Python numpy 模块，ceil() 实例源码

def make_grid(I, ncols=8):
    assert isinstance(I, np.ndarray), 'plugin error, should pass numpy array here'
    assert I.ndim == 4 and I.shape[1] == 3
    nimg = I.shape[0]
    H = I.shape[2]
    W = I.shape[3]
    ncols = min(nimg, ncols)
    nrows = int(np.ceil(float(nimg) / ncols))
    canvas = np.zeros((3, H * nrows, W * ncols))
    i = 0
    for y in range(nrows):
        for x in range(ncols):
            if i >= nimg:
                break
            canvas[:, y * H:(y + 1) * H, x * W:(x + 1) * W] = I[i]
            i = i + 1
    return canvas

def saveHintonPlot(self, matrix, num_tests, max_weight=None, ax=None):
        """Draw Hinton diagram for visualizing a weight matrix."""
        fig,ax = plt.subplots(1,1)

        if not max_weight:
            max_weight = 2**np.ceil(np.log(np.abs(matrix).max())/np.log(2))

        ax.patch.set_facecolor('gray')
        ax.set_aspect('equal', 'box')
        ax.xaxis.set_major_locator(plt.NullLocator())
        ax.yaxis.set_major_locator(plt.NullLocator())

        for (x, y), w in np.ndenumerate(matrix):
            color = 'white' if w > 0 else 'black'
            size = np.sqrt(np.abs(0.5*w/num_tests)) # Need to scale so that it is between 0 and 0.5
            rect = plt.Rectangle([x - size / 2, y - size / 2], size, size,
                                 facecolor=color, edgecolor=color)
            ax.add_patch(rect)

        ax.autoscale_view()
        ax.invert_yaxis()
        plt.savefig(self.figures_path + self.save_prefix + '-Hinton.eps')
        plt.close()

def fftfilt(b, x, *n):
    N_x = len(x)
    N_b = len(b)
    N = 2**np.arange(np.ceil(np.log2(N_b)),np.floor(np.log2(N_x)))
    cost = np.ceil(N_x / (N - N_b + 1)) * N * (np.log2(N) + 1)
    N_fft = int(N[np.argmin(cost)])
    N_fft = int(N_fft)    
    # Compute the block length:
    L = int(N_fft - N_b + 1)
    # Compute the transform of the filter:
    H = np.fft.fft(b,N_fft)
    y = np.zeros(N_x, x.dtype)
    i = 0
    while i <= N_x:
        il = np.min([i+L,N_x])
        k = np.min([i+N_fft,N_x])
        yt = np.fft.ifft(np.fft.fft(x[i:il],N_fft)*H,N_fft) # Overlap..
        y[i:k] = y[i:k] + yt[:k-i]                          # and add
        i += L
    return y

def laplace_gpu(y_gpu, mode='valid'):

  shape = np.array(y_gpu.shape).astype(np.uint32)
  dtype = y_gpu.dtype
  block_size = (16,16,1)
  grid_size = (int(np.ceil(float(shape[1])/block_size[0])),
               int(np.ceil(float(shape[0])/block_size[1])))
  shared_size = int((2+block_size[0])*(2+block_size[1])*dtype.itemsize)

  preproc = _generate_preproc(dtype, shape)
  mod = SourceModule(preproc + kernel_code, keep=True)

  if mode == 'valid':
    laplace_fun_gpu = mod.get_function("laplace_valid")
    laplace_gpu = cua.empty((y_gpu.shape[0]-2, y_gpu.shape[1]-2), y_gpu.dtype)

  if mode == 'same':
    laplace_fun_gpu = mod.get_function("laplace_same")
    laplace_gpu = cua.empty((y_gpu.shape[0], y_gpu.shape[1]), y_gpu.dtype)

  laplace_fun_gpu(laplace_gpu.gpudata, y_gpu.gpudata,
                  block=block_size, grid=grid_size, shared=shared_size)

