Python numpy.transpose() Examples

def similarity_label(self, words, normalization=True):
        """
        you can calculate more than one word at the same time.
        """
        if self.model==None:
            raise Exception('no model.')
        if isinstance(words, string_types):
            words=[words]
        vectors=np.transpose(self.model.wv.__getitem__(words))
        if normalization:
            unit_vector=unitvec(vectors,ax=0) # 这样写比原来那样速度提升一倍
            #unit_vector=np.zeros((len(vectors),len(words)))
            #for i in range(len(words)):
            #    unit_vector[:,i]=matutils.unitvec(vectors[:,i])
            dists=np.dot(self.Label_vec_u, unit_vector)
        else:
            dists=np.dot(self.Label_vec, vectors)
        return dists 

def superpose_array(refArray, array, check=False):
    """
    Superpose arrays by calculating the rotation matrix and the
    translations that minimize the root mean square deviation between and
    array of vectors and a reference array.

    :Parameters:
        #. refArray (numpy.ndarray): the NX3 reference array to superpose to.
        #. array (numpy.ndarray): the NX3 array to calculate the
           transformation of.
        #. check (boolean): whether to check arguments before generating
           points.

    :Returns:
        #. superposedArray (numpy.ndarray): the NX3 array to superposed array.
    """
    rotationMatrix, _,_,_ = get_superposition_transformation(refArray=refArray, array=array, check=check)
    return np.dot( rotationMatrix, np.transpose(array).\
                   reshape(1,3,-1)).transpose().reshape(-1,3) 

def collectdata(self,):
        print 'Start Collect Data...'

        train_x_path = os.path.join(self.input_dir, 'unlabeled_X.bin')

        train_xf = open(train_x_path, 'rb')
        train_x = np.fromfile(train_xf, dtype=np.uint8)
        train_x = np.reshape(train_x, (-1, 3, 96, 96))
        train_x = np.transpose(train_x, (0, 3, 2, 1))

        idx = 0
        for i in xrange(train_x.shape[0]):
            if not self.skipimg:
                transform_and_save(img_arr=train_x[i], output_filename=os.path.join(self.unlabeldir, str(idx) + '.jpg'))
            self.trainpairlist[os.path.join('images', 'unlabeled', str(idx) + '.jpg')] = 'labels/11.txt'
            idx += 1

        print 'Finished Collect Data...' 

def train_lr_rfeinman(densities_pos, densities_neg, uncerts_pos, uncerts_neg):
    """
    TODO
    :param densities_pos:
    :param densities_neg:
    :param uncerts_pos:
    :param uncerts_neg:
    :return:
    """
    values_neg = np.concatenate(
        (densities_neg.reshape((1, -1)),
         uncerts_neg.reshape((1, -1))),
        axis=0).transpose([1, 0])
    values_pos = np.concatenate(
        (densities_pos.reshape((1, -1)),
         uncerts_pos.reshape((1, -1))),
        axis=0).transpose([1, 0])

    values = np.concatenate((values_neg, values_pos))
    labels = np.concatenate(
        (np.zeros_like(densities_neg), np.ones_like(densities_pos)))

    lr = LogisticRegressionCV(n_jobs=-1).fit(values, labels)

    return values, labels, lr 

def intersection(boxes1, boxes2):
  """Compute pairwise intersection areas between boxes.

  Args:
    boxes1: a numpy array with shape [N, 4] holding N boxes
    boxes2: a numpy array with shape [M, 4] holding M boxes

  Returns:
    a numpy array with shape [N*M] representing pairwise intersection area
  """
  [y_min1, x_min1, y_max1, x_max1] = np.split(boxes1, 4, axis=1)
  [y_min2, x_min2, y_max2, x_max2] = np.split(boxes2, 4, axis=1)

  all_pairs_min_ymax = np.minimum(y_max1, np.transpose(y_max2))
  all_pairs_max_ymin = np.maximum(y_min1, np.transpose(y_min2))
  intersect_heights = np.maximum(
      np.zeros(all_pairs_max_ymin.shape),
      all_pairs_min_ymax - all_pairs_max_ymin)
  all_pairs_min_xmax = np.minimum(x_max1, np.transpose(x_max2))
  all_pairs_max_xmin = np.maximum(x_min1, np.transpose(x_min2))
  intersect_widths = np.maximum(
      np.zeros(all_pairs_max_xmin.shape),
      all_pairs_min_xmax - all_pairs_max_xmin)
  return intersect_heights * intersect_widths 

