def CrossEntropyLoss(self, y_out, y_true):
if not isinstance(y_true, Parameter):
y_true = Parameter(data=y_true, eval_grad=False, graph=self.graph)
batch_size = Parameter((1, 1),
init_zeros=True,
eval_grad=False,
graph=self.graph) # mini-batch size
batch_size.data.fill(float(y_true.shape[-1]))
neg_one = Parameter((1, 1),
init_zeros=True,
eval_grad=False,
graph=self.graph)
neg_one.data.fill(
-1.0
) # just a -1 to make the l.grad look same in all the loss defs (dl/dl = 1)
# KL(P || Q): Summation(P*log(P)){result: 0} - Summation(P*log(Q))
l = self.graph.multiply(
self.graph.sum(self.graph.multiply(y_true, self.graph.log(y_out))),
neg_one)
# avg_loss = (1/m)*sigma{i = 1,..,m}(loss[i])
l = self.graph.divide(l, batch_size)
l.grad[0, 0] = 1.0 # dl/dl = 1.0
return l
def load(self,
file=None): # model.load() - loads the state of net from a file
print('loading model...')
if file == None:
file = self.__class__.__name__ + '.npy'
load_dict = np.load(file, allow_pickle=True).item()
module_layers_stored = load_dict['module_layers']
module_params_stored = load_dict['module_params']
module_layers_actual = self.get_module_layers()
module_params_actual = self.get_module_params()
for layer_name, layer_stored in module_layers_stored.items():
if layer_name in module_layers_actual:
for param_name, param in layer_stored.items():
layer_actual = module_layers_actual[layer_name]
setattr(layer_actual, str(param_name),
Parameter(data=param))
for param_name, param in module_params_stored.items():
if param_name in module_params_actual:
setattr(self, str(param_name), Parameter(data=param))
print('Successfully loaded model from {}'.format(file))
def forward(self, x):
if not isinstance(x, Parameter):
x = Parameter(data=x, eval_grad=False, graph=self.graph)
if self.graph.grad_mode: # training
# useful: https://arxiv.org/abs/1502.03167
batch_size = Parameter((1, 1), init_zeros=True, eval_grad=False, graph=self.graph) # mini-batch size/channel size
batch_size.data.fill(float(x.shape[self.axis]))
# calculate mean and variance
mean = self.graph.divide(self.graph.sum(x, axis=self.axis), batch_size)
centered = self.graph.subtract(x, mean, axis=self.axis)
var = self.graph.divide(self.graph.sum(self.graph.power(centered, 2), axis=self.axis), batch_size)
self.m = self.momentum * self.m + (1 - self.momentum) * mean.data
self.v = self.momentum * self.v + (1 - self.momentum) * var.data
# normalize the data to zero mean and unit variance
normalized = self.graph.multiply(centered, self.graph.power(var, -0.5), axis=self.axis)
else: # testing
centered = np.subtract(x.data, self.m)
normalized = np.multiply(centered, np.power(self.v + 1e-6, -0.5))
normalized = Parameter(data=normalized, eval_grad=False, graph=self.graph)
# scale and shift
out = self.graph.multiply(normalized, self.gamma, axis=(-1,)) # scale
if self.bias: # shift
out = self.graph.add(out, self.beta, axis=(-1,))
return out
def init_params(self):
self.W_ih = Parameter((self.hidden_size, self.input_size),
