def test_sgd_parser():
N_weights = 6
W0 = 0.1 * npr.randn(N_weights)
N_data = 12
batch_size = 4
num_epochs = 4
batch_idxs = BatchList(N_data, batch_size)
parser = VectorParser()
parser.add_shape('first', [2,])
parser.add_shape('second', [1,])
parser.add_shape('third', [3,])
N_weight_types = 3
alphas = 0.1 * npr.rand(len(batch_idxs) * num_epochs, N_weight_types)
betas = 0.5 + 0.2 * npr.rand(len(batch_idxs) * num_epochs, N_weight_types)
meta = 0.1 * npr.randn(N_weights*2)
A = npr.randn(N_data, N_weights)
def loss_fun(W, meta, i=None):
idxs = batch_idxs.all_idxs if i is None else batch_idxs[i % len(batch_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 full_loss(params):
(W0, alphas, betas, meta) = params
result = sgd_parsed(grad(loss_fun), kylist(W0, alphas, betas, meta), parser)
return loss_fun(result, meta)
d_num = nd(full_loss, (W0, alphas, betas, meta))
d_an_fun = grad(full_loss)
d_an = d_an_fun([W0, alphas, betas, meta])
for i, (an, num) in enumerate(zip(d_an, d_num[0])):
assert np.allclose(an, num, rtol=1e-3, atol=1e-4), \
"Type {0}, diffs are: {1}".format(i, an - num)
def make_parabola(d):
parser = VectorParser()
parser.add_shape('weights', d)
dimscale = np.exp(np.linspace(-3, 3, d))
offset = npr.randn(d)
def loss(w, X=0.0, T=0.0, L2_reg=0.0):
return np.dot((w - offset) * dimscale, (w - offset))
return parser, loss
def make_parabola(d):
parser = VectorParser()
parser.add_shape('weights', d)
dimscale = np.exp(np.linspace(-3, 3, d))
offset = npr.randn(d)
def loss(w, X=0.0, T=0.0, L2_reg=0.0):
return np.dot((w - offset) * dimscale, (w - offset))
return parser, loss
def make_toy_funs():
parser = VectorParser()
parser.add_shape('weights', 2)
def rosenbrock(x):
return sum(100.0*(x[1:]-x[:-1]**2.0)**2.0 + (1-x[:-1])**2.0)
def loss(W_vect, X=0.0, T=0.0, L2_reg=0.0):
return 500 * logit(rosenbrock(W_vect) / 500)
return parser, loss
def make_toy_funs():
parser = VectorParser()
parser.add_shape('weights', 2)
def rosenbrock(x):
return sum(100.0 * (x[1:] - x[:-1]**2.0)**2.0 + (1 - x[:-1])**2.0)
def loss(W_vect, X=0.0, T=0.0, L2_reg=0.0):
return 500 * logit(rosenbrock(W_vect) / 500)
return parser, loss
def make_toy_funs():
parser = VectorParser()
parser.add_shape('weights', 2)
def rosenbrock(w):
x = w[1:]
y = w[:-1]
return sum(100.0 * (x - y**2.0)**2.0 + (1 - y)**2.0 + 200.0 * y)
def loss(W_vect, X=0.0, T=0.0, L2_reg=0.0):
return 800 * logit(rosenbrock(W_vect) / 500)
return parser, loss
def make_toy_funs():
parser = VectorParser()
parser.add_shape("weights", 2)
def rosenbrock(w):
x = w[1:]
y = w[:-1]
return sum(100.0 * (x - y ** 2.0) ** 2.0 + (1 - y) ** 2.0 + 200.0 * y)
def loss(W_vect, X=0.0, T=0.0, L2_reg=0.0):
return 800 * logit(rosenbrock(W_vect) / 500)
return parser, loss
def run():
train_data, valid_data, tests_data = load_data_dicts(N_train, N_valid, N_tests)
parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weight_types = len(parser.names)
hyperparams = VectorParser()
hyperparams['log_param_scale'] = np.full(N_weight_types, init_log_param_scale)
hyperparams['log_alphas'] = np.full((N_iters, N_weight_types), init_log_alphas)
hyperparams['invlogit_betas'] = np.full((N_iters, N_weight_types), init_invlogit_betas)
fixed_hyperparams = VectorParser()
fixed_hyperparams['log_L2_reg'] = np.full(N_weight_types, init_log_L2_reg)
def primal_optimizer(hyperparam_vect, i_hyper):
def indexed_loss_fun(w, L2_vect, i_iter):
rs = RandomState((seed, i_hyper, i_iter)) # Deterministic seed needed for backwards pass.
idxs = rs.randint(N_train, size=batch_size)
return loss_fun(w, train_data['X'][idxs], train_data['T'][idxs], L2_vect)
learning_curve_dict = defaultdict(list)
def callback(x, v, g, i_iter):
if i_iter % thin == 0:
learning_curve_dict['learning_curve'].append(loss_fun(x, **train_data))
learning_curve_dict['grad_norm'].append(np.linalg.norm(g))
learning_curve_dict['weight_norm'].append(np.linalg.norm(x))
learning_curve_dict['velocity_norm'].append(np.linalg.norm(v))
cur_hyperparams = hyperparams.new_vect(hyperparam_vect)
rs = RandomState((seed, i_hyper))
W0 = fill_parser(parser, np.exp(cur_hyperparams['log_param_scale']))
W0 *= rs.randn(W0.size)
alphas = np.exp(cur_hyperparams['log_alphas'])
betas = logit(cur_hyperparams['invlogit_betas'])
L2_reg = fill_parser(parser, np.exp(fixed_hyperparams['log_L2_reg']))
W_opt = sgd_parsed(grad(indexed_loss_fun), kylist(W0, alphas, betas, L2_reg),
parser, callback=callback)
return W_opt, learning_curve_dict
def hyperloss(hyperparam_vect, i_hyper):
W_opt, _ = primal_optimizer(hyperparam_vect, i_hyper)
return loss_fun(W_opt, **train_data)
hyperloss_grad = grad(hyperloss)
initial_hypergrad = hyperloss_grad( hyperparams.vect, 0)
parsed_init_hypergrad = hyperparams.new_vect(initial_hypergrad.copy())
avg_hypergrad = initial_hypergrad.copy()
for i in xrange(1, N_meta_iter):
avg_hypergrad += hyperloss_grad( hyperparams.vect, i)
print i
parsed_avg_hypergrad = hyperparams.new_vect(avg_hypergrad)
parser.vect = None # No need to pickle zeros
return parser, parsed_init_hypergrad, parsed_avg_hypergrad
def make_transform(layer_corr):
diag = np.eye(N_scripts)
full = np.full((N_scripts, N_scripts), 1.0 / N_scripts)
transform_parser = VectorParser()
for i_layer, corr in enumerate(layer_corr):
transform_parser[i_layer] = (1 - corr) * diag + corr * full
return transform_parser
def run(superparams):
alpha, log_scale_init, offset_init_std = superparams
RS = RandomState((seed, "top_rs"))
all_alphabets = omniglot.load_data()
RS.shuffle(all_alphabets)
train_alphabets = all_alphabets[:-N_test_alphabets]
tests_alphabets = all_alphabets[-N_test_alphabets:]
w_parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weights = w_parser.vect.size
hyperparams_0 = VectorParser()
hyperparams_0['log_scale'] = log_scale_init * np.ones(N_weights)
hyperparams_0['offset'] = offset_init_std * RS.randn(N_weights)
def reg_loss_fun(W, data, hyperparam_vect, reg_penalty):
hyperparams = hyperparams_0.new_vect(hyperparam_vect)
Z = np.exp(hyperparams['log_scale']) * W + hyperparams['offset']
return loss_fun(Z, **data) + np.dot(W, W) * reg_penalty
def hyperloss(hyperparam_vect, i_hyper, alphabets, verbose=True, report_train_loss=False):
RS = RandomState((seed, i_hyper, "hyperloss"))
alphabet = shuffle_alphabet(RS.choice(alphabets), RS)
N_train = alphabet['X'].shape[0] - N_valid_dpts
train_data = dictslice(alphabet, slice(None, N_train))
if report_train_loss:
valid_data = dictslice(alphabet, slice(None, N_valid_dpts))
else:
valid_data = dictslice(alphabet, slice(N_train, None))
def primal_loss(W, hyperparam_vect, i_primal, reg_penalty=True):
RS = RandomState((seed, i_hyper, i_primal))
idxs = RS.permutation(N_train)[:batch_size]
minibatch = dictslice(train_data, idxs)
loss = reg_loss_fun(W, minibatch, hyperparam_vect, reg_penalty)
if verbose and i_primal % 30 == 0:
print "Iter {0}, loss, {1}".format(i_primal, getval(loss))
return loss
W0 = np.zeros(N_weights)
W_final = sgd(grad(primal_loss), hyperparam_vect, W0, alpha, beta, N_iters, callback=None)
return reg_loss_fun(W_final, valid_data, hyperparam_vect, reg_penalty=False)
results = defaultdict(list)
def record_results(hyperparam_vect, i_hyper, g):
# print "Meta iter {0}. Recording results".format(i_hyper)
RS = RandomState((seed, i_hyper, "evaluation"))
new_seed = RS.int32()
def loss_fun(alphabets, report_train_loss):
return np.mean([hyperloss(hyperparam_vect, new_seed, alphabets=alphabets,
verbose=False, report_train_loss=report_train_loss)
for i in range(N_alphabets_eval)])
cur_hyperparams = hyperparams_0.new_vect(hyperparam_vect.copy())
if i_hyper % N_hyper_thin == 0:
# Storing O(N_weights) is a bit expensive so we thin it out and store in low precision
for field in cur_hyperparams.names:
results[field].append(cur_hyperparams[field].astype(np.float16))
results['train_loss'].append(loss_fun(train_alphabets, report_train_loss=True))
results['valid_loss'].append(loss_fun(train_alphabets, report_train_loss=False))
record_results(hyperparams_0.vect, 0, None)
return [results['train_loss'][0], results['valid_loss'][0]]
def run():
train_data, valid_data, tests_data = load_data_dicts(
N_train, N_valid, N_tests)
parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weight_types = len(parser.names)
hyperparams = VectorParser()
hyperparams['log_L2_reg'] = np.full(N_weight_types, init_log_L2_reg)
hyperparams['log_param_scale'] = np.full(N_weight_types,
init_log_param_scale)
hyperparams['log_alphas'] = np.full(N_iters, init_log_alphas)
hyperparams['invlogit_betas'] = np.full(N_iters, init_invlogit_betas)
def primal_optimizer(hyperparam_vect, i_hyper):
def indexed_loss_fun(w, L2_vect, i_iter):
seed = i_hyper * 10**6 + i_iter
idxs = npr.RandomState(seed).randint(N_train, size=batch_size)
return loss_fun(w, train_data['X'][idxs], train_data['T'][idxs],
L2_vect)
learning_curve = []
def callback(x, v, g, i_iter):
if i_iter % N_batches == 0:
learning_curve.append(loss_fun(x, **train_data))
cur_hyperparams = hyperparams.new_vect(hyperparam_vect)