def resample(imgs, spacing, new_spacing,order=2):
    if len(imgs.shape)==3:
        new_shape = np.round(imgs.shape * spacing / new_spacing)
        true_spacing = spacing * imgs.shape / new_shape
        resize_factor = new_shape / imgs.shape
        imgs = zoom(imgs, resize_factor, mode = 'nearest',order=order)
        return imgs, true_spacing
    elif len(imgs.shape)==4:
        n = imgs.shape[-1]
        newimg = []
        for i in range(n):
            slice = imgs[:,:,:,i]
            newslice,true_spacing = resample(slice,spacing,new_spacing)
            newimg.append(newslice)
        newimg=np.transpose(np.array(newimg),[1,2,3,0])
        return newimg,true_spacing
    else:
        raise ValueError('wrong shape') 

def pre_processing(obs, cuda):
    mean = np.array([0.485, 0.456, 0.406]).reshape([1, 1, 3])
    std = np.array([0.229, 0.224, 0.225]).reshape([1, 1, 3])
    obs = obs / 255
    obs = (obs - mean) / std
    obs = np.transpose(obs, (2, 0, 1))
    obs = np.expand_dims(obs, 0)
    obs = np.array(obs)
    if cuda:
        torch_device = torch.device('cuda:0')
    else:
        torch_device = torch.device('cpu')
    obs_tensor = torch.tensor(obs, dtype=torch.float32, device=torch_device, requires_grad=True)
    return obs_tensor

# generate the entire images 

def train(self, inputs_list, targets_list):
        inputs = np.array(inputs_list, ndmin=2).T
        targets = np.array(targets_list, ndmin=2).T

        hidden_inputs = np.dot(self.wih, inputs)
        hidden_outputs = self.activation_function(hidden_inputs)

        final_inputs = np.dot(self.who, hidden_outputs)
        final_outputs = self.activation_function(final_inputs)

        output_errors = targets - final_outputs
        hidden_errors = np.dot(self.who.T, output_errors)

        self.who += self.lr * np.dot((output_errors *
                                      final_outputs *
                                      (1.0 - final_outputs)), np.transpose(hidden_outputs))
        self.wih += self.lr * np.dot((hidden_errors *
                                      hidden_outputs *
                                      (1.0 - hidden_outputs)), np.transpose(inputs))
        
        pass

    # query 

def intersection(boxes1, boxes2):
  """Compute pairwise intersection areas between boxes.

  Args:
    boxes1: a numpy array with shape [N, 4] holding N boxes
    boxes2: a numpy array with shape [M, 4] holding M boxes

  Returns:
    a numpy array with shape [N*M] representing pairwise intersection area
  """
  [y_min1, x_min1, y_max1, x_max1] = np.split(boxes1, 4, axis=1)
  [y_min2, x_min2, y_max2, x_max2] = np.split(boxes2, 4, axis=1)

  all_pairs_min_ymax = np.minimum(y_max1, np.transpose(y_max2))
  all_pairs_max_ymin = np.maximum(y_min1, np.transpose(y_min2))
  intersect_heights = np.maximum(
      np.zeros(all_pairs_max_ymin.shape),
      all_pairs_min_ymax - all_pairs_max_ymin)
  all_pairs_min_xmax = np.minimum(x_max1, np.transpose(x_max2))
  all_pairs_max_xmin = np.maximum(x_min1, np.transpose(x_min2))
  intersect_widths = np.maximum(
      np.zeros(all_pairs_max_xmin.shape),
      all_pairs_min_xmax - all_pairs_max_xmin)
  return intersect_heights * intersect_widths 

