def run():
train_images, train_labels, _, _, _ = load_data(normalize=True)
train_images = train_images[:N_real_data, :]
train_labels = train_labels[:N_real_data, :]
batch_idxs = BatchList(N_fake_data, batch_size)
parser, _, loss_fun, frac_err = make_nn_funs(layer_sizes, L2_reg, return_parser=True)
N_weights = parser.N
fake_data = npr.randn(*(train_images[:N_fake_data, :].shape)) * init_fake_data_scale
fake_labels = one_hot(np.array(range(N_fake_data)) % N_classes, N_classes) # One of each.
def indexed_loss_fun(x, meta_params, idxs): # To be optimized by SGD.
return loss_fun(x, X=meta_params[idxs], T=fake_labels[idxs])
def meta_loss_fun(x): # To be optimized in the outer loop.
return loss_fun(x, X=train_images, T=train_labels)
log_alphas = np.full(N_iters, log_alpha_0)
betas = np.full(N_iters, beta_0)
npr.seed(0)
v0 = npr.randn(N_weights) * velocity_scale
x0 = npr.randn(N_weights) * np.exp(log_param_scale)
output = []
for i in range(N_meta_iter):
results = sgd2(indexed_loss_fun, meta_loss_fun, batch_idxs, N_iters,
x0, v0, np.exp(log_alphas), betas, fake_data)
learning_curve = results['learning_curve']
validation_loss = results['M_final']
output.append((learning_curve, validation_loss, fake_data))
fake_data -= results['dMd_meta'] * data_stepsize # Update data with one gradient step.
print "Meta iteration {0} Valiation loss {1}".format(i, validation_loss)
return output
def run():
(train_images, train_labels), (val_images, val_labels), (test_images, test_labels) \
= load_data_subset(N_train_data, N_val_data, N_test_data)
batch_idxs = BatchList(N_train_data, batch_size)
parser, _, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weights = parser.N
hyperparser = WeightsParser()
hyperparser.add_weights('log_L2_reg', (N_weights, ))
metas = np.zeros(hyperparser.N)
print "Number of hyperparameters to be trained:", hyperparser.N
npr.seed(0)
hyperparser.set(metas, 'log_L2_reg', log_L2_reg_scale + np.ones(N_weights))
def indexed_loss_fun(x, meta_params, idxs): # To be optimized by SGD.
L2_reg = np.exp(hyperparser.get(meta_params, 'log_L2_reg'))
return loss_fun(x,
X=train_images[idxs],
T=train_labels[idxs],
L2_reg=L2_reg)
def meta_loss_fun(x, meta_params): # To be optimized in the outer loop.
L2_reg = np.exp(hyperparser.get(meta_params, 'log_L2_reg'))
log_prior = -meta_L2_reg * np.dot(L2_reg.ravel(), L2_reg.ravel())
return loss_fun(x, X=val_images, T=val_labels) - log_prior
def test_loss_fun(x): # To measure actual performance.
return loss_fun(x, X=test_images, T=test_labels)
log_alphas = np.full(N_iters, log_alpha_0)
betas = np.full(N_iters, beta_0)
v0 = npr.randn(N_weights) * velocity_scale
x0 = npr.randn(N_weights) * np.exp(log_param_scale)
output = []
for i in range(N_meta_iter):
results = sgd2(indexed_loss_fun, meta_loss_fun, batch_idxs, N_iters,
x0, v0, np.exp(log_alphas), betas, metas)
learning_curve = results['learning_curve']
validation_loss = results['M_final']
test_loss = test_loss_fun(results['x_final'])
weightparser = parser.new_vect(results['x_final'])
l2parser = parser.new_vect(np.exp(hyperparser.get(metas,
'log_L2_reg')))
output.append((learning_curve, validation_loss, test_loss,
weightparser[('weights', 0)], l2parser[('weights', 0)]))
metas -= results['dMd_meta'] * meta_stepsize
print "Meta iteration {0} Valiation loss {1} Test loss {2}"\
.format(i, validation_loss, test_loss)
return output
def run():
(train_images, train_labels), (val_images, val_labels), (test_images, test_labels) = load_data_subset(
N_train_data, N_val_data, N_test_data
)
batch_idxs = BatchList(N_train_data, batch_size)
parser, _, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weights = parser.N
hyperparser = WeightsParser()
hyperparser.add_weights("log_L2_reg", (N_weights,))
metas = np.zeros(hyperparser.N)
print "Number of hyperparameters to be trained:", hyperparser.N
npr.seed(0)
hyperparser.set(metas, "log_L2_reg", log_L2_reg_scale + np.ones(N_weights))
def indexed_loss_fun(x, meta_params, idxs): # To be optimized by SGD.
