def adam(x, dx, config=None):
"""
Uses the Adam update rule, which incorporates moving averages of both the
gradient and its square and a bias correction term.
config format:
- learning_rate: Scalar learning rate.
- beta1: Decay rate for moving average of first moment of gradient.
- beta2: Decay rate for moving average of second moment of gradient.
- epsilon: Small scalar used for smoothing to avoid dividing by zero.
- m: Moving average of gradient.
- v: Moving average of squared gradient.
- t: Iteration number.
"""
if config is None: config = {}
config.setdefault('learning_rate', 1e-3)
config.setdefault('beta1', 0.9)
config.setdefault('beta2', 0.999)
config.setdefault('epsilon', 1e-8)
config.setdefault('m', np.zeros_like(x))
config.setdefault('v', np.zeros_like(x))
config.setdefault('t', 0)
v = config['v']
m = config['m']
t = config['t']
#############################################################################
# TODO: Implement the Adam update formula, storing the next value of x in #
# the next_x variable. Don't forget to update the m, v, and t variables #
# stored in config. #
#############################################################################
t = t + 1
m = config['beta1'] * m + (1 - config['beta1']) * dx
v = config['beta2'] * v + (1 - config['beta2']) * (dx**2)
m_ = m / (1 - config['beta1']**t)
v_ = v / (1 - config['beta2']**t)
next_x = x - config['learning_rate'] * m_ / (np.sqrt(v_) + config['epsilon']
)
#############################################################################
# END OF YOUR CODE #
#############################################################################
config['v'] = v
config['m'] = m
config['t'] = t
return next_x, config
def adam(x, dx, config=None):
"""
Uses the Adam update rule, which incorporates moving averages of both the
gradient and its square and a bias correction term.
config format:
- learning_rate: Scalar learning rate.
- beta1: Decay rate for moving average of first moment of gradient.
- beta2: Decay rate for moving average of second moment of gradient.
- epsilon: Small scalar used for smoothing to avoid dividing by zero.
- m: Moving average of gradient.
- v: Moving average of squared gradient.
- t: Iteration number.
"""
if config is None: config = {}
config.setdefault('learning_rate', 1e-3)
config.setdefault('beta1', 0.9)
config.setdefault('beta2', 0.999)
config.setdefault('epsilon', 1e-8)
config.setdefault('m', np.zeros_like(x))
config.setdefault('v', np.zeros_like(x))
config.setdefault('t', 0)
v = config['v']
m = config['m']
t = config['t']
#############################################################################
# TODO: Implement the Adam update formula, storing the next value of x in #
# the next_x variable. Don't forget to update the m, v, and t variables #
# stored in config. #
#############################################################################
t = t + 1
m = config['beta1'] * m + (1 - config['beta1']) * dx
v = config['beta2'] * v + (1 - config['beta2']) * (dx**2)
m_ = m / (1 - config['beta1']**t)
v_ = v / (1 - config['beta2']**t)
next_x = x - config['learning_rate'] * m_ / (np.sqrt(v_) +
config['epsilon'])
#############################################################################
# END OF YOUR CODE #
#############################################################################
config['v'] = v
config['m'] = m
config['t'] = t
return next_x, config
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 word_embedding_backward(dout, x, W):
"""
Backward pass for word embeddings. We cannot back-propagate into the words
since they are integers, so we only return gradient for the word embedding
matrix.
HINT: Look up the function np.add.at
Inputs:
- dout: Upstream gradients of shape (N, T, D)
- cache: Values from the forward pass
Returns:
- dW: Gradient of word embedding matrix, of shape (V, D).
"""
dW = None
##############################################################################
# TODO: Implement the backward pass for word embeddings. #
# #
# HINT: Look up the function np.add.at #
##############################################################################
N, T, D = dout.shape
dW = np.zeros_like(W)
for n in range(N):
for t in range(T):
dW[x[n, t], :] += dout[n, t, :]
##############################################################################
# END OF YOUR CODE #
##############################################################################
return dW
def rmsprop(x, dx, config=None):
"""
Uses the RMSProp update rule, which uses a moving average of squared gradient
values to set adaptive per-parameter learning rates.
config format:
- learning_rate: Scalar learning rate.
- decay_rate: Scalar between 0 and 1 giving the decay rate for the squared
gradient cache.
- epsilon: Small scalar used for smoothing to avoid dividing by zero.
- cache: Moving average of second moments of gradients.