def decode_topk(self, sess, latest_tokens, enc_top_states, dec_init_states):
    """Return the topK results and new decoder states."""
    feed = {
        self._enc_top_states: enc_top_states,
        self._dec_in_state:
            np.squeeze(np.array(dec_init_states)),
        self._abstracts:
            np.transpose(np.array([latest_tokens])),
        self._abstract_lens: np.ones([len(dec_init_states)], np.int32)}

    results = sess.run(
        [self._topk_ids, self._topk_log_probs, self._dec_out_state],
        feed_dict=feed)

    ids, probs, states = results[0], results[1], results[2]
    new_states = [s for s in states]
    return ids, probs, new_states 

def _write_map_files(b_in, b_out, transform):
  cats = get_categories()

  env = utils.Foo(padding=10, resolution=5, num_point_threshold=2,
                  valid_min=-10, valid_max=200, n_samples_per_face=200)
  robot = utils.Foo(radius=15, base=10, height=140, sensor_height=120,
                    camera_elevation_degree=-15)
  
  building_loader = factory.get_dataset('sbpd')
  for flip in [False, True]:
    b = nav_env.Building(b_out, robot, env, flip=flip,
                         building_loader=building_loader)
    logging.info("building_in: %s, building_out: %s, transform: %d", b_in,
                 b_out, transform)
    maps = _get_semantic_maps(b_in, transform, b.map, flip, cats)
    maps = np.transpose(np.array(maps), axes=[1,2,0])

    #  Load file from the cache.
    file_name = '{:s}_{:d}_{:d}_{:d}_{:d}_{:d}_{:d}.pkl'
    file_name = file_name.format(b.building_name, b.map.size[0], b.map.size[1],
                                 b.map.origin[0], b.map.origin[1],
                                 b.map.resolution, flip)
    out_file = os.path.join(DATA_DIR, 'processing', 'class-maps', file_name)
    logging.info('Writing semantic maps to %s.', out_file)
    save_variables(out_file, [maps, cats], ['maps', 'cats'], overwrite=True) 

def measure_cost(repeat, scipy_trans_lhs, scipy_dns_lhs, func_name, *args, **kwargs):
    """Measure time cost of running a function
    """
    mx.nd.waitall()
    args_list = []
    for arg in args:
        args_list.append(arg)
    start = time.time()
    if scipy_trans_lhs:
        args_list[0] = np.transpose(args_list[0]) if scipy_dns_lhs else sp.spmatrix.transpose(args_list[0])
    for _ in range(repeat):
        func_name(*args_list, **kwargs)
    mx.nd.waitall()
    end = time.time()
    diff = end - start
    return diff / repeat 

def jacobian(self, p, into=None):
        # transpose to be 3 x 2 x n
        p = np.transpose(np.reshape(p, (-1, 3, 2)), (1,2,0))
        # First, get the two legs...
        (dx_ab, dy_ab) = p[1] - p[0]
        (dx_ac, dy_ac) = p[2] - p[0]
        (dx_bc, dy_bc) = p[2] - p[1]
        # now, the area is half the z-value of the cross-product...
        sarea0 = 0.5 * (dx_ab*dy_ac - dx_ac*dy_ab)
        # but we want to abs it
        dsarea0 = np.sign(sarea0)
        z = np.transpose([[-dy_bc,dx_bc], [dy_ac,-dx_ac], [-dy_ab,dx_ab]], (2,0,1))
        z = times(0.5*dsarea0, z)
        m = numel(p)
        n = p.shape[2]
        ii = (np.arange(n) * np.ones([6, n])).T.flatten()
        z = sps.csr_matrix((z.flatten(), (ii, np.arange(len(ii)))), shape=(n, m))
        return safe_into(into, z) 

def from_logeccen(logecc, vmin=0, vmax=90, offset=0.75):
    '''
    from_logeccen(logecc) yields a rescaled linear-space version of the log-eccentricity value (or
      values) logecc.
    from_logeccen(logxy_matrix) rescales all the (x,y) points in the given matrix to have
      linearly-spaced eccentricity values.