graph=self.graph)
self.W_hh = Parameter((self.hidden_size, self.hidden_size),
graph=self.graph)
self.b_ih = Parameter((self.hidden_size, 1),
graph=self.graph) # not much use
self.b_hh = Parameter((self.hidden_size, 1), graph=self.graph)
def init_params(self):
self.W_ih = Parameter(
(4 * self.hidden_size, self.input_size),
graph=self.graph) # input to hidden weight volume
self.W_hh = Parameter(
(4 * self.hidden_size, self.hidden_size),
graph=self.graph) # hidden to hidden weight volume
self.b_ih = Parameter((4 * self.hidden_size, 1),
graph=self.graph) # input to hidden bias vector
self.b_hh = Parameter((4 * self.hidden_size, 1),
graph=self.graph) # hidden to hidden bias vector
def init_params(self):
root_k = np.sqrt(1. / self.input_features)
self.W = Parameter((self.output_features, self.input_features),
uniform=True,
low=-root_k,
high=root_k,
graph=self.graph) # weight volume
self.b = Parameter((self.output_features, 1),
uniform=True,
low=-root_k,
high=root_k,
graph=self.graph) # bias vector
def forward(self, x, hidden):
h, c = hidden
if not isinstance(x, Parameter):
x = Parameter(data=x, eval_grad=False, graph=self.graph)
i_h = self.graph.dot(self.W_ih, x)
if self.bias:
i_h = self.graph.add(i_h, self.b_ih, axis=(-1, ))
h_h = self.graph.dot(self.W_hh, h)
if self.bias:
h_h = self.graph.add(h_h, self.b_hh, axis=(-1, ))
gates = self.graph.add(i_h, h_h)
# forget, input, gate(also called cell gate - different from cell state), output gates of the lstm cell
# useful: http://colah.github.io/posts/2015-08-Understanding-LSTMs/
f, i, g, o = self.graph.split(gates, sections=4, axis=0)
f = self.graph.sigmoid(f)
i = self.graph.sigmoid(i)
g = self.graph.tanh(g)
o = self.graph.sigmoid(o)
c = self.graph.add(self.graph.multiply(f, c),
self.graph.multiply(i, g))
h = self.graph.multiply(o, self.graph.tanh(c))
return (h, c)
def JSDivLoss(self, y_out, y_true):
if not isinstance(y_true, Parameter):
y_true = Parameter(data=y_true, eval_grad=False, graph=self.graph)
batch_size = Parameter((1, 1),
init_zeros=True,
eval_grad=False,
graph=self.graph) # mini-batch size
batch_size.data.fill(float(y_true.shape[-1]))
two = Parameter((1, 1),
init_zeros=True,
eval_grad=False,
graph=self.graph)
two.data.fill(2.0) # just a 2 :p
neg_one = Parameter((1, 1),
init_zeros=True,
eval_grad=False,
graph=self.graph)
neg_one.data.fill(
-1.0
) # just a -1 to make the l.grad look same in all the loss defs (dl/dl = 1)
# mean probability: (P + Q)/2
y_mean = self.graph.divide(self.graph.add(y_out, y_true), two)
# KL(P || (P + Q)/2): Summation(P*log(P)){result: 0} - Summation(P*log((P+Q)/2))
kl_1 = self.graph.multiply(
self.graph.sum(self.graph.multiply(y_true,
self.graph.log(y_mean))),
neg_one)
# KL(Q || (P + Q)/2): Summation(Q*log(Q)) - Summation(Q*log((P+Q)/2))
kl_2 = self.graph.add(self.graph.multiply(y_out,
self.graph.log(y_out)),
self.graph.multiply(
self.graph.multiply(y_out,
self.graph.log(y_mean)),
neg_one)) # !!!!!
