我们从Python开源项目中，提取了以下13个代码示例，用于说明如何使用keras.initializations.uniform()。
def test_uniform(tensor_shape): _runner(initializations.uniform, tensor_shape, target_mean=0., target_max=0.05, target_min=-0.05)
def create_critic_network(self, state_size,action_dim): print("Now we build the model") S = Input(shape=[state_size]) A = Input(shape=[action_dim],name='action2') w = Dense(HIDDEN1_UNITS, init='he_uniform',activation='relu')(S) h = merge([w,A],mode='concat') h3 = Dense(HIDDEN2_UNITS, init='he_uniform',activation='relu')(h) V = Dense(action_dim,init=lambda shape, name: uniform(shape, scale=3e-3, name=name),activation='linear')(h3) model = Model(input=[S,A],output=V) adam = Adam(lr=self.LEARNING_RATE) model.compile(loss='mse', optimizer=adam) return model, A, S
def create_actor_network(self, state_size,action_dim): print("Now we build the model") model = Sequential() S = Input(shape=[state_size]) h0 = Dense(100, init='he_uniform',activation='relu')(S) h1 = Dense(100, init='he_uniform',activation='relu')(h0) V = Dense(8, init=lambda shape, name: uniform(shape, scale=3e-3, name=name),activation='tanh')(h1) model = Model(input=S,output=V) return model, model.trainable_weights, S
def unitary_ASB2016_init(shape, name=None): assert shape[0]==shape[1] N=shape[1] theta = initializations.uniform((3,N),scale=np.pi,name='{}_theta'.format(name)) reflection = initializations.glorot_uniform((2,2*N),name='{}_reflection'.format(name)) idxperm = np.random.permutation(N) idxpermaug = np.concatenate((idxperm,N+idxperm)) Iaug=augLeft(np.concatenate((np.eye(N),np.zeros((N,N))),axis=0),module=np).astype(np.float32) Uaug=times_unitary_ASB2016(Iaug,N,[theta,reflection,idxpermaug]) return Uaug,theta,reflection,idxpermaug
def buildConvolution(self, name): filters = self.params.get('filters') nb_filter = self.params.get('nb_filter') assert filters assert nb_filter convs = [] for fsz in filters: layer_name = '%s-conv-%d' % (name, fsz) conv = Convolution1D( nb_filter=nb_filter, filter_length=fsz, border_mode='valid', #activation='relu', subsample_length=1, init='glorot_uniform', #init=init, #init=lambda shape, name: initializations.uniform(shape, scale=0.01, name=name), W_constraint=maxnorm(self.params.get('w_maxnorm')), b_constraint=maxnorm(self.params.get('b_maxnorm')), #W_regularizer=regularizers.l2(self.params.get('w_l2')), #b_regularizer=regularizers.l2(self.params.get('b_l2')), #input_shape=(self.q_length, self.wdim), name=layer_name ) convs.append(conv) self.layers['%s-convolution' % name] = convs
def glorot_uniform_sigm(shape, name=None, dim_ordering='th'): """ Glorot style weight initializer for sigmoid activations. Like keras.initializations.glorot_uniform(), but with uniform random interval like in Deeplearning.net tutorials. They claim that the initialization random interval should be +/- sqrt(6 / (fan_in + fan_out)) (like Keras' glorot_uniform()) when tanh activations are used, +/- 4 sqrt(6 / (fan_in + fan_out)) when sigmoid activations are used. See: http://deeplearning.net/tutorial/mlp.html#going-from-logistic-regression-to-mlp """ fan_in, fan_out = get_fans(shape, dim_ordering=dim_ordering) s = 4. * np.sqrt(6. / (fan_in + fan_out)) return uniform(shape, s, name=name)
def build(self, input_shape): self.input_spec = [InputSpec(shape=input_shape)] if self.stateful: self.reset_states() else: # initial states: all-zero tensor of shape (output_dim) self.states = [None] input_dim = input_shape[2] self.input_dim = input_dim self.W = self.init((input_dim, self.output_dim), name='{}_W'.format(self.name)) #self.b = K.zeros((self.N,), name='{}_b'.format(self.name)) self.b = initializations.uniform((self.N,),scale=0.01,name='{}_b'.format(self.name)) self.baug=K.tile(self.b,[2]) h0 = self.h0_mean+initializations.uniform((2*self.N,),scale=0.01).get_value() self.h0 = K.variable(h0,name='{}_h0'.format(self.name)) if ('full' in self.unitary_impl): # we're using a full unitary recurrence matrix if (self.inner_init=='svd'): # use SVD to initialize U self.U = unitary_svd_init((self.N, self.N),name='{}_U'.format(self.name)) elif (self.inner_init=='ASB2016'): # use parameterization of [ASB2016] to initialize U Uaug,_,_,_ = unitary_ASB2016_init((self.N,self.N)) Uaug=Uaug.eval() self.U=K.variable(np.concatenate((Uaug[:self.N,:self.N],Uaug[:self.N,self.N:]),axis=0),name='{}_U'.format(self.name)) self.Uaug=augRight(self.U,module=K) elif (self.unitary_impl=='ASB2016'): # we're using the parameterization of [Arjovsky, Shah, Bengio 2016] self.Uaug,self.theta,self.reflection,_ = unitary_ASB2016_init((self.N, self.N),name=self.name) # set the trainable weights if ('full' in self.unitary_impl): self.trainable_weights = [self.W, self.U, self.b, self.h0] elif (self.unitary_impl=='ASB2016'): self.trainable_weights = [self.W, self.theta, self.reflection, self.b, self.h0] self.regularizers = [] #if self.W_regularizer: # self.W_regularizer.set_param(self.W) # self.regularizers.append(self.W_regularizer) #if self.U_regularizer: # self.U_regularizer.set_param(self.U) # self.regularizers.append(self.U_regularizer) #if self.b_regularizer: # self.b_regularizer.set_param(self.b) # self.regularizers.append(self.b_regularizer) if self.initial_weights is not None: self.set_weights(self.initial_weights) del self.initial_weights