    from_logeccen is the inverse of to_logeccen.
    '''
    if pimms.is_matrix(logecc):
        xy = np.asarray(logecc)
        trq = xy.shape[0] != 2
        xy = np.transpose(xy) if trq else np.asarray(xy)
        r = np.sqrt(np.sum(xy**2, axis=0))
        esc = from_logeccen(r, vmin=vmin, vmax=vmax, offset=offset)
        ecc = zinv(r)
        xy = xy * [ecc,ecc] * [esc,esc]
        return xy.T if trq else xy
    else:
        logecc = np.asarray(logecc)
        (vmin,vmax,offset) = [np.asarray(u) for u in (vmin,vmax,offset)]
        (vmin, vmax) = [np.log(u + offset) for u in (vmin, vmax)]
        logecc = logecc*(vmax - vmin) + vmin
        return np.exp(logecc) - offset 

def load_batch_norm(idx, variables, weights, assign_ops, offset):
  """Loads kernel, gamma, beta, mean, variance for Batch Normalization"""
  kernel = variables[idx]
  gamma, beta, mean, variance = variables[idx + 1:idx + 5]
  batch_norm_vars = [beta, gamma, mean, variance]

  for var in batch_norm_vars:
    shape = var.shape.as_list()
    num_params = np.prod(shape)
    var_weights = weights[offset:offset + num_params].reshape(shape)
    offset += num_params
    assign_ops.append(tf.assign(var, var_weights))

  shape = kernel.shape.as_list()
  num_params = np.prod(shape)
  var_weights = weights[offset:offset + num_params].reshape((shape[3], shape[2], shape[0], shape[1]))
  var_weights = np.transpose(var_weights, (2, 3, 1, 0))
  offset += num_params
  assign_ops.append(tf.assign(kernel, var_weights))
  return assign_ops, offset 

def __call__(self, x, y=None):
        if y is not None: x = (x,y)
        x = np.asarray(x)
        if len(x.shape) == 1: return self([x])[0]
        x = np.transpose(x) if x.shape[0] == 2 else x
        if not x.flags['WRITEABLE']: x = np.array(x)
        crd = self.coordinates
        sig = self.sigma
        wts = self._weight
        res = np.zeros(x.shape[0])
        for (sh, qd, bi) in zip(self.spatial_hashes, self.bin_query_distances, self.sigma_bins):
            neis = sh.query_ball_point(x, qd)
            res += [
                np.sum(w * np.exp(-0.5 * d2/s**2))
                for (ni,pt) in zip(neis,x)
                for ii in [bi[ni]]
                for (w,s,d2) in [(wts[ii], sig[ii], np.sum((crd[ii] - pt)**2, axis=1))]]
        return res 

def visualize_sampling(self, permutations):
        max_length = len(permutations[0])
        grid = np.zeros([max_length,max_length]) # initialize heatmap grid to 0

        transposed_permutations = np.transpose(permutations)
        for t, cities_t in enumerate(transposed_permutations): # step t, cities chosen at step t
            city_indices, counts = np.unique(cities_t,return_counts=True,axis=0)
            for u,v in zip(city_indices, counts):
                grid[t][u]+=v # update grid with counts from the batch of permutations

        # plot heatmap
        fig = plt.figure()
        rcParams.update({'font.size': 22})
        ax = fig.add_subplot(1,1,1)
        ax.set_aspect('equal')
        plt.imshow(grid, interpolation='nearest', cmap='gray')
        plt.colorbar()
        plt.title('Sampled permutations')
        plt.ylabel('Time t')
        plt.xlabel('City i')
        plt.show() 

def test_spacetodepth():
    n, c, h, w = shape = (1, 1, 4, 6)
    input1 = np.random.rand(n, c, h, w).astype("float32")
    blocksize = 2
    inputs = [helper.make_tensor_value_info("input1", TensorProto.FLOAT, shape=shape)]

    outputs = [helper.make_tensor_value_info("output", TensorProto.FLOAT, shape=(1, 4, 2, 3))]

    nodes = [helper.make_node("SpaceToDepth", ["input1"], ["output"], block_size=blocksize)]

    graph = helper.make_graph(nodes,
                              "spacetodepth_test",
                              inputs,
                              outputs)

    spacetodepth_model = helper.make_model(graph)

    bkd_rep = backend.prepare(spacetodepth_model)
    output = bkd_rep.run([input1])