"""
if config is None: config = {}
config.setdefault('learning_rate', 1e-2)
config.setdefault('decay_rate', 0.99)
config.setdefault('epsilon', 1e-8)
config.setdefault('cache', np.zeros_like(x))
cache = config['cache']
#############################################################################
# TODO: Implement the RMSprop update formula, storing the next value of x #
# in the next_x variable. Don't forget to update cache value stored in #
# config['cache']. #
#############################################################################
cache = cache * config['decay_rate'] + dx ** 2 * ( 1 - config['decay_rate'] )
next_x = x - config['learning_rate'] * dx / ( np.sqrt(cache) + config['epsilon'] )
#############################################################################
# END OF YOUR CODE #
#############################################################################
config['cache'] = cache
return next_x, config
def train(net_shapes, net_params, optimizer, utility, pool):
# pass seed instead whole noise matrix to parallel will save your time
noise_seed = np.random.randint(
0, 2**32 - 1, size=N_KID,
dtype=np.uint32) #.repeat(2) # mirrored sampling
noise_seed = np.concatenate([noise_seed, noise_seed
]).reshape([2, N_KID]).T.reshape([N_KID * 2])
# serially train with GPU
rewards = []
for k_id in range(N_KID * 2):
reward = get_reward(net_shapes, net_params, env, CONFIG['ep_max_step'],
CONFIG['continuous_a'], [noise_seed[k_id], k_id])
rewards = np.array(rewards)
kids_rank = np.argsort(rewards)[::-1] # rank kid id by reward
cumulative_update = np.zeros_like(net_params) # initialize update values
for ui, k_id in enumerate(kids_rank):
np.random.seed(noise_seed[k_id]) # reconstruct noise using seed
cumulative_update += utility[ui] * sign(k_id) * np.random.randn(
net_params.size)
gradients = optimizer.get_gradients(cumulative_update /
(2 * N_KID * SIGMA))
return net_params + gradients, rewards
def sgd_momentum(w, dw, config=None):
"""
Performs stochastic gradient descent with momentum.
config format:
- learning_rate: Scalar learning rate.
- momentum: Scalar between 0 and 1 giving the momentum value.
Setting momentum = 0 reduces to sgd.
- velocity: A numpy array of the same shape as w and dw used to store a moving
average of the gradients.
"""
if config is None: config = {}
config.setdefault('learning_rate', 1e-2)
config.setdefault('momentum', 0.9)
v = config.get('velocity', np.zeros_like(w))
#############################################################################
# TODO: Implement the momentum update formula. Store the updated value in #
# the next_w variable. You should also use and update the velocity v. #
#############################################################################
v = v * config['momentum'] - dw * config['learning_rate']
next_w = w + v
#############################################################################
# END OF YOUR CODE #
#############################################################################
config['velocity'] = v
return next_w, config
def sgd_momentum(w, dw, config=None):
"""
Performs stochastic gradient descent with momentum.
config format:
- learning_rate: Scalar learning rate.
- momentum: Scalar between 0 and 1 giving the momentum value.
Setting momentum = 0 reduces to sgd.
- velocity: A numpy array of the same shape as w and dw used to store a moving
average of the gradients.
"""
if config is None: config = {}
config.setdefault('learning_rate', 1e-2)
config.setdefault('momentum', 0.9)
v = config.get('velocity', np.zeros_like(w))
#############################################################################
# TODO: Implement the momentum update formula. Store the updated value in #
# the next_w variable. You should also use and update the velocity v. #
#############################################################################
v = v * config['momentum'] - dw * config['learning_rate']
next_w = w + v
#############################################################################
# END OF YOUR CODE #
#############################################################################
config['velocity'] = v
return next_w, config
def rmsprop(x, dx, config=None):
"""
Uses the RMSProp update rule, which uses a moving average of squared gradient
values to set adaptive per-parameter learning rates.
config format:
- learning_rate: Scalar learning rate.
- decay_rate: Scalar between 0 and 1 giving the decay rate for the squared
gradient cache.
- epsilon: Small scalar used for smoothing to avoid dividing by zero.
- cache: Moving average of second moments of gradients.
