Python theano.tensor 模块，square() 实例源码

def rbf_kernel(X0):
    XY = T.dot(X0, X0.transpose())
    x2 = T.reshape(T.sum(T.square(X0), axis=1), (X0.shape[0], 1))
    X2e = T.repeat(x2, X0.shape[0], axis=1)
    H = T.sub(T.add(X2e, X2e.transpose()), 2 * XY)

    V = H.flatten()

    # median distance
    h = T.switch(T.eq((V.shape[0] % 2), 0),
        # if even vector
        T.mean(T.sort(V)[ ((V.shape[0] // 2) - 1) : ((V.shape[0] // 2) + 1) ]),
        # if odd vector
        T.sort(V)[V.shape[0] // 2])

    h = T.sqrt(0.5 * h / T.log(X0.shape[0].astype('float32') + 1.0)) / 2.

    Kxy = T.exp(-H / h ** 2 / 2.0)
    neighbors = T.argsort(H, axis=1)[:, 1]

    return Kxy, neighbors, h

def adam_updates(params, cost, lr=0.001, mom1=0.9, mom2=0.999):
    updates = []
    grads = T.grad(cost, params)
    t = th.shared(np.cast[th.config.floatX](1.))
    for p, g in zip(params, grads):
        v = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        mg = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        v_t = mom1*v + (1. - mom1)*g
        mg_t = mom2*mg + (1. - mom2)*T.square(g)
        v_hat = v_t / (1. - mom1 ** t)
        mg_hat = mg_t / (1. - mom2 ** t)
        g_t = v_hat / T.sqrt(mg_hat + 1e-8)
        p_t = p - lr * g_t
        updates.append((v, v_t))
        updates.append((mg, mg_t))
        updates.append((p, p_t))
    updates.append((t, t+1))
    return updates

def adam_updates(params, cost, lr=0.001, mom1=0.9, mom2=0.999):
    updates = []
    grads = T.grad(cost, params)
    t = th.shared(np.cast[th.config.floatX](1.))
    for p, g in zip(params, grads):
        v = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        mg = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        v_t = mom1*v + (1. - mom1)*g
        mg_t = mom2*mg + (1. - mom2)*T.square(g)
        v_hat = v_t / (1. - mom1 ** t)
        mg_hat = mg_t / (1. - mom2 ** t)
        g_t = v_hat / T.sqrt(mg_hat + 1e-8)
        p_t = p - lr * g_t
        updates.append((v, v_t))
        updates.append((mg, mg_t))
        updates.append((p, p_t))
    updates.append((t, t+1))
    return updates

def get_output_for(self, input, deterministic=False, **kwargs):
        if deterministic:
            norm_features = (input-self.avg_batch_mean.dimshuffle(*self.dimshuffle_args)) / T.sqrt(1e-6 + self.avg_batch_var).dimshuffle(*self.dimshuffle_args)
        else:
            batch_mean = T.mean(input,axis=self.axes_to_sum).flatten()
            centered_input = input-batch_mean.dimshuffle(*self.dimshuffle_args)
            batch_var = T.mean(T.square(centered_input),axis=self.axes_to_sum).flatten()
            batch_stdv = T.sqrt(1e-6 + batch_var)
            norm_features = centered_input / batch_stdv.dimshuffle(*self.dimshuffle_args)

            # BN updates
            new_m = 0.9*self.avg_batch_mean + 0.1*batch_mean
            new_v = 0.9*self.avg_batch_var + T.cast((0.1*input.shape[0])/(input.shape[0]-1),th.config.floatX)*batch_var
            self.bn_updates = [(self.avg_batch_mean, new_m), (self.avg_batch_var, new_v)]

        if hasattr(self, 'g'):
            activation = norm_features*self.g.dimshuffle(*self.dimshuffle_args)
        else:
            activation = norm_features
        if hasattr(self, 'b'):
            activation += self.b.dimshuffle(*self.dimshuffle_args)

        return self.nonlinearity(activation)

def get_output_for(self, input, init=False, deterministic=False, **kwargs):
        if input.ndim > 2:
            # if the input has more than two dimensions, flatten it into a
            # batch of feature vectors.
            input = input.flatten(2)

        activation = T.dot(input, self.W)

        if init:
            ma = T.mean(activation, axis=0)
            activation -= ma.dimshuffle('x',0)
            stdv = T.sqrt(T.mean(T.square(activation),axis=0))
            activation /= stdv.dimshuffle('x',0)
            self.init_updates = [(self.weight_scale, self.weight_scale/stdv), (self.b, -ma/stdv)]
        else:
            activation += self.b.dimshuffle('x', 0)

        return self.nonlinearity(activation)