L2_reg = np.exp(hyperparser.get(meta_params, "log_L2_reg"))
return loss_fun(x, X=train_images[idxs], T=train_labels[idxs], L2_reg=L2_reg)
def meta_loss_fun(x, meta_params): # To be optimized in the outer loop.
L2_reg = np.exp(hyperparser.get(meta_params, "log_L2_reg"))
log_prior = -meta_L2_reg * np.dot(L2_reg.ravel(), L2_reg.ravel())
return loss_fun(x, X=val_images, T=val_labels) - log_prior
def test_loss_fun(x): # To measure actual performance.
return loss_fun(x, X=test_images, T=test_labels)
log_alphas = np.full(N_iters, log_alpha_0)
betas = np.full(N_iters, beta_0)
v0 = npr.randn(N_weights) * velocity_scale
x0 = npr.randn(N_weights) * np.exp(log_param_scale)
output = []
for i in range(N_meta_iter):
results = sgd2(indexed_loss_fun, meta_loss_fun, batch_idxs, N_iters, x0, v0, np.exp(log_alphas), betas, metas)
learning_curve = results["learning_curve"]
validation_loss = results["M_final"]
test_loss = test_loss_fun(results["x_final"])
output.append(
(
learning_curve,
validation_loss,
test_loss,
parser.get(results["x_final"], (("weights", 0))),
parser.get(np.exp(hyperparser.get(metas, "log_L2_reg")), (("weights", 0))),
)
)
metas -= results["dMd_meta"] * meta_stepsize
print "Meta iteration {0} Valiation loss {1} Test loss {2}".format(i, validation_loss, test_loss)
return output
def test_sgd2():
N_weights = 5
W0 = 0.1 * npr.randn(N_weights)
V0 = 0.1 * npr.randn(N_weights)
N_data = 12
batch_size = 4
num_epochs = 3
batch_idxs = BatchList(N_data, batch_size)
N_iter = num_epochs * len(batch_idxs)
alphas = 0.1 * npr.rand(len(batch_idxs) * num_epochs)
betas = 0.5 + 0.2 * npr.rand(len(batch_idxs) * num_epochs)
meta = 0.1 * npr.randn(N_weights * 2)
A = npr.randn(N_data, N_weights)
def loss_fun(W, meta, idxs):
sub_A = A[idxs, :]
return np.dot(
np.dot(W + meta[:N_weights] + meta[N_weights:],
np.dot(sub_A.T, sub_A)), W)
def meta_loss_fun(w, meta):
return np.dot(w, w) + np.dot(meta, meta)
def full_loss(W0, V0, alphas, betas, meta):
result = sgd2(loss_fun, meta_loss_fun, batch_idxs, N_iter, W0, V0,
alphas, betas, meta)
return result['L_final']
def meta_loss(W0, V0, alphas, betas, meta):
result = sgd2(loss_fun, meta_loss_fun, batch_idxs, N_iter, W0, V0,
alphas, betas, meta)
return result['M_final']
result = sgd2(loss_fun, meta_loss_fun, batch_idxs, N_iter, W0, V0, alphas,
betas, meta)
d_an = (result['dLd_x'], result['dLd_v'], result['dLd_alphas'],