# JS(P, Q) = 1/2*(KL(P || (P + Q)/2) + KL(Q || (P + Q)/2))
l = self.graph.divide(self.graph.add(kl_1, kl_2), two)
# avg_loss = (1/m)*sigma{i = 1,..,m}(loss[i])
l = self.graph.divide(l, batch_size)
l.grad[0, 0] = 1.0 # dl/dl = 1.0
return l
def BCELoss(self, y_out, y_true):
if not isinstance(y_true, Parameter):
y_true = Parameter(data=y_true, eval_grad=False, graph=self.graph)
batch_size = Parameter((1, 1),
init_zeros=True,
eval_grad=False,
graph=self.graph) # mini-batch size
batch_size.data.fill(float(y_true.shape[-1]))
neg_one = Parameter((1, 1),
init_zeros=True,
eval_grad=False,
graph=self.graph)
neg_one.data.fill(
-1.0
) # just a -1 to make the l.grad look same in all the loss defs (dl/dl = 1)
one = Parameter((1, 1),
init_zeros=True,
eval_grad=False,
graph=self.graph)
one.data.fill(1.0)
# class 2 output: 1 - c1
c2 = self.graph.multiply(self.graph.subtract(y_out, one), neg_one)
# class 2 target: 1 - t1
t2 = self.graph.multiply(self.graph.subtract(y_true, one), neg_one)
# -Summation(t1*log(c1))
l1 = self.graph.multiply(
self.graph.sum(self.graph.multiply(y_true, self.graph.log(y_out))),
neg_one)
# -Summation((1 - t1)*log(1 - c1))
l2 = self.graph.multiply(
self.graph.sum(self.graph.multiply(t2, self.graph.log(c2))),
neg_one)
# loss = -Summation(t1*log(c1)) -Summation((1 - t1)*log(1 - c1))
l = self.graph.add(l1, l2)
# avg_loss = (1/m)*sigma{i = 1,..,m}(loss[i])
l = self.graph.divide(l, batch_size)
l.grad[0, 0] = 1.0 # dl/dl = 1.0
return l
def forward(self, x):
if not isinstance(x, Parameter):
x = Parameter(data=x, eval_grad=False, graph=self.graph)
out = self.graph.dropout(x, p=self.p)
return out
def forward(self, x):
if not isinstance(x, Parameter):
x = Parameter(data=x, eval_grad=False, graph=self.graph)
out = self.graph.max_pool2d(x,
k=self.kernel_size,
s=self.stride,
p=self.padding)
return out
def MSELoss(self, y_out, y_true):
if not isinstance(y_true, Parameter):
y_true = Parameter(data=y_true, eval_grad=False, graph=self.graph)
batch_size = Parameter((1, 1),
init_zeros=True,
eval_grad=False,
graph=self.graph) # mini-batch size
batch_size.data.fill(float(y_true.shape[-1]))
# L = (y_out - y_true)^2
l = self.graph.sum(
self.graph.multiply(self.graph.subtract(y_out, y_true),
self.graph.subtract(y_out, y_true)))
# avg_loss = (1/m)*sigma{i = 1,..,m}(loss[i])
l = self.graph.divide(l, batch_size)
l.grad[0, 0] = 1.0 # dl/dl = 1.0
return l
def forward(self, x):
if not isinstance(x, Parameter):
x = Parameter(data=x, eval_grad=False, graph=self.graph)
# convolution operation
out = self.graph.conv2d_faster(x, self.K, self.stride, self.padding)
if self.bias: # adding bias
out = self.graph.add(out, self.b, axis=(-3, -2, -1))
return out
def init_params(self):
root_k = np.sqrt(1. / self.hidden_size)
self.W_ih = Parameter((self.hidden_size, self.input_size),
uniform=True,
low=-root_k,
high=root_k,
graph=self.graph)
self.W_hh = Parameter((self.hidden_size, self.hidden_size),
uniform=True,
low=-root_k,
high=root_k,
graph=self.graph)
self.b_ih = Parameter((self.hidden_size, 1),
uniform=True,
low=-root_k,
high=root_k,
graph=self.graph) # not much use
self.b_hh = Parameter((self.hidden_size, 1),
uniform=True,
low=-root_k,
high=root_k,
graph=self.graph)
def init_params(self):
root_k = np.sqrt(1. / self.hidden_size)
self.W_ih = Parameter(
(4 * self.hidden_size, self.input_size),
uniform=True,
low=-root_k,
high=root_k,
graph=self.graph) # input to hidden weight volume
self.W_hh = Parameter(
(4 * self.hidden_size, self.hidden_size),
uniform=True,
low=-root_k,
high=root_k,
graph=self.graph) # hidden to hidden weight volume
self.b_ih = Parameter((4 * self.hidden_size, 1),
uniform=True,
low=-root_k,
high=root_k,
graph=self.graph) # input to hidden bias vector
self.b_hh = Parameter((4 * self.hidden_size, 1),
uniform=True,
low=-root_k,
high=root_k,
graph=self.graph) # hidden to hidden bias vector