    tmp = np.reshape(input1, [n, c,
                    h // blocksize, blocksize,
                    w // blocksize, blocksize])
    tmp = np.transpose(tmp, [0, 3, 5, 1, 2, 4])
    numpy_op = np.reshape(tmp, [n, c * (blocksize**2),
                    h // blocksize,
                    w // blocksize])

    npt.assert_almost_equal(output[0], numpy_op) 

def nparray_and_transpose(data_a_b_c):
  """Convert the list of items in data to a numpy array, and transpose it
  Args:
    data: data_asbsc: a nested, nested list of length a, with sublist length
      b, with sublist length c.
  Returns:
    a numpy 3-tensor with dimensions a x c x b
"""
  data_axbxc = np.array([datum_b_c for datum_b_c in data_a_b_c])
  data_axcxb = np.transpose(data_axbxc, axes=[0,2,1])
  return data_axcxb 

def nb_greedy_FPS(xyz,K):
    start_element = 0
    sample_num = xyz.shape[0]
    sum_vec = np.zeros((sample_num,1),dtype = np.float32)
    xyz_sq = xyz**2
    for j in range(sample_num):
        sum_vec[j,0] = np.sum(xyz_sq[j,:])
    pairwise_distance = sum_vec + np.transpose(sum_vec) - 2*np.dot(xyz, np.transpose(xyz))
    
    candidates_ind = np.zeros((sample_num,),dtype = np.bool_)
    candidates_ind[start_element] = True
    remain_ind = np.ones((sample_num,),dtype = np.bool_)
    remain_ind[start_element] = False
    all_ind = np.arange(sample_num)
    
    for i in range(1,K):
        if i == 1:
            min_remain_pt_dis = pairwise_distance[:,start_element]
            min_remain_pt_dis = min_remain_pt_dis[remain_ind]
        else:
            cur_dis = pairwise_distance[remain_ind,:]
            cur_dis = cur_dis[:,candidates_ind]
            min_remain_pt_dis = np.zeros((cur_dis.shape[0],),dtype = np.float32)
            for j in range(cur_dis.shape[0]):
                min_remain_pt_dis[j] = np.min(cur_dis[j,:])
        next_ind_in_remain = np.argmax(min_remain_pt_dis)
        next_ind = all_ind[remain_ind][next_ind_in_remain]
        candidates_ind[next_ind] = True
        remain_ind[next_ind] = False
        
    return candidates_ind 

def GenerateSample(filename, code_shape, layer_depth):
  # {0, +1} binary codes.
  # No conversion since the output file is expected to store
  # codes using {0, +1} codes (and not {-1, +1}).
  code = synthetic_model.GenerateSingleCode(code_shape)
  code = np.round(code)

  # Reformat the code so as to be compatible with what is generated
  # by the image encoder.
  # The image encoder generates a tensor of size:
  # iteration_count x batch_size x height x width x iteration_depth.
  # Here: batch_size = 1
  if code_shape[-1] % layer_depth != 0:
    raise ValueError('Number of layers is not an integer')
  height = code_shape[0]
  width = code_shape[1]
  code = code.reshape([1, height, width, -1, layer_depth])
  code = np.transpose(code, [3, 0, 1, 2, 4])

  int_codes = code.astype(np.int8)
  exported_codes = np.packbits(int_codes.reshape(-1))

  output = io.BytesIO()
  np.savez_compressed(output, shape=int_codes.shape, codes=exported_codes)
  with tf.gfile.FastGFile(filename, 'wb') as code_file:
    code_file.write(output.getvalue()) 

def spectrogram_from_file(filename, step=10, window=20, max_freq=None,
                          eps=1e-14, overwrite=False, save_feature_as_csvfile=False):
    """ Calculate the log of linear spectrogram from FFT energy
    Params:
        filename (str): Path to the audio file
        step (int): Step size in milliseconds between windows
        window (int): FFT window size in milliseconds
        max_freq (int): Only FFT bins corresponding to frequencies between
            [0, max_freq] are returned
        eps (float): Small value to ensure numerical stability (for ln(x))
    """