result['dLd_betas'], result['dLd_meta'])
d_num = nd(full_loss, W0, V0, alphas, betas, meta)
for i, (an, num) in enumerate(zip(d_an, d_num)):
assert np.allclose(an, num, rtol=1e-3, atol=1e-4), \
"Type {0}, diffs are: {1}".format(i, an - num)
print "Type {0}, diffs are: {1}".format(i, an - num)
d_an = (result['dMd_x'], result['dMd_v'], result['dMd_alphas'],
result['dMd_betas'], result['dMd_meta'])
d_num = nd(meta_loss, W0, V0, alphas, betas, meta)
for i, (an, num) in enumerate(zip(d_an, d_num)):
assert np.allclose(an, num, rtol=1e-3, atol=1e-4), \
"Type {0}, diffs are: {1}".format(i, an - num)
print "Type {0}, diffs are: {1}".format(i, an - num)
def run():
val_images, val_labels, test_images, test_labels, _ = load_data(
normalize=True)
val_images = val_images[:N_val_data, :]
val_labels = val_labels[:N_val_data, :]
truedatasize = np.std(val_images)
test_images = test_images[:N_test_data, :]
test_labels = test_labels[:N_test_data, :]
batch_idxs = BatchList(N_fake_data, batch_size)
parser, _, loss_fun, frac_err = make_nn_funs(layer_sizes,
L2_reg,
return_parser=True)
N_weights = parser.N
fake_data = npr.randn(
*(val_images[:N_fake_data, :].shape)) * init_fake_data_scale
fake_labels = one_hot(np.array(range(N_fake_data)) % N_classes,
N_classes) # One of each.
def indexed_loss_fun(x, meta_params, idxs): # To be optimized by SGD.
return loss_fun(x, X=meta_params[idxs], T=fake_labels[idxs])
def meta_loss_fun(x, meta_params): # To be optimized in the outer loop.
log_prior = -fake_data_L2_reg * np.dot(meta_params.ravel(),
meta_params.ravel())
return loss_fun(x, X=val_images, T=val_labels) - log_prior
def test_loss_fun(x): # To measure actual performance.
return loss_fun(x, X=test_images, T=test_labels)
log_alphas = np.full(N_iters, log_alpha_0)
betas = np.full(N_iters, beta_0)
npr.seed(0)
v0 = npr.randn(N_weights) * velocity_scale
x0 = npr.randn(N_weights) * np.exp(log_param_scale)
output = []
for i in range(N_meta_iter):
results = sgd2(indexed_loss_fun, meta_loss_fun, batch_idxs, N_iters,
x0, v0, np.exp(log_alphas), betas, fake_data)
learning_curve = results['learning_curve']
validation_loss = results['M_final']
fakedatasize = np.std(fake_data) / truedatasize
test_loss = test_loss_fun(results['x_final'])
output.append((learning_curve, validation_loss, test_loss, fake_data,
fakedatasize))
fake_data -= results[
'dMd_meta'] * data_stepsize # Update data with one gradient step.