def adam_updates(params, cost, lr=0.001, mom1=0.9, mom2=0.999):
    updates = []
    grads = T.grad(cost, params)
    t = th.shared(np.cast[th.config.floatX](1.))
    for p, g in zip(params, grads):
        v = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        mg = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        v_t = mom1*v + (1. - mom1)*g
        mg_t = mom2*mg + (1. - mom2)*T.square(g)
        v_hat = v_t / (1. - mom1 ** t)
        mg_hat = mg_t / (1. - mom2 ** t)
        g_t = v_hat / T.sqrt(mg_hat + 1e-8)
        p_t = p - lr * g_t
        updates.append((v, v_t))
        updates.append((mg, mg_t))
        updates.append((p, p_t))
    updates.append((t, t+1))
    return updates

def get_output_for(self, input, deterministic=False, **kwargs):
        if deterministic:
            norm_features = (input-self.avg_batch_mean.dimshuffle(*self.dimshuffle_args)) / T.sqrt(1e-6 + self.avg_batch_var).dimshuffle(*self.dimshuffle_args)
        else:
            batch_mean = T.mean(input,axis=self.axes_to_sum).flatten()
            centered_input = input-batch_mean.dimshuffle(*self.dimshuffle_args)
            batch_var = T.mean(T.square(centered_input),axis=self.axes_to_sum).flatten()
            batch_stdv = T.sqrt(1e-6 + batch_var)
            norm_features = centered_input / batch_stdv.dimshuffle(*self.dimshuffle_args)

            # BN updates
            new_m = 0.9*self.avg_batch_mean + 0.1*batch_mean
            new_v = 0.9*self.avg_batch_var + T.cast((0.1*input.shape[0])/(input.shape[0]-1),th.config.floatX)*batch_var
            self.bn_updates = [(self.avg_batch_mean, new_m), (self.avg_batch_var, new_v)]

        if hasattr(self, 'g'):
            activation = norm_features*self.g.dimshuffle(*self.dimshuffle_args)
        else:
            activation = norm_features
        if hasattr(self, 'b'):
            activation += self.b.dimshuffle(*self.dimshuffle_args)

        return self.nonlinearity(activation)

def l2normalize(layer, train_scale=True):
    W_param = layer.W
    s = W_param.get_value().shape
    if len(s)==4:
        axes_to_sum = (1,2,3)
        dimshuffle_args = [0,'x','x','x']
        k = s[0]
    else:
        axes_to_sum = 0
        dimshuffle_args = ['x',0]
        k = s[1]
    layer.W_scale = layer.add_param(lasagne.init.Constant(1.),
                          (k,), name="W_scale", trainable=train_scale, regularizable=False)
    layer.W = W_param * (layer.W_scale/T.sqrt(1e-6 + T.sum(T.square(W_param),axis=axes_to_sum))).dimshuffle(*dimshuffle_args)
    return layer

# fully connected layer with weight normalization

def get_output_for(self, input, init=False, deterministic=False, **kwargs):
        if input.ndim > 2:
            # if the input has more than two dimensions, flatten it into a
            # batch of feature vectors.
            input = input.flatten(2)

        activation = T.dot(input, self.W)

        if init:
            ma = T.mean(activation, axis=0)
            activation -= ma.dimshuffle('x',0)
            stdv = T.sqrt(T.mean(T.square(activation),axis=0))
            activation /= stdv.dimshuffle('x',0)
            self.init_updates = [(self.weight_scale, self.weight_scale/stdv), (self.b, -ma/stdv)]
        else:
            activation += self.b.dimshuffle('x', 0)

        return self.nonlinearity(activation)

def kl_sym(self, old_dist_info_vars, new_dist_info_vars):
        old_means = old_dist_info_vars["mean"]
        old_log_stds = old_dist_info_vars["log_std"]
        new_means = new_dist_info_vars["mean"]
        new_log_stds = new_dist_info_vars["log_std"]
        """
        Compute the KL divergence of two multivariate Gaussian distribution with
        diagonal covariance matrices
        """
        old_std = TT.exp(old_log_stds)
        new_std = TT.exp(new_log_stds)
        # means: (N*A)
        # std: (N*A)
        # formula:
        # { (\mu_1 - \mu_2)^2 + \sigma_1^2 - \sigma_2^2 } / (2\sigma_2^2) +
        # ln(\sigma_2/\sigma_1)
        numerator = TT.square(old_means - new_means) + \
                    TT.square(old_std) - TT.square(new_std)
        denominator = 2 * TT.square(new_std) + 1e-8
        return TT.sum(
            numerator / denominator + new_log_stds - old_log_stds, axis=-1)

def kl(self, old_dist_info, new_dist_info):
        old_means = old_dist_info["mean"]
        old_log_stds = old_dist_info["log_std"]
        new_means = new_dist_info["mean"]
        new_log_stds = new_dist_info["log_std"]
        """
        Compute the KL divergence of two multivariate Gaussian distribution with
        diagonal covariance matrices
        """
        old_std = np.exp(old_log_stds)
        new_std = np.exp(new_log_stds)
        # means: (N*A)
        # std: (N*A)
        # formula:
        # { (\mu_1 - \mu_2)^2 + \sigma_1^2 - \sigma_2^2 } / (2\sigma_2^2) +
        # ln(\sigma_2/\sigma_1)
        numerator = np.square(old_means - new_means) + \
                    np.square(old_std) - np.square(new_std)
        denominator = 2 * np.square(new_std) + 1e-8
        return np.sum(
            numerator / denominator + new_log_stds - old_log_stds, axis=-1)

def rbf_kernel(X0):
    XY = T.dot(X0, X0.transpose())
    x2 = T.reshape(T.sum(T.square(X0), axis=1), (X0.shape[0], 1))
    X2e = T.repeat(x2, X0.shape[0], axis=1)
    H = T.sub(T.add(X2e, X2e.transpose()), 2 * XY)