    csvfilename = filename.replace(".wav", ".csv")
    if (os.path.isfile(csvfilename) is False) or overwrite:
        with soundfile.SoundFile(filename) as sound_file:
            audio = sound_file.read(dtype='float32')
            sample_rate = sound_file.samplerate
            if audio.ndim >= 2:
                audio = np.mean(audio, 1)
            if max_freq is None:
                max_freq = sample_rate / 2
            if max_freq > sample_rate / 2:
                raise ValueError("max_freq must not be greater than half of "
                                 " sample rate")
            if step > window:
                raise ValueError("step size must not be greater than window size")
            hop_length = int(0.001 * step * sample_rate)
            fft_length = int(0.001 * window * sample_rate)

            pxx, freqs = spectrogram(
                audio, fft_length=fft_length, sample_rate=sample_rate,
                hop_length=hop_length)

            ind = np.where(freqs <= max_freq)[0][-1] + 1
            res = np.transpose(np.log(pxx[:ind, :] + eps))
            if save_feature_as_csvfile:
                np.savetxt(csvfilename, res)
            return res
    else:
        return np.loadtxt(csvfilename) 

def make_data_iter_plan(self):
        "make a random data iteration plan"
        # truncate each bucket into multiple of batch-size
        bucket_n_batches = []
        for i in range(len(self.data)):
            bucket_n_batches.append(np.floor((self.data[i]) / self.batch_size))
            self.data[i] = self.data[i][:int(bucket_n_batches[i]*self.batch_size)]

        bucket_plan = np.hstack([np.zeros(n, int)+i for i, n in enumerate(bucket_n_batches)])
        np.random.shuffle(bucket_plan)

        bucket_idx_all = [np.random.permutation(len(x)) for x in self.data]

        self.bucket_plan = bucket_plan
        self.bucket_idx_all = bucket_idx_all
        self.bucket_curr_idx = [0 for x in self.data]

        self.data_buffer = []
        self.label_buffer = []
        for i_bucket in range(len(self.data)):
            if not self.model_parallel:
                data = np.zeros((self.batch_size, self.buckets[i_bucket]))
                label = np.zeros((self.batch_size, self.buckets[i_bucket]))
                self.data_buffer.append(data)
                self.label_buffer.append(label)
            else:
                data = np.zeros((self.buckets[i_bucket], self.batch_size))
                self.data_buffer.append(data)

        if self.model_parallel:
            # Transpose data if model parallel
            for i in range(len(self.data)):
                bucket_data = self.data[i]
                self.data[i] = np.transpose(bucket_data) 

def cumsum(x, axis=0, exclusive=False):
  """TPU hack for tf.cumsum.

  This is equivalent to tf.cumsum and is faster on TPU as of 04/2018 unless
  the axis dimension is very large.

  Args:
    x: a Tensor
    axis: an integer
    exclusive: a boolean

  Returns:
    Tensor of the same shape as x.
  """
  if not is_on_tpu():
    return tf.cumsum(x, axis=axis, exclusive=exclusive)
  x_shape = shape_list(x)
  rank = len(x_shape)
  length = x_shape[axis]
  my_range = tf.range(length)
  comparator = tf.less if exclusive else tf.less_equal
  mask = tf.cast(
      comparator(tf.expand_dims(my_range, 1), tf.expand_dims(my_range, 0)),
      x.dtype)
  ret = tf.tensordot(x, mask, axes=[[axis], [0]])
  if axis != rank - 1:
    ret = tf.transpose(
        ret,
        list(range(axis)) + [rank - 1] + list(range(axis, rank - 1)))
  return ret 

