def softmax_loss(x, y):
"""
Computes the loss and gradient for softmax classification.
Inputs:
- x: Input data, of shape (N, C) where x[i, j] is the score for the jth class
for the ith input.
- y: Vector of labels, of shape (N,) where y[i] is the label for x[i] and
0 <= y[i] < C
Returns a tuple of:
- loss: Scalar giving the loss
- dx: Gradient of the loss with respect to x
"""
#np.expand_dims(correct_class_scores, axis = 1)
#probs = np.exp(x - np.max(x, axis=1, keepdims=True))
#print "x.shape", x.shape
#Somehow Buggy. Max doesn't work.
probs = np.exp(x - np.expand_dims(np.max(x, axis=1), axis = 1))
probs /= np.expand_dims(np.sum(probs, axis=1), axis = 1)
N = x.shape[0]
loss = -np.sum(np.log(probs[np.arange(N), y])) / N
dx = probs.copy()
dx[np.arange(N), y] -= 1
dx /= N
return loss, dx
def softmax_loss(x, y):
"""
Computes the loss and gradient for softmax classification.
Inputs:
- x: Input data, of shape (N, C) where x[i, j] is the score for the jth class
for the ith input.
- y: Vector of labels, of shape (N,) where y[i] is the label for x[i] and
0 <= y[i] < C
Returns a tuple of:
- loss: Scalar giving the loss
- dx: Gradient of the loss with respect to x
"""
#np.expand_dims(correct_class_scores, axis = 1)
#probs = np.exp(x - np.max(x, axis=1, keepdims=True))
#print "x.shape", x.shape
#Somehow Buggy. Max doesn't work.
probs = np.exp(x - np.max(x, axis=1))
#probs /= np.expand_dims(np.sum(probs, axis=1), axis = 1)
probs /= np.expand_dims(np.sum(probs, axis=1), axis=1)
N = x.shape[0]
loss = -np.sum(np.log(probs[np.arange(N), y])) / N
dx = probs.copy()
dx[np.arange(N), y] -= 1
dx /= N
return loss, dx
def affine_forward(x, w, b): """ Computes the forward pass for an affine (fully-connected) layer. The input x has shape (N, d_1, ..., d_k) and contains a minibatch of N examples, where each example x[i] has shape (d_1, ..., d_k). We will reshape each input into a vector of dimension D = d_1 * ... * d_k, and then transform it to an output vector of dimension M. Inputs: - x: A numpy array containing input data, of shape (N, d_1, ..., d_k) - w: A numpy array of weights, of shape (D, M) - b: A numpy array of biases, of shape (M,) Returns a tuple of: - out: output, of shape (N, M) - cache: (x, w, b) """ x_plain = np.reshape(x, (x.shape[0], -1)) # Note: GPU has no automatically broadcast feature? out = np.dot(x_plain, w) + np.repeat(np.expand_dims(b, axis=0), x_plain.shape[0], axis = 0) cache = (x, w, b) return out, cache
def svm_loss(x, y, mode):
"""
Computes the loss and gradient using for multiclass SVM classification.
Inputs:
- x: Input data, of shape (N, C) where x[i, j] is the score for the jth class
for the ith input.
- y: Vector of labels, of shape (N,) where y[i] is the label for x[i] and
0 <= y[i] < C
Returns a tuple of:
- loss: Scalar giving the loss
- dx: Gradient of the loss with respect to x
"""
if mode == 'cpu':
np.set_policy(policy.OnlyNumpyPolicy())
else:
np.set_policy(policy.PreferMXNetPolicy())
N = x.shape[0]
correct_class_scores = x[np.arange(N), y]
#margins = np.maximum(0, x - correct_class_scores[:, np.newaxis] + 1.0)
margins = np.maximum(0, x - np.expand_dims(correct_class_scores, axis = 1) + 1.0)
margins[np.arange(N), y] = 0
loss = np.sum(margins) / N
num_pos = np.sum(margins > 0, axis=1)
dx = np.zeros_like(x)
dx[margins > 0] = 1
dx[np.arange(N), y] -= num_pos
dx /= N
return loss, dx
def softmax_loss(x, y):
"""
Computes the loss and gradient for softmax classification.
Inputs:
- x: Input data, of shape (N, C) where x[i, j] is the score for the jth class
for the ith input.