W0 = fill_parser(parser, np.exp(cur_hyperparams['log_param_scale']))
W0 *= npr.RandomState(i_hyper).randn(W0.size)
alphas = np.exp(cur_hyperparams['log_alphas'])
betas = logit(cur_hyperparams['invlogit_betas'])
L2_reg = fill_parser(parser, np.exp(cur_hyperparams['log_L2_reg']))
V0 = np.zeros(W0.size)
W_opt = sgd4(grad(indexed_loss_fun), kylist(W0, alphas, betas, L2_reg),
callback)
return W_opt, learning_curve
def hyperloss(hyperparam_vect, i_hyper):
W_opt, _ = primal_optimizer(hyperparam_vect, i_hyper)
return loss_fun(W_opt, **valid_data)
hyperloss_grad = grad(hyperloss)
meta_results = defaultdict(list)
def meta_callback(hyperparam_vect, i_hyper, g):
print "Epoch {0}".format(i_hyper)
x, learning_curve = primal_optimizer(hyperparam_vect, i_hyper)
cur_hyperparams = hyperparams.new_vect(hyperparam_vect.copy())
for field in cur_hyperparams.names:
meta_results[field].append(cur_hyperparams[field])
meta_results['train_loss'].append(loss_fun(x, **train_data))
meta_results['valid_loss'].append(loss_fun(x, **valid_data))
meta_results['tests_loss'].append(loss_fun(x, **tests_data))
meta_results['learning_curves'].append(learning_curve)
final_result = rms_prop(hyperloss_grad, hyperparams.vect, meta_callback,
N_meta_iter, meta_alpha)
parser.vect = None # No need to pickle zeros
return meta_results, parser
def run():
N_iters = N_epochs
parser, loss_fun = make_toy_funs()
N_weight_types = len(parser.names)
N_weights = parser.vect.size
hyperparams = VectorParser()
hyperparams['log_alphas'] = np.full(N_iters, init_log_alphas)
hyperparams['invlogit_betas'] = np.full(N_iters, init_invlogit_betas)
hyperparams['V0'] = np.full(N_weights, init_V0)
all_learning_curves = []
all_param_curves = []
all_x = []
def hyperloss_grad(hyperparam_vect, ii):
learning_curve = []
params_curve = []
def callback(x, i):
params_curve.append(x)
learning_curve.append(loss_fun(x))
def indexed_loss_fun(w, log_L2_reg, j):
return loss_fun(w)
cur_hyperparams = hyperparams.new_vect(hyperparam_vect)
W0 = init_params
V0 = cur_hyperparams['V0']
alphas = np.exp(cur_hyperparams['log_alphas'])
betas = logit(cur_hyperparams['invlogit_betas'])
log_L2_reg = 0.0
results = sgd3(indexed_loss_fun, loss_fun, W0, V0,
alphas, betas, log_L2_reg, callback=callback)
hypergrads = hyperparams.copy()
hypergrads['V0'] = results['dMd_v'] * 0
hypergrads['log_alphas'] = results['dMd_alphas'] * alphas
hypergrads['invlogit_betas'] = (results['dMd_betas'] *
d_logit(cur_hyperparams['invlogit_betas']))
all_x.append(results['x_final'])
all_learning_curves.append(learning_curve)
all_param_curves.append(params_curve)
return hypergrads.vect
add_fields = ['train_loss', 'valid_loss', 'tests_loss', 'iter_num']
meta_results = {field : [] for field in add_fields + hyperparams.names}
def meta_callback(hyperparam_vect, i, g):
if i % N_meta_thin == 0:
print "Meta iter {0}".format(i)
x = all_x[-1]
cur_hyperparams = hyperparams.new_vect(hyperparam_vect.copy())
for field in cur_hyperparams.names:
meta_results[field].append(cur_hyperparams[field])
meta_results['train_loss'].append(loss_fun(x))
meta_results['iter_num'].append(i)
final_result = simple_sgd(hyperloss_grad, hyperparams.vect,
meta_callback, N_meta_iter, meta_alpha, meta_gamma)
meta_results['all_learning_curves'] = all_learning_curves
meta_results['all_param_curves'] = all_param_curves
parser.vect = None # No need to pickle zeros
return meta_results, parser
def make_transform(N_scripts, corr):
uncorrelated_mat = np.eye(N_scripts)
fully_correlated_mat = np.full((N_scripts, N_scripts), 1.0 / N_scripts)
transform_mat = (1 - corr) * uncorrelated_mat + corr * fully_correlated_mat
transform_parser = VectorParser()
for i_layer in range(N_layers):
transform_parser[i_layer] = transform_mat
return transform_parser
def build_hypervect(init_log_alphas, init_invlogit_betas,
init_log_param_scale):
hyperparams = VectorParser()
hyperparams['log_param_scale'] = np.full(N_weight_types,
init_log_param_scale)
hyperparams['log_alphas'] = np.full((N_iters, N_weight_types),
init_log_alphas)
hyperparams['invlogit_betas'] = np.full((N_iters, N_weight_types),
init_invlogit_betas)
return hyperparams
def test_sgd_parser():
N_weights = 6
W0 = 0.1 * npr.randn(N_weights)
N_data = 12
batch_size = 4
num_epochs = 4
batch_idxs = BatchList(N_data, batch_size)
parser = VectorParser()
parser.add_shape('first', [
2,
])
parser.add_shape('second', [
1,
])
parser.add_shape('third', [
3,
])
N_weight_types = 3
alphas = 0.1 * npr.rand(len(batch_idxs) * num_epochs, N_weight_types)
betas = 0.5 + 0.2 * npr.rand(len(batch_idxs) * num_epochs, N_weight_types)
meta = 0.1 * npr.randn(N_weights * 2)
A = npr.randn(N_data, N_weights)
def loss_fun(W, meta, i=None):
idxs = batch_idxs.all_idxs if i is None else batch_idxs[
i % len(batch_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 full_loss(params):
(W0, alphas, betas, meta) = params
result = sgd_parsed(grad(loss_fun), kylist(W0, alphas, betas, meta),
parser)
return loss_fun(result, meta)
d_num = nd(full_loss, (W0, alphas, betas, meta))
d_an_fun = grad(full_loss)
d_an = d_an_fun([W0, alphas, betas, meta])
for i, (an, num) in enumerate(zip(d_an, d_num[0])):
assert np.allclose(an, num, rtol=1e-3, atol=1e-4), \
"Type {0}, diffs are: {1}".format(i, an - num)
def run():
parser, loss_fun = make_parabola(dimension)
N_weight_types = len(parser.names)
hyperparams = VectorParser()
hyperparams['log_L2_reg'] = np.full(N_weight_types, init_log_L2_reg)
hyperparams['log_param_scale'] = np.full(N_weight_types, init_log_param_scale)
hyperparams['log_alphas'] = np.full(N_iters, init_log_alphas)
hyperparams['invlogit_betas'] = np.full(N_iters, init_invlogit_betas)
def primal_optimizer(hyperparam_vect, i_hyper):
learning_curve = []
def callback(x, i_iter):
learning_curve.append(loss_fun(x))
cur_hyperparams = hyperparams.new_vect(hyperparam_vect)
W0 = fill_parser(parser, np.exp(cur_hyperparams['log_param_scale']))
W0 *= npr.RandomState(hash(i_hyper)).randn(W0.size)
alphas = np.exp(cur_hyperparams['log_alphas'])
betas = logit(cur_hyperparams['invlogit_betas'])
L2_reg = fill_parser(parser, np.exp(cur_hyperparams['log_L2_reg']))
W_opt = sgd4(grad(loss_fun), kylist(W0, alphas, betas, L2_reg), callback)
callback(W_opt, N_iters)
return W_opt, learning_curve
def hyperloss(hyperparam_vect, i_hyper):
W_opt, _ = primal_optimizer(hyperparam_vect, i_hyper)
return loss_fun(W_opt)
hyperloss_grad = grad(hyperloss)
meta_results = defaultdict(list)
def meta_callback(hyperparam_vect, i_hyper):
print "Meta Epoch {0}".format(i_hyper)
x, learning_curve = primal_optimizer(hyperparam_vect, i_hyper)
cur_hyperparams = hyperparams.new_vect(hyperparam_vect.copy())
for field in cur_hyperparams.names:
meta_results[field].append(cur_hyperparams[field])
meta_results['train_loss'].append(loss_fun(x))
meta_results['learning_curves'].append(learning_curve)
final_result = rms_prop(hyperloss_grad, hyperparams.vect,
meta_callback, N_meta_iter, meta_alpha, gamma=0.0)
meta_callback(final_result, N_meta_iter)
parser.vect = None # No need to pickle zeros
return meta_results, parser
def run():
N_iters = N_epochs
parser, loss_fun = make_toy_funs()
N_weights = parser.vect.size
hyperparams = VectorParser()
hyperparams['log_alphas'] = np.full(N_iters, init_log_alphas)
hyperparams['invlogit_betas'] = np.full(N_iters, init_invlogit_betas)
hyperparams['V0'] = np.full(N_weights, init_V0)
forward_path = []
forward_learning_curve = []
def fwd_callback(x, i):
print type(x[0])
forward_path.append(x.copy())
forward_learning_curve.append(loss_fun(x))
reverse_path = []
reverse_learning_curve = []
def reverse_callback(x, i):
reverse_path.append(x.copy())
reverse_learning_curve.append(loss_fun(x))
def indexed_loss_fun(w, log_L2_reg, j):
return loss_fun(w)
cur_hyperparams = hyperparams
W0 = init_params
V0 = cur_hyperparams['V0']
alphas = np.exp(cur_hyperparams['log_alphas'])
betas = logit(cur_hyperparams['invlogit_betas'])
log_L2_reg = 0.0
sgd3_naive(indexed_loss_fun,
W0,
V0,
alphas,
betas,
log_L2_reg,
fwd_callback=fwd_callback,
reverse_callback=reverse_callback)
return forward_path, forward_learning_curve, reverse_path, reverse_learning_curve
def run():
train_data, valid_data, tests_data = load_data_dicts(
N_train, N_valid, N_tests)
parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weight_types = len(parser.names)
N_weights = len(parser.vect)
hyperparams = VectorParser()
rs = RandomState((seed))
hyperparams['log_L2_reg'] = np.full(N_weights, init_log_L2_reg)\
+ rs.randn(N_weights) * init_log_L2_reg_noise
hyperparams['log_param_scale'] = np.full(N_weight_types,
init_log_param_scale)
hyperparams['log_alphas'] = np.full((N_iters, N_weight_types),
init_log_alphas)
hyperparams['invlogit_betas'] = np.full((N_iters, N_weight_types),
init_invlogit_betas)
cur_primal_results = {}
def primal_optimizer(hyperparam_vect, i_hyper):
def indexed_loss_fun(w, L2_vect, i_iter):
rs = RandomState(
(seed, i_hyper,
i_iter)) # Deterministic seed needed for backwards pass.
