def embeddings_set_grads_update_d(folder, gamma):
'''
Update the value of the direction for each input to be gamma (given) times the old direction
minus grad_now
'''
grad_now_files = sorted(glob.glob(folder + '/*.grad_now.npy'))
grad_d_files = sorted(glob.glob(folder + '/*.grad_d.npy'))
for grad_now_file, grad_d_file in zip(grad_now_files, grad_d_files):
grad_now = load(grad_now_file)
grad_d = load(grad_d_file)
save(grad_d_file, -(grad_now + gamma * grad_d))
def embeddings_set_grads_update_d(folder, gamma):
'''
Update the value of the direction for each input to be gamma (given) times the old direction
minus grad_new
'''
global time_acc
start = time.time()
grad_new_files = sorted(glob.glob(folder + '/*.grad_new.npy'))
grad_d_files = sorted(glob.glob(folder + '/*.grad_d.npy'))
for grad_new_file, grad_d_file in zip(grad_new_files, grad_d_files):
grad_new = load(grad_new_file)
grad_d = load(grad_d_file)
save(grad_d_file, gamma * grad_d - grad_new)
end = time.time()
time_acc['embeddings_set_grads_update_d'] += [end - start]
def embeddings_get_grads_gamma(folder):
'''
Get the sum over the inputs of the inner product of grad_old and grad_new
'''
global time_acc
start = time.time()
gamma = 0
grad_new_files = sorted(glob.glob(folder + '/*.grad_new.npy'))
grad_old_files = sorted(glob.glob(folder + '/*.grad_old.npy'))
for grad_new_file, grad_old_file in zip(grad_new_files, grad_old_files):
grad_new = load(grad_new_file)
grad_old = load(grad_old_file)
gamma += (grad_new * grad_old).sum()
end = time.time()
time_acc['embeddings_get_grads_gamma'] += [end - start]
return gamma
def embeddings_set_grads(folder):
'''
Sets the grads and other local statistics often needed for optimisation locally for
each node. This is currently only implemented locally, but could easly be adapted
to the MapReduce framework to be done on remote nodes in parallel. There's no real
need to do this in parallel though, as the computaions taking place are not that
time consuming.
'''
global time_acc
start = time.time()
input_files = sorted(glob.glob(folder + '/*.grad_latest.npy'))
for file_name in input_files:
grads = load(file_name)
#print 'grads'
#print grads
# Save grad new as the latest grad evaluated
new_file = splitext(splitext(file_name)[0])[0] + '.grad_new.npy'
save(new_file, grads)
# Init the old grad to be grad new
new_file = splitext(splitext(file_name)[0])[0] + '.grad_old.npy'
save(new_file, grads)
# Save the direction as the negative grad
new_file = splitext(splitext(file_name)[0])[0] + '.grad_d.npy'
save(new_file, -1 * grads)
end = time.time()
time_acc['embeddings_set_grads'] += [end - start]
def embeddings_get_grads_mu(folder):
'''
Get the sum over the inputs of the inner product of the direction and grad_new
'''
global time_acc
start = time.time()
mu = 0
grad_new_files = sorted(glob.glob(folder + '/*.grad_new.npy'))
grad_d_files = sorted(glob.glob(folder + '/*.grad_d.npy'))
for grad_new_file, grad_d_file in zip(grad_new_files, grad_d_files):
grad_new = load(grad_new_file)
grad_d = load(grad_d_file)
mu += (grad_new * grad_d).sum()
end = time.time()
time_acc['embeddings_get_grads_mu'] += [end - start]
return mu
def embeddings_get_grads_max_gradnow(folder):
'''
Get the max abs element of the direction over all input files
'''
max_gradnow = 0
grad_now_files = sorted(glob.glob(folder + '/*.grad_now.npy'))
for grad_now_file in grad_now_files:
grad_now = load(grad_now_file)
max_gradnow = max(max_gradnow, numpy.max(numpy.abs(grad_now)))
return max_gradnow
def embeddings_get_grads_current_grad(folder):
'''
Get the sum over the inputs of the inner product of grad_now with itself
'''
current_grad = 0
grad_now_files = sorted(glob.glob(folder + '/*.grad_now.npy'))
for grad_now_file in grad_now_files:
grad_now = load(grad_now_file)
current_grad += numpy.sum(numpy.abs(grad_now))
return current_grad
def embeddings_get_grads_theta(folder):
'''
Get the sum over the inputs of the inner product of the direction and grad_latest
'''
global time_acc
start = time.time()
theta = 0
grad_new_files = sorted(glob.glob(folder + '/*.grad_new.npy'))
grad_latest_files = sorted(glob.glob(folder + '/*.grad_latest.npy'))
grad_d_files = sorted(glob.glob(folder + '/*.grad_d.npy'))
for grad_latest_file, grad_d_file, grad_new_file in zip(
grad_latest_files, grad_d_files, grad_new_files):
grad_latest = load(grad_latest_file)
grad_new = load(grad_new_file)
grad_d = load(grad_d_file)
theta += (grad_d * (grad_latest - grad_new)).sum()
end = time.time()
time_acc['embeddings_get_grads_theta'] += [end - start]
return theta
