numpy.float32的典型用法

import numpy as np
from numpy import float32

def draw_image(self, img, color=[0, 255, 0], alpha=1.0, copy=True, from_img=None):
        if copy:
            img = np.copy(img)

        orig_dtype = img.dtype
        if alpha != 1.0 and img.dtype != np.float32:
            img = img.astype(np.float32, copy=False)

        for rect in self:
            if from_img is not None:
                rect.resize(from_img, img).draw_on_image(img, color=color, alpha=alpha, copy=False)
            else:
                rect.draw_on_image(img, color=color, alpha=alpha, copy=False)

        if orig_dtype != img.dtype:
            img = img.astype(orig_dtype, copy=False)

        return img

import numpy as np
from numpy import float32

def generate_moving_mnist(self, num_digits=2):
    '''
    Get random trajectories for the digits and generate a video.
    '''
    data = np.zeros((self.n_frames_total, self.image_size_, self.image_size_), dtype=np.float32)
    for n in range(num_digits):
      # Trajectory
      start_y, start_x = self.get_random_trajectory(self.n_frames_total)
      ind = random.randint(0, self.mnist.shape[0] - 1)
      digit_image = self.mnist[ind]
      for i in range(self.n_frames_total):
        top    = start_y[i]
        left   = start_x[i]
        bottom = top + self.digit_size_
        right  = left + self.digit_size_
        # Draw digit
        data[i, top:bottom, left:right] = np.maximum(data[i, top:bottom, left:right], digit_image)

    data = data[..., np.newaxis]
    return data 

import numpy as np
from numpy import float32

def wav_format(self, input_wave_file, output_wave_file, target_phrase):
        pop_size = 100
        elite_size = 10
        mutation_p = 0.005
        noise_stdev = 40
        noise_threshold = 1
        mu = 0.9
        alpha = 0.001
        max_iters = 3000
        num_points_estimate = 100
        delta_for_gradient = 100
        delta_for_perturbation = 1e3
        input_audio = load_wav(input_wave_file).astype(np.float32)
        pop = np.expand_dims(input_audio, axis=0)
        pop = np.tile(pop, (pop_size, 1))
        output_wave_file = output_wave_file
        target_phrase = target_phrase
        funcs = setup_graph(pop, np.array([toks.index(x) for x in target_phrase])) 

import numpy as np
from numpy import float32

def get_rois_blob(im_rois, im_scale_factors):
    """Converts RoIs into network inputs.
    Arguments:
        im_rois (ndarray): R x 4 matrix of RoIs in original image coordinates
        im_scale_factors (list): scale factors as returned by _get_image_blob
    Returns:
        blob (ndarray): R x 5 matrix of RoIs in the image pyramid
    """
    rois_blob_real = []

    for i in range(len(im_scale_factors)):
        rois, levels = _project_im_rois(im_rois, np.array([im_scale_factors[i]]))
        rois_blob = np.hstack((levels, rois))
        rois_blob_real.append(rois_blob.astype(np.float32, copy=False))

    return rois_blob_real 

import numpy as np
from numpy import float32

def generate_anchors_pre(height, width, feat_stride, anchor_scales=(8,16,32), anchor_ratiOS=(0.5,1,2)):
  """ A wrapper function to generate anchors given different scales
    Also return the number of anchors in variable 'length'
  """
  anchors = generate_anchors(ratios=np.array(anchor_ratios), scales=np.array(anchor_scales))
  A = anchors.shape[0]
  shift_x = np.arange(0, width) * feat_stride
  shift_y = np.arange(0, height) * feat_stride
  shift_x, shift_y = np.meshgrid(shift_x, shift_y)
  shifts = np.vstack((shift_x.ravel(), shift_y.ravel(), shift_x.ravel(), shift_y.ravel())).transpose()
  K = shifts.shape[0]
  # width changes faster, so here it is H, W, C
  anchors = anchors.reshape((1, A, 4)) + shifts.reshape((1, K, 4)).transpose((1, 0, 2))
  anchors = anchors.reshape((K * A, 4)).astype(np.float32, copy=False)
  length = np.int32(anchors.shape[0])

  return anchors, length 

import numpy as np
from numpy import float32

def draw_heatmap(img, heatmap, alpha=0.5):
    """Draw a heatmap overlay over an image."""
    assert len(heatmap.shape) == 2 or \
        (len(heatmap.shape) == 3 and heatmap.shape[2] == 1)
    assert img.dtype in [np.uint8, np.int32, np.int64]
    assert heatmap.dtype in [np.float32, np.float64]