def _codify__(self, engine, name='constraint', addDependencies=True):
        assert isinstance(name, basestring), LOGGER.error("name must be a string")
        assert re.match('[a-zA-Z_][a-zA-Z0-9_]*$', name) is not None, LOGGER.error("given name '%s' can't be used as a variable name"%name)
        klass        = self.__class__.__name__
        dependencies = ['import numpy as np','from fullrmc.Constraints import {klass}s'.format(klass=klass)]
        code         = []
        if addDependencies:
            code.extend(dependencies)
        x = list(self.experimentalData[:,0])
        y = list(self.experimentalData[:,1])
        code.append("x = {x}".format(x=x))
        code.append("y = {y}".format(y=y))
        code.append("d = np.transpose([x,y]).astype(np.float32)")
        dw = self.dataWeights
        if dw is not None:
            dw = list(dw)
        code.append("dw = {dw}".format(dw=dw))
        sfp = self._shapeFuncParams
        if isinstance(sfp, np.ndarray):
            sfp = list(sfp)
        code.append("sfp = {sfp}".format(sfp=sfp))
        wf = self.windowFunction
        if isinstance(wf, np.ndarray):
            code.append("wf = np.array({wf})".format(wf=list(wf)))
        else:
            code.append("wf = {wf}".format(wf=wf))
        code.append("{name} = {klass}s.{klass}\
(experimentalData=d, dataWeights=dw, weighting='{weighting}', atomsWeight={atomsWeight}, \
scaleFactor={scaleFactor}, adjustScaleFactor={adjustScaleFactor}, \
shapeFuncParams=sfp, windowFunction=wf, limits={limits})".format(name=name, klass=klass,
                weighting=self.weighting, atomsWeight=self.atomsWeight,
                scaleFactor=self.scaleFactor, adjustScaleFactor=self.adjustScaleFactor,
                shapeFuncParams=sfp, limits=self.limits))
        code.append("{engine}.add_constraints([{name}])".format(engine=engine, name=name))
        # return
        return dependencies, '\n'.join(code) 

def download_svhn(download_path):
    data_dir = osp.join(download_path, 'svhn')

    import scipy.io as sio

    # svhn file loader
    def svhn_loader(url, path):
        cmd = ['curl', url, '-o', path]
        subprocess.call(cmd)
        m = sio.loadmat(path)
        return m['X'], m['y']

    if check_file(data_dir):
        print('SVHN was downloaded.')
        return

    data_url = 'http://ufldl.stanford.edu/housenumbers/train_32x32.mat'
    train_image, train_label = svhn_loader(data_url, osp.join(data_dir, 'train_32x32.mat'))

    data_url = 'http://ufldl.stanford.edu/housenumbers/test_32x32.mat'
    test_image, test_label = svhn_loader(data_url, osp.join(data_dir, 'test_32x32.mat'))

    prepare_h5py(np.transpose(train_image, (3, 0, 1, 2)),
                 np.transpose(test_image, (3, 0, 1, 2)), data_dir)

    cmd = ['rm', '-f', osp.join(data_dir, '*.mat')]
    subprocess.call(cmd) 

def __init__(self, engine, weighting="atomicNumber",
                       qmin=0.001, qmax=1, dq=0.005,
                       rmin=0.00, rmax=100, dr=1):
        # get qmin
        assert is_number(qmin), LOGGER.error("qmin must be a number")
        qmin = FLOAT_TYPE(qmin)
        assert qmin>0, LOGGER.error("qmin '%s' must be bigger than 0"%qmin)
        # get qmin
        assert is_number(qmax), LOGGER.error("qmax must be a number")
        qmax = FLOAT_TYPE(qmax)
        assert qmax>qmin, LOGGER.error("qmax '%s' must be bigger than qmin '%s'"%(qmin,qmax))
        # get dq
        assert is_number(dq), LOGGER.error("dq must be a number")
        dq = FLOAT_TYPE(dq)
        assert dq>0, LOGGER.error("dq '%s' must be bigger than 0"%dq)
        # import StructureFactorConstraint
        from fullrmc.Constraints.StructureFactorConstraints import StructureFactorConstraint
        # overload constraint
        class _ShapeFunctionStructureFactor(StructureFactorConstraint):
            """This overloading is needed to avoid the constraint being
            saved upon setting the different properties"""
            def _get_repository(self, *args, **kwargs):
                return None
            def _dump_to_repository(self, *args, **kwargs):
                return
        # create StructureFactorConstraint
        Q = np.arange(qmin, qmax, dq)
        D = np.transpose([Q, np.zeros(len(Q))]).astype(FLOAT_TYPE)
        #self._SFC = StructureFactorConstraint(rmin=rmin, rmax=rmax, dr=dr, experimentalData=D, weighting="atomicNumber")
        self._SFC = _ShapeFunctionStructureFactor(rmin=rmin, rmax=rmax, dr=dr, experimentalData=D, weighting="atomicNumber")
        self._SFC._set_engine(engine)
        self._SFC.listen(message="engine set")
        # set parameters
        self._rmin = FLOAT_TYPE(rmin)
        self._rmax = FLOAT_TYPE(rmax)
        self._dr   = FLOAT_TYPE(dr)
        self._qmin = FLOAT_TYPE(qmin)
        self._qmax = FLOAT_TYPE(qmax)
        self._dq   = FLOAT_TYPE(dq)
        self._weighting = weighting 