"""
if config is None: config = {}
config.setdefault('learning_rate', 1e-2)
config.setdefault('decay_rate', 0.99)
config.setdefault('epsilon', 1e-8)
config.setdefault('cache', np.zeros_like(x))
cache = config['cache']
#############################################################################
# TODO: Implement the RMSprop update formula, storing the next value of x #
# in the next_x variable. Don't forget to update cache value stored in #
# config['cache']. #
#############################################################################
cache = cache * config['decay_rate'] + dx**2 * (1 - config['decay_rate'])
next_x = x - config['learning_rate'] * dx / (np.sqrt(cache) +
config['epsilon'])
#############################################################################
# END OF YOUR CODE #
#############################################################################
config['cache'] = cache
return next_x, config
def _setStudLayerWeights(self, l, refParam, student, weightName):
"""Set param/biases of layer l for student to match refParam"""
refShape = refParam.shape
studentP = student.getLayerWeights(l, weightName)
newP = np.zeros_like(studentP)
if len(refShape) > 1:
newP[:refShape[0], :refShape[1]] = refParam
else:
newP[:refShape[0]] = refParam
student.setLayerWeights(l, weightName, newP)
def discount_rewards(self, rs):
drs = np.zeros_like(rs).asnumpy()
s = 0
for t in reversed(range(0, len(rs))):
# Reset the running sum at a game boundary.
if rs[t] != 0:
s = 0
s = s * self.gamma + rs[t]
drs[t] = s
drs -= np.mean(drs)
drs /= np.std(drs)
return drs
def discount_rewards(self, rs):
drs = np.zeros_like(rs).asnumpy()
s = 0
for t in reversed(xrange(0, len(rs))):
# Reset the running sum at a game boundary.
if rs[t] != 0:
s = 0
s = s * self.gamma + rs[t]
drs[t] = s
drs -= np.mean(drs)
drs /= np.std(drs)
return drs
def adam(x, dx, config=None):
"""
Uses the Adam update rule, which incorporates moving averages of both the
gradient and its square and a bias correction term.
config format:
- learning_rate: Scalar learning rate.
- beta1: Decay rate for moving average of first moment of gradient.
- beta2: Decay rate for moving average of second moment of gradient.
- epsilon: Small scalar used for smoothing to avoid dividing by zero.
- m: Moving average of gradient.
- v: Moving average of squared gradient.
- t: Iteration number.
"""
if config is None: config = {}
config.setdefault('learning_rate', 1e-3)
config.setdefault('beta1', 0.9)
config.setdefault('beta2', 0.999)
config.setdefault('epsilon', 1e-8)
config.setdefault('m', np.zeros_like(x))
config.setdefault('v', np.zeros_like(x))
config.setdefault('t', 0)
next_x = None
beta1, beta2, eps = config['beta1'], config['beta2'], config['epsilon']
t, m, v = config['t'], config['m'], config['v']
m = beta1 * m + (1 - beta1) * dx
v = beta2 * v + (1 - beta2) * (dx * dx)
t += 1
alpha = config['learning_rate'] * np.sqrt(1 - beta2 ** t) / (1 - beta1 ** t)
x -= alpha * (m / (np.sqrt(v) + eps))
config['t'] = t
config['m'] = m
config['v'] = v
next_x = x
return next_x, config
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]
#TODO: Support broadcast case: (X,) (X, Y)
#margins = np.maximum(0, x - correct_class_scores + 1.0)
margins = np.transpose(
np.maximum(0,
np.transpose(x) - np.transpose(correct_class_scores) + 1.0))
#margins[np.arange(N), y] = 0
#loss = np.sum(margins) / N
loss = (np.sum(margins) - np.sum(margins[np.arange(N), y])) / N
margins[np.arange(N), y] = 0
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 sgd_momentum(w, dw, config=None):
"""
Performs stochastic gradient descent with momentum.
config format:
- learning_rate: Scalar learning rate.
- momentum: Scalar between 0 and 1 giving the momentum value.
Setting momentum = 0 reduces to sgd.
- velocity: A numpy array of the same shape as w and dw used to store a moving
average of the gradients.
"""
if config is None: config = {}
config.setdefault('learning_rate', 1e-2)
config.setdefault('momentum', 0.9)
v = config.get('velocity', np.zeros_like(w))
v = v * config['momentum'] - dw * config['learning_rate']
next_w = w + v
config['velocity'] = v
return next_w, config
def sgd_momentum(w, dw, config=None):
"""
Performs stochastic gradient descent with momentum.
config format:
- learning_rate: Scalar learning rate.
- momentum: Scalar between 0 and 1 giving the momentum value.
Setting momentum = 0 reduces to sgd.