- y: Vector of labels, of shape (N,) where y[i] is the label for x[i] and
0 <= y[i] < C
Returns a tuple of:
- loss: Scalar giving the loss
"""
#TODO: Missing Max Operator
probs = np.exp(x - np.expand_dims(np.max(x, axis=1), axis=1))
probs = probs / np.expand_dims(np.sum(probs, axis=1), axis=1)
N = x.shape[0]
loss = -np.sum(np.log(probs[np.arange(N), y])) / N
return loss
def softmax_loss(x, y):
"""
Computes the loss and gradient for softmax classification.
Inputs:
- x: Input data, of shape (N, C) where x[i, j] is the score for the jth class
for the ith input.
- y: Vector of labels, of shape (N,) where y[i] is the label for x[i] and
0 <= y[i] < C
Returns a tuple of:
- loss: Scalar giving the loss
"""
#TODO: Missing Max Operator
probs = np.exp(x - np.expand_dims(np.max(x, axis=1), axis = 1))
probs = probs / np.expand_dims(np.sum(probs, axis=1), axis = 1)
N = x.shape[0]
loss = -np.sum(np.log(probs[np.arange(N), y])) / N
return loss
def preprocess(self, img):
""" Preprocess a 210x160x3 uint8 frame into a 6400 (80x80) (1 x input_size) float vector."""
# Crop, down-sample, erase background and set foreground to 1.
# Ref: https://gist.github.com/karpathy/a4166c7fe253700972fcbc77e4ea32c5
img = img[35:195]
img = img[::2, ::2, 0]
img[img == 144] = 0
img[img == 109] = 0
img[img != 0] = 1
curr = np.expand_dims(img.astype(numpy.float).ravel(), axis=0)
# Subtract the last preprocessed image.
diff = curr - self.prev if self.prev is not None else np.zeros((1, curr.shape[1]))
self.prev = curr
return diff
def run_episode(self):
"""Run an episode using the current model to generate training data.
Specifically, this involves repeatedly getting an observation from the environment,
performing a forward pass using the single observation to get a distribution over actions
(in binary case a probability of a single action), and choosing an action.
Finally, rewards are discounted when the episode completes.
Returns
-------
(xs, ys, rs) : tuple
The N x input_size observations, N x 1 action labels, and N x 1 discounted rewards
obtained from running the episode's N steps.
"""
observation = self.env.reset()
self.model.preprocessor.reset()
self.episode_reward = 0
xs, ys, rs = [], [], []
done = False
game_number = 1
game_start = time.time()
while not done:
if self.render:
self.env.render()
x = self.model.preprocessor.preprocess(observation)
p = self.model.forward(x)
a, y = self.model.choose_action(p.asnumpy().ravel()[0])
observation, r, done, info = self.env.step(a)
xs.append(x.asnumpy().ravel())
ys.append(y)
rs.append(r)
self.episode_reward += r
if self._game_complete(r):
game_time = time.time() - game_start
if self.verbose:
print('game %d complete (%.2fs), reward: %f' %
(game_number, game_time, r))
game_number += 1
game_start = time.time()
# Episode finished.
self.running_reward = self.episode_reward if self.running_reward is None else (
0.99 * self.running_reward + 0.01 * self.episode_reward)
xs = np.vstack(xs)
ys = np.vstack(ys)
rs = np.expand_dims(self.model.discount_rewards(rs), axis=1)
return xs, ys, rs
def preprocess(self, img):
""" Preprocess a 210x160x3 uint8 frame into a 6400 (80x80) (1 x input_size) float vector."""
# Crop, down-sample, erase background and set foreground to 1.
# Ref: https://gist.github.com/karpathy/a4166c7fe253700972fcbc77e4ea32c5
img = img[35:195]
img = img[::2, ::2, 0]
img[img == 144] = 0
img[img == 109] = 0
img[img != 0] = 1
curr = np.expand_dims(img.astype(numpy.float).ravel(), axis=0)
# Subtract the last preprocessed image.
diff = curr - self.prev if self.prev is not None else np.zeros(
(1, curr.shape[1]))
self.prev = curr
return diff
def run_episode(self):
"""Run an episode using the current model to generate training data.
Specifically, this involves repeatedly getting an observation from the environment,
performing a forward pass using the single observation to get a distribution over actions
(in binary case a probability of a single action), and choosing an action.
Finally, rewards are discounted when the episode completes.
Returns
-------
(xs, ys, rs) : tuple
The N x input_size observations, N x 1 action labels, and N x 1 discounted rewards
obtained from running the episode's N steps.
