def main():
train_X = np.asarray([[0.2, -0.3], [0.1, -0.9], [0.3, 0.5]])
train_Y = np.asarray([[0.0, 1.0, 0.0], [0.0, 0.0, 1.0], [1.0, 0.0, 0.0]])
# For comparison
net = ParticleDipoleNetwork(cost="mse", particle_input=ParticleDipoleInput(2))
net.append(ParticleDipole(2, 5, activation="sigmoid"))
net.append(ParticleDipole(5, 3, activation="sigmoid"))
print(net.predict(train_X))
print(net.cost(train_X, train_Y))
net2 = ParticleDipoleTreeNetwork(cost="mse", particle_input=ParticleDipoleTreeInput(2))
net2.append(ParticleDipoleTree(2, 5, activation="sigmoid"))
net2.append(ParticleDipoleTree(5, 3, activation="sigmoid"))
# Make sure we have the same coordinates and charges
net2.particle_input.copy_pos_neg_positions(net.particle_input.rx_pos, net.particle_input.ry_pos, net.particle_input.rz_pos,
net.particle_input.rx_neg, net.particle_input.ry_neg, net.particle_input.rz_neg)
for l in range(len(net.layers)):
net2.layers[l].copy_pos_neg_positions(net.layers[l].q, net.layers[l].b,
net.layers[l].rx_pos, net.layers[l].ry_pos, net.layers[l].rz_pos,
net.layers[l].rx_neg, net.layers[l].ry_neg, net.layers[l].rz_neg)
print(net2.predict(train_X))
print(net2.cost(train_X, train_Y))
def main():
train_X = np.asarray([[0.2, -0.3], [0.1, -0.9]])
train_Y = np.asarray([[0.0, 1.0, 0.0], [0.0, 0.0, 1.0]])
net = ParticleDipoleNetwork(cost="mse", particle_input=ParticleDipoleInput(2))
net.append(ParticleDipole(2, 5, activation="sigmoid", k_eq=0.1))
net.append(ParticleDipole(5, 3, activation="sigmoid", k_eq=0.1))
print(net.predict(train_X))
print(net.cost(train_X, train_Y))
def main():
train_X = np.asarray([[0.2, -0.3], [0.1, -0.9], [0.3, 0.5]])
train_Y = np.asarray([[0.0, 1.0, 0.0], [0.0, 0.0, 1.0], [1.0, 0.0, 0.0]])
# For comparison
net = ParticleDipoleNetwork(cost="mse",
particle_input=ParticleDipoleInput(2))
net.append(ParticleDipole(2, 5, activation="sigmoid"))
net.append(ParticleDipole(5, 3, activation="sigmoid"))
print(net.predict(train_X))
print(net.cost(train_X, train_Y))
net2 = ParticleDipoleTreeNetwork(cost="mse",
particle_input=ParticleDipoleTreeInput(2))
net2.append(ParticleDipoleTree(2, 5, activation="sigmoid"))
net2.append(ParticleDipoleTree(5, 3, activation="sigmoid"))
# Make sure we have the same coordinates and charges
net2.particle_input.copy_pos_neg_positions(net.particle_input.rx_pos,
net.particle_input.ry_pos,
net.particle_input.rz_pos,
net.particle_input.rx_neg,
net.particle_input.ry_neg,
net.particle_input.rz_neg)
for l in range(len(net.layers)):
net2.layers[l].copy_pos_neg_positions(
net.layers[l].q, net.layers[l].b, net.layers[l].rx_pos,
net.layers[l].ry_pos, net.layers[l].rz_pos, net.layers[l].rx_neg,
net.layers[l].ry_neg, net.layers[l].rz_neg)
print(net2.predict(train_X))
print(net2.cost(train_X, train_Y))
def main():
train_X = np.asarray([[0.2, -0.3], [0.1, -0.9]])
train_Y = np.asarray([[0.0, 1.0, 0.0], [0.0, 0.0, 1.0]])
net = ParticleDipoleNetwork(cost="mse",
particle_input=ParticleDipoleInput(2))