    if img.shape[0:2] != heatmap.shape[0:2]:
        heatmap_rs = np.clip(heatmap * 255, 0, 255).astype(np.uint8)
        heatmap_rs = ia.imresize_single_image(
            heatmap_rs[..., np.newaxis],
            img.shape[0:2],
            interpolation="nearest"
        )
        heatmap = np.squeeze(heatmap_rs) / 255.0

    cmap = plt.get_cmap('jet')
    heatmap_cmapped = cmap(heatmap)
    heatmap_cmapped = np.delete(heatmap_cmapped, 3, 2)
    heatmap_cmapped = heatmap_cmapped * 255
    mix = (1-alpha) * img + alpha * heatmap_cmapped
    mix = np.clip(mix, 0, 255).astype(np.uint8)
    return mix 

import numpy as np
from numpy import float32

def maybe_cast_to_float64(da):
    """Cast DataArrays to np.float64 if they are of type np.float32.

    Parameters
    ----------
    da : xr.DataArray
        Input DataArray

    Returns
    -------
    DataArray

    """
    if da.dtype == np.float32:
        logging.warning('Datapoints were stored using the np.float32 datatype.'
                        'For accurate reduction operations using bottleneck, '
                        'datapoints are being cast to the np.float64 datatype.'
                        ' For more information see: https://GitHub.com/pydata/'
                        'xarray/issues/1346')
        return da.astype(np.float64)
    else:
        return da 

import numpy as np
from numpy import float32

def in_top_k(predictions, targets, k):
    '''Returns whether the `targets` are in the top `k` `predictions`

    # Arguments
        predictions: A tensor of shape batch_size x classess and type float32.
        targets: A tensor of shape batch_size and type int32 or int64.
        k: An int, number of top elements to consider.

    # Returns
        A tensor of shape batch_size and type int. output_i is 1 if
        targets_i is within top-k values of predictions_i
    '''
    predictions_top_k = T.argsort(predictions)[:, -k:]
    result, _ = theano.map(lambda prediction, target: any(equal(prediction, target)), sequences=[predictions_top_k, targets]

import numpy as np
from numpy import float32

def ctc_path_probs(predict, Y, alpha=1e-4):
    smoothed_predict = (1 - alpha) * predict[:, Y] + alpha * np.float32(1.) / Y.shape[0]
    L = T.log(smoothed_predict)
    zeros = T.zeros_like(L[0])
    log_first = zeros

    f_skip_idxs = ctc_create_skip_idxs(Y)
    b_skip_idxs = ctc_create_skip_idxs(Y[::-1])  # there should be a shortcut to calculating this

    def step(log_f_curr, log_b_curr, f_active, log_f_prev, b_active, log_b_prev):
        f_active_next, log_f_next = ctc_update_log_p(f_skip_idxs, zeros, f_active, log_f_curr, log_f_prev)
        b_active_next, log_b_next = ctc_update_log_p(b_skip_idxs, zeros, b_active, log_b_curr, log_b_prev)
        return f_active_next, log_f_next, b_active_next, log_b_next

    [f_active, log_f_probs, b_active, log_b_probs], _ = theano.scan(
        step, sequences=[L, L[::-1, ::-1]], outputs_info=[np.int32(1), log_first, np.int32(1), log_first])

    idxs = T.arange(L.shape[1]).dimshuffle('x', 0)
    mask = (idxs < f_active.dimshuffle(0, 'x')) & (idxs < b_active.dimshuffle(0, 'x'))[::-1, ::-1]
    log_probs = log_f_probs + log_b_probs[::-1, ::-1] - L
    return log_probs, mask 

import numpy as np
from numpy import float32

def rmsprop(self, cost, params, lr=0.001, rho=0.9, eps=1e-6,consider_constant=None):
        """
        RMSProp.
        """
        lr = theano.shared(np.float32(lr).astype(floatX))

        gradients = self.get_gradients(cost, params,consider_constant)
        accumulators = [theano.shared(np.zeros_like(p.get_value()).astype(np.float32)) for p in params]

        updates = []

        for param, gradient, accumulator in zip(params, gradients, accumulators):
            new_accumulator = rho * accumulator + (1 - rho) * gradient ** 2
            updates.append((accumulator, new_accumulator))