print "Meta iteration {0} Valiation loss {1} Test loss {2}"\
.format(i, validation_loss, test_loss)
return output
def test_sgd2():
N_weights = 5
W0 = 0.1 * npr.randn(N_weights)
V0 = 0.1 * npr.randn(N_weights)
N_data = 12
batch_size = 4
num_epochs = 3
batch_idxs = BatchList(N_data, batch_size)
N_iter = num_epochs * len(batch_idxs)
alphas = 0.1 * npr.rand(len(batch_idxs) * num_epochs)
betas = 0.5 + 0.2 * npr.rand(len(batch_idxs) * num_epochs)
meta = 0.1 * npr.randn(N_weights * 2)
A = npr.randn(N_data, N_weights)
def loss_fun(W, meta, idxs):
sub_A = A[idxs, :]
return np.dot(np.dot(W + meta[:N_weights] + meta[N_weights:], np.dot(sub_A.T, sub_A)), W)
def meta_loss_fun(w, meta):
return np.dot(w, w) + np.dot(meta, meta)
def full_loss(W0, V0, alphas, betas, meta):
result = sgd2(loss_fun, meta_loss_fun, batch_idxs, N_iter, W0, V0, alphas, betas, meta)
return result['L_final']
def meta_loss(W0, V0, alphas, betas, meta):
result = sgd2(loss_fun, meta_loss_fun, batch_idxs, N_iter, W0, V0, alphas, betas, meta)
return result['M_final']
result = sgd2(loss_fun, meta_loss_fun, batch_idxs, N_iter, W0, V0, alphas, betas, meta)
d_an = (result['dLd_x'], result['dLd_v'], result['dLd_alphas'], result['dLd_betas'], result['dLd_meta'])
d_num = nd(full_loss, W0, V0, alphas, betas, meta)
for i, (an, num) in enumerate(zip(d_an, d_num)):
assert np.allclose(an, num, rtol=1e-3, atol=1e-4), \
"Type {0}, diffs are: {1}".format(i, an - num)
print "Type {0}, diffs are: {1}".format(i, an - num)
d_an = (result['dMd_x'], result['dMd_v'], result['dMd_alphas'], result['dMd_betas'], result['dMd_meta'])
d_num = nd(meta_loss, W0, V0, alphas, betas, meta)
for i, (an, num) in enumerate(zip(d_an, d_num)):
assert np.allclose(an, num, rtol=1e-3, atol=1e-4), \
"Type {0}, diffs are: {1}".format(i, an - num)
print "Type {0}, diffs are: {1}".format(i, an - num)
def run():
train_images, train_labels, _, _, _ = load_data(normalize=True)
train_images = train_images[:N_real_data, :]
train_labels = train_labels[:N_real_data, :]
batch_idxs = BatchList(N_fake_data, batch_size)
parser, _, loss_fun, frac_err = make_nn_funs(layer_sizes,
L2_reg,
return_parser=True)
N_weights = parser.N
#fake_data = npr.randn(*(train_images[:N_fake_data, :].shape))
fake_data = np.zeros(train_images[:N_fake_data, :].shape)
one_hot = lambda x, K: np.array(x[:, None] == np.arange(K)[None, :],
dtype=int)
fake_labels = one_hot(np.array(range(0, 10)), 10) # One of each label.
def indexed_loss_fun(x, meta_params, idxs): # To be optimized by SGD.
return loss_fun(x, X=meta_params[idxs], T=fake_labels[idxs])
def meta_loss_fun(x): # To be optimized in the outer loop.
return loss_fun(x, X=train_images, T=train_labels)
log_alphas = np.full(N_iters, log_alpha_0)
betas = np.full(N_iters, beta_0)
npr.seed(0)
v0 = npr.randn(N_weights) * velocity_scale
x0 = npr.randn(N_weights) * np.exp(log_param_scale)
output = []
for i in range(N_meta_iter):
print "Meta iteration {0}".format(i)
results = sgd2(indexed_loss_fun, meta_loss_fun, batch_idxs, N_iters,
x0, v0, np.exp(log_alphas), betas, fake_data)
learning_curve = results['learning_curve']
output.append((learning_curve, fake_data))
fake_data -= results[
'dMd_meta'] * data_stepsize # Update data with one gradient step.
return output
def run():
train_images, train_labels, _, _, _ = load_data(normalize=True)
train_images = train_images[:N_real_data, :]
train_labels = train_labels[:N_real_data, :]
batch_idxs = BatchList(N_fake_data, batch_size)
parser, _, loss_fun, frac_err = make_nn_funs(layer_sizes, L2_reg, return_parser=True)
N_weights = parser.N
# fake_data = npr.randn(*(train_images[:N_fake_data, :].shape))
fake_data = np.zeros(train_images[:N_fake_data, :].shape)
one_hot = lambda x, K: np.array(x[:, None] == np.arange(K)[None, :], dtype=int)
fake_labels = one_hot(np.array(range(0, 10)), 10) # One of each label.
