def adam(nn, X_train, y_train, val_set=None, alpha=0.001, mb_size=256, n_iter=2000, print_after=100):
M = {k: np.zeros_like(v) for k, v in nn.model.items()}
R = {k: np.zeros_like(v) for k, v in nn.model.items()}
beta1 = .9
beta2 = .999
minibatches = get_minibatch(X_train, y_train, mb_size)
if val_set:
X_val, y_val = val_set
for iter in range(1, n_iter + 1):
t = iter
idx = np.random.randint(0, len(minibatches))
X_mini, y_mini = minibatches[idx]
grad, loss = nn.train_step(X_mini, y_mini)
if iter % print_after == 0:
if val_set:
val_acc = util.accuracy(y_val, nn.predict(X_val))
print('Iter-{} loss: {:.4f} validation: {:4f}'.format(iter, loss, val_acc))
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
for k in grad:
M[k] = util.exp_running_avg(M[k], grad[k], beta1)
R[k] = util.exp_running_avg(R[k], grad[k]**2, beta2)
m_k_hat = M[k] / (1. - beta1**(t))
r_k_hat = R[k] / (1. - beta2**(t))
nn.model[k] -= alpha * m_k_hat / (np.sqrt(r_k_hat) + c.eps)
return nn
def adam_rnn(nn,
X_train,
y_train,
alpha=0.001,
mb_size=256,
n_iter=2000,
print_after=100):
M = {k: np.zeros_like(v) for k, v in nn.model.items()}
R = {k: np.zeros_like(v) for k, v in nn.model.items()}
beta1 = .9
beta2 = .999
minibatches = get_minibatch(X_train, y_train, mb_size, shuffle=False)
idx = 0
state = nn.initial_state()
smooth_loss = -np.log(1.0 / len(set(X_train)))
for iter in range(1, n_iter + 1):
t = iter
if idx >= len(minibatches):
idx = 0
state = nn.initial_state()
X_mini, y_mini = minibatches[idx]
idx += 1
if iter % print_after == 0:
print(
"========================================================================="
)
print('Iter-{} loss: {:.4f}'.format(iter, smooth_loss))
print(
"========================================================================="
)
sample = nn.sample(X_mini[0], state, 100)
print(sample)
print(
"========================================================================="
)
print()
print()
grad, loss, state = nn.train_step(X_mini, y_mini, state)
smooth_loss = 0.999 * smooth_loss + 0.001 * loss
for k in grad:
M[k] = util.exp_running_avg(M[k], grad[k], beta1)
R[k] = util.exp_running_avg(R[k], grad[k]**2, beta2)
m_k_hat = M[k] / (1. - beta1**(t))
r_k_hat = R[k] / (1. - beta2**(t))
nn.model[k] -= alpha * m_k_hat / (np.sqrt(r_k_hat) + c.eps)
return nn
def rmsprop(nn, X_train, y_train, val_set=None, alpha=1e-3, mb_size=256, n_iter=2000, print_after=100):
cache = {k: np.zeros_like(v) for k, v in nn.model.items()}
gamma = .9
minibatches = get_minibatch(X_train, y_train, mb_size)
if val_set:
X_val, y_val = val_set
for iter in range(1, n_iter + 1):
idx = np.random.randint(0, len(minibatches))
X_mini, y_mini = minibatches[idx]
grad, loss = nn.train_step(X_mini, y_mini)
if iter % print_after == 0:
if val_set:
val_acc = util.accuracy(y_val, nn.predict(X_val))
print('Iter-{} loss: {:.4f} validation: {:4f}'.format(iter, loss, val_acc))
print('grad:',grad)
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
print('grad:',grad)
for k in grad:
cache[k] = util.exp_running_avg(cache[k], grad[k]**2, gamma)
nn.model[k] -= alpha * grad[k] / (np.sqrt(cache[k]) + c.eps)
return nn
def rmsprop(nn, X_train, y_train, val_set=None, alpha=1e-3, mb_size=256, n_iter=2000, print_after=100):
cache = {k: np.zeros_like(v) for k, v in nn.model.items()}
gamma = .9
minibatches = get_minibatch(X_train, y_train, mb_size)
if val_set:
X_val, y_val = val_set
for iter in range(1, n_iter + 1):
idx = np.random.randint(0, len(minibatches))
X_mini, y_mini = minibatches[idx]
grad, loss = nn.train_step(X_mini, y_mini)
if iter % print_after == 0:
if val_set:
val_acc = util.accuracy(y_val, nn.predict(X_val))
print('Iter-{} loss: {:.4f} validation: {:4f}'.format(iter, loss, val_acc))
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
for k in grad:
cache[k] = util.exp_running_avg(cache[k], grad[k]**2, gamma)
nn.model[k] -= alpha * grad[k] / (np.sqrt(cache[k]) + c.eps)
return nn
def adam_rnn(nn, X_train, y_train, alpha=0.001, mb_size=256, n_iter=2000, print_after=100):
M = {k: np.zeros_like(v) for k, v in nn.model.items()}
R = {k: np.zeros_like(v) for k, v in nn.model.items()}
beta1 = .9
beta2 = .999
minibatches = get_minibatch(X_train, y_train, mb_size, shuffle=False)
idx = 0
state = nn.initial_state()
smooth_loss = -np.log(1.0 / len(set(X_train)))
for iter in range(1, n_iter + 1):
t = iter
if idx >= len(minibatches):
idx = 0
state = nn.initial_state()
X_mini, y_mini = minibatches[idx]
idx += 1
if iter % print_after == 0:
print("=========================================================================")
print('Iter-{} loss: {:.4f}'.format(iter, smooth_loss))
print("=========================================================================")
sample = nn.sample(X_mini[0], state, 100)
print(sample)
print("=========================================================================")
print()
print()
grad, loss, state = nn.train_step(X_mini, y_mini, state)
smooth_loss = 0.999 * smooth_loss + 0.001 * loss
for k in grad:
M[k] = util.exp_running_avg(M[k], grad[k], beta1)
R[k] = util.exp_running_avg(R[k], grad[k]**2, beta2)
m_k_hat = M[k] / (1. - beta1**(t))
r_k_hat = R[k] / (1. - beta2**(t))
nn.model[k] -= alpha * m_k_hat / (np.sqrt(r_k_hat) + c.eps)
return nn
def bn_forward(X, gamma, beta, cache, momentum=.9, train=True):
running_mean, running_var = cache
if train:
mu = np.mean(X, axis=0)
var = np.var(X, axis=0)
X_norm = (X - mu) / np.sqrt(var + c.eps)
out = gamma * X_norm + beta
cache = (X, X_norm, mu, var, gamma, beta)
running_mean = util.exp_running_avg(running_mean, mu, momentum)
running_var = util.exp_running_avg(running_var, var, momentum)
else:
X_norm = (X - running_mean) / np.sqrt(running_var + c.eps)
out = gamma * X_norm + beta
cache = None
return out, cache, running_mean, running_var
def bn_forward(X, gamma, beta, cache, momentum=.9, train=True):
running_mean, running_var = cache
if train:
mu = np.mean(X, axis=0)
var = np.var(X, axis=0)
X_norm = (X - mu) / np.sqrt(var + c.eps)
out = gamma * X_norm + beta
cache = (X, X_norm, mu, var, gamma, beta)
running_mean = util.exp_running_avg(running_mean, mu, momentum)
running_var = util.exp_running_avg(running_var, var, momentum)
else:
X_norm = (X - running_mean) / np.sqrt(running_var + c.eps)
out = gamma * X_norm + beta
cache = None
return out, cache, running_mean, running_var
