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():
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['invlogit_betas'] = np.full(N_iters, init_invlogit_betas)
# 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 = []
def callback(x, 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(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']))
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)
return loss_fun(W_opt, **train_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 = 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, gamma=0.0)
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)
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():
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
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 == N_iters or i_iter == 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))
learning_curve_dict['iteration'].append(i_iter + 1)
print "iteration", i_iter
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
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)
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])
meta_callback(hyperparams.vect, 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)
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)
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_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)
for name in parser.names:
hyperparams[("rescale", name)] = np.full(N_iters, init_rescales)
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)
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],
)
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
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():
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
meta_alpha = 0.04
N_meta_iter = 50
seed = 0
# --
# 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 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
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():
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