    V = H.flatten()

    # median distance
    h = T.switch(T.eq((V.shape[0] % 2), 0),
        # if even vector
        T.mean(T.sort(V)[ ((V.shape[0] // 2) - 1) : ((V.shape[0] // 2) + 1) ]),
        # if odd vector
        T.sort(V)[V.shape[0] // 2])

    h = T.sqrt(0.5 * h / T.log(X0.shape[0].astype('float32') + 1.0)) / 2.

    Kxy = T.exp(-H / h ** 2 / 2.0)
    neighbors = T.argsort(H, axis=1)[:, 1]

    return Kxy, neighbors, h

def kl_sym(self, old_dist_info_vars, new_dist_info_vars):
        old_means = old_dist_info_vars["mean"]
        old_log_stds = old_dist_info_vars["log_std"]
        new_means = new_dist_info_vars["mean"]
        new_log_stds = new_dist_info_vars["log_std"]
        """
        Compute the KL divergence of two multivariate Gaussian distribution with
        diagonal covariance matrices
        """
        old_std = TT.exp(old_log_stds)
        new_std = TT.exp(new_log_stds)
        # means: (N*A)
        # std: (N*A)
        # formula:
        # { (\mu_1 - \mu_2)^2 + \sigma_1^2 - \sigma_2^2 } / (2\sigma_2^2) +
        # ln(\sigma_2/\sigma_1)
        numerator = TT.square(old_means - new_means) + \
                    TT.square(old_std) - TT.square(new_std)
        denominator = 2 * TT.square(new_std) + 1e-8
        return TT.sum(
            numerator / denominator + new_log_stds - old_log_stds, axis=-1)

def kl(self, old_dist_info, new_dist_info):
        old_means = old_dist_info["mean"]
        old_log_stds = old_dist_info["log_std"]
        new_means = new_dist_info["mean"]
        new_log_stds = new_dist_info["log_std"]
        """
        Compute the KL divergence of two multivariate Gaussian distribution with
        diagonal covariance matrices
        """
        old_std = np.exp(old_log_stds)
        new_std = np.exp(new_log_stds)
        # means: (N*A)
        # std: (N*A)
        # formula:
        # { (\mu_1 - \mu_2)^2 + \sigma_1^2 - \sigma_2^2 } / (2\sigma_2^2) +
        # ln(\sigma_2/\sigma_1)
        numerator = np.square(old_means - new_means) + \
                    np.square(old_std) - np.square(new_std)
        denominator = 2 * np.square(new_std) + 1e-8
        return np.sum(
            numerator / denominator + new_log_stds - old_log_stds, axis=-1)

def get_output_for(self, input, deterministic=False, **kwargs):
        if deterministic:
            norm_features = (input-self.avg_batch_mean.dimshuffle(*self.dimshuffle_args)) / T.sqrt(1e-6 + self.avg_batch_var).dimshuffle(*self.dimshuffle_args)
        else:
            batch_mean = T.mean(input,axis=self.axes_to_sum).flatten()
            centered_input = input-batch_mean.dimshuffle(*self.dimshuffle_args)
            batch_var = T.mean(T.square(centered_input),axis=self.axes_to_sum).flatten()
            batch_stdv = T.sqrt(1e-6 + batch_var)
            norm_features = centered_input / batch_stdv.dimshuffle(*self.dimshuffle_args)

            # BN updates
            new_m = 0.9*self.avg_batch_mean + 0.1*batch_mean
            new_v = 0.9*self.avg_batch_var + T.cast((0.1*input.shape[0])/(input.shape[0]-1),th.config.floatX)*batch_var
            self.bn_updates = [(self.avg_batch_mean, new_m), (self.avg_batch_var, new_v)]

        if hasattr(self, 'g'):
            activation = norm_features*self.g.dimshuffle(*self.dimshuffle_args)
        else:
            activation = norm_features
        if hasattr(self, 'b'):
            activation += self.b.dimshuffle(*self.dimshuffle_args)

        return self.nonlinearity(activation)