net.append(ParticleDipole(2, 5, activation="sigmoid", k_eq=0.1))
net.append(ParticleDipole(5, 3, activation="sigmoid", k_eq=0.1))
print(net.predict(train_X))
print(net.cost(train_X, train_Y))
def fd():
train_X = np.asarray([[0.2, -0.3], [0.1, -0.9]])
train_Y = np.asarray([[0.0, 1.0, 0.0], [0.0, 0.0, 1.0]])
net = ParticleDipoleNetwork(
cost="categorical_cross_entropy",
particle_input=ParticleDipoleInput(2),
regularizer=ParticleDipoleRegularize(coeff_lambda=0.03))
net.append(
ParticleDipole(2, 5, activation="sigmoid", k_eq=0.1, k_bond=10.0))
net.append(
ParticleDipole(5, 6, activation="sigmoid", k_eq=0.1, k_bond=10.0))
net.append(
ParticleDipole(6, 3, activation="softmax", k_eq=0.1, k_bond=10.0))
db, dq, drx_pos, dry_pos, drz_pos, drx_neg, dry_neg, drz_neg = net.cost_gradient(
train_X, train_Y)
h = 0.00001
print("analytic b")
print(db)
fd_b = []
for l in range(len(net.layers)):
lb = []
for c in range(len(net.layers[l].b)):
for b in range(len(net.layers[l].b[c])):
orig = net.layers[l].b[c][b]
net.layers[l].b[c][b] += h
fp = net.cost(train_X, train_Y)
net.layers[l].b[c][b] -= 2 * h
fm = net.cost(train_X, train_Y)
lb.append((fp - fm) / (2 * h))
net.layers[l].b[c][b] = orig
fd_b.append(lb)
print("numerical b")
print(fd_b)
print("analytic q")
for x in dq:
print(x)
fd_q = []
for l in range(len(net.layers)):
lq = []
for i in range(len(net.layers[l].q)):
orig = net.layers[l].q[i]
net.layers[l].q[i] += h
fp = net.cost(train_X, train_Y)
net.layers[l].q[i] -= 2 * h
fm = net.cost(train_X, train_Y)
lq.append((fp - fm) / (2 * h))
net.layers[l].q[i] = orig
fd_q.append(lq)
print("numerical q")
for x in fd_q:
print(x)
print("analytic x pos")
for layer in drx_pos:
print(layer)
print("analytic y pos")
for layer in dry_pos:
print(layer)
print("analytic z pos")
for layer in drz_pos:
print(layer)
fd_r_x = []
fd_r_y = []
fd_r_z = []
# input first
layer = net.particle_input
lr_x = []
lr_y = []
lr_z = []
for i in range(layer.output_size):
# x
orig = layer.rx_pos[i]
layer.rx_pos[i] += h
fp = net.cost(train_X, train_Y)
layer.rx_pos[i] -= 2 * h
fm = net.cost(train_X, train_Y)
lr_x.append((fp - fm) / (2 * h))
layer.rx_pos[i] = orig
# y
orig = layer.ry_pos[i]
layer.ry_pos[i] += h
fp = net.cost(train_X, train_Y)
layer.ry_pos[i] -= 2 * h
fm = net.cost(train_X, train_Y)
lr_y.append((fp - fm) / (2 * h))
layer.ry_pos[i] = orig
# z
orig = layer.rz_pos[i]
layer.rz_pos[i] += h
fp = net.cost(train_X, train_Y)
layer.rz_pos[i] -= 2 * h
fm = net.cost(train_X, train_Y)
lr_z.append((fp - fm) / (2 * h))
layer.rz_pos[i] = orig
fd_r_x.append(lr_x)
fd_r_y.append(lr_y)
fd_r_z.append(lr_z)
# layers
for layer in net.layers:
lr_x = []
lr_y = []
lr_z = []
for i in range(layer.output_size):
# x
orig = layer.rx_pos[i]
layer.rx_pos[i] += h
fp = net.cost(train_X, train_Y)
layer.rx_pos[i] -= 2 * h
fm = net.cost(train_X, train_Y)
lr_x.append((fp - fm) / (2 * h))
layer.rx_pos[i] = orig
# y
orig = layer.ry_pos[i]
layer.ry_pos[i] += h
