len(set(Y)) Y = (indicator(Y)).astype(np.float32) N, D = X.shape model = ANN(D, K, [1024, 1024]) init = tf.global_variables_initializer() with tf.Session() as session: session.run(init) model.set_session(session) model.fit(X, Y)
def parity(B=12, learning_rate=10e-5, epochs=200):
X, Y = all_parity_pairs(B)
N, t = X.shape
Y_t = np.zeros(X.shape, dtype=np.int32)
for n in xrange(N):
ones_count = 0
for i in xrange(t):
if X[n,i] == 1:
ones_count += 1
if ones_count % 2 == 1:
Y_t[n,i] = 1
X = X.reshape(N, t, 1).astype(np.float32)
rnn = SimpleRNN(4)
rnn.fit(X, Y_t, learning_rate=learning_rate, epochs=epochs, activation=T.nnet.sigmoid, show_fig=True)
def parity(B=12, learning_rate=10e-5, epochs=200): X, Y = all_parity_pairs(B) N, t = X.shape Y_t = np.zeros(X.shape, dtype=np.int32) for n in range(N): ones_count = 0 for i in range(t): if X[n,i] == 1: ones_count += 1 if ones_count % 2 == 1: Y_t[n,i] = 1 X = X.reshape(N, T, 1).astype(np.float32) rnn = SimpleRNN(4) rnn.fit(X, Y_t, learning_rate=learning_rate, epochs=epochs, activation=T.nnet.sigmoid, show_fig=True)
def parity(B=12, learning_rate=10e-5, epochs=200):
X, Y = all_parity_pairs(B)
N, t = X.shape
# we want every time step to have a label
Y_t = np.zeros(X.shape, dtype=np.int32)
for n in xrange(N):
ones_count = 0
for i in xrange(t):
if X[n,i] == 1:
ones_count += 1
if ones_count % 2 == 1:
Y_t[n,i] = 1
# for x, y in zip(X, Y_t):
# print "x:", x, "y:", y
X = X.reshape(N, t, 1).astype(np.float32)
rnn = SimpleRNN(4)
rnn.fit(X, Y_t, learning_rate=learning_rate, epochs=epochs, activation=T.nnet.sigmoid, show_fig=True)
def parity(B=4, learning_rate=1e-5, epochs=2):
#X, Y_t = all_parity_pairs_with_sequence_labels(B)
X, Y = all_parity_pairs(B)
N, t = X.shape
Y_t = np.zeros(X.shape, dtype=np.int32)
for n in range(N):
ones_count = 0
for i in range(t):
if X[n, i] == 1:
ones_count += 1
if ones_count % 2 == 1:
Y_t[n, i] = 1
X = X.reshape(N, t, 1).astype(np.float32)
#rnn = SimpleRNN(20)
rnn = SimpleRNN(4)
rnn.fit(X, Y_t, learning_rate=learning_rate, epochs=epochs, activation=T.nnet.relu, show_fig=False)
def deep(): X, Y = all_parity_pairs(12) model = ANN([1024]*2) model.fit(X, Y, learning_rate=10e-4, print_period=10, epochs=100, show_fig=True)
def wide():
X, Y = all_parity_pairs(12)
model = ANN([2048])
def deep():
# Challenge - find a deeper, slimmer network to solve the problem
X, Y = all_parity_pairs(12)
model = ANN([1024]*2)
def wide():
X, Y = all_parity_pairs(12)
model = ANN([2048])
model.fit(X, Y, learning_rate=1e-4, print_period=10, epochs=300, show_fig=True)
def deep():
# Challenge - find a deeper, slimmer network to solve the problem
X, Y = all_parity_pairs(12)
model = ANN([1024]*2)
model.fit(X, Y, learning_rate=1e-3, print_period=10, epochs=100, show_fig=True)