def adam_updates(params, cost, lr=0.001, mom1=0.9, mom2=0.999):
    updates = []
    grads = T.grad(cost, params)
    t = th.shared(np.cast[th.config.floatX](1.))
    for p, g in zip(params, grads):
        v = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        mg = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        v_t = mom1*v + (1. - mom1)*g
        mg_t = mom2*mg + (1. - mom2)*T.square(g)
        v_hat = v_t / (1. - mom1 ** t)
        mg_hat = mg_t / (1. - mom2 ** t)
        g_t = v_hat / T.sqrt(mg_hat + 1e-8)
        p_t = p - lr * g_t
        updates.append((v, v_t))
        updates.append((mg, mg_t))
        updates.append((p, p_t))
    updates.append((t, t+1))
    return updates

def get_output_for(self, input, deterministic=False, **kwargs):
        if deterministic:
            norm_features = (input-self.avg_batch_mean.dimshuffle(*self.dimshuffle_args)) / T.sqrt(1e-6 + self.avg_batch_var).dimshuffle(*self.dimshuffle_args)
        else:
            batch_mean = T.mean(input,axis=self.axes_to_sum).flatten()
            centered_input = input-batch_mean.dimshuffle(*self.dimshuffle_args)
            batch_var = T.mean(T.square(centered_input),axis=self.axes_to_sum).flatten()
            batch_stdv = T.sqrt(1e-6 + batch_var)
            norm_features = centered_input / batch_stdv.dimshuffle(*self.dimshuffle_args)

            # BN updates
            new_m = 0.9*self.avg_batch_mean + 0.1*batch_mean
            new_v = 0.9*self.avg_batch_var + T.cast((0.1*input.shape[0])/(input.shape[0]-1),th.config.floatX)*batch_var
            self.bn_updates = [(self.avg_batch_mean, new_m), (self.avg_batch_var, new_v)]

        if hasattr(self, 'g'):
            activation = norm_features*self.g.dimshuffle(*self.dimshuffle_args)
        else:
            activation = norm_features
        if hasattr(self, 'b'):
            activation += self.b.dimshuffle(*self.dimshuffle_args)

        return self.nonlinearity(activation)

def get_output_for(self, input, init=False, deterministic=False, **kwargs):
        if input.ndim > 2:
            # if the input has more than two dimensions, flatten it into a
            # batch of feature vectors.
            input = input.flatten(2)

        activation = T.dot(input, self.W)

        if init:
            ma = T.mean(activation, axis=0)
            activation -= ma.dimshuffle('x',0)
            stdv = T.sqrt(T.mean(T.square(activation),axis=0))
            activation /= stdv.dimshuffle('x',0)
            self.init_updates = [(self.weight_scale, self.weight_scale/stdv), (self.b, -ma/stdv)]
        else:
            activation += self.b.dimshuffle('x', 0)

        return self.nonlinearity(activation)

def get_output(self, input_):
        """
        This function overrides the parents' one.
        Creates symbolic function to compute output from an input.

        Parameters
        ----------
        input_: TensorVariable

        Returns
        -------
        TensorVariable
        """
        mu = input_[:, :self.output_shape[0]]
        logvar = input_[:, self.output_shape[0]:]
        return 0.5 * T.mean(T.square(mu) + T.exp(logvar) - logvar - 1)

def get_output(self, input1_, input2_):
        """
        This function overrides the parents' one.
        Creates symbolic function to compute output from an input.
        http://stats.stackexchange.com/questions/66271/kullback-leibler-divergence-of-two-normal-distributions

        Parameters
        ----------
        input_: TensorVariable

        Returns
        -------
        TensorVariable
        """
        mu1 = input1_[:, :self.output_shape[0]]
        logvar1 = input1_[:, self.output_shape[0]:]
        mu2 = input2_[:, :self.output_shape[0]]
        logvar2 = input2_[:, self.output_shape[0]:]
        return 0.5 * T.mean((T.square(mu1 - mu2) + T.exp(logvar1) + T.exp(logvar2)) * (1 / T.exp(logvar1) + 1 / T.exp(logvar2))) - 2

# TODO: Fix L1, L2 to work!

def adam_updates(params, cost, lr=0.001, mom1=0.9, mom2=0.999):
    updates = []
    grads = T.grad(cost, params)
    t = th.shared(np.cast[th.config.floatX](1.))
    for p, g in zip(params, grads):
        v = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        mg = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        v_t = mom1*v + (1. - mom1)*g
        mg_t = mom2*mg + (1. - mom2)*T.square(g)
        v_hat = v_t / (1. - mom1 ** t)
        mg_hat = mg_t / (1. - mom2 ** t)
        g_t = v_hat / T.sqrt(mg_hat + 1e-8)
        p_t = p - lr * g_t
        updates.append((v, v_t))
        updates.append((mg, mg_t))
        updates.append((p, p_t))
    updates.append((t, t+1))
    return updates

def get_output_for(self, input, deterministic=False, **kwargs):
        if deterministic:
            norm_features = (input-self.avg_batch_mean.dimshuffle(*self.dimshuffle_args)) / T.sqrt(1e-6 + self.avg_batch_var).dimshuffle(*self.dimshuffle_args)
        else:
            batch_mean = T.mean(input,axis=self.axes_to_sum).flatten()
            centered_input = input-batch_mean.dimshuffle(*self.dimshuffle_args)
            batch_var = T.mean(T.square(centered_input),axis=self.axes_to_sum).flatten()
            batch_stdv = T.sqrt(1e-6 + batch_var)
            norm_features = centered_input / batch_stdv.dimshuffle(*self.dimshuffle_args)