idxs = rs.randint(N_train, size=batch_size)
return loss_fun(w, train_data['X'][idxs], train_data['T'][idxs],
L2_vect)
learning_curve_dict = defaultdict(list)
def callback(x, v, g, i_iter):
if i_iter % thin == 0:
learning_curve_dict['learning_curve'].append(
loss_fun(x, **train_data))
learning_curve_dict['grad_norm'].append(np.linalg.norm(g))
learning_curve_dict['weight_norm'].append(np.linalg.norm(x))
learning_curve_dict['velocity_norm'].append(np.linalg.norm(v))
cur_hyperparams = hyperparams.new_vect(hyperparam_vect)
rs = RandomState((seed, i_hyper))
W0 = fill_parser(parser, np.exp(cur_hyperparams['log_param_scale']))
W0 *= rs.randn(W0.size)
alphas = np.exp(cur_hyperparams['log_alphas'])
betas = logit(cur_hyperparams['invlogit_betas'])
L2_reg = np.exp(cur_hyperparams['log_L2_reg'])
W_opt = sgd_parsed(grad(indexed_loss_fun),
kylist(W0, alphas, betas, L2_reg),
parser,
callback=callback)
cur_primal_results['weights'] = getval(W_opt).copy()
cur_primal_results['learning_curve'] = getval(learning_curve_dict)
return W_opt, learning_curve_dict
def hyperloss(hyperparam_vect, i_hyper):
W_opt, _ = primal_optimizer(hyperparam_vect, i_hyper)
return loss_fun(W_opt, **valid_data)
hyperloss_grad = grad(hyperloss)
meta_results = defaultdict(list)
old_metagrad = [np.ones(hyperparams.vect.size)]
def meta_callback(hyperparam_vect, i_hyper, metagrad=None):
#x, learning_curve_dict = primal_optimizer(hyperparam_vect, i_hyper)
x, learning_curve_dict = cur_primal_results[
'weights'], cur_primal_results['learning_curve']
cur_hyperparams = hyperparams.new_vect(hyperparam_vect.copy())
for field in cur_hyperparams.names:
meta_results[field] = cur_hyperparams[field]
meta_results['train_loss'].append(loss_fun(x, **train_data))
meta_results['valid_loss'].append(loss_fun(x, **valid_data))
meta_results['tests_loss'].append(loss_fun(x, **tests_data))
meta_results['test_err'].append(frac_err(x, **tests_data))
meta_results['learning_curves'].append(learning_curve_dict)
meta_results['example_weights'] = x
if metagrad is not None:
meta_results['meta_grad_magnitude'].append(
np.linalg.norm(metagrad))
meta_results['meta_grad_angle'].append(np.dot(old_metagrad[0], metagrad) \
/ (np.linalg.norm(metagrad)*
np.linalg.norm(old_metagrad[0])))
old_metagrad[0] = metagrad
print "Meta Epoch {0} Train loss {1:2.4f} Valid Loss {2:2.4f}" \
" Test Loss {3:2.4f} Test Err {4:2.4f}".format(
i_hyper, meta_results['train_loss'][-1], meta_results['valid_loss'][-1],
meta_results['tests_loss'][-1], meta_results['test_err'][-1])
initial_hypergrad = hyperloss_grad(hyperparams.vect, 0)
parsed_init_hypergrad = hyperparams.new_vect(initial_hypergrad.copy())
final_result = adam(hyperloss_grad, hyperparams.vect, meta_callback,
N_meta_iter, meta_alpha)
meta_callback(final_result, N_meta_iter)
parser.vect = None # No need to pickle zeros
return meta_results, parser, parsed_init_hypergrad
def run():
train_data, valid_data, tests_data = load_data_dicts(N_train, N_valid, N_tests)
parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weight_types = len(parser.names)
hyperparams = VectorParser()
hyperparams['log_param_scale'] = np.full(N_weight_types, init_log_param_scale)
hyperparams['log_alphas'] = np.full((N_iters, N_weight_types), init_log_alphas)
hyperparams['invlogit_betas'] = np.full((N_iters, N_weight_types), init_invlogit_betas)
fixed_hyperparams = VectorParser()
fixed_hyperparams['log_L2_reg'] = np.full(N_weight_types, init_log_L2_reg)
def primal_optimizer(hyperparam_vect, i_hyper):
def indexed_loss_fun(w, L2_vect, i_iter):
rs = RandomState((seed, i_hyper, i_iter)) # Deterministic seed needed for backwards pass.
idxs = rs.randint(N_train, size=batch_size)
return loss_fun(w, train_data['X'][idxs], train_data['T'][idxs], L2_vect)
learning_curve_dict = defaultdict(list)
def callback(x, v, g, i_iter):
if i_iter % thin == 0 or i_iter == 0 or i_iter == N_iters - 1:
learning_curve_dict['learning_curve'].append(loss_fun(x, **train_data))
learning_curve_dict['grad_norm'].append(np.linalg.norm(g))
learning_curve_dict['weight_norm'].append(np.linalg.norm(x))
learning_curve_dict['velocity_norm'].append(np.linalg.norm(v))
learning_curve_dict['iteration'].append(i_iter)
init_hyperparams = hyperparams.new_vect(hyperparam_vect)
rs = RandomState((seed, i_hyper))
W0 = fill_parser(parser, np.exp(init_hyperparams['log_param_scale']))
W0 *= rs.randn(W0.size)
alphas = np.exp(init_hyperparams['log_alphas'])
betas = logit(init_hyperparams['invlogit_betas'])
L2_reg = fill_parser(parser, np.exp(fixed_hyperparams['log_L2_reg']))
W_opt = sgd_parsed(grad(indexed_loss_fun), kylist(W0, alphas, betas, L2_reg),
parser, callback=callback)
return W_opt, learning_curve_dict
meta_results = defaultdict(list)
old_metagrad = [np.ones(hyperparams.vect.size)]
def meta_callback(hyperparam_vect, i_hyper, metagrad=None):
x, learning_curve_dict = primal_optimizer(hyperparam_vect, i_hyper)
cur_hyperparams = hyperparams.new_vect(hyperparam_vect.copy())
for field in cur_hyperparams.names:
meta_results[field].append(cur_hyperparams[field])
meta_results['train_loss'].append(loss_fun(x, **train_data))
meta_results['valid_loss'].append(loss_fun(x, **valid_data))
meta_results['tests_loss'].append(loss_fun(x, **tests_data))
meta_results['test_err'].append(frac_err(x, **tests_data))
meta_results['learning_curves'].append(learning_curve_dict)
if metagrad is not None:
meta_results['meta_grad_magnitude'].append(np.linalg.norm(metagrad))
meta_results['meta_grad_angle'].append(np.dot(old_metagrad[0], metagrad) \
/ (np.linalg.norm(metagrad)*
np.linalg.norm(old_metagrad[0])))
old_metagrad[0] = metagrad
print "Meta Epoch {0} Train loss {1:2.4f} Valid Loss {2:2.4f}" \
" Test Loss {3:2.4f} Test Err {4:2.4f}".format(
i_hyper, meta_results['train_loss'][-1], meta_results['valid_loss'][-1],
meta_results['train_loss'][-1], meta_results['test_err'][-1])
# Now do a line search along that direction.
log_stepsizes = np.linspace(-init_log_alphas, init_log_alphas*4, N_points_in_line_search)
for log_stepsizes in log_stepsizes:
hyperparams['log_alphas'] = np.full((N_iters, N_weight_types), log_stepsizes)
meta_callback(hyperparams.vect, 0) # Use the same random seed every time.