  return laplace_gpu

def split(args):
    if args.skip or args.is_multi_genome:
        return {'chunks': [{'__mem_gb': cr_constants.MIN_MEM_GB}]}

    chunks = []
    min_clusters = cr_constants.MIN_N_CLUSTERS
    max_clusters = args.max_clusters if args.max_clusters is not None else cr_constants.MAX_N_CLUSTERS_DEFAULT
    matrix_mem_gb = np.ceil(MEM_FACTOR * cr_matrix.GeneBCMatrix.get_mem_gb_from_matrix_h5(args.matrix_h5))
    for n_clusters in xrange(min_clusters, max_clusters + 1):
        chunk_mem_gb = max(matrix_mem_gb, cr_constants.MIN_MEM_GB)
        chunks.append({
            'n_clusters': n_clusters,
            '__mem_gb': chunk_mem_gb,
        })

    return {'chunks': chunks}

def expand_to_chunk_size(self, chunk_size, offset=Vec(0,0,0, dtype=int)):
    """
    Align a potentially non-axis aligned bbox to the grid by growing it
    to the nearest grid lines.

    Required:
      chunk_size: arraylike (x,y,z), the size of chunks in the 
                    dataset e.g. (64,64,64)
    Optional:
      offset: arraylike (x,y,z), the starting coordinate of the dataset
    """
    chunk_size = np.array(chunk_size, dtype=np.float32)
    result = self.clone()
    result = result - offset
    result.minpt = np.floor(result.minpt / chunk_size) * chunk_size
    result.maxpt = np.ceil(result.maxpt / chunk_size) * chunk_size 
    return result + offset

def shrink_to_chunk_size(self, chunk_size, offset=Vec(0,0,0, dtype=int)):
    """
    Align a potentially non-axis aligned bbox to the grid by shrinking it
    to the nearest grid lines.

    Required:
      chunk_size: arraylike (x,y,z), the size of chunks in the 
                    dataset e.g. (64,64,64)
    Optional:
      offset: arraylike (x,y,z), the starting coordinate of the dataset
    """
    chunk_size = np.array(chunk_size, dtype=np.float32)
    result = self.clone()
    result = result - offset
    result.minpt = np.ceil(result.minpt / chunk_size) * chunk_size
    result.maxpt = np.floor(result.maxpt / chunk_size) * chunk_size 
    return result + offset

def _draw_single_box(image, xmin, ymin, xmax, ymax, display_str, font, color='black', thickness=4):
  draw = ImageDraw.Draw(image)
  (left, right, top, bottom) = (xmin, xmax, ymin, ymax)
  draw.line([(left, top), (left, bottom), (right, bottom),
             (right, top), (left, top)], width=thickness, fill=color)
  text_bottom = bottom
  # Reverse list and print from bottom to top.
  text_width, text_height = font.getsize(display_str)
  margin = np.ceil(0.05 * text_height)
  draw.rectangle(
      [(left, text_bottom - text_height - 2 * margin), (left + text_width,
                                                        text_bottom)],
      fill=color)
  draw.text(
      (left + margin, text_bottom - text_height - margin),
      display_str,
      fill='black',
      font=font)

  return image

def resize_image(image,target_shape, pad_value = 0):
    assert isinstance(target_shape, list) or isinstance(target_shape, tuple)
    add_shape, subs_shape = [], []

    image_shape = image.shape
    shape_difference = np.asarray(target_shape, dtype=int) - np.asarray(image_shape,dtype=int)
    for diff in shape_difference:
        if diff < 0:
            subs_shape.append(np.s_[int(np.abs(np.ceil(diff/2))):int(np.floor(diff/2))])
            add_shape.append((0, 0))
        else:
            subs_shape.append(np.s_[:])
            add_shape.append((int(np.ceil(1.0*diff/2)),int(np.floor(1.0*diff/2))))
    output = np.pad(image, tuple(add_shape), 'constant', constant_values=(pad_value, pad_value))
    output = output[subs_shape]
    return output

def logTickValues(self, minVal, maxVal, size, stdTicks):