            # BN updates
            new_m = 0.9*self.avg_batch_mean + 0.1*batch_mean
            new_v = 0.9*self.avg_batch_var + T.cast((0.1*input.shape[0])/(input.shape[0]-1.), th.config.floatX)*batch_var
            self.bn_updates = [(self.avg_batch_mean, new_m), (self.avg_batch_var, new_v)]

        if hasattr(self, 'g'):
            activation = norm_features*self.g.dimshuffle(*self.dimshuffle_args)
        else:
            activation = norm_features
        if hasattr(self, 'b'):
            activation += self.b.dimshuffle(*self.dimshuffle_args)

        return self.nonlinearity(activation)

def adam_updates(params, cost, lr=0.001, mom1=0.9, mom2=0.999):
    updates = []
    grads = T.grad(cost, params)
    t = th.shared(np.cast[th.config.floatX](1.))
    for p, g in zip(params, grads):
        v = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        mg = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        v_t = mom1*v + (1. - mom1)*g
        mg_t = mom2*mg + (1. - mom2)*T.square(g)
        v_hat = v_t / (1. - mom1 ** t)
        mg_hat = mg_t / (1. - mom2 ** t)
        g_t = v_hat / T.sqrt(mg_hat + 1e-8)
        p_t = p - lr * g_t
        updates.append((v, v_t))
        updates.append((mg, mg_t))
        updates.append((p, p_t))
    updates.append((t, t+1))
    return updates

def adam_conditional_updates(params, cost, mincost, lr=0.001, mom1=0.9, mom2=0.999): # if cost is less than mincost, don't do update
    updates = []
    grads = T.grad(cost, params)
    t = th.shared(np.cast[th.config.floatX](1.))
    for p, g in zip(params, grads):
        v = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        mg = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        v_t = mom1*v + (1. - mom1)*g
        mg_t = mom2*mg + (1. - mom2)*T.square(g)
        v_hat = v_t / (1. - mom1 ** t)
        mg_hat = mg_t / (1. - mom2 ** t)
        g_t = v_hat / T.sqrt(mg_hat + 1e-8)
        p_t = p - lr * g_t
        updates.append((v, ifelse(cost<mincost,v,v_t)))
        updates.append((mg, ifelse(cost<mincost,mg,mg_t)))
        updates.append((p, ifelse(cost<mincost,p,p_t)))
    updates.append((t, ifelse(cost<mincost,t,t+1)))
    return updates

def get_output_for(self, input, deterministic=False, **kwargs):
        if deterministic:
            norm_features = (input-self.avg_batch_mean.dimshuffle(*self.dimshuffle_args)) / T.sqrt(1e-6 + self.avg_batch_var).dimshuffle(*self.dimshuffle_args)
        else:
            batch_mean = T.mean(input,axis=self.axes_to_sum).flatten()
            centered_input = input-batch_mean.dimshuffle(*self.dimshuffle_args)
            batch_var = T.mean(T.square(centered_input),axis=self.axes_to_sum).flatten()
            batch_stdv = T.sqrt(1e-6 + batch_var)
            norm_features = centered_input / batch_stdv.dimshuffle(*self.dimshuffle_args)

            # BN updates
            new_m = 0.9*self.avg_batch_mean + 0.1*batch_mean
            new_v = 0.9*self.avg_batch_var + T.cast((0.1*input.shape[0])/(input.shape[0]-1),th.config.floatX)*batch_var
            self.bn_updates = [(self.avg_batch_mean, new_m), (self.avg_batch_var, new_v)]

        if hasattr(self, 'g'):
            activation = norm_features*self.g.dimshuffle(*self.dimshuffle_args)
        else:
            activation = norm_features
        if hasattr(self, 'b'):
            activation += self.b.dimshuffle(*self.dimshuffle_args)
    if self.nonlinearity is not None:
            return self.nonlinearity(activation)
    else:
        return activation

def adam_updates(params, cost, lr=0.001, mom1=0.9, mom2=0.999):
    updates = []
    grads = T.grad(cost, params)
    t = th.shared(np.cast[th.config.floatX](1.))
    for p, g in zip(params, grads):
        v = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        mg = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        v_t = mom1*v + (1. - mom1)*g
        mg_t = mom2*mg + (1. - mom2)*T.square(g)
        v_hat = v_t / (1. - mom1 ** t)
        mg_hat = mg_t / (1. - mom2 ** t)
        g_t = v_hat / T.sqrt(mg_hat + 1e-8)
        p_t = p - lr * g_t
        updates.append((v, v_t))
        updates.append((mg, mg_t))
        updates.append((p, p_t))
    updates.append((t, t+1))
    return updates

def get_output_for(self, input, deterministic=False, set_bn_updates=True, **kwargs):
        if deterministic:
            norm_features = (input-self.avg_batch_mean.dimshuffle(*self.dimshuffle_args)) / T.sqrt(1e-6 + self.avg_batch_var).dimshuffle(*self.dimshuffle_args)
        else:
            batch_mean = T.mean(input,axis=self.axes_to_sum).flatten()
            centered_input = input-batch_mean.dimshuffle(*self.dimshuffle_args)
            batch_var = T.mean(T.square(centered_input),axis=self.axes_to_sum).flatten()
            batch_stdv = T.sqrt(1e-6 + batch_var)
            norm_features = centered_input / batch_stdv.dimshuffle(*self.dimshuffle_args)