parser.vect = None # No need to pickle zeros
return meta_results, parser, log_stepsizes
def run():
train_data, valid_data, tests_data = load_data_dicts(
N_train, N_valid, N_tests)
parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weight_types = len(parser.names)
hyperparams = VectorParser()
hyperparams['log_L2_reg'] = np.full(N_weight_types, init_log_L2_reg)
hyperparams['log_param_scale'] = np.full(N_weight_types,
init_log_param_scale)
hyperparams['log_alphas'] = np.full(N_iters, init_log_alphas)
hyperparams['invlogit_betas'] = np.full(N_iters, init_invlogit_betas)
fixed_hyperparams = VectorParser()
fixed_hyperparams['log_param_scale'] = np.full(N_iters,
init_log_param_scale)
# TODO: memoize
def primal_optimizer(hyperparam_vect, i_hyper):
def indexed_loss_fun(w, L2_vect, i_iter):
rs = npr.RandomState(
npr.RandomState(global_seed + i_hyper).randint(1000))
seed = i_hyper * 10**6 + i_iter # Deterministic seed needed for backwards pass.
idxs = rs.randint(N_train, size=batch_size)
return loss_fun(w, train_data['X'][idxs], train_data['T'][idxs],
L2_vect)
learning_curve_dict = defaultdict(list)
def callback(x, v, g, i_iter):
if i_iter % N_batches == 0:
learning_curve_dict['learning_curve'].append(
loss_fun(x, **train_data))
learning_curve_dict['grad_norm'].append(np.linalg.norm(g))
learning_curve_dict['weight_norm'].append(np.linalg.norm(x))
learning_curve_dict['velocity_norm'].append(np.linalg.norm(v))
cur_hyperparams = hyperparams.new_vect(hyperparam_vect)
W0 = fill_parser(parser, np.exp(fixed_hyperparams['log_param_scale']))
W0 *= npr.RandomState(global_seed + i_hyper).randn(W0.size)
alphas = np.exp(cur_hyperparams['log_alphas'])
betas = logit(cur_hyperparams['invlogit_betas'])
L2_reg = fill_parser(parser, np.exp(cur_hyperparams['log_L2_reg']))
W_opt = sgd4(grad(indexed_loss_fun), kylist(W0, alphas, betas, L2_reg),
callback)
#callback(W_opt, N_iters)
return W_opt, learning_curve_dict
def hyperloss(hyperparam_vect, i_hyper):
W_opt, _ = primal_optimizer(hyperparam_vect, i_hyper)
return loss_fun(W_opt, **valid_data)
hyperloss_grad = grad(hyperloss)
meta_results = defaultdict(list)
def meta_callback(hyperparam_vect, i_hyper):
print "Meta Epoch {0}".format(i_hyper)
x, learning_curve_dict = primal_optimizer(hyperparam_vect, i_hyper)
cur_hyperparams = hyperparams.new_vect(hyperparam_vect.copy())
for field in cur_hyperparams.names:
meta_results[field].append(cur_hyperparams[field])
meta_results['train_loss'].append(loss_fun(x, **train_data))
meta_results['valid_loss'].append(loss_fun(x, **valid_data))
meta_results['tests_loss'].append(loss_fun(x, **tests_data))
meta_results['learning_curves'].append(learning_curve_dict)
final_result = rms_prop(hyperloss_grad,
hyperparams.vect,
meta_callback,
N_meta_iter,
meta_alpha,
gamma=0.0)
meta_callback(final_result, N_meta_iter)
parser.vect = None # No need to pickle zeros
return meta_results, parser
def run():
train_data, valid_data, tests_data = load_data_dicts(N_train, N_valid, N_tests)
parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weight_types = len(parser.names)
hyperparams = VectorParser()
hyperparams['log_param_scale'] = np.full(N_weight_types, init_log_param_scale)
hyperparams['log_alphas'] = np.full((N_iters, N_weight_types), init_log_alphas)
hyperparams['invlogit_betas'] = np.full((N_iters, N_weight_types), init_invlogit_betas)
fixed_hyperparams = VectorParser()
fixed_hyperparams['log_L2_reg'] = np.full(N_weight_types, init_log_L2_reg)
cur_primal_results = {}
def primal_optimizer(hyperparam_vect, i_hyper):
def indexed_loss_fun(w, L2_vect, i_iter):
rs = RandomState((seed, i_hyper, i_iter)) # Deterministic seed needed for backwards pass.
idxs = rs.randint(N_train, size=batch_size)
return loss_fun(w, train_data['X'][idxs], train_data['T'][idxs], L2_vect)
learning_curve_dict = defaultdict(list)
def callback(x, v, g, i_iter):
if i_iter % thin == 0:
learning_curve_dict['learning_curve'].append(loss_fun(x, **train_data))
learning_curve_dict['grad_norm'].append(np.linalg.norm(g))
learning_curve_dict['weight_norm'].append(np.linalg.norm(x))
learning_curve_dict['velocity_norm'].append(np.linalg.norm(v))
cur_hyperparams = hyperparams.new_vect(hyperparam_vect)
rs = RandomState((seed, i_hyper))
W0 = fill_parser(parser, np.exp(cur_hyperparams['log_param_scale']))
W0 *= rs.randn(W0.size)
alphas = np.exp(cur_hyperparams['log_alphas'])
betas = logit(cur_hyperparams['invlogit_betas'])
L2_reg = fill_parser(parser, np.exp(fixed_hyperparams['log_L2_reg']))
W_opt = sgd_parsed(grad(indexed_loss_fun), kylist(W0, alphas, betas, L2_reg),
parser, callback=callback)
cur_primal_results['weights'] = getval(W_opt).copy()
cur_primal_results['learning_curve'] = getval(learning_curve_dict)
return W_opt, learning_curve_dict
def hyperloss(hyperparam_vect, i_hyper):
W_opt, _ = primal_optimizer(hyperparam_vect, i_hyper)
return loss_fun(W_opt, **train_data)
hyperloss_grad = grad(hyperloss)
meta_results = defaultdict(list)
old_metagrad = [np.ones(hyperparams.vect.size)]
def meta_callback(hyperparam_vect, i_hyper, metagrad=None):
x, learning_curve_dict = cur_primal_results['weights'], cur_primal_results['learning_curve']
cur_hyperparams = hyperparams.new_vect(hyperparam_vect.copy())
for field in cur_hyperparams.names:
meta_results[field].append(cur_hyperparams[field])
meta_results['train_loss'].append(loss_fun(x, **train_data))
meta_results['valid_loss'].append(loss_fun(x, **valid_data))
meta_results['tests_loss'].append(loss_fun(x, **tests_data))
meta_results['test_err'].append(frac_err(x, **tests_data))
meta_results['learning_curves'].append(learning_curve_dict)
meta_results['example_weights'] = x
if metagrad is not None:
meta_results['meta_grad_magnitude'].append(np.linalg.norm(metagrad))
meta_results['meta_grad_angle'].append(np.dot(old_metagrad[0], metagrad) \
/ (np.linalg.norm(metagrad)*
np.linalg.norm(old_metagrad[0])))
old_metagrad[0] = metagrad
print "Meta Epoch {0} Train loss {1:2.4f} Valid Loss {2:2.4f}" \
" Test Loss {3:2.4f} Test Err {4:2.4f}".format(
i_hyper, meta_results['train_loss'][-1], meta_results['valid_loss'][-1],
meta_results['train_loss'][-1], meta_results['test_err'][-1])
initial_hypergrad = hyperloss_grad( hyperparams.vect, 0)
hypergrads = np.zeros((N_meta_iter, len(initial_hypergrad)))
for i in xrange(N_meta_iter):
hypergrads[i] = hyperloss_grad( hyperparams.vect, i)
print i
avg_hypergrad = np.mean(hypergrads, axis=0)
parsed_avg_hypergrad = hyperparams.new_vect(avg_hypergrad)
parser.vect = None # No need to pickle zeros
return parser, parsed_avg_hypergrad
def run():
train_data, valid_data, tests_data = load_data_dicts(
N_train, N_valid, N_tests)
parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weight_types = len(parser.names)
hyperparams = VectorParser()
hyperparams['log_param_scale'] = np.full(N_weight_types,
init_log_param_scale)
hyperparams['log_alphas'] = np.full((N_iters, N_weight_types),
init_log_alphas)
hyperparams['invlogit_betas'] = np.full((N_iters, N_weight_types),
init_invlogit_betas)
fixed_hyperparams = VectorParser()
fixed_hyperparams['log_L2_reg'] = np.full(N_weight_types, init_log_L2_reg)
def primal_optimizer(hyperparam_vect, i_hyper):
def indexed_loss_fun(w, L2_vect, i_iter):
rs = RandomState(
(seed, i_hyper,
i_iter)) # Deterministic seed needed for backwards pass.