"""
observation = self.env.reset()
self.model.preprocessor.reset()
self.episode_reward = 0
xs, ys, rs = [], [], []
done = False
game_number = 1
game_start = time.time()
while not done:
if self.render:
self.env.render()
x = self.model.preprocessor.preprocess(observation)
p = self.model.forward(x)
a, y = self.model.choose_action(p.asnumpy().ravel()[0])
observation, r, done, info = self.env.step(a)
xs.append(x.asnumpy().ravel())
ys.append(y)
rs.append(r)
self.episode_reward += r
if self._game_complete(r):
game_time = time.time() - game_start
if self.verbose:
print('game %d complete (%.2fs), reward: %f' % (game_number, game_time, r))
game_number += 1
game_start = time.time()
# Episode finished.
self.running_reward = self.episode_reward if self.running_reward is None else (
0.99*self.running_reward + 0.01*self.episode_reward)
xs = np.vstack(xs)
ys = np.vstack(ys)
rs = np.expand_dims(self.model.discount_rewards(rs), axis=1)
return xs, ys, rs
def test_sum_forward(): np_x = py_np.zeros((2, 10)) np_w = py_np.zeros((10, 3)) np_b = py_np.zeros(3) x = NumpyVarToMinpy(np_x) w = NumpyVarToMinpy(np_w) b = NumpyVarToMinpy(np_b) x_plain = np.reshape(x, (x.shape[0], -1)) out0 = np.dot(x_plain, w) out = out0 + np.repeat(np.expand_dims(b, axis=0), out0.shape[0], axis = 0) np_out = MinpyVarToNumpy(out) var = py_np.random.randn(2, 3) tmp = NumpyVarToMinpy(var) sum_tmp = np.sum(tmp, axis = 0) sum_py = MinpyVarToNumpy(sum_tmp)
def batchnorm(x,
gamma,
beta,
mode='train',
eps=1e-5,
momentum=0.9,
running_mean=None,
running_var=None):
"""
Forward pass for batch normalization.
During training the sample mean and (uncorrected) sample variance are
computed from minibatch statistics and used to normalize the incoming data.
During training we also keep an exponentially decaying running mean of the mean
and variance of each feature, and these averages are used to normalize data
at test-time.
At each timestep we update the running averages for mean and variance using
an exponential decay based on the momentum parameter:
running_mean = momentum * running_mean + (1 - momentum) * sample_mean
running_var = momentum * running_var + (1 - momentum) * sample_var
Note that the batch normalization paper suggests a different test-time
behavior: they compute sample mean and variance for each feature using a
large number of training images rather than using a running average. For
this implementation we have chosen to use running averages instead since
they do not require an additional estimation step; the torch7 implementation
of batch normalization also uses running averages.
Input:
- x: Data of shape (N, D)
- gamma: Scale parameter of shape (D,)
- beta: Shift paremeter of shape (D,)
- mode: 'train' or 'test'
- eps: Constant for numeric stability
- momentum: Constant for running mean / variance.
- running_mean: Array of shape (D,) giving running mean of features
- running_var Array of shape (D,) giving running variance of features
Returns a tuple of:
- out: of shape (N, D)
- running_mean: updated running_mean
- running_var: updated running_var
"""
# TODO: fix NDArray type system
N, D = x.shape
N, D = int(N), int(D)
if running_mean is None:
running_mean = np.zeros(D)
if running_var is None:
running_var = np.zeros(D)
out = None
if mode == 'train':
mean = np.sum(x, axis=0) / N
x_mean = (x - np.expand_dims(mean, axis=0))
sqr_x_mean = x_mean**2
var = np.sum(sqr_x_mean, axis=0) / N
sqrt_var = np.sqrt(var + eps)
inv_sqrt_var = 1.0 / sqrt_var
x_hat = x_mean * np.expand_dims(inv_sqrt_var, axis=0)
out = gamma * x_hat + beta
running_mean = momentum * running_mean + (1.0 - momentum) * mean
running_var = momentum * running_var + (1.0 - momentum) * var
elif mode == 'test':
x_hat = (x - running_mean) / np.sqrt(running_var + eps)
out = gamma * x_hat + beta
else:
raise ValueError('Invalid forward batchnorm mode "%s"' % mode)
# return the updated running means
return out, running_mean, running_var
def batchnorm(x,
gamma,
beta,
mode='train',
eps=1e-5,
momentum=0.9,
running_mean=None,
running_var=None):
"""
Forward pass for batch normalization.
During training the sample mean and (uncorrected) sample variance are
computed from minibatch statistics and used to normalize the incoming data.
During training we also keep an exponentially decaying running mean of the mean
and variance of each feature, and these averages are used to normalize data
at test-time.