- velocity: A numpy array of the same shape as w and dw used to store a moving
average of the gradients.
"""
if config is None: config = {}
config.setdefault('learning_rate', 1e-2)
config.setdefault('momentum', 0.9)
v = config.get('velocity', np.zeros_like(w))
v = v * config['momentum'] - dw * config['learning_rate']
next_w = w + v
config['velocity'] = v
return next_w, config
def test_numeric():
# 'newaxis', 'ndarray', 'flatiter', 'nditer', 'nested_iters', 'ufunc',
# 'arange', 'array', 'zeros', 'count_nonzero', 'empty', 'broadcast',
# 'dtype', 'fromstring', 'fromfile', 'frombuffer', 'int_asbuffer',
# 'where', 'argwhere', 'copyto', 'concatenate', 'fastCopyAndTranspose',
# 'lexsort', 'set_numeric_ops', 'can_cast', 'promote_types',
# 'min_scalar_type', 'result_type', 'asarray', 'asanyarray',
# 'ascontiguousarray', 'asfortranarray', 'isfortran', 'empty_like',
# 'zeros_like', 'ones_like', 'correlate', 'convolve', 'inner', 'dot',
# 'einsum', 'outer', 'vdot', 'alterdot', 'restoredot', 'roll',
# 'rollaxis', 'moveaxis', 'cross', 'tensordot', 'array2string',
# 'get_printoptions', 'set_printoptions', 'array_repr', 'array_str',
# 'set_string_function', 'little_endian', 'require', 'fromiter',
# 'array_equal', 'array_equiv', 'indices', 'fromfunction', 'isclose', 'load',
# 'loads', 'isscalar', 'binary_repr', 'base_repr', 'ones', 'identity',
# 'allclose', 'compare_chararrays', 'putmask', 'seterr', 'geterr',
# 'setbufsize', 'getbufsize', 'seterrcall', 'geterrcall', 'errstate',
# 'flatnonzero', 'Inf', 'inf', 'infty', 'Infinity', 'nan', 'NaN', 'False_',
# 'True_', 'bitwise_not', 'full', 'full_like', 'matmul'
x = np.arange(6)
x = x.reshape((2, 3))
np.zeros_like(x)
y = np.arange(3, dtype=np.float)
np.zeros_like(y)
np.ones(5)
np.ones((5,), dtype=np.int)
np.ones((2, 1))
s = (2,2)
np.ones(s)
x = np.arange(6)
x = x.reshape((2, 3))
np.ones_like(x)
y = np.arange(3, dtype=np.float)
np.ones_like(y)
np.full((2, 2), np.inf)
x = np.arange(6, dtype=np.int)
np.full_like(x, 1)
np.full_like(x, 0.1)
np.full_like(y, 0.1)
np.count_nonzero(np.eye(4))
np.count_nonzero([[0,1,7,0,0],[3,0,0,2,19]])
np.count_nonzero([[0,1,7,0,0],[3,0,0,2,19]], axis=0)
np.count_nonzero([[0,1,7,0,0],[3,0,0,2,19]], axis=1)
a = [1, 2]
np.asarray(a)
a = np.array([1, 2])
np.asarray(a) is a
a = np.array([1, 2], dtype=np.float32)
np.asarray(a, dtype=np.float32) is a
np.asarray(a, dtype=np.float64) is a
np.asarray(a) is a
np.asanyarray(a) is a
a = [1, 2]
np.asanyarray(a)
np.asanyarray(a) is a
x = np.arange(6).reshape(2,3)
np.ascontiguousarray(x, dtype=np.float32)
x = np.arange(6).reshape(2,3)
y = np.asfortranarray(x)
x = np.arange(6).reshape(2,3)
y = np.require(x, dtype=np.float32, requirements=['A', 'O', 'W', 'F'])
a = np.array([[1, 2, 3], [4, 5, 6]], order='C')
np.isfortran(a)
b = np.array([[1, 2, 3], [4, 5, 6]], order='FORTRAN')
np.isfortran(b)
a = np.array([[1, 2, 3], [4, 5, 6]], order='C')
np.isfortran(a)
b = a.T
np.isfortran(b)
np.isfortran(np.array([1, 2], order='FORTRAN'))
x = np.arange(6).reshape(2,3)
np.argwhere(x>1)
x = np.arange(-2, 3)
np.flatnonzero(x)
np.correlate([1, 2, 3], [0, 1, 0.5])
np.correlate([1, 2, 3], [0, 1, 0.5], "same")
np.correlate([1, 2, 3], [0, 1, 0.5], "full")
np.correlate([1+1j, 2, 3-1j], [0, 1, 0.5j], 'full')
np.correlate([0, 1, 0.5j], [1+1j, 2, 3-1j], 'full')
np.convolve([1, 2, 3], [0, 1, 0.5])
np.convolve([1,2,3],[0,1,0.5], 'same')
np.convolve([1,2,3],[0,1,0.5], 'valid')
rl = np.outer(np.ones((5,)), np.linspace(-2, 2, 5))
# im = np.outer(1j*np.linspace(2, -2, 5), np.ones((5,)))
# grid = rl + im
x = np.array(['a', 'b', 'c'], dtype=object)
np.outer(x, [1, 2, 3])
a = np.arange(60.).reshape(3,4,5)
b = np.arange(24.).reshape(4,3,2)
c = np.tensordot(a,b, axes=([1,0],[0,1]))