idxs = rs.randint(N_train, size=batch_size)
return loss_fun(w, train_data['X'][idxs], train_data['T'][idxs],
L2_vect)
learning_curve_dict = defaultdict(list)
def callback(x, v, g, i_iter):
if i_iter % thin == 0:
learning_curve_dict['learning_curve'].append(
loss_fun(x, **train_data))
learning_curve_dict['grad_norm'].append(np.linalg.norm(g))
learning_curve_dict['weight_norm'].append(np.linalg.norm(x))
learning_curve_dict['velocity_norm'].append(np.linalg.norm(v))
cur_hyperparams = hyperparams.new_vect(hyperparam_vect)
rs = RandomState((seed, i_hyper))
W0 = fill_parser(parser, np.exp(cur_hyperparams['log_param_scale']))
W0 *= rs.randn(W0.size)
alphas = np.exp(cur_hyperparams['log_alphas'])
betas = logit(cur_hyperparams['invlogit_betas'])
L2_reg = fill_parser(parser, np.exp(fixed_hyperparams['log_L2_reg']))
W_opt = sgd_parsed(grad(indexed_loss_fun),
kylist(W0, alphas, betas, L2_reg),
parser,
callback=callback)
return W_opt, learning_curve_dict
def hyperloss(hyperparam_vect, i_hyper):
W_opt, _ = primal_optimizer(hyperparam_vect, i_hyper)
return loss_fun(W_opt, **train_data)
hyperloss_grad = grad(hyperloss)
initial_hypergrad = hyperloss_grad(hyperparams.vect, 0)
parsed_init_hypergrad = hyperparams.new_vect(initial_hypergrad.copy())
avg_hypergrad = initial_hypergrad.copy()
for i in xrange(1, N_meta_iter):
avg_hypergrad += hyperloss_grad(hyperparams.vect, i)
print i
parsed_avg_hypergrad = hyperparams.new_vect(avg_hypergrad)
parser.vect = None # No need to pickle zeros
return parser, parsed_init_hypergrad, parsed_avg_hypergrad
def run():
RS = RandomState((seed, "top_rs"))
all_data = omniglot.load_flipped_alphabets()
train_data, tests_data = random_partition(all_data, RS, [12, 3])
w_parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weights = w_parser.vect.size
script_parser = VectorParser()
for i_script in range(N_scripts):
script_parser[i_script] = np.zeros(N_weights)
transform_parser = make_transform([0] * N_layers)
def get_layers(vect):
layers = []
for i_layer in range(N_layers):
weights_by_scripts = vect.reshape((N_scripts, N_weights))
weights_idxs, _ = w_parser.idxs_and_shapes[('weights', i_layer)]
biases_idxs, _ = w_parser.idxs_and_shapes[('biases', i_layer)]
assert weights_idxs.stop == biases_idxs.start
layer_idxs = slice(weights_idxs.start, biases_idxs.stop)
layers.append(weights_by_scripts[:, layer_idxs])
return layers
def transform_weights(z_vect, transform_vect):
z_layers = get_layers(z_vect)
transform = transform_parser.new_vect(transform_vect)
w_layers = [np.dot(transform[i], z) for i, z in enumerate(z_layers)]
return np.concatenate(w_layers, axis=1).ravel()
def likelihood_loss(w_vect, data):
w = script_parser.new_vect(w_vect)
return sum([
loss_fun(w[i], **script_data) for i, script_data in enumerate(data)
])
def regularization(z_vect):
return np.dot(z_vect, z_vect) * np.exp(log_L2)
def train_z(data, transform_vect, RS):
def primal_loss(z_vect,
transform_vect,
i_primal,
record_results=False):
w_vect = transform_weights(z_vect, transform_vect)
loss = likelihood_loss(w_vect, data)
reg = regularization(z_vect)
if record_results and i_primal % N_thin == 0:
print "Iter {0}: train: {1}".format(i_primal,
getval(loss) / N_scripts)
return loss + reg
z_vect_0 = RS.randn(script_parser.vect.size) * np.exp(log_init_scale)
return sgd(grad(primal_loss), transform_vect, z_vect_0, alpha, beta,
N_iters)
def train_sharing():
def hyperloss(transform_vect, i_hyper):
RS = RandomState((seed, i_hyper, "hyperloss"))
cur_train_data, cur_valid_data = random_partition(
train_data, RS, [10, 2])
z_vect_final = train_z(cur_train_data, transform_vect, RS)
w_vect_final = transform_weights(z_vect_final, transform_vect)
return likelihood_loss(w_vect_final, cur_valid_data) / N_scripts
hypergrad = grad(hyperloss)
cur_transform_vect = make_transform([init_script_corr] * N_layers).vect
for i_hyper in range(N_meta_iter):
print "Hyper iter {0}".format(i_hyper)
grad_transform = hypergrad(cur_transform_vect, i_hyper)
cur_transform_vect = cur_transform_vect - grad_transform * meta_alpha
return cur_transform_vect
transform_vects, train_losses, tests_losses = {}, {}, {}
transform_vects['no_sharing'] = make_transform([0, 0, 0]).vect
transform_vects['full_sharing'] = make_transform([1, 0, 0]).vect
transform_vects['learned_sharing'] = train_sharing()
for name in transform_vects.keys():
RS = RandomState("final_training")
tv = transform_vects[name]
trained_z = train_z(train_data, tv, RS)
trained_w = transform_weights(trained_z, tv)
train_losses[name] = likelihood_loss(trained_w, train_data) / N_scripts
tests_losses[name] = likelihood_loss(trained_w, tests_data) / N_scripts
print "{0} : train: {1}, test: {2}".format(name, train_losses[name],
tests_losses[name])
return transform_parser, transform_vects, train_losses, tests_losses
def run():
"""Three different parsers:
w_parser[('biases', i_layer)] : neural net weights/biases per layer for a single script
script_parser[i_script] : weights vector for each script
transform_parser[i_layer] : transform matrix (scripts x scripts) for each alphabet"""
RS = RandomState((seed, "top_rs"))
train_data, valid_data, tests_data = omniglot.load_data_split(
[11, 2, 2], RS, num_alphabets=N_scripts)
w_parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weights = w_parser.vect.size
transform_parser = make_transform(N_scripts, script_corr_init)
script_parser = VectorParser()
for i_script in range(N_scripts):
script_parser[i_script] = np.zeros(N_weights)
def get_layers(vect):
layers = []
for i_layer in range(N_layers):
weights_by_scripts = vect.reshape((N_scripts, N_weights))
weights_idxs, _ = w_parser.idxs_and_shapes[('weights', i_layer)]
biases_idxs, _ = w_parser.idxs_and_shapes[('biases', i_layer)]
assert weights_idxs.stop == biases_idxs.start
layer_idxs = slice(weights_idxs.start, biases_idxs.stop)
layers.append(weights_by_scripts[:, layer_idxs])
return layers
def transform_weights(z_vect, transform_vect):
z_layers = get_layers(z_vect)
transform = transform_parser.new_vect(transform_vect)
w_layers = [np.dot(transform[i], z) for i, z in enumerate(z_layers)]
return np.concatenate(w_layers, axis=1).ravel()
def total_loss(w_vect, data):
w = script_parser.new_vect(w_vect)
return sum([loss_fun(w[i], **script_data) for i, script_data in enumerate(data)])
def regularization(z_vect):
return np.dot(z_vect, z_vect) * np.exp(log_L2_init)
results = defaultdict(list)
def hyperloss(transform_vect, i_hyper, record_results=True):
RS = RandomState((seed, i_hyper, "hyperloss"))
def primal_loss(z_vect, transform_vect, i_primal, record_results=False):
w_vect = transform_weights(z_vect, transform_vect)
loss = total_loss(w_vect, train_data)
reg = regularization(z_vect)
if VERBOSE and record_results and i_primal % N_thin == 0:
print "Iter {0}: train: {1}, valid: {2}, reg: {3}".format(
i_primal,
getval(loss) / N_scripts,
total_loss(getval(w_vect), valid_data) / N_scripts,
getval(reg))
return loss + reg
z_vect_0 = RS.randn(script_parser.vect.size) * np.exp(log_initialization_scale)
z_vect_final = sgd(grad(primal_loss), transform_vect, z_vect_0,
alpha, beta, N_iters, callback=None)
w_vect_final = transform_weights(z_vect_final, transform_vect)
valid_loss = total_loss(w_vect_final, valid_data)
if record_results:
results['valid_loss'].append(getval(valid_loss) / N_scripts)
results['train_loss'].append(total_loss(w_vect_final, train_data) / N_scripts)
results['tests_loss'].append(total_loss(w_vect_final, tests_data) / N_scripts)
return valid_loss
grad_transform = grad(hyperloss)(transform_parser.vect, 0, record_results=False)
for i, d in enumerate(line_search_dists):
new_transform_vect = transform_parser.vect - d * grad_transform
hyperloss(new_transform_vect, 0, record_results=True)
print "Hyper iter {0}".format(i)
print "Results", {k : v[-1] for k, v in results.iteritems()}
grad_transform_dict = transform_parser.new_vect(grad_transform).as_dict()
return results, grad_transform_dict
def run():
(train_images, train_labels),\
(valid_images, valid_labels),\
(tests_images, tests_labels) = load_data_subset(N_train, N_valid, N_tests)
batch_idxs = BatchList(N_train, batch_size)
N_iters = N_epochs * len(batch_idxs)
parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weight_types = len(parser.names)
hyperparams = VectorParser()
hyperparams['log_L2_reg'] = np.full(N_weight_types, init_log_L2_reg)
hyperparams['log_param_scale'] = np.full(N_weight_types,
init_log_param_scale)
hyperparams['log_alphas'] = np.full(N_iters, init_log_alphas)
hyperparams['invlogit_betas'] = np.full(N_iters, init_invlogit_betas)
def indexed_loss_fun(w, log_L2_reg, i):
idxs = batch_idxs[i % len(batch_idxs)]
partial_vects = [
np.full(parser[name].size, np.exp(log_L2_reg[i]))
for i, name in enumerate(parser.names)
]
L2_reg_vect = np.concatenate(partial_vects, axis=0)
return loss_fun(w,
X=train_images[idxs],
T=train_labels[idxs],
L2_reg=L2_reg_vect)
def train_loss_fun(w, log_L2_reg=0.0):
return loss_fun(w, X=train_images, T=train_labels)
def valid_loss_fun(w, log_L2_reg=0.0):
return loss_fun(w, X=valid_images, T=valid_labels)
def tests_loss_fun(w, log_L2_reg=0.0):
return loss_fun(w, X=tests_images, T=tests_labels)
all_learning_curves = []
all_x = []
def hyperloss_grad(hyperparam_vect, i):
learning_curve = []
def callback(x, i):
if i % len(batch_idxs) == 0:
learning_curve.append(
loss_fun(x, X=train_images, T=train_labels))
npr.seed(i)
N_weights = parser.vect.size
V0 = np.zeros(N_weights)
cur_hyperparams = hyperparams.new_vect(hyperparam_vect)
layer_param_scale = [
np.full(parser[name].size,
np.exp(cur_hyperparams['log_param_scale'][i]))
for i, name in enumerate(parser.names)
]
W0 = npr.randn(N_weights) * np.concatenate(layer_param_scale, axis=0)
alphas = np.exp(cur_hyperparams['log_alphas'])
betas = logit(cur_hyperparams['invlogit_betas'])
log_L2_reg = cur_hyperparams['log_L2_reg']
results = sgd3(indexed_loss_fun,
valid_loss_fun,
W0,
V0,
alphas,
betas,
log_L2_reg,
callback=callback)
hypergrads = hyperparams.copy()
hypergrads['log_L2_reg'] = results['dMd_meta']
weights_grad = parser.new_vect(W0 * results['dMd_x'])
hypergrads['log_param_scale'] = [
np.sum(weights_grad[name]) for name in parser.names
]
hypergrads['log_alphas'] = results['dMd_alphas'] * alphas
hypergrads['invlogit_betas'] = (
results['dMd_betas'] * d_logit(cur_hyperparams['invlogit_betas']))
all_x.append(results['x_final'])
all_learning_curves.append(learning_curve)
return hypergrads.vect
add_fields = ['train_loss', 'valid_loss', 'tests_loss']
meta_results = {field: [] for field in add_fields + hyperparams.names}
def meta_callback(hyperparam_vect, i):
print "Meta iter {0}".format(i)
x = all_x[-1]
cur_hyperparams = hyperparams.new_vect(hyperparam_vect.copy())
log_L2_reg = cur_hyperparams['log_L2_reg']
for field in cur_hyperparams.names:
meta_results[field].append(cur_hyperparams[field])
meta_results['train_loss'].append(train_loss_fun(x))
meta_results['valid_loss'].append(valid_loss_fun(x))
meta_results['tests_loss'].append(tests_loss_fun(x))
final_result = rms_prop(hyperloss_grad, hyperparams.vect, meta_callback,
N_meta_iter, meta_alpha)
meta_results['all_learning_curves'] = all_learning_curves
parser.vect = None # No need to pickle zeros
return meta_results, parser
def run():
train_data, valid_data, tests_data = load_data_dicts(N_train, N_valid, N_tests)
parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes) #only uses two different regularization hyperparameters, one for each layer?