        ## start with the tick spacing given by tickValues().
        ## Any level whose spacing is < 1 needs to be converted to log scale

        ticks = []
        for (spacing, t) in stdTicks:
            if spacing >= 1.0:
                ticks.append((spacing, t))

        if len(ticks) < 3:
            v1 = int(np.floor(minVal))
            v2 = int(np.ceil(maxVal))
            #major = list(range(v1+1, v2))

            minor = []
            for v in range(v1, v2):
                minor.extend(v + np.log10(np.arange(1, 10)))
            minor = [x for x in minor if x>minVal and x<maxVal]
            ticks.append((None, minor))
        return ticks

def renderSymbol(symbol, size, pen, brush, device=None):
    """
    Render a symbol specification to QImage.
    Symbol may be either a QPainterPath or one of the keys in the Symbols dict.
    If *device* is None, a new QPixmap will be returned. Otherwise,
    the symbol will be rendered into the device specified (See QPainter documentation
    for more information).
    """
    ## Render a spot with the given parameters to a pixmap
    penPxWidth = max(np.ceil(pen.widthF()), 1)
    if device is None:
        device = QtGui.QImage(int(size+penPxWidth), int(size+penPxWidth), QtGui.QImage.Format_ARGB32)
        device.fill(0)
    p = QtGui.QPainter(device)
    try:
        p.setRenderHint(p.Antialiasing)
        p.translate(device.width()*0.5, device.height()*0.5)
        drawSymbol(p, symbol, size, pen, brush)
    finally:
        p.end()
    return device

def logTickValues(self, minVal, maxVal, size, stdTicks):

        ## start with the tick spacing given by tickValues().
        ## Any level whose spacing is < 1 needs to be converted to log scale

        ticks = []
        for (spacing, t) in stdTicks:
            if spacing >= 1.0:
                ticks.append((spacing, t))

        if len(ticks) < 3:
            v1 = int(np.floor(minVal))
            v2 = int(np.ceil(maxVal))
            #major = list(range(v1+1, v2))

            minor = []
            for v in range(v1, v2):
                minor.extend(v + np.log10(np.arange(1, 10)))
            minor = [x for x in minor if x>minVal and x<maxVal]
            ticks.append((None, minor))
        return ticks

def processBlocks(lines,header,obstimes,svset,headlines,sats):
    obstypes = header['# / TYPES OF OBSERV'][1:]
    blocks = Panel4D(labels=obstimes,
                     items=list(svset),
                     major_axis=obstypes,
                     minor_axis=['data','lli','ssi'])
    ttime1 = 0
    ttime2 = 0
    for i in range(len(headlines)):
        linesinblock = len(sats[i])*int(np.ceil(header['# / TYPES OF OBSERV'][0]/5))
        block = ''.join(lines[headlines[i]+1:headlines[i]+linesinblock+1])
        t1 = time.time()
        bdf = _block2df(block,obstypes,sats[i],len(sats[i]))
        ttime1 += (time.time()-t1)
        t2 = time.time()
        blocks.loc[obstimes[i],sats[i]] = bdf
        ttime2 += (time.time()-t2)            
    print("{0:.2f} seconds for _block2df".format(ttime1))
    print("{0:.2f} seconds for panel assignments".format(ttime2))
    return blocks

def processBlocks(lines,header,obstimes,svset,headlines,sats):
    obstypes = header['# / TYPES OF OBSERV'][1:]
    blocks = Panel4D(labels=obstimes,
                     items=list(svset),
                     major_axis=obstypes,
                     minor_axis=['data','lli','ssi'])
    ttime1 = 0
    ttime2 = 0
    for i in range(len(headlines)):
        linesinblock = len(sats[i])*int(np.ceil(header['# / TYPES OF OBSERV'][0]/5))
        block = ''.join(lines[headlines[i]+1:headlines[i]+linesinblock+1])
        t1 = time.time()
        bdf = _block2df(block,obstypes,sats[i],len(sats[i]))
        ttime1 += (time.time()-t1)
        t2 = time.time()
        blocks.loc[obstimes[i],sats[i]] = bdf
        ttime2 += (time.time()-t2)            
    print("{0:.2f} seconds for _block2df".format(ttime1))
    print("{0:.2f} seconds for panel assignments".format(ttime2))
    return blocks