c.shape
# A slower but equivalent way of computing the same...
d = np.zeros((5,2))
a = np.array(range(1, 9))
A = np.array(('a', 'b', 'c', 'd'), dtype=object)
x = np.arange(10)
np.roll(x, 2)
x2 = np.reshape(x, (2,5))
np.roll(x2, 1)
np.roll(x2, 1, axis=0)
np.roll(x2, 1, axis=1)
a = np.ones((3,4,5,6))
np.rollaxis(a, 3, 1).shape
np.rollaxis(a, 2).shape
np.rollaxis(a, 1, 4).shape
x = np.zeros((3, 4, 5))
np.moveaxis(x, 0, -1).shape
np.moveaxis(x, -1, 0).shape
np.transpose(x).shape
np.moveaxis(x, [0, 1], [-1, -2]).shape
np.moveaxis(x, [0, 1, 2], [-1, -2, -3]).shape
x = [1, 2, 3]
y = [4, 5, 6]
np.cross(x, y)
x = [1, 2]
y = [4, 5, 6]
np.cross(x, y)
x = [1, 2, 0]
y = [4, 5, 6]
np.cross(x, y)
x = [1,2]
y = [4,5]
np.cross(x, y)
x = np.array([[1,2,3], [4,5,6]])
y = np.array([[4,5,6], [1,2,3]])
np.cross(x, y)
np.cross(x, y, axisc=0)
x = np.array([[1,2,3], [4,5,6], [7, 8, 9]])
y = np.array([[7, 8, 9], [4,5,6], [1,2,3]])
np.cross(x, y)
np.cross(x, y, axisa=0, axisb=0)
# np.array_repr(np.array([1,2]))
# np.array_repr(np.ma.array([0.]))
# np.array_repr(np.array([], np.int32))
x = np.array([1e-6, 4e-7, 2, 3])
# np.array_repr(x, precision=6, suppress_small=True)
# np.array_str(np.arange(3))
a = np.arange(10)
x = np.arange(4)
np.set_string_function(lambda x:'random', repr=False)
grid = np.indices((2, 3))
grid.shape
grid[0] # row indices
grid[1] # column indices
x = np.arange(20).reshape(5, 4)
row, col = np.indices((2, 3))
x[row, col]
np.fromfunction(lambda i, j: i == j, (3, 3), dtype=int)
np.fromfunction(lambda i, j: i + j, (3, 3), dtype=int)
np.isscalar(3.1)
np.isscalar([3.1])
np.isscalar(False)
# np.binary_repr(3)
# np.binary_repr(-3)
# np.binary_repr(3, width=4)
# np.binary_repr(-3, width=3)
# np.binary_repr(-3, width=5)
# np.base_repr(5)
# np.base_repr(6, 5)
# np.base_repr(7, base=5, padding=3)
# np.base_repr(10, base=16)
# np.base_repr(32, base=16)
np.identity(3)
np.allclose([1e10,1e-7], [1.00001e10,1e-8])
np.allclose([1e10,1e-8], [1.00001e10,1e-9])
np.allclose([1e10,1e-8], [1.0001e10,1e-9])
# np.allclose([1.0, np.nan], [1.0, np.nan])
# np.allclose([1.0, np.nan], [1.0, np.nan], equal_nan=True)
np.isclose([1e10,1e-7], [1.00001e10,1e-8])
np.isclose([1e10,1e-8], [1.00001e10,1e-9])
np.isclose([1e10,1e-8], [1.0001e10,1e-9])
# np.isclose([1.0, np.nan], [1.0, np.nan])
# np.isclose([1.0, np.nan], [1.0, np.nan], equal_nan=True)
np.array_equal([1, 2], [1, 2])
np.array_equal(np.array([1, 2]), np.array([1, 2]))
np.array_equal([1, 2], [1, 2, 3])
np.array_equal([1, 2], [1, 4])
np.array_equiv([1, 2], [1, 2])
np.array_equiv([1, 2], [1, 3])
np.array_equiv([1, 2], [[1, 2], [1, 2]])
np.array_equiv([1, 2], [[1, 2, 1, 2], [1, 2, 1, 2]])
np.array_equiv([1, 2], [[1, 2], [1, 3]])
def train(self):
"""Trains the model for `num_episodes` iterations.