fp = net.cost(train_X, train_Y)
layer.ry_pos[i] -= 2 * h
fm = net.cost(train_X, train_Y)
lr_y.append((fp - fm) / (2 * h))
layer.ry_pos[i] = orig
# z
orig = layer.rz_pos[i]
layer.rz_pos[i] += h
fp = net.cost(train_X, train_Y)
layer.rz_pos[i] -= 2 * h
fm = net.cost(train_X, train_Y)
lr_z.append((fp - fm) / (2 * h))
layer.rz_pos[i] = orig
fd_r_x.append(lr_x)
fd_r_y.append(lr_y)
fd_r_z.append(lr_z)
print("numerical r x pos")
for f in fd_r_x:
print(f)
print("numerical r y pos")
for f in fd_r_y:
print(f)
print("numerical r z pos")
for f in fd_r_z:
print(f)
print("analytic x neg")
for layer in drx_neg:
print(layer)
print("analytic y neg")
for layer in dry_neg:
print(layer)
print("analytic z neg")
for layer in drz_neg:
print(layer)
fd_r_x = []
fd_r_y = []
fd_r_z = []
# input first
layer = net.particle_input
lr_x = []
lr_y = []
lr_z = []
for i in range(layer.output_size):
# x
orig = layer.rx_neg[i]
layer.rx_neg[i] += h
fp = net.cost(train_X, train_Y)
layer.rx_neg[i] -= 2 * h
fm = net.cost(train_X, train_Y)
lr_x.append((fp - fm) / (2 * h))
layer.rx_neg[i] = orig
# y
orig = layer.ry_neg[i]
layer.ry_neg[i] += h
fp = net.cost(train_X, train_Y)
layer.ry_neg[i] -= 2 * h
fm = net.cost(train_X, train_Y)
lr_y.append((fp - fm) / (2 * h))
layer.ry_neg[i] = orig
# z
orig = layer.rz_neg[i]
layer.rz_neg[i] += h
fp = net.cost(train_X, train_Y)
layer.rz_neg[i] -= 2 * h
fm = net.cost(train_X, train_Y)
lr_z.append((fp - fm) / (2 * h))
layer.rz_neg[i] = orig
fd_r_x.append(lr_x)
fd_r_y.append(lr_y)
fd_r_z.append(lr_z)
# layers
for layer in net.layers:
lr_x = []
lr_y = []
lr_z = []
for i in range(layer.output_size):
# x
orig = layer.rx_neg[i]
layer.rx_neg[i] += h
fp = net.cost(train_X, train_Y)
layer.rx_neg[i] -= 2 * h
fm = net.cost(train_X, train_Y)
lr_x.append((fp - fm) / (2 * h))
layer.rx_neg[i] = orig
# y
orig = layer.ry_neg[i]
layer.ry_neg[i] += h
fp = net.cost(train_X, train_Y)
layer.ry_neg[i] -= 2 * h
fm = net.cost(train_X, train_Y)
lr_y.append((fp - fm) / (2 * h))
layer.ry_neg[i] = orig
# z
orig = layer.rz_neg[i]
layer.rz_neg[i] += h
fp = net.cost(train_X, train_Y)
layer.rz_neg[i] -= 2 * h
fm = net.cost(train_X, train_Y)
lr_z.append((fp - fm) / (2 * h))
layer.rz_neg[i] = orig
fd_r_x.append(lr_x)
fd_r_y.append(lr_y)
fd_r_z.append(lr_z)
print("numerical r x neg")
for f in fd_r_x:
print(f)
print("numerical r y neg")
for f in fd_r_y:
print(f)
print("numerical r z neg")
for f in fd_r_z:
print(f)
def fd():
train_X = np.asarray([[0.2, -0.3], [0.1, -0.9]])
train_Y = np.asarray([[0.0, 1.0, 0.0], [0.0, 0.0, 1.0]])
net = ParticleDipoleNetwork(cost="categorical_cross_entropy", particle_input=ParticleDipoleInput(2),
regularizer=ParticleDipoleRegularize(coeff_lambda=0.03))
net.append(ParticleDipole(2, 5, activation="sigmoid", k_eq=0.1, k_bond=10.0))
net.append(ParticleDipole(5, 6, activation="sigmoid", k_eq=0.1, k_bond=10.0))
net.append(ParticleDipole(6, 3, activation="softmax", k_eq=0.1, k_bond=10.0))