def reset(self):
        """ Resets the state of the generator"""
        self.step = 0
        Y = np.argmax(self.Y,1)
        labels = np.unique(Y)
        idx = []
        smallest = len(Y)
        for i,label in enumerate(labels):
            where = np.where(Y==label)[0]
            if smallest > len(where): 
                self.slabel = i
                smallest = len(where)
            idx.append(where)
        self.idx = idx
        self.labels = labels
        self.n_per_class = int(self.batch_size // len(labels))
        self.n_batches = int(np.ceil((smallest//self.n_per_class)))+1
        self.update_probabilities()

def __init__(self, X, Y, batch_size,cropsize=0, truncate=False, sequential=False,
                 random=True, val=False, class_weights=None):

        assert len(X) == len(Y), 'X and Y must be the same length {}!={}'.format(len(X),len(Y))
        if sequential: print('Using sequential mode')
        print ('starting normal generator')
        self.X = X
        self.Y = Y
        self.rnd_idx = np.arange(len(Y))
        self.Y_last_epoch = []
        self.val = val
        self.step = 0
        self.i = 0
        self.cropsize=cropsize
        self.truncate = truncate
        self.random = False if sequential or val else random
        self.batch_size = int(batch_size)
        self.sequential = sequential
        self.c_weights = class_weights if class_weights else dict(zip(np.unique(np.argmax(Y,1)),np.ones(len(np.argmax(Y,1)))))
        assert set(np.argmax(Y,1)) == set([int(x) for x in self.c_weights.keys()]), 'not all labels in class weights'
        self.n_batches = int(len(X)//batch_size if truncate else np.ceil(len(X)/batch_size))
        if self.random: self.randomize()

def calc_row_col(self, num_ex, num_items):
        num_rows_per_ex = int(np.ceil(num_items / self.max_num_col))
        if num_items > self.max_num_col:
            num_col = self.max_num_col
            num_row = num_rows_per_ex * num_ex
        else:
            num_row = num_ex
            num_col = num_items

        def calc(ii, jj):
            col = jj % self.max_num_col
            row = num_rows_per_ex * ii + int(jj / self.max_num_col)

            return row, col

        return num_row, num_col, calc

def card_strength(self, include_gem=True):
        # Base attribute value from naked card
        base_attr = np.array([getattr(self, attr.lower()) for attr in attr_list], dtype=float)
        # Bonus from bond
        bond_bonus = np.array([self.bond*(attr==self.main_attr) for attr in attr_list], dtype=float)
        # Compute card-only attribute: base+bond
        card_only_attr = base_attr + bond_bonus
        if not include_gem:
            strength = np.array(card_only_attr, dtype=int).tolist()
        else:
            gem_type_list = ['Kiss', 'Perfume', 'Ring', 'Cross']
            gem_matrix = {gem_type:np.zeros(3) for gem_type in gem_type_list}
            for gem in self.equipped_gems:
                gem_type = gem.name.split()[1]
                if gem_type in gem_type_list:
                    gem_matrix[gem_type][attr_list.index(gem.attribute)] = gem.value / 100**(gem.effect=='attr_boost')
            strength = card_only_attr.copy()
            for gem_type in gem_type_list:
                if gem_type in ['Kiss', 'Perfume']:
                    strength += gem_matrix[gem_type]
                elif gem_type in ['Ring', 'Cross']:
                    strength += np.ceil(card_only_attr*gem_matrix[gem_type])
            strength = np.array(strength, dtype=int)
        return {k.lower()+'*':v for k,v in zip(attr_list, strength)}

def vis_square(data):
    """Take an array of shape (n, height, width) or (n, height, width, 3)
       and visualize each (height, width) thing in a grid of size approx. sqrt(n) by sqrt(n)"""