def indexed_loss_fun(x, meta_params, idxs): # To be optimized by SGD.
return loss_fun(x, X=meta_params[idxs], T=fake_labels[idxs])
def meta_loss_fun(x): # To be optimized in the outer loop.
return loss_fun(x, X=train_images, T=train_labels)
log_alphas = np.full(N_iters, log_alpha_0)
betas = np.full(N_iters, beta_0)
npr.seed(0)
v0 = npr.randn(N_weights) * velocity_scale
x0 = npr.randn(N_weights) * np.exp(log_param_scale)
output = []
for i in range(N_meta_iter):
print "Meta iteration {0}".format(i)
results = sgd2(
indexed_loss_fun, meta_loss_fun, batch_idxs, N_iters, x0, v0, np.exp(log_alphas), betas, fake_data
)
learning_curve = results["learning_curve"]
output.append((learning_curve, fake_data))
fake_data -= results["dMd_meta"] * data_stepsize # Update data with one gradient step.
return output
def run():
val_images, val_labels, test_images, test_labels, _ = load_data(
normalize=True)
val_images = val_images[:N_val_data, :]
val_labels = val_labels[:N_val_data, :]
true_data_scale = np.std(val_images)
test_images = test_images[:N_test_data, :]
test_labels = test_labels[:N_test_data, :]
batch_idxs = BatchList(N_fake_data, batch_size)
parser, _, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weights = len(parser.vect)
npr.seed(0)
init_fake_data = npr.randn(
*(val_images[:N_fake_data, :].shape)) * init_fake_data_scale
one_hot = lambda x, K: np.array(x[:, None] == np.arange(K)[None, :],
dtype=int)
fake_labels = one_hot(np.array(range(N_fake_data)) % N_classes,
N_classes) # One of each.
hyperparser = WeightsParser()
hyperparser.add_weights('log_L2_reg', (1, ))
hyperparser.add_weights('fake_data', init_fake_data.shape)
metas = np.zeros(hyperparser.N)
print "Number of hyperparameters to be trained:", hyperparser.N
hyperparser.set(metas, 'log_L2_reg', init_log_L2_reg)
hyperparser.set(metas, 'fake_data', init_fake_data)
def indexed_loss_fun(x, meta_params, idxs): # To be optimized by SGD.
L2_reg = np.exp(hyperparser.get(meta_params, 'log_L2_reg')[0])
fake_data = hyperparser.get(meta_params, 'fake_data')
return loss_fun(x,
X=fake_data[idxs],
T=fake_labels[idxs],
L2_reg=L2_reg)
def meta_loss_fun(x, meta_params): # To be optimized in the outer loop.
fake_data = hyperparser.get(meta_params, 'fake_data')
log_prior = -fake_data_L2_reg * np.dot(fake_data.ravel(),
fake_data.ravel())
return loss_fun(x, X=val_images, T=val_labels) - log_prior
def test_loss_fun(x): # To measure actual performance.