db, dq, drx_pos, dry_pos, drz_pos, drx_neg, dry_neg, drz_neg = net.cost_gradient(train_X, train_Y)
h = 0.00001
print("analytic b")
print(db)
fd_b = []
for l in range(len(net.layers)):
lb = []
for c in range(len(net.layers[l].b)):
for b in range(len(net.layers[l].b[c])):
orig = net.layers[l].b[c][b]
net.layers[l].b[c][b] += h
fp = net.cost(train_X, train_Y)
net.layers[l].b[c][b] -= 2*h
fm = net.cost(train_X, train_Y)
lb.append((fp - fm) / (2*h))
net.layers[l].b[c][b] = orig
fd_b.append(lb)
print("numerical b")
print(fd_b)
print("analytic q")
for x in dq:
print(x)
fd_q = []
for l in range(len(net.layers)):
lq = []
for i in range(len(net.layers[l].q)):
orig = net.layers[l].q[i]
net.layers[l].q[i] += h
fp = net.cost(train_X, train_Y)
net.layers[l].q[i] -= 2*h
fm = net.cost(train_X, train_Y)
lq.append((fp - fm) / (2*h))
net.layers[l].q[i] = orig
fd_q.append(lq)
print("numerical q")
for x in fd_q:
print(x)
print("analytic x pos")
for layer in drx_pos:
print(layer)
print("analytic y pos")
for layer in dry_pos:
print(layer)
print("analytic z pos")
for layer in drz_pos:
print(layer)
fd_r_x = []
fd_r_y = []
fd_r_z = []
# input first
layer = net.particle_input
lr_x = []
lr_y = []
lr_z = []
for i in range(layer.output_size):
# x
orig = layer.rx_pos[i]
layer.rx_pos[i] += h
fp = net.cost(train_X, train_Y)
layer.rx_pos[i] -= 2*h
fm = net.cost(train_X, train_Y)
lr_x.append((fp - fm) / (2*h))
layer.rx_pos[i] = orig
# y
orig = layer.ry_pos[i]
layer.ry_pos[i] += h
fp = net.cost(train_X, train_Y)
layer.ry_pos[i] -= 2*h
fm = net.cost(train_X, train_Y)
lr_y.append((fp - fm) / (2*h))
layer.ry_pos[i] = orig
# z
orig = layer.rz_pos[i]
layer.rz_pos[i] += h
fp = net.cost(train_X, train_Y)
layer.rz_pos[i] -= 2*h
fm = net.cost(train_X, train_Y)
lr_z.append((fp - fm) / (2*h))
layer.rz_pos[i] = orig
fd_r_x.append(lr_x)
fd_r_y.append(lr_y)
fd_r_z.append(lr_z)
# layers
for layer in net.layers:
lr_x = []
lr_y = []
lr_z = []
for i in range(layer.output_size):
# x
orig = layer.rx_pos[i]
layer.rx_pos[i] += h
fp = net.cost(train_X, train_Y)
layer.rx_pos[i] -= 2*h
fm = net.cost(train_X, train_Y)
lr_x.append((fp - fm) / (2*h))
layer.rx_pos[i] = orig
# y
orig = layer.ry_pos[i]
layer.ry_pos[i] += h
fp = net.cost(train_X, train_Y)
layer.ry_pos[i] -= 2*h
fm = net.cost(train_X, train_Y)
lr_y.append((fp - fm) / (2*h))
layer.ry_pos[i] = orig
# z
orig = layer.rz_pos[i]
layer.rz_pos[i] += h
fp = net.cost(train_X, train_Y)
layer.rz_pos[i] -= 2*h
fm = net.cost(train_X, train_Y)
lr_z.append((fp - fm) / (2*h))
layer.rz_pos[i] = orig
fd_r_x.append(lr_x)
fd_r_y.append(lr_y)
fd_r_z.append(lr_z)
print("numerical r x pos")
for f in fd_r_x:
print(f)
print("numerical r y pos")
for f in fd_r_y:
print(f)
print("numerical r z pos")
for f in fd_r_z:
print(f)
print("analytic x neg")
for layer in drx_neg:
print(layer)
print("analytic y neg")
for layer in dry_neg:
print(layer)
print("analytic z neg")
for layer in drz_neg:
print(layer)
fd_r_x = []
fd_r_y = []
fd_r_z = []