def embeddings_set_grads_update_grad_now(folder):
'''
Set grad_now to be grad_latest (a temp grad that keeps changing every evaluation)
'''
input_files = sorted(glob.glob(folder + '/*.grad_latest.npy'))
for file_name in input_files:
grads = load(file_name)
# Save grad old as latest grad new
new_file = splitext(splitext(file_name)[0])[0] + '.grad_now.npy'
save(new_file, grads)
def embeddings_set_grads_update_X(folder, step_size):
'''
Update the value of the local embeddings and variances themselves to be X + alpha * direction
'''
grad_d_files = sorted(glob.glob(folder + '/*.grad_d.npy'))
X_mu_files = sorted(glob.glob(folder + '/*.embedding.npy'))
X_S_files = sorted(glob.glob(folder + '/*.variance.npy'))
for grad_d_file, X_mu_file, X_S_file in zip(grad_d_files, X_mu_files,
X_S_files):
grad_d = load(grad_d_file)
grad_d_X_mu = grad_d[0]
grad_d_X_S = grad_d[1]
X_mu = load(X_mu_file)
X_S = load(X_S_file)
#print 'X_mu'
#print X_mu
#print 'X_S'
#print X_S
save(X_mu_file, X_mu + step_size * grad_d_X_mu)
save(X_S_file, X_S + step_size * grad_d_X_S)
def embeddings_set_grads_reset_d(folder):
'''
Reset the direction to be the negative of grad_new
'''
global time_acc
start = time.time()
input_files = sorted(glob.glob(folder + '/*.grad_new.npy'))
for file_name in input_files:
grads = load(file_name)
# Save the direction as the negative grad
new_file = splitext(splitext(file_name)[0])[0] + '.grad_d.npy'
save(new_file, -1 * grads)
end = time.time()
time_acc['embeddings_set_grads_reset_d'] += [end - start]
def embeddings_get_grads_kappa(folder):
'''
Get the sum over the inputs of the inner product of the direction with itself
'''
global time_acc
start = time.time()
kappa = 0
grad_d_files = sorted(glob.glob(folder + '/*.grad_d.npy'))
for grad_d_file in grad_d_files:
grad_d = load(grad_d_file)
kappa += (grad_d * grad_d).sum()
end = time.time()
time_acc['embeddings_get_grads_kappa'] += [end - start]
return kappa
def embeddings_get_grads_current_grad(folder):
'''
Get the sum over the inputs of the inner product of grad_new with itself
'''
global time_acc
start = time.time()
current_grad = 0
grad_new_files = sorted(glob.glob(folder + '/*.grad_new.npy'))
for grad_new_file in grad_new_files:
grad_new = load(grad_new_file)
current_grad += (grad_new * grad_new).sum()
end = time.time()
time_acc['embeddings_get_grads_current_grad'] += [end - start]
return current_grad
def embeddings_set_grads_update_grad_new(folder):
'''
Set grad_new to be grad_latest (a temp grad that keeps changing every evaluation)
'''
global time_acc
start = time.time()
input_files = sorted(glob.glob(folder + '/*.grad_latest.npy'))
for file_name in input_files:
grads = load(file_name)
# Save grad old as latest grad new
new_file = splitext(splitext(file_name)[0])[0] + '.grad_new.npy'
save(new_file, grads)
end = time.time()
time_acc['embeddings_set_grads_update_grad_new'] += [end - start]
def embeddings_set_grads_update_grad_old(folder):
'''
Set grad_old to be grad_new
'''
global time_acc
start = time.time()
input_files = sorted(glob.glob(folder + '/*.grad_new.npy'))
for file_name in input_files:
grads = load(file_name)
# Save grad old as latest grad new
new_file = splitext(splitext(file_name)[0])[0] + '.grad_old.npy'
save(new_file, grads)
end = time.time()
time_acc['embeddings_set_grads_update_grad_old'] += [end - start]
def embeddings_get_grads_max_d(folder, alpha):
'''
Get the max abs element of the direction over all input files
'''
global time_acc
start = time.time()
max_d = 0
grad_d_files = sorted(glob.glob(folder + '/*.grad_d.npy'))
for grad_d_file in grad_d_files:
grad_d = load(grad_d_file)
max_d = max(max_d, numpy.max(numpy.abs(alpha * grad_d)))
end = time.time()
time_acc['embeddings_get_grads_max_d'] += [end - start]
return max_d
def embeddings_set_grads_update_X(folder, alpha):
'''
Update the value of the local embeddings and variances themselves to be X + alpha * direction
'''
global time_acc
start = time.time()
grad_d_files = sorted(glob.glob(folder + '/*.grad_d.npy'))
X_mu_files = sorted(glob.glob(folder + '/*.embedding.npy'))
X_S_files = sorted(glob.glob(folder + '/*.variance.npy'))
for grad_d_file, X_mu_file, X_S_file in zip(grad_d_files, X_mu_files,
X_S_files):
grad_d = load(grad_d_file)
grad_d_X_mu = grad_d[0]
grad_d_X_S = grad_d[1]
X_mu = load(X_mu_file)
X_S = load(X_S_file)
#print 'X_mu'
#print X_mu
#print 'X_S'
#print X_S
save(X_mu_file, X_mu + alpha * grad_d_X_mu)
save(X_S_file, X_S + alpha * grad_d_X_S)
end = time.time()
time_acc['embeddings_set_grads_update_X'] += [end - start]