def diagonal_conv_gru(x,
                      kernel_size,
                      filters,
                      dropout=0.0,
                      name=None,
                      reuse=None):
  """Diagonal Convolutional GRU as in https://arxiv.org/abs/1702.08727."""

  # Let's make a shorthand for conv call first.
  def do_conv(args, name, bias_start):
    return conv(
        args,
        filters,
        kernel_size,
        padding="SAME",
        bias_initializer=tf.constant_initializer(bias_start),
        name=name)

  # Here comes the GRU gate.
  with tf.variable_scope(
      name, default_name="diagonal_conv_gru", values=[x], reuse=reuse):
    reset, reset_cost = hard_sigmoid(do_conv(x, "reset", 0.5))
    gate, gate_cost = hard_sigmoid(do_conv(x, "gate", 0.7))
    candidate = tf.tanh(do_conv(reset * x, "candidate", 0.0))

    if dropout > 0.0:
      candidate = tf.nn.dropout(candidate, 1.0 - dropout)

    # Diagonal shift.
    shift_filters = filters // 3
    base_filter = ([[0, 1, 0]] * (filters - 2 * shift_filters) +
                   [[1, 0, 0]] * shift_filters + [[0, 0, 1]] * shift_filters)
    shift_filter = tf.constant(np.transpose(base_filter), dtype=tf.float32)
    shift_filter = tf.expand_dims(tf.expand_dims(shift_filter, 0), 3)
    x_shifted = tf.nn.depthwise_conv2d(
        x, shift_filter, [1, 1, 1, 1], padding="SAME")

    # Return the gated result and cost.
    total_cost_avg = 0.5 * (reset_cost + gate_cost)
    return gate * x_shifted + (1 - gate) * candidate, total_cost_avg 

def running_global_pool_1d(inputs, pooling_type="MAX"):
  """Same global pool, but only for the elements up to the current element.

  Useful for outputs where the state of future elements is not known.
  Takes no mask as all elements up to the current element are assumed to exist.
  Currently only supports maximum. Equivalent to using a lower triangle bias.

  Args:
    inputs: A tensor of shape [batch_size, sequence_length, input_dims]
      containing the sequences of input vectors.
    pooling_type: Pooling type to use. Currently only supports 'MAX'.

  Returns:
    A tensor of shape [batch_size, sequence_length, input_dims] containing the
    running 'totals'.
  """
  del pooling_type
  with tf.name_scope("running_global_pool", values=[inputs]):
    scan_fct = tf.maximum
    # Permute inputs so seq_length is first.
    elems = tf.transpose(inputs, [1, 0, 2])
    # Perform scan.
    cumulatives = tf.scan(scan_fct, elems, swap_memory=True)
    # Permute output to get back to original order.
    output = tf.transpose(cumulatives, [1, 0, 2])
  return output 

def decoder(model, z):
        params = model.arg_params
        decoder_n = np.shape(params['decoder_z_bias'].asnumpy())[0]
        decoder_z = np.dot(params['decoder_z_weight'].asnumpy(),np.transpose(z)) \
                    + np.reshape(params['decoder_z_bias'].asnumpy(),(decoder_n,1))
        act_z = np.tanh(decoder_z)
        decoder_x = np.transpose(np.dot(params['decoder_x_weight'].asnumpy(),act_z)) + params['decoder_x_bias'].asnumpy()
        reconstructed_x = 1/(1+np.exp(-decoder_x))
        return reconstructed_x