# input first
layer = net.particle_input
lr_x = []
lr_y = []
lr_z = []
for i in range(layer.output_size):
# x
orig = layer.rx_neg[i]
layer.rx_neg[i] += h
fp = net.cost(train_X, train_Y)
layer.rx_neg[i] -= 2*h
fm = net.cost(train_X, train_Y)
lr_x.append((fp - fm) / (2*h))
layer.rx_neg[i] = orig
# y
orig = layer.ry_neg[i]
layer.ry_neg[i] += h
fp = net.cost(train_X, train_Y)
layer.ry_neg[i] -= 2*h
fm = net.cost(train_X, train_Y)
lr_y.append((fp - fm) / (2*h))
layer.ry_neg[i] = orig
# z
orig = layer.rz_neg[i]
layer.rz_neg[i] += h
fp = net.cost(train_X, train_Y)
layer.rz_neg[i] -= 2*h
fm = net.cost(train_X, train_Y)
lr_z.append((fp - fm) / (2*h))
layer.rz_neg[i] = orig
fd_r_x.append(lr_x)
fd_r_y.append(lr_y)
fd_r_z.append(lr_z)
# layers
for layer in net.layers:
lr_x = []
lr_y = []
lr_z = []
for i in range(layer.output_size):
# x
orig = layer.rx_neg[i]
layer.rx_neg[i] += h
fp = net.cost(train_X, train_Y)
layer.rx_neg[i] -= 2*h
fm = net.cost(train_X, train_Y)
lr_x.append((fp - fm) / (2*h))
layer.rx_neg[i] = orig
# y
orig = layer.ry_neg[i]
layer.ry_neg[i] += h
fp = net.cost(train_X, train_Y)
layer.ry_neg[i] -= 2*h
fm = net.cost(train_X, train_Y)
lr_y.append((fp - fm) / (2*h))
layer.ry_neg[i] = orig
# z
orig = layer.rz_neg[i]
layer.rz_neg[i] += h
fp = net.cost(train_X, train_Y)
layer.rz_neg[i] -= 2*h
fm = net.cost(train_X, train_Y)
lr_z.append((fp - fm) / (2*h))
layer.rz_neg[i] = orig
fd_r_x.append(lr_x)
fd_r_y.append(lr_y)
fd_r_z.append(lr_z)
print("numerical r x neg")
for f in fd_r_x:
print(f)
print("numerical r y neg")
for f in fd_r_y:
print(f)
print("numerical r z neg")
for f in fd_r_z:
print(f)
Y.append(y)
Y = np.asarray(Y)
# Data subset
n_sub = 333
X_sub = X[:n_sub, :]
Y_sub = Y[:n_sub, :]
s = 4.0
cut = 3.0
max_level = 5
mac = 0.0
n = 128
n_min = 10
net = ParticleDipoleNetwork(cost="categorical_cross_entropy",
particle_input=ParticleDipoleInput(784, s=s))
net.append(ParticleDipole(784, n, activation="sigmoid", s=s))
net.append(ParticleDipole(n, 10, activation="softmax", s=s))
print("starting predict")
times = []
nt = 3
for _ in range(nt):
ts = time.time()
c = net.cost(X_sub, Y_sub)
# c = 1.0
# net.cost_gradient(X_sub, Y_sub)
t = time.time() - ts
print("Cost: {} time: {}".format(c, t))
times.append(t)
print("Mean: " + str(sum(times) / nt))
Y.append(y)
Y = np.asarray(Y)
# Data subset
n_sub = 333
X_sub = X[:n_sub, :]
Y_sub = Y[:n_sub, :]
s = 4.0
cut = 3.0
max_level = 5
mac = 0.0
n = 128
n_min = 10
net = ParticleDipoleNetwork(cost="categorical_cross_entropy", particle_input=ParticleDipoleInput(784, s=s))
net.append(ParticleDipole(784, n, activation="sigmoid", s=s))
net.append(ParticleDipole(n, 10, activation="softmax", s=s))
print("starting predict")
times = []
nt = 3
for _ in range(nt):
ts = time.time()
c = net.cost(X_sub, Y_sub)
# c = 1.0
# net.cost_gradient(X_sub, Y_sub)
t = time.time() - ts
print("Cost: {} time: {}".format(c, t))
times.append(t)
print("Mean: " + str(sum(times) / nt))