            # BN updates
            if set_bn_updates:
                new_m = 0.9*self.avg_batch_mean + 0.1*batch_mean
                new_v = 0.9*self.avg_batch_var + T.cast((0.1*input.shape[0])/(input.shape[0]-1),th.config.floatX)*batch_var
                self.bn_updates = [(self.avg_batch_mean, new_m), (self.avg_batch_var, new_v)]

        if hasattr(self, 'g'):
            activation = norm_features*self.g.dimshuffle(*self.dimshuffle_args)
        else:
            activation = norm_features
        if hasattr(self, 'b'):
            activation += self.b.dimshuffle(*self.dimshuffle_args)

        return self.nonlinearity(activation)

def get_output_for(self, input, init=False, deterministic=False, **kwargs):
        if input.ndim > 2:
            # if the input has more than two dimensions, flatten it into a
            # batch of feature vectors.
            input = input.flatten(2)

        activation = T.dot(input, self.W)

        if init:
            ma = T.mean(activation, axis=0)
            activation -= ma.dimshuffle('x',0)
            stdv = T.sqrt(T.mean(T.square(activation),axis=0))
            activation /= stdv.dimshuffle('x',0)
            self.init_updates = [(self.weight_scale, self.weight_scale/stdv), (self.b, -ma/stdv)]
        else:
            activation += self.b.dimshuffle('x', 0)

        return self.nonlinearity(activation)


# comes from Ishamel code base

def __init__(self, dim, eps=1e-6, init_count=0, init_mean=0., init_meansq=1.):
        '''
        Args:
            dim: dimension of the space of points to be standardized
            eps: small constant to add to denominators to prevent division by 0
            init_count, init_mean, init_meansq: initial values for accumulators

        Note:
            if init_count is 0, then init_mean and init_meansq have no effect beyond
            the first call to update(), which will ignore their values and
            replace them with values from a new batch of data.
        '''
        self._eps = eps
        self._dim = dim
        with variable_scope(type(self).__name__) as self.__varscope:
            self._count = get_variable('count', np.array(float(init_count)), trainable=False)
            self._mean_1_D = get_variable('mean_1_D', np.full((1, self._dim), init_mean), broadcastable=(True,False), trainable=False)
            self._meansq_1_D = get_variable('meansq_1_D', np.full((1, self._dim), init_meansq), broadcastable=(True,False), trainable=False)
        self._stdev_1_D = tensor.sqrt(tensor.nnet.relu(self._meansq_1_D - tensor.square(self._mean_1_D)))
        # Relu ensures inside is nonnegative. maybe the better choice would have been to
        # add self._eps inside the square root, but I'm keeping things this way to preserve
        # backwards compatibility with existing saved models.

        self.get_mean = self._mean_1_D.get_value
        self.get_stdev = theano.function([], self._stdev_1_D[0,:]) # TODO: return with shape (1,D)

def adam(cost, params, lr, beta1=0.9, beta2=0.999, eps=1e-8):
    updates = []
    grads = tensor.grad(cost, params); assert len(params) == len(grads)
    t0 = theano.shared(np.array(0., dtype=theano.config.floatX))
    t = t0 + 1
    corr1 = (1 - beta1**t)
    corr2 = (1 - beta2**t)
    alpha = lr * tensor.sqrt(corr2) / corr1
    for p, g in zip(params, grads):
        m = theano.shared(value=np.zeros(p.get_value().shape, dtype=theano.config.floatX), broadcastable=p.broadcastable)
        v = theano.shared(value=np.zeros(p.get_value().shape, dtype=theano.config.floatX), broadcastable=p.broadcastable)
        m_t = beta1 * m + (1 - beta1) * g
        v_t = beta2 * v + (1 - beta2) * tensor.square(g)
        p_t = p - alpha * m_t/(tensor.sqrt(v_t) + eps)
        updates.append((m, m_t))
        updates.append((v, v_t))
        updates.append((p, p_t))
    updates.append((t0, t))
    return updates