def TestLoss(self, y_out):
batch_size = Parameter((1, 1),
init_zeros=True,
eval_grad=False,
graph=self.graph) # mini-batch size
batch_size.data.fill(float(y_out.shape[-1]))
# a test loss score function that measures the sum of elements of each output vector as the loss of that sample
# helps identify leaks in between samples in a batch
l = self.graph.sum(y_out)
l = self.graph.divide(l, batch_size)
l.grad[0, 0] = 1.0
return l
def forward(self, x):
# making the input compatible with graph operations
if not isinstance(x, Parameter):
x = Parameter(data=x, eval_grad=False, graph=self.graph)
# flatten the input if it came from layers like Conv2d
if len(x.shape) > 2:
x = self.graph.reshape(x)
# y = Wx + b
out = self.graph.dot(self.W, x) # matmul
if self.bias: # adding bias
out = self.graph.add(out, self.b, axis=(-1, ))
return out
def forward(self, x, hidden):
h = hidden
if not isinstance(x, Parameter):
x = Parameter(data=x, eval_grad=False, graph=self.graph)
i_h = self.graph.dot(self.W_ih, x)
if self.bias:
i_h = self.graph.add(i_h, self.b_ih, axis=(-1, ))
h_h = self.graph.dot(self.W_hh, h)
if self.bias:
h_h = self.graph.add(h_h, self.b_hh, axis=(-1, ))
h = self.graph.add(i_h, h_h)
h = self.graph.tanh(h)
return h
def init_params(self):
root_k = np.sqrt(1. / (self.output_channels * self.kernel_size[0] * self.kernel_size[1]))
self.K = Parameter((self.input_channels, *self.filter_size), uniform=True, low=-root_k, high=root_k, graph=self.graph)
self.b = Parameter((self.output_channels, 1, 1, 1), uniform=True, low=-root_k, high=root_k, graph=self.graph)
def init_params(self):
shape = (*self.hidden_shape, 1)
self.gamma = Parameter(shape, init_ones=True, graph=self.graph)
self.beta = Parameter(shape, init_zeros=True, graph=self.graph)
self.m = np.sum(np.zeros(shape), axis=self.axis, keepdims=True) / shape[self.axis] # moving mean
self.v = np.sum(np.ones(shape), axis=self.axis, keepdims=True) / shape[self.axis] # moving variance
def init_params(self):
self.K = Parameter((self.input_channels, *self.filter_size),
graph=self.graph)
self.b = Parameter((self.output_channels, 1, 1, 1),
init_zeros=True,
graph=self.graph)
class Environment(gym.Env):
params = [
Parameter('param_' + str(x), min_value=-10, max_value=10)
for x in range(MAX_RESULTS)
]
def __init__(self, agent):
super(Environment, self).__init__()
self._agent = agent
self._epoch_num = 0
self._last_render = None
self._state = {}
self.last = {
'reward': 0.0,
'observation': None,
'policy': None,
'params': {},
'state': None,
'state_v': None
}
self.action_space = spaces.MultiDiscrete(
[p.space_size() for p in Environment.params if p.trainable])
self.observation_space = spaces.Box(low=0,
high=1,
shape=featurizer.shape,
dtype=np.float32)
self.reward_range = reward.range
@staticmethod
def policy_size():
return len(list(p for p in Environment.params if p.trainable))
@staticmethod
def policy_to_params(policy):
num = len(policy)
params = {}
assert len(Environment.params) == num
for i in range(num):
param = Environment.params[i]
params[param.name] = param.to_param_value(policy[i])
return params
def _apply_policy(self, policy):
new_params = Environment.policy_to_params(policy)
self.last['policy'] = policy
self.last['params'] = new_params
self._agent.on_ai_policy(new_params)
def set_state(self, state):
self._state = state
return self._state
def get_state(self):
return self._state
def step(self, policy):
self._apply_policy(policy)
self._epoch_num += 1
f = reward.RewardFunction()
self.last['reward'] = f(self._state)
self.last['state'] = self._state
self.last['state_v'] = featurizer.featurize(self._state)
self._agent.on_ai_step()
return self.last['state_v'], self.last[
'reward'], not self._agent.is_training(), {}
def reset(self):
self._epoch_num = 0
def render(self, mode='human', close=False, force=False):
return