            new_param = param - lr * gradient / T.sqrt(new_accumulator + eps)
            updates.append((param, new_param))

        return updates

import numpy as np
from numpy import float32

def adadelta(self, cost, params, rho=0.95, epsilon=1e-6,consider_constant=None):
        """
        Adadelta. Based on:
        Http://www.matthewzeiler.com/pubs/GoogleTR2012/googleTR2012.pdf
        """
        rho = theano.shared(np.float32(rho).astype(floatX))
        epsilon = theano.shared(np.float32(epsilon).astype(floatX))

        gradients = self.get_gradients(cost, params,consider_constant)
        accu_gradients = [theano.shared(np.zeros_like(param.get_value(borrow=True)).astype(floatX)) for param in params]
        accu_deltas = [theano.shared(np.zeros_like(param.get_value(borrow=True)).astype(floatX)) for param in params]

        updates = []
        for param, gradient, accu_gradient, accu_delta in zip(params, gradients, accu_gradients, accu_deltas):
            new_accu_gradient = rho * accu_gradient + (1. - rho) * gradient ** 2.
            delta_x = - T.sqrt((accu_delta + epsilon) / (new_accu_gradient + epsilon)) * gradient
            new_accu_delta = rho * accu_delta + (1. - rho) * delta_x ** 2.
            updates.append((accu_gradient, new_accu_gradient))
            updates.append((accu_delta, new_accu_delta))
            updates.append((param, param + delta_x))
        return updates

import numpy as np
from numpy import float32

def adagrad(self, cost, params, lr=1.0, epsilon=1e-6,consider_constant=None):
        """
        Adagrad. Based on http://www.ark.cs.cmu.edu/cdyer/adagrad.pdf
        """
        lr = theano.shared(np.float32(lr).astype(floatX))
        epsilon = theano.shared(np.float32(epsilon).astype(floatX))

        gradients = self.get_gradients(cost, params,consider_constant)
        gsums = [theano.shared(np.zeros_like(param.get_value(borrow=True)).astype(floatX)) for param in params]

        updates = []
        for param, gradient, gsum in zip(params, gradients, gsums):
            new_gsum = gsum + gradient ** 2.
            updates.append((gsum, new_gsum))
            updates.append((param, param - lr * gradient / (T.sqrt(gsum + epsilon))))
        return updates 

import numpy as np
from numpy import float32

def sgd(self, cost, params,constraints={}, lr=0.01):
        """
        Stochatic gradient descent.
        """
        updates = []
        
        lr = theano.shared(np.float32(lr).astype(floatX))
        gradients = self.get_gradients(cost, params)
        
        for p, g in zip(params, gradients):
            v=-lr*g;
            new_p=p+v;
            # apply constraints
            if p in constraints:
                c=constraints[p];
                new_p=c(new_p);
            updates.append((p, new_p))

        return updates

import numpy as np
from numpy import float32

def sgdmomentum(self, cost, params,constraints={}, lr=0.01,consider_constant=None, momentum=0.):
        """
        Stochatic gradient descent with momentum. Momentum has to be in [0, 1)
        """
        # Check that the momentum is a correct value
        assert 0 <= momentum < 1

        lr = theano.shared(np.float32(lr).astype(floatX))
        momentum = theano.shared(np.float32(momentum).astype(floatX))

        gradients = self.get_gradients(cost, params)
        velocities = [theano.shared(np.zeros_like(param.get_value(borrow=True)).astype(floatX)) for param in params]

        updates = []
        for param, gradient, velocity in zip(params, gradients, velocities):
            new_velocity = momentum * velocity - lr * gradient
            updates.append((velocity, new_velocity))
            new_p=param+new_velocity;
            # apply constraints
            if param in constraints:
                c=constraints[param];
                new_p=c(new_p);
            updates.append((param, new_p))
        return updates 

import numpy as np
from numpy import float32

def set_values(name, param, pretrained):
    """
    Initialize a network parameter with pretrained values.
    We check that sizes are compatible.
    """
    param_value = param.get_value()
    if pretrained.size != param_value.size:
        raise Exception(
            "Size mismatch for parameter %s. Expected %i, found %i."
            % (name, param_value.size, pretrained.size)
        )
    param.set_value(np.reshape(
        pretrained, param_value.shape
    ).astype(np.float32))