def kl_sym(self, old_dist_info_vars, new_dist_info_vars):
        old_means = old_dist_info_vars["mean"]
        old_log_stds = old_dist_info_vars["log_std"]
        new_means = new_dist_info_vars["mean"]
        new_log_stds = new_dist_info_vars["log_std"]
        """
        Compute the KL divergence of two multivariate Gaussian distribution with
        diagonal covariance matrices
        """
        old_std = TT.exp(old_log_stds)
        new_std = TT.exp(new_log_stds)
        # means: (N*A)
        # std: (N*A)
        # formula:
        # { (\mu_1 - \mu_2)^2 + \sigma_1^2 - \sigma_2^2 } / (2\sigma_2^2) +
        # ln(\sigma_2/\sigma_1)
        numerator = TT.square(old_means - new_means) + \
            TT.square(old_std) - TT.square(new_std)
        denominator = 2 * TT.square(new_std) + 1e-8
        return TT.sum(
            numerator / denominator + new_log_stds - old_log_stds, axis=-1)

def kl(self, old_dist_info, new_dist_info):
        old_means = old_dist_info["mean"]
        old_log_stds = old_dist_info["log_std"]
        new_means = new_dist_info["mean"]
        new_log_stds = new_dist_info["log_std"]
        """
        Compute the KL divergence of two multivariate Gaussian distribution with
        diagonal covariance matrices
        """
        old_std = np.exp(old_log_stds)
        new_std = np.exp(new_log_stds)
        # means: (N*A)
        # std: (N*A)
        # formula:
        # { (\mu_1 - \mu_2)^2 + \sigma_1^2 - \sigma_2^2 } / (2\sigma_2^2) +
        # ln(\sigma_2/\sigma_1)
        numerator = np.square(old_means - new_means) + \
            np.square(old_std) - np.square(new_std)
        denominator = 2 * np.square(new_std) + 1e-8
        return np.sum(
            numerator / denominator + new_log_stds - old_log_stds, axis=-1)

def adam_updates(params, cost, lr=0.001, mom1=0.9, mom2=0.999):
    updates = []
    grads = T.grad(cost, params)
    t = th.shared(np.cast[th.config.floatX](1.))
    for p, g in zip(params, grads):
        v = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        mg = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        v_t = mom1*v + (1. - mom1)*g
        mg_t = mom2*mg + (1. - mom2)*T.square(g)
        v_hat = v_t / (1. - mom1 ** t)
        mg_hat = mg_t / (1. - mom2 ** t)
        g_t = v_hat / T.sqrt(mg_hat + 1e-8)
        p_t = p - lr * g_t
        updates.append((v, v_t))
        updates.append((mg, mg_t))
        updates.append((p, p_t))
    updates.append((t, t+1))
    return updates

def get_output_for(self, input, deterministic=False, set_bn_updates=True, **kwargs):
        if deterministic:
            norm_features = (input-self.avg_batch_mean.dimshuffle(*self.dimshuffle_args)) / T.sqrt(1e-6 + self.avg_batch_var).dimshuffle(*self.dimshuffle_args)
        else:
            batch_mean = T.mean(input,axis=self.axes_to_sum).flatten()
            centered_input = input-batch_mean.dimshuffle(*self.dimshuffle_args)
            batch_var = T.mean(T.square(centered_input),axis=self.axes_to_sum).flatten()
            batch_stdv = T.sqrt(1e-6 + batch_var)
            norm_features = centered_input / batch_stdv.dimshuffle(*self.dimshuffle_args)

            # BN updates
            if set_bn_updates:
                new_m = 0.9*self.avg_batch_mean + 0.1*batch_mean
                new_v = 0.9*self.avg_batch_var + T.cast((0.1*input.shape[0])/(input.shape[0]-1),th.config.floatX)*batch_var
                self.bn_updates = [(self.avg_batch_mean, new_m), (self.avg_batch_var, new_v)]

        if hasattr(self, 'g'):
            activation = norm_features*self.g.dimshuffle(*self.dimshuffle_args)
        else:
            activation = norm_features
        if hasattr(self, 'b'):
            activation += self.b.dimshuffle(*self.dimshuffle_args)

        return self.nonlinearity(activation)

def get_output_for(self, input, init=False, deterministic=False, **kwargs):
        if input.ndim > 2:
            # if the input has more than two dimensions, flatten it into a
            # batch of feature vectors.
            input = input.flatten(2)

        activation = T.dot(input, self.W)

        if init:
            ma = T.mean(activation, axis=0)
            activation -= ma.dimshuffle('x',0)
            stdv = T.sqrt(T.mean(T.square(activation),axis=0))
            activation /= stdv.dimshuffle('x',0)
            self.init_updates = [(self.weight_scale, self.weight_scale/stdv), (self.b, -ma/stdv)]
        else:
            activation += self.b.dimshuffle('x', 0)

        return self.nonlinearity(activation)