On each iteration, runs an episode (see `.run_episode()`) to generate three matrices of
observations, labels and rewards (xs, ys, rs) containing data for the _entire_ episode.
Then the parameter gradients are found using these episode matrices.
Specifically, auto-grad is performed on `loss_func`, which does a single forward pass
with the episode's observations `xs` then computes the loss using the output of the forward
pass and the episode's labels `ys` and discounted rewards `rs`.
This two-step approach of generating episode data then doing a single forward/backward pass
is done to conserve memory during the auto-grad computation.
"""
# Accumulate gradients since updates are only performed every `update_every` iterations.
grad_buffer = self._init_grad_buffer()
for episode_number in xrange(1, self.num_episodes):
episode_start = time.time()
# Generate an episode of training data.
xs, ys, rs = self.run_episode()
# Performs a forward pass and computes loss using an entire episode's data.
def loss_func(*params):
ps = self.model.forward(xs)
return self.model.loss(ps, ys, rs)
# Compute gradients with auto-grad on `loss_func` (duplicated from `Solver`).
param_arrays = list(self.model.params.values())
param_keys = list(self.model.params.keys())
grad_and_loss_func = core.grad_and_loss(loss_func, argnum=range(len(param_arrays)))
backward_start = time.time()
grad_arrays, loss = grad_and_loss_func(*param_arrays)
backward_time = time.time() - backward_start
grads = dict(zip(param_keys, grad_arrays))
# Accumulate gradients until an update is performed.
for k, v in grads.iteritems():
grad_buffer[k] += v
# Misc. diagnostic info.
self.loss_history.append(loss.asnumpy())
episode_time = time.time() - episode_start
if self.verbose:
print('Backward pass complete (%.2fs)' % backward_time)
if self.verbose or episode_number % self.print_every == 0:
print('Episode %d complete (%.2fs), loss: %s, reward: %s, running reward: %s' %
(episode_number, episode_time, loss, self.episode_reward, self.running_reward))
# Perform parameter update and reset the `grad_buffer` when appropriate.
if episode_number % self.update_every == 0:
for p, w in self.model.params.items():
dw = grad_buffer[p]
config = self.optim_configs[p]
next_w, next_config = self.update_rule(w, dw, config)
self.model.params[p] = next_w
self.optim_configs[p] = next_config
grad_buffer[p] = np.zeros_like(w)
# Save model parameters to `save_dir` when appropriate..
if episode_number % self.save_every == 0:
if self.verbose:
print('Saving model parameters...')