return loss_fun(x, X=test_images, T=test_labels)
log_alphas = np.full(N_iters, log_alpha_0)
betas = np.full(N_iters, beta_0)
output = []
velocity = np.zeros(hyperparser.N)
for i in range(N_meta_iter):
print "L2 reg is ", np.exp(hyperparser.get(metas,
'log_L2_reg')[0]), "| ",
npr.seed(0)
v0 = npr.randn(N_weights) * velocity_scale
x0 = npr.randn(N_weights) * np.exp(log_param_scale)
results = sgd2(indexed_loss_fun, meta_loss_fun, batch_idxs, N_iters,
x0, v0, np.exp(log_alphas), betas, metas)
learning_curve = results['learning_curve']
validation_loss = results['M_final']
test_err = frac_err(results['x_final'], test_images, test_labels)
fake_data_scale = np.std(hyperparser.get(
metas, 'fake_data')) / true_data_scale
test_loss = test_loss_fun(results['x_final'])
output.append(
(learning_curve, validation_loss, test_loss, fake_data_scale,
np.exp(hyperparser.get(metas, 'log_L2_reg')[0]), test_err))
# Do meta-SGD with momentum
g = results['dMd_meta']
velocity = meta_momentum * velocity - (1.0 - meta_momentum) * g
metas += velocity * meta_stepsize
print "Meta iteration {0} Validation loss {1} Test loss {2} Test err {3}"\
.format(i, validation_loss, test_loss, test_err)
return output, hyperparser.get(metas, 'fake_data')
def run():
val_images, val_labels, test_images, test_labels, _ = load_data(normalize=True)
val_images = val_images[:N_val_data, :]
val_labels = val_labels[:N_val_data, :]
true_data_scale = np.std(val_images)
test_images = test_images[:N_test_data, :]
test_labels = test_labels[:N_test_data, :]
batch_idxs = BatchList(N_fake_data, batch_size)
parser, _, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weights = len(parser.vect)
npr.seed(0)
init_fake_data = npr.randn(*(val_images[:N_fake_data, :].shape)) * init_fake_data_scale
one_hot = lambda x, K : np.array(x[:,None] == np.arange(K)[None, :], dtype=int)
fake_labels = one_hot(np.array(range(N_fake_data)) % N_classes, N_classes) # One of each.
hyperparser = WeightsParser()
hyperparser.add_weights('log_L2_reg', (1,))
hyperparser.add_weights('fake_data', init_fake_data.shape)
metas = np.zeros(hyperparser.N)
print "Number of hyperparameters to be trained:", hyperparser.N
hyperparser.set(metas, 'log_L2_reg', init_log_L2_reg)
hyperparser.set(metas, 'fake_data', init_fake_data)
def indexed_loss_fun(x, meta_params, idxs): # To be optimized by SGD.
L2_reg=np.exp(hyperparser.get(meta_params, 'log_L2_reg')[0])
fake_data=hyperparser.get(meta_params, 'fake_data')
return loss_fun(x, X=fake_data[idxs], T=fake_labels[idxs], L2_reg=L2_reg)
def meta_loss_fun(x, meta_params): # To be optimized in the outer loop.
fake_data=hyperparser.get(meta_params, 'fake_data')
log_prior = -fake_data_L2_reg * np.dot(fake_data.ravel(), fake_data.ravel())
return loss_fun(x, X=val_images, T=val_labels) - log_prior
def test_loss_fun(x): # To measure actual performance.
return loss_fun(x, X=test_images, T=test_labels)
log_alphas = np.full(N_iters, log_alpha_0)
betas = np.full(N_iters, beta_0)
output = []
for i in range(N_meta_iter):
print "L2 reg is ", np.exp(hyperparser.get(metas, 'log_L2_reg')[0]), "| ",
npr.seed(0)
v0 = npr.randn(N_weights) * velocity_scale
x0 = npr.randn(N_weights) * np.exp(log_param_scale)
results = sgd2(indexed_loss_fun, meta_loss_fun, batch_idxs, N_iters,
x0, v0, np.exp(log_alphas), betas, metas)
learning_curve = results['learning_curve']
validation_loss = results['M_final']
test_err = frac_err(results['x_final'], test_images, test_labels)
fake_data_scale = np.std(hyperparser.get(metas, 'fake_data')) / true_data_scale
test_loss = test_loss_fun(results['x_final'])
output.append((learning_curve, validation_loss, test_loss, fake_data_scale,
np.exp(hyperparser.get(metas, 'log_L2_reg')[0]), test_err))
metas -= results['dMd_meta'] * meta_stepsize
print "Meta iteration {0} Validation loss {1} Test loss {2} Test err {3}"\
.format(i, validation_loss, test_loss, test_err)
return output, hyperparser.get(metas, 'fake_data')
def meta_loss(W0, V0, alphas, betas, meta):
result = sgd2(loss_fun, meta_loss_fun, batch_idxs, N_iter, W0, V0, alphas, betas, meta)
return result['M_final']