At each timestep we update the running averages for mean and variance using
an exponential decay based on the momentum parameter:
running_mean = momentum * running_mean + (1 - momentum) * sample_mean
running_var = momentum * running_var + (1 - momentum) * sample_var
Note that the batch normalization paper suggests a different test-time
behavior: they compute sample mean and variance for each feature using a
large number of training images rather than using a running average. For
this implementation we have chosen to use running averages instead since
they do not require an additional estimation step; the torch7 implementation
of batch normalization also uses running averages.
Input:
- x: Data of shape (N, D)
- gamma: Scale parameter of shape (D,)
- beta: Shift paremeter of shape (D,)
- mode: 'train' or 'test'
- eps: Constant for numeric stability
- momentum: Constant for running mean / variance.
- running_mean: Array of shape (D,) giving running mean of features
- running_var Array of shape (D,) giving running variance of features
Returns a tuple of:
- out: of shape (N, D)
- running_mean: updated running_mean
- running_var: updated running_var
"""
# TODO: fix NDArray type system
N, D = x.shape
N, D = int(N), int(D)
if running_mean is None:
running_mean = np.zeros(D)
if running_var is None:
running_var = np.zeros(D)
out = None
if mode == 'train':
mean = np.sum(x, axis=0) / N
x_mean = (x - np.expand_dims(mean, axis=0))
sqr_x_mean = x_mean ** 2
var = np.sum(sqr_x_mean, axis=0) / N
sqrt_var = np.sqrt(var + eps)
inv_sqrt_var = 1.0 / sqrt_var
x_hat = x_mean * np.expand_dims(inv_sqrt_var, axis=0)
out = gamma * x_hat + beta
running_mean = momentum * running_mean + (1.0 - momentum) * mean
running_var = momentum * running_var + (1.0 - momentum) * var
elif mode == 'test':
x_hat = (x - running_mean) / np.sqrt(running_var + eps)
out = gamma * x_hat + beta
else:
raise ValueError('Invalid forward batchnorm mode "%s"' % mode)
# return the updated running means
return out, running_mean, running_var
w3 = np.random.randn(150, 1)
v1, v2, v3 = 0, 0, 0
# 2.8 Open a figure
plt.figure()
# 3. Create the model and train
for iter in range(number_of_epoch):
current_x, current_label = sklearn.utils.shuffle(data_x, label_x)
for i in range(0, current_x.shape[0], 100):
current_x_batch = current_x[i:i + 100]
current_label_batch = np.expand_dims(current_label[i:i + 100],
axis=1)
layer_1 = current_x_batch.dot(w1)
layer_1_act = tanh(layer_1)
layer_2 = layer_1_act.dot(w2)
layer_2_act = tanh(layer_2)
layer_3 = layer_2_act.dot(w3)
layer_3_act = sigmoid(layer_3)
cost = np.square(current_label_batch -
layer_3_act) / (current_x.shape[0] * 2)
grad_3_part_1 = current_label_batch - layer_3_act
grad_3_part_2 = d_sigmoid(layer_3)
def dU(beta):
return mp.dot(X.T, (mp.exp(mp.dot(X,beta))/(1+mp.exp(mp.dot(X,beta))) - y)) + beta/alpha
D = X.shape[1]
q = mp.zeros((D, 1), dtype=mp.float32)
out = mp.zeros((n_iter, D), dtype=mp.float32)
for i in range(n_iter):
q = hmc(U, dU, epsilon, L, q)
out[i,:] = mp.ravel(q)
return out
with cpu() if args.mode == 'cpu' else gpu(0):
with open('params.json') as params_file:
out = {}
params = json.load(params_file)
X_train, y_train, X_test, y_test = get_data()
X_train = mp.array(X_train)
y_train = mp.array(y_train)
X_test = mp.array(X_test)
y_test = mp.array(y_test)
y_train = mp.expand_dims(y_train, 1)
z = lr_hmc(y_train, X_train, params['epsilon'], params['n_leaps'], params['alpha'], 1) # Warm-up
t = time.perf_counter()
z = lr_hmc(y_train, X_train, params['epsilon'], params['n_leaps'], params['alpha'], params['n_iter'])
t = time.perf_counter() - t
out[f'minpy-{args.mode}'] = t
coef_ = mp.mean(z[params['burn_in']:], 0)
acc = mp.mean((sigmoid(mp.dot(X_test, coef_)) > 0.5) == y_test)[0]
assert acc > 0.8
print(json.dumps(out))