file_name = os.path.join(self.save_dir, 'params_%d.p' % episode_number)
with open(file_name, 'w') as f:
pickle.dump({k: v.asnumpy() for k, v in self.model.params.iteritems()}, f)
if self.verbose:
print('Wrote parameter file %s' % file_name)
def __init__(self, params, learning_rate, momentum=0.9):
self.v = np.zeros_like(params)
self.lr, self.momentum = learning_rate, momentum
def _init_grad_buffer(self):
return {k: np.zeros_like(v) for k, v in self.model.params.iteritems()}
def test_numeric():
# 'newaxis', 'ndarray', 'flatiter', 'nditer', 'nested_iters', 'ufunc',
# 'arange', 'array', 'zeros', 'count_nonzero', 'empty', 'broadcast',
# 'dtype', 'fromstring', 'fromfile', 'frombuffer', 'int_asbuffer',
# 'where', 'argwhere', 'copyto', 'concatenate', 'fastCopyAndTranspose',
# 'lexsort', 'set_numeric_ops', 'can_cast', 'promote_types',
# 'min_scalar_type', 'result_type', 'asarray', 'asanyarray',
# 'ascontiguousarray', 'asfortranarray', 'isfortran', 'empty_like',
# 'zeros_like', 'ones_like', 'correlate', 'convolve', 'inner', 'dot',
# 'einsum', 'outer', 'vdot', 'alterdot', 'restoredot', 'roll',
# 'rollaxis', 'moveaxis', 'cross', 'tensordot', 'array2string',
# 'get_printoptions', 'set_printoptions', 'array_repr', 'array_str',
# 'set_string_function', 'little_endian', 'require', 'fromiter',
# 'array_equal', 'array_equiv', 'indices', 'fromfunction', 'isclose', 'load',
# 'loads', 'isscalar', 'binary_repr', 'base_repr', 'ones', 'identity',
# 'allclose', 'compare_chararrays', 'putmask', 'seterr', 'geterr',
# 'setbufsize', 'getbufsize', 'seterrcall', 'geterrcall', 'errstate',
# 'flatnonzero', 'Inf', 'inf', 'infty', 'Infinity', 'nan', 'NaN', 'False_',
# 'True_', 'bitwise_not', 'full', 'full_like', 'matmul'
x = np.arange(6)
x = x.reshape((2, 3))
np.zeros_like(x)
y = np.arange(3, dtype=np.float)
np.zeros_like(y)
np.ones(5)
np.ones((5, ), dtype=np.int)
np.ones((2, 1))
s = (2, 2)
np.ones(s)
x = np.arange(6)
x = x.reshape((2, 3))
np.ones_like(x)
y = np.arange(3, dtype=np.float)
np.ones_like(y)
np.full((2, 2), np.inf)
x = np.arange(6, dtype=np.int)
np.full_like(x, 1)
np.full_like(x, 0.1)
np.full_like(y, 0.1)
np.count_nonzero(np.eye(4))
np.count_nonzero([[0, 1, 7, 0, 0], [3, 0, 0, 2, 19]])
np.count_nonzero([[0, 1, 7, 0, 0], [3, 0, 0, 2, 19]], axis=0)
np.count_nonzero([[0, 1, 7, 0, 0], [3, 0, 0, 2, 19]], axis=1)
a = [1, 2]
np.asarray(a)
a = np.array([1, 2])
np.asarray(a) is a
a = np.array([1, 2], dtype=np.float32)
np.asarray(a, dtype=np.float32) is a
np.asarray(a, dtype=np.float64) is a
np.asarray(a) is a
np.asanyarray(a) is a
a = [1, 2]
np.asanyarray(a)
np.asanyarray(a) is a
x = np.arange(6).reshape(2, 3)
np.ascontiguousarray(x, dtype=np.float32)
x = np.arange(6).reshape(2, 3)
y = np.asfortranarray(x)
x = np.arange(6).reshape(2, 3)
y = np.require(x, dtype=np.float32, requirements=['A', 'O', 'W', 'F'])
a = np.array([[1, 2, 3], [4, 5, 6]], order='C')
np.isfortran(a)
b = np.array([[1, 2, 3], [4, 5, 6]], order='FORTRAN')
np.isfortran(b)
a = np.array([[1, 2, 3], [4, 5, 6]], order='C')
np.isfortran(a)
b = a.T
np.isfortran(b)
np.isfortran(np.array([1, 2], order='FORTRAN'))
x = np.arange(6).reshape(2, 3)
np.argwhere(x > 1)
x = np.arange(-2, 3)
np.flatnonzero(x)
np.correlate([1, 2, 3], [0, 1, 0.5])
np.correlate([1, 2, 3], [0, 1, 0.5], "same")
np.correlate([1, 2, 3], [0, 1, 0.5], "full")
np.correlate([1 + 1j, 2, 3 - 1j], [0, 1, 0.5j], 'full')
np.correlate([0, 1, 0.5j], [1 + 1j, 2, 3 - 1j], 'full')
np.convolve([1, 2, 3], [0, 1, 0.5])
np.convolve([1, 2, 3], [0, 1, 0.5], 'same')
np.convolve([1, 2, 3], [0, 1, 0.5], 'valid')
rl = np.outer(np.ones((5, )), np.linspace(-2, 2, 5))
# im = np.outer(1j*np.linspace(2, -2, 5), np.ones((5,)))
# grid = rl + im
x = np.array(['a', 'b', 'c'], dtype=object)
np.outer(x, [1, 2, 3])
a = np.arange(60.).reshape(3, 4, 5)
b = np.arange(24.).reshape(4, 3, 2)
c = np.tensordot(a, b, axes=([1, 0], [0, 1]))