N_weight_types = len(parser.names) # = 2
print(parser.names)
hyperparams = VectorParser()
hyperparams['log_L2_reg'] = np.full(N_weight_types, init_log_L2_reg)
hyperparams['log_param_scale'] = np.full(N_weight_types, init_log_param_scale)
hyperparams['log_alphas'] = np.full(N_iters, init_log_alphas)
hyperparams['invlogit_betas'] = np.full(N_iters, init_invlogit_betas)
fixed_hyperparams = VectorParser()
fixed_hyperparams['log_param_scale'] = np.full(N_iters, init_log_param_scale) #don't update scale
#TODO: remove scale from gradient, then?
exact_metagrad = VectorParser()
exact_metagrad['log_L2_reg'] = fill_parser(parser, hyperparams['log_L2_reg']) #np.zeros(N_weight_types)
exact_metagrad['log_param_scale'] = fill_parser(parser, fixed_hyperparams['log_param_scale']) #np.zeros(N_weight_types)
exact_metagrad['log_alphas'] = np.zeros(N_iters)
exact_metagrad['invlogit_betas'] = np.zeros(N_iters)
exact_metagrad2 = VectorParser()
exact_metagrad2['log_L2_reg'] = np.zeros(N_weight_types)
exact_metagrad2['log_param_scale'] = np.zeros(N_weight_types)
exact_metagrad2['log_alphas'] = np.zeros(N_iters)
exact_metagrad2['invlogit_betas'] = np.zeros(N_iters)
#exact_metagrad = exact_metagradV.vect
#print(hyperparams.vect)
#exact_metagrad = [np.zeros(N_weight_types), np.zeros(N_weight_types), np.zeros(N_iters), np.zeros(N_iters)] #initialize
# TODO: memoize
def primal_optimizer(hyperparam_vect, i_hyper):
def indexed_loss_fun(w, L2_vect, i_iter):
rs = npr.RandomState(npr.RandomState(global_seed + i_hyper + i_iter * 10000).randint(1000))
seed = i_hyper * 10**6 + i_iter # Deterministic seed needed for backwards pass.
idxs = rs.randint(N_train, size=batch_size)
return loss_fun(w, train_data['X'][idxs], train_data['T'][idxs], L2_vect)
learning_curve_dict = defaultdict(list)
def callback(x, v, g, i_iter):
if i_iter % thin == 0: # N_batches=10 times
learning_curve_dict['learning_curve'].append(loss_fun(x, **train_data))
learning_curve_dict['grad_norm'].append(np.linalg.norm(g))
learning_curve_dict['weight_norm'].append(np.linalg.norm(x))
learning_curve_dict['velocity_norm'].append(np.linalg.norm(v))
cur_hyperparams = hyperparams.new_vect(hyperparam_vect)
# TODO: why doesn't the following line work with N_iter=1?
W0 = fill_parser(parser, np.exp(fixed_hyperparams['log_param_scale'])) #don't update scale
W0 *= npr.RandomState(global_seed + i_hyper).randn(W0.size)
# TODO: Put on proper scale; no SGD on log/invlogit scale
alphas = np.exp(cur_hyperparams['log_alphas'])
betas = logit(cur_hyperparams['invlogit_betas'])
# TODO: check this
L2_reg = fill_parser(parser, np.exp(cur_hyperparams['log_L2_reg']))
W_opt = sgd4(grad(indexed_loss_fun), kylist(W0, alphas, betas, L2_reg), exact_metagrad, callback)
#W_opt = sgd4(grad(indexed_loss_fun), kylist(W0, alphas, betas, L2_reg), callback)
#callback(W_opt, N_iters)
return W_opt, learning_curve_dict
def hyperloss(hyperparam_vect, i_hyper):
W_opt, _ = primal_optimizer(hyperparam_vect, i_hyper)
return loss_fun(W_opt, **valid_data)
hyperloss_grad = grad(hyperloss)
# TODO: This is where the chain rule happens, dhyperloss/dW_opt x dW_opt/dhyperparam_vect; first term is SGD
meta_results = defaultdict(list)
old_metagrad = [np.ones(hyperparams.vect.size)]
#def meta_callback(hyperparam_vect, i_hyper, metagrad):
def meta_callback(hyperparam_vect, i_hyper, metagrad, exact_metagrad=exact_metagrad):
x, learning_curve_dict = primal_optimizer(hyperparam_vect, i_hyper)
cur_hyperparams = hyperparams.new_vect(hyperparam_vect.copy())
for field in cur_hyperparams.names:
meta_results[field].append(cur_hyperparams[field])
# these are the unregularized losses below; default sets L2_reg=0.0
meta_results['train_loss'].append(loss_fun(x, **train_data))
meta_results['valid_loss'].append(loss_fun(x, **valid_data))
meta_results['tests_loss'].append(loss_fun(x, **tests_data))
meta_results['train_err'].append(frac_err(x, **train_data))
meta_results['valid_err'].append(frac_err(x, **valid_data))
meta_results['test_err'].append(frac_err(x, **tests_data))
meta_results['learning_curves'].append(learning_curve_dict)
print("metagrad", len(metagrad))
meta_results['meta_grad_magnitude'].append(np.linalg.norm(metagrad))
meta_results['meta_grad_angle'].append(np.dot(old_metagrad[0], metagrad) \
/ (np.linalg.norm(metagrad)*
np.linalg.norm(old_metagrad[0])))
#Michael: added comparisons with exact metagrad here
#(2) Angle condition: More strongly, is the cosine of the angle between the two strictly bounded away from 0?
#(3) Length: Since hypergradient optimization procedures do not necessarily use a proper line search, it may also be important for the approximate hypergradient to have a length comparable to the true hypergradient.
exact_metagrad2['log_L2_reg'] = [sum(exact_metagrad['log_L2_reg'][0:7840]), sum(exact_metagrad['log_L2_reg'][7840:7850])]
exact_metagrad2['log_param_scale'] = [sum(exact_metagrad['log_param_scale'][0:7840]), sum(exact_metagrad['log_param_scale'][7840:7850])]
exact_metagrad2['log_alphas'] = exact_metagrad['log_alphas']
exact_metagrad2['invlogit_betas'] = exact_metagrad['invlogit_betas']
meta_results['exact_meta_grad_magnitude'].append(np.linalg.norm(exact_metagrad2.vect))
meta_results['DrMAD_exact_angle'].append(np.dot(exact_metagrad2.vect, metagrad) \
/ (np.linalg.norm(metagrad)*
np.linalg.norm(exact_metagrad2.vect)))
#TODO: do the above for parameters separately? E.g. check log_alphas separately
old_metagrad[0] = metagrad
print "Meta Epoch {0} Train loss {1:2.4f} Valid Loss {2:2.4f}" \
" Test Loss {3:2.4f} Test Err {4:2.4f}".format(
i_hyper, meta_results['train_loss'][-1], meta_results['valid_loss'][-1],
meta_results['tests_loss'][-1], meta_results['test_err'][-1]) #Michael: train->tests
# final_result = adam(hyperloss_grad, hyperparams.vect,
# meta_callback, N_meta_iter, meta_alpha)
final_result = adam(hyperloss_grad, hyperparams.vect, exact_metagrad,
meta_callback, N_meta_iter, meta_alpha)
#write modified adam to ignore exact hypergrad in sgd4_mad_with_exact
#meta_callback(final_result, N_meta_iter)
parser.vect = None # No need to pickle zeros
return meta_results, parser
if callback:
callback(x, i, g)
m = b1t * g + (1 - b1t) * m
v = b2 * (g**2) + (1 - b2) * v
mhat = m / (1 - (1 - b1)**(i + 1))
vhat = v / (1 - (1 - b2)**(i + 1))
x -= step_size * mhat / (np.sqrt(vhat) + eps)
return x
# --
# Make NN functions
parser = VectorParser()
for i, shape in enumerate(zip(layer_sizes[:-1], layer_sizes[1:])):
parser.add_shape(('weights', i), shape)
parser.add_shape(('biases', i), (1, shape[1]))
def pred_fun(W_vect, X):
"""Outputs normalized log-probabilities."""