# comes from Ishamel code base

def adam_updates(params, cost, lr=0.001, mom1=0.9, mom2=0.999):
    updates = []
    grads = T.grad(cost, params)
    t = th.shared(np.cast[th.config.floatX](1.))
    for p, g in zip(params, grads):
        value = p.get_value(borrow=True)
        v = th.shared(np.zeros(value.shape, dtype=value.dtype),
                               broadcastable=p.broadcastable)
        mg = th.shared(np.zeros(value.shape, dtype=value.dtype),
                               broadcastable=p.broadcastable)

        v_t = mom1*v + (1. - mom1)*g
        mg_t = mom2*mg + (1. - mom2)*T.square(g)
        v_hat = v_t / (1. - mom1 ** t)
        mg_hat = mg_t / (1. - mom2 ** t)
        g_t = v_hat / T.sqrt(mg_hat + 1e-8)
        p_t = p - lr * g_t
        updates.append((v, v_t))
        updates.append((mg, mg_t))
        updates.append((p, p_t))
    updates.append((t, t+1))
    return updates

def get_output_for(self, input, deterministic=False, **kwargs):
        if deterministic:
            norm_features = (input-self.avg_batch_mean.dimshuffle(*self.dimshuffle_args)) / T.sqrt(1e-6 + self.avg_batch_var).dimshuffle(*self.dimshuffle_args)
        else:
            batch_mean = T.mean(input,axis=self.axes_to_sum).flatten()
            centered_input = input-batch_mean.dimshuffle(*self.dimshuffle_args)
            batch_var = T.mean(T.square(centered_input),axis=self.axes_to_sum).flatten()
            batch_stdv = T.sqrt(1e-6 + batch_var)
            norm_features = centered_input / batch_stdv.dimshuffle(*self.dimshuffle_args)

            # BN updates
            new_m = 0.9*self.avg_batch_mean + 0.1*batch_mean
            new_v = 0.9*self.avg_batch_var + T.cast((0.1*input.shape[0])/(input.shape[0]-1),th.config.floatX)*batch_var
            self.bn_updates = [(self.avg_batch_mean, new_m), (self.avg_batch_var, new_v)]

        if hasattr(self, 'g'):
            activation = norm_features*self.g.dimshuffle(*self.dimshuffle_args)
        else:
            activation = norm_features
        if hasattr(self, 'b'):
            activation += self.b.dimshuffle(*self.dimshuffle_args)

        return self.nonlinearity(activation)

def get_output_for(self, input, init=False, deterministic=False, **kwargs):
        if input.ndim > 2:
            # if the input has more than two dimensions, flatten it into a
            # batch of feature vectors.
            input = input.flatten(2)

        activation = T.dot(input, self.W)

        if init:
            ma = T.mean(activation, axis=0)
            activation -= ma.dimshuffle('x',0)
            stdv = T.sqrt(T.mean(T.square(activation),axis=0))
            activation /= stdv.dimshuffle('x',0)
            self.init_updates = [(self.weight_scale, self.weight_scale/stdv), (self.b, -ma/stdv)]
        else:
            activation += self.b.dimshuffle('x', 0)

        return self.nonlinearity(activation)

def adam_updates(params, cost, lr=0.001, mom1=0.9, mom2=0.999):
    updates = []
    grads = T.grad(cost, params)
    t = th.shared(np.cast[th.config.floatX](1.))
    for p, g in zip(params, grads):
        v = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        mg = th.shared(np.cast[th.config.floatX](p.get_value() * 0.))
        v_t = mom1*v + (1. - mom1)*g
        mg_t = mom2*mg + (1. - mom2)*T.square(g)
        v_hat = v_t / (1. - mom1 ** t)
        mg_hat = mg_t / (1. - mom2 ** t)
        g_t = v_hat / T.sqrt(mg_hat + 1e-8)
        p_t = p - lr * g_t
        updates.append((v, v_t))
        updates.append((mg, mg_t))
        updates.append((p, p_t))
    updates.append((t, t+1))
    return updates

def get_output_for(self, input, deterministic=False, **kwargs):
        if deterministic:
            norm_features = (input-self.avg_batch_mean.dimshuffle(*self.dimshuffle_args)) / T.sqrt(1e-6 + self.avg_batch_var).dimshuffle(*self.dimshuffle_args)
        else:
            batch_mean = T.mean(input,axis=self.axes_to_sum).flatten()
            centered_input = input-batch_mean.dimshuffle(*self.dimshuffle_args)
            batch_var = T.mean(T.square(centered_input),axis=self.axes_to_sum).flatten()
            batch_stdv = T.sqrt(1e-6 + batch_var)
            norm_features = centered_input / batch_stdv.dimshuffle(*self.dimshuffle_args)

            # BN updates
            new_m = 0.9*self.avg_batch_mean + 0.1*batch_mean
            new_v = 0.9*self.avg_batch_var + T.cast((0.1*input.shape[0])/(input.shape[0]-1),th.config.floatX)*batch_var
            self.bn_updates = [(self.avg_batch_mean, new_m), (self.avg_batch_var, new_v)]

        if hasattr(self, 'g'):
            activation = norm_features*self.g.dimshuffle(*self.dimshuffle_args)
        else:
            activation = norm_features
        if hasattr(self, 'b'):
            activation += self.b.dimshuffle(*self.dimshuffle_args)

        return self.nonlinearity(activation)