c.shape
# A slower but equivalent way of computing the same...
d = np.zeros((5, 2))
a = np.array(range(1, 9))
A = np.array(('a', 'b', 'c', 'd'), dtype=object)
x = np.arange(10)
np.roll(x, 2)
x2 = np.reshape(x, (2, 5))
np.roll(x2, 1)
np.roll(x2, 1, axis=0)
np.roll(x2, 1, axis=1)
a = np.ones((3, 4, 5, 6))
np.rollaxis(a, 3, 1).shape
np.rollaxis(a, 2).shape
np.rollaxis(a, 1, 4).shape
x = np.zeros((3, 4, 5))
np.moveaxis(x, 0, -1).shape
np.moveaxis(x, -1, 0).shape
np.transpose(x).shape
np.moveaxis(x, [0, 1], [-1, -2]).shape
np.moveaxis(x, [0, 1, 2], [-1, -2, -3]).shape
x = [1, 2, 3]
y = [4, 5, 6]
np.cross(x, y)
x = [1, 2]
y = [4, 5, 6]
np.cross(x, y)
x = [1, 2, 0]
y = [4, 5, 6]
np.cross(x, y)
x = [1, 2]
y = [4, 5]
np.cross(x, y)
x = np.array([[1, 2, 3], [4, 5, 6]])
y = np.array([[4, 5, 6], [1, 2, 3]])
np.cross(x, y)
np.cross(x, y, axisc=0)
x = np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]])
y = np.array([[7, 8, 9], [4, 5, 6], [1, 2, 3]])
np.cross(x, y)
np.cross(x, y, axisa=0, axisb=0)
# np.array_repr(np.array([1,2]))
# np.array_repr(np.ma.array([0.]))
# np.array_repr(np.array([], np.int32))
x = np.array([1e-6, 4e-7, 2, 3])
# np.array_repr(x, precision=6, suppress_small=True)
# np.array_str(np.arange(3))
a = np.arange(10)
x = np.arange(4)
np.set_string_function(lambda x: 'random', repr=False)
grid = np.indices((2, 3))
grid.shape
grid[0] # row indices
grid[1] # column indices
x = np.arange(20).reshape(5, 4)
row, col = np.indices((2, 3))
x[row, col]
np.fromfunction(lambda i, j: i == j, (3, 3), dtype=int)
np.fromfunction(lambda i, j: i + j, (3, 3), dtype=int)
np.isscalar(3.1)
np.isscalar([3.1])
np.isscalar(False)
# np.binary_repr(3)
# np.binary_repr(-3)
# np.binary_repr(3, width=4)
# np.binary_repr(-3, width=3)
# np.binary_repr(-3, width=5)
# np.base_repr(5)
# np.base_repr(6, 5)
# np.base_repr(7, base=5, padding=3)
# np.base_repr(10, base=16)
# np.base_repr(32, base=16)
np.identity(3)
np.allclose([1e10, 1e-7], [1.00001e10, 1e-8])
np.allclose([1e10, 1e-8], [1.00001e10, 1e-9])
np.allclose([1e10, 1e-8], [1.0001e10, 1e-9])
# np.allclose([1.0, np.nan], [1.0, np.nan])
# np.allclose([1.0, np.nan], [1.0, np.nan], equal_nan=True)
np.isclose([1e10, 1e-7], [1.00001e10, 1e-8])
np.isclose([1e10, 1e-8], [1.00001e10, 1e-9])
np.isclose([1e10, 1e-8], [1.0001e10, 1e-9])
# np.isclose([1.0, np.nan], [1.0, np.nan])
# np.isclose([1.0, np.nan], [1.0, np.nan], equal_nan=True)
np.array_equal([1, 2], [1, 2])
np.array_equal(np.array([1, 2]), np.array([1, 2]))
np.array_equal([1, 2], [1, 2, 3])
np.array_equal([1, 2], [1, 4])
np.array_equiv([1, 2], [1, 2])
np.array_equiv([1, 2], [1, 3])
np.array_equiv([1, 2], [[1, 2], [1, 2]])
np.array_equiv([1, 2], [[1, 2, 1, 2], [1, 2, 1, 2]])
np.array_equiv([1, 2], [[1, 2], [1, 3]])