W = parser.new_vect(W_vect)
cur_units = X
N_iter = len(layer_sizes) - 1
for i in range(N_iter):
cur_W = W[('weights', i)]
cur_B = W[('biases', i)]
cur_units = np.dot(cur_units, cur_W) + cur_B
if i == (N_iter - 1):
def run():
train_data, valid_data, tests_data = load_data_dicts(
N_train, N_valid, N_tests)
parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weight_types = len(parser.names)
def build_hypervect(init_log_alphas, init_invlogit_betas,
init_log_param_scale):
hyperparams = VectorParser()
hyperparams['log_param_scale'] = np.full(N_weight_types,
init_log_param_scale)
hyperparams['log_alphas'] = np.full((N_iters, N_weight_types),
init_log_alphas)
hyperparams['invlogit_betas'] = np.full((N_iters, N_weight_types),
init_invlogit_betas)
return hyperparams
hyperparams = build_hypervect(
init_log_alphas, init_invlogit_betas,
init_log_param_scale) # Build just for parser.
fixed_hyperparams = VectorParser()
fixed_hyperparams['log_L2_reg'] = np.full(N_weight_types, init_log_L2_reg)
def whetlab_optimize(loss, max_iters, callback):
for i in xrange(max_iters):
params = scientist.suggest()
hyperparams = build_hypervect(**params)
cur_loss = loss(hyperparams.vect, 0) # No randomness
scientist.update(params, -cur_loss)
if callback: callback(hyperparams.vect, 0)
def primal_optimizer(hyperparam_vect, i_hyper):
def indexed_loss_fun(w, L2_vect, i_iter):
rs = RandomState(
(seed, i_hyper,
i_iter)) # Deterministic seed needed for backwards pass.
idxs = rs.randint(N_train, size=batch_size)
return loss_fun(w, train_data['X'][idxs], train_data['T'][idxs],
L2_vect)
learning_curve_dict = defaultdict(list)
def callback(x, v, g, i_iter):
if i_iter % thin == 0:
learning_curve_dict['learning_curve'].append(
loss_fun(x, **train_data))
learning_curve_dict['grad_norm'].append(np.linalg.norm(g))
learning_curve_dict['weight_norm'].append(np.linalg.norm(x))
learning_curve_dict['velocity_norm'].append(np.linalg.norm(v))
cur_hyperparams = hyperparams.new_vect(hyperparam_vect)
rs = RandomState((seed, i_hyper))
W0 = fill_parser(parser, np.exp(cur_hyperparams['log_param_scale']))
W0 *= rs.randn(W0.size)
alphas = np.exp(cur_hyperparams['log_alphas'])
betas = logit(cur_hyperparams['invlogit_betas'])
L2_reg = fill_parser(parser, np.exp(fixed_hyperparams['log_L2_reg']))
W_opt = sgd_parsed(grad(indexed_loss_fun),
kylist(W0, alphas, betas, L2_reg),
parser,
callback=callback)
return W_opt, learning_curve_dict
def hyperloss(hyperparam_vect, i_hyper):
W_opt, _ = primal_optimizer(hyperparam_vect, i_hyper)
return loss_fun(W_opt, **train_data)
meta_results = defaultdict(list)
old_metagrad = [np.ones(hyperparams.vect.size)]
def meta_callback(hyperparam_vect, i_hyper, metagrad=None):
x, learning_curve_dict = primal_optimizer(hyperparam_vect, i_hyper)
cur_hyperparams = hyperparams.new_vect(hyperparam_vect.copy())
for field in cur_hyperparams.names:
meta_results[field].append(cur_hyperparams[field])
meta_results['train_loss'].append(loss_fun(x, **train_data))
meta_results['valid_loss'].append(loss_fun(x, **valid_data))
meta_results['tests_loss'].append(loss_fun(x, **tests_data))
meta_results['test_err'].append(frac_err(x, **tests_data))
meta_results['learning_curves'].append(learning_curve_dict)
if metagrad is not None:
meta_results['meta_grad_magnitude'].append(
np.linalg.norm(metagrad))
meta_results['meta_grad_angle'].append(np.dot(old_metagrad[0], metagrad) \
/ (np.linalg.norm(metagrad)*
np.linalg.norm(old_metagrad[0])))
old_metagrad[0] = metagrad
print "Meta Epoch {0} Train loss {1:2.4f} Valid Loss {2:2.4f}" \
" Test Loss {3:2.4f} Test Err {4:2.4f}".format(
i_hyper, meta_results['train_loss'][-1], meta_results['valid_loss'][-1],
meta_results['train_loss'][-1], meta_results['test_err'][-1])
whetlab_optimize(hyperloss, N_meta_iter, meta_callback)
best_params = scientist.best()
print "best params:", best_params
parser.vect = None # No need to pickle zeros
return meta_results, parser, best_params
# Helpers
def fill_parser(parser, items):
partial_vects = [np.full(parser[name].size, items[i]) for i, name in enumerate(parser.names)]
return np.concatenate(partial_vects, axis=0)
# --
# Run
train_data, valid_data, test_data = load_data_dicts(N_train, N_valid, N_tests)
parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weight_types = len(parser.names)
hyperparams = VectorParser()
hyperparams['log_param_scale'] = np.full(N_weight_types, init_log_param_scale)
hyperparams['log_alphas'] = np.full((N_iters, N_weight_types), init_log_alphas)
hyperparams['invlogit_betas'] = np.full((N_iters, N_weight_types), init_invlogit_betas)
fixed_hyperparams = VectorParser()
fixed_hyperparams['log_L2_reg'] = np.full(N_weight_types, init_log_L2_reg)
cur_primal_results = {}
def primal_optimizer(hyperparam_vect, i_hyper):
def indexed_loss_fun(w, L2_vect, i_iter):
rs = RandomState((seed, i_hyper, i_iter)) # Deterministic seed needed for backwards pass.
idxs = rs.randint(N_train, size=batch_size)
return loss_fun(w, train_data['X'][idxs], train_data['T'][idxs], L2_vect)
def run():
train_data, valid_data, tests_data = load_data_dicts(N_train, N_valid, N_tests)
parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weight_types = len(parser.names)
hyperparams = VectorParser()
hyperparams['log_param_scale'] = np.full(N_weight_types, init_log_param_scale)
hyperparams['log_alphas'] = np.full((N_iters, N_weight_types), init_log_alphas)
hyperparams['invlogit_betas'] = np.full((N_iters, N_weight_types), init_invlogit_betas)
fixed_hyperparams = VectorParser()
fixed_hyperparams['log_L2_reg'] = np.full(N_weight_types, init_log_L2_reg)
def primal_optimizer(hyperparam_vect, i_hyper):
def indexed_loss_fun(w, L2_vect, i_iter):
rs = RandomState((seed, i_hyper, i_iter)) # Deterministic seed needed for backwards pass.
idxs = rs.randint(N_train, size=batch_size)
return loss_fun(w, train_data['X'][idxs], train_data['T'][idxs], L2_vect)
learning_curve_dict = defaultdict(list)
def callback(x, v, g, i_iter):
if i_iter % thin == 0:
learning_curve_dict['learning_curve'].append(loss_fun(x, **train_data))
learning_curve_dict['grad_norm'].append(np.linalg.norm(g))
learning_curve_dict['weight_norm'].append(np.linalg.norm(x))
learning_curve_dict['velocity_norm'].append(np.linalg.norm(v))
init_hyperparams = hyperparams.new_vect(hyperparam_vect)
rs = RandomState((seed, i_hyper))
W0 = fill_parser(parser, np.exp(init_hyperparams['log_param_scale']))
W0 *= rs.randn(W0.size)
alphas = np.exp(init_hyperparams['log_alphas'])
betas = logit(init_hyperparams['invlogit_betas'])
L2_reg = fill_parser(parser, np.exp(fixed_hyperparams['log_L2_reg']))
W_opt = sgd_parsed(grad(indexed_loss_fun), kylist(W0, alphas, betas, L2_reg),
parser, callback=callback)
return W_opt, learning_curve_dict
def hyperloss(hyperparam_vect, i_hyper):
W_opt, _ = primal_optimizer(hyperparam_vect, i_hyper)
return loss_fun(W_opt, **train_data)
hyperloss_grad = grad(hyperloss)
meta_results = defaultdict(list)
old_metagrad = [np.ones(hyperparams.vect.size)]
def meta_callback(hyperparam_vect, i_hyper, metagrad=None):
x, learning_curve_dict = primal_optimizer(hyperparam_vect, i_hyper)
cur_hyperparams = hyperparams.new_vect(hyperparam_vect.copy())
for field in cur_hyperparams.names:
meta_results[field].append(cur_hyperparams[field])
meta_results['train_loss'].append(loss_fun(x, **train_data))
meta_results['valid_loss'].append(loss_fun(x, **valid_data))
meta_results['tests_loss'].append(loss_fun(x, **tests_data))
meta_results['test_err'].append(frac_err(x, **tests_data))
meta_results['learning_curves'].append(learning_curve_dict)
if metagrad is not None:
meta_results['meta_grad_magnitude'].append(np.linalg.norm(metagrad))
meta_results['meta_grad_angle'].append(np.dot(old_metagrad[0], metagrad) \
/ (np.linalg.norm(metagrad)*
np.linalg.norm(old_metagrad[0])))
old_metagrad[0] = metagrad
print "Meta Epoch {0} Train loss {1:2.4f} Valid Loss {2:2.4f}" \
" Test Loss {3:2.4f} Test Err {4:2.4f}".format(
i_hyper, meta_results['train_loss'][-1], meta_results['valid_loss'][-1],
meta_results['train_loss'][-1], meta_results['test_err'][-1])
# Average many gradient evaluations at the initial point.
hypergrads = np.zeros((N_gradients_in_average, hyperparams.vect.size))
for i in xrange(N_gradients_in_average):
hypergrads[i] = hyperloss_grad(hyperparams.vect, i)
print i
first_gradient = hypergrads[0]
avg_gradient = np.mean(hypergrads, axis=0)
# Now do a line search along that direction.
parsed_avg_grad = hyperparams.new_vect(avg_gradient)
stepsize_scale = stepsize_search_rescale/np.max(np.exp(parsed_avg_grad['log_alphas'].ravel()))
stepsizes = np.linspace(-stepsize_scale, stepsize_scale, N_points_in_line_search)
for i, stepsize in enumerate(stepsizes):
cur_hypervect = hyperparams.vect - stepsize * avg_gradient
meta_callback(cur_hypervect, 0) # Use the same random seed every time.
parser.vect = None # No need to pickle zeros
return meta_results, parser, first_gradient, parsed_avg_grad, stepsizes
def run(script_corr):
"""Three different parsers:
w_parser[('biases', i_layer)] : neural net weights/biases per layer for a single script
script_parser[i_script] : weights vector for each script
transform_parser[i_layer] : transform matrix (scripts x scripts) for each alphabet"""
RS = RandomState((seed, "top_rs"))
train_data, valid_data, tests_data = omniglot.load_data_split(
[11, 2, 2], RS, num_alphabets=N_scripts)
w_parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weights = w_parser.vect.size
uncorrelated_mat = np.eye(N_scripts)
fully_correlated_mat = np.full((N_scripts, N_scripts), 1.0 / N_scripts)
transform_mat = (1 - script_corr
) * uncorrelated_mat + script_corr * fully_correlated_mat
transform_mat = transform_mat
transform_parser = VectorParser()
for i_layer in range(N_layers):
if i_layer == N_layers - 1:
transform_parser[i_layer] = uncorrelated_mat
else:
transform_parser[i_layer] = transform_mat
script_parser = VectorParser()
for i_script in range(N_scripts):
script_parser[i_script] = np.zeros(N_weights)
def transform_weights(all_z_vect, transform_vect, i_script_out):
all_z = script_parser.new_vect(all_z_vect)
transform = transform_parser.new_vect(transform_vect)
W = OrderedDict(
) # Can't use parser because setting plain array ranges with funkyyak nodes not yet supported
for k in w_parser.idxs_and_shapes.keys():
W[k] = 0.0
for i_layer in range(N_layers):
script_weightings = transform[i_layer][i_script_out, :]
for i_script in range(N_scripts):
z_i_script = w_parser.new_vect(all_z[i_script])
script_weighting = script_weightings[i_script]
W[('biases',
i_layer)] += z_i_script[('biases',
i_layer)] * script_weighting
W[('weights',
i_layer)] += z_i_script[('weights',
i_layer)] * script_weighting
return np.concatenate([v.ravel() for v in W.values()])
def loss_from_latents(z_vect, transform_vect, i_script, data):
w_vect = transform_weights(z_vect, transform_vect, i_script)
return loss_fun(w_vect, **data)
def regularization(z_vect):
return np.dot(z_vect, z_vect) * np.exp(log_L2_init)
results = defaultdict(list)
def hyperloss(transform_vect, i_hyper, record_results=False):
def primal_stochastic_loss(z_vect, transform_vect, i_primal):
RS = RandomState((seed, i_hyper, i_primal))
loss = 0.0
for _ in range(N_scripts_per_iter):
i_script = RS.randint(N_scripts)
N_train = train_data[i_script]['X'].shape[0]
idxs = RS.permutation(N_train)[:batch_size]
minibatch = dictslice(train_data[i_script], idxs)
loss += loss_from_latents(z_vect, transform_vect, i_script,
minibatch)
reg = regularization(z_vect)
if i_primal % 20 == 0:
print "Iter {0}, loss {1}, reg {2}".format(
i_primal, getval(loss), getval(reg))
print "Full losses: train: {0}, valid: {1}".format(
total_loss(train_data, getval(z_vect)),
total_loss(valid_data, getval(z_vect)))
return loss + reg
def total_loss(data, z_vect):
return np.mean([
loss_from_latents(z_vect, transform_vect, i_script,
data[i_script])
for i_script in range(N_scripts)
])
z_vect_0 = RS.randn(
script_parser.vect.size) * np.exp(log_initialization_scale)
z_vect_final = sgd(grad(primal_stochastic_loss),
transform_vect,
z_vect_0,
alpha,
beta,
N_iters,
callback=None)
valid_loss = total_loss(valid_data, z_vect_final)
if record_results:
results['valid_loss'].append(valid_loss)
results['train_loss'].append(total_loss(train_data, z_vect_final))
# results['tests_loss'].append(total_loss(tests_data, z_vect_final))
return valid_loss
hyperloss(transform_parser.vect, 0, record_results=True)
return results['train_loss'][-1], results['valid_loss'][-1]
def run():
train_data, valid_data, tests_data = load_data_dicts(
N_train, N_valid, N_tests)
parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weight_types = len(parser.names)
rs = RandomState((seed))
init_fake_data = rs.randn(*(train_data['X'].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_train)) % N_classes,
N_classes) # One of each.
hyperparams = VectorParser()
hyperparams['fake_data'] = init_fake_data
fixed_hyperparams = VectorParser()
fixed_hyperparams['log_param_scale'] = np.full(N_weight_types,
init_log_param_scale)
fixed_hyperparams['log_alphas'] = np.full((N_iters, N_weight_types),
init_log_alphas)
fixed_hyperparams['invlogit_betas'] = np.full((N_iters, N_weight_types),
init_invlogit_betas)
fixed_hyperparams['log_L2_reg'] = np.full(N_weight_types, init_log_L2_reg)
cur_primal_results = {}
loss_meta_parser = VectorParser()
loss_meta_parser['']
def primal_optimizer(hyperparam_vect, i_hyper):
def indexed_loss_fun(w, meta_vect, i_iter):
(train_data, train_labels, L2_vect) = meta
return loss_fun(w, train_data, train_labels, L2_vect)
#return loss_fun(w, train_data['X'], train_data['T'], L2_vect + np.sum(fake_data.ravel()))
learning_curve_dict = defaultdict(list)
def callback(x, v, g, i_iter):
if i_iter % thin == 0:
# learning_curve_dict['learning_curve'].append(loss_fun(x, getval(cur_hyperparams['fake_data']), fake_labels))
learning_curve_dict['grad_norm'].append(np.linalg.norm(g))
learning_curve_dict['weight_norm'].append(np.linalg.norm(x))
learning_curve_dict['velocity_norm'].append(np.linalg.norm(v))
cur_hyperparams = hyperparams.new_vect(hyperparam_vect)
fake_data = cur_hyperparams['fake_data']
rs = RandomState((seed, i_hyper))
W0 = fill_parser(parser, np.exp(fixed_hyperparams['log_param_scale']))
W0 *= rs.randn(W0.size)
alphas = np.exp(fixed_hyperparams['log_alphas'])
betas = logit(fixed_hyperparams['invlogit_betas'])
L2_reg = fill_parser(parser, np.exp(fixed_hyperparams['log_L2_reg']))
meta = kylist(fake_data, fake_labels, L2_reg)
W_opt = sgd_parsed(grad(indexed_loss_fun),
kylist(W0, alphas, betas, meta),
parser,
callback=callback)
cur_primal_results['weights'] = getval(W_opt).copy()
cur_primal_results['learning_curve'] = getval(learning_curve_dict)
return W_opt, learning_curve_dict
def hyperloss(hyperparam_vect, i_hyper):
W_opt, _ = primal_optimizer(hyperparam_vect, i_hyper)
return loss_fun(W_opt, **valid_data)
hyperloss_grad = grad(hyperloss)
meta_results = defaultdict(list)
old_metagrad = [np.ones(hyperparams.vect.size)]
def meta_callback(hyperparam_vect, i_hyper, metagrad=None):
x, learning_curve_dict = cur_primal_results[
'weights'], cur_primal_results['learning_curve']
cur_hyperparams = hyperparams.new_vect(hyperparam_vect.copy())
for field in cur_hyperparams.names:
meta_results[field].append(cur_hyperparams[field])
#meta_results['train_loss'].append(loss_fun(x, getval(cur_hyperparams['fake_data']), fake_labels))
meta_results['train_loss'].append(0)
meta_results['valid_loss'].append(loss_fun(x, **valid_data))
meta_results['tests_loss'].append(loss_fun(x, **tests_data))
meta_results['test_err'].append(frac_err(x, **tests_data))
meta_results['learning_curves'].append(learning_curve_dict)
meta_results['example_weights'] = x
if metagrad is not None:
print metagrad
meta_results['meta_grad_magnitude'].append(
np.linalg.norm(metagrad))
meta_results['meta_grad_angle'].append(np.dot(old_metagrad[0], metagrad) \
/ (np.linalg.norm(metagrad)*
np.linalg.norm(old_metagrad[0])))
old_metagrad[0] = metagrad
print "Meta Epoch {0} Train loss {1:2.4f} Valid Loss {2:2.4f}" \
" Test Loss {3:2.4f} Test Err {4:2.4f}".format(
i_hyper, meta_results['train_loss'][-1], meta_results['valid_loss'][-1],
meta_results['tests_loss'][-1], meta_results['test_err'][-1])
final_result = adam(hyperloss_grad, hyperparams.vect, meta_callback,
N_meta_iter, meta_alpha)
meta_callback(final_result, N_meta_iter)
parser.vect = None # No need to pickle zeros
return meta_results, parser