    # normalize data for display
    data = (data - data.min()) / (data.max() - data.min())

    # force the number of filters to be square
    n = int(np.ceil(np.sqrt(data.shape[0])))
    padding = (((0, n ** 2 - data.shape[0]),
               (0, 1), (0, 1))                 # add some space between filters
               + ((0, 0),) * (data.ndim - 3))  # don't pad the last dimension (if there is one)
    data = np.pad(data, padding, mode='constant', constant_values=1)  # pad with ones (white)

    # tile the filters into an image
    data = data.reshape((n, n) + data.shape[1:]).transpose((0, 2, 1, 3) + tuple(range(4, data.ndim + 1)))
    data = data.reshape((n * data.shape[1], n * data.shape[3]) + data.shape[4:])
    plt.imshow(data, interpolation='nearest'); plt.axis('off')

def process(self, wave):
        wave.check_mono()
        if wave.sample_rate != self.sr:
            raise Exception("Wrong sample rate")                              
        n = int(np.ceil(2 * wave.num_frames / float(self.w_len)))
        m = (n + 1) * self.w_len / 2 
        swindow = self.make_signal_window(n)
        win_ratios = [self.window / swindow[t * self.w_len / 2 : 
            t * self.w_len / 2 + self.w_len] 
            for t in range(n)]
        wave = wave.zero_pad(0, int(m - wave.num_frames))
        wave = audio.Wave(signal.hilbert(wave), wave.sample_rate)        
        result = np.zeros((self.n_bins, n))

        for b in range(self.n_bins): 
            w = self.widths[b]
            wc = 1 / np.square(w + 1)
            filter = self.filters[b]
            band = fftfilt(filter, wave.zero_pad(0, int(2 * w))[:,0])
            band = band[int(w) : int(w + m), np.newaxis]    
            for t in range(n):
                frame = band[t * self.w_len / 2:
                             t * self.w_len / 2 + self.w_len,:] * win_ratios[t]
                result[b, t] =  wc * np.real(np.conj(np.dot(frame.conj().T, frame)))
        return audio.Spectrogram(result, self.sr, self.w_len, self.w_len / 2)

def n2mfrow(nr_plots):
    """
    Compute the rows and columns given the number
    of plots.

    This is a port of grDevices::n2mfrow from R
    """
    if nr_plots <= 3:
        nrow, ncol = nr_plots, 1
    elif nr_plots <= 6:
        nrow, ncol = (nr_plots + 1) // 2, 2
    elif nr_plots <= 12:
        nrow, ncol = (nr_plots + 2) // 3, 3
    else:
        nrow = int(np.ceil(np.sqrt(nr_plots)))
        ncol = int(np.ceil(nr_plots/nrow))
    return (nrow, ncol)

def get_padding_type(kernel_params, input_shape, output_shape):
    '''Translates Caffe's numeric padding to one of ('SAME', 'VALID').
    Caffe supports arbitrary padding values, while TensorFlow only
    supports 'SAME' and 'VALID' modes. So, not all Caffe paddings
    can be translated to TensorFlow. There are some subtleties to
    how the padding edge-cases are handled. These are described here:
    https://github.com/Yangqing/caffe2/blob/master/caffe2/proto/caffe2_legacy.proto
    '''
    k_h, k_w, s_h, s_w, p_h, p_w = kernel_params
    s_o_h = np.ceil(input_shape.height / float(s_h))
    s_o_w = np.ceil(input_shape.width / float(s_w))
    if (output_shape.height == s_o_h) and (output_shape.width == s_o_w):
        return 'SAME'
    v_o_h = np.ceil((input_shape.height - k_h + 1.0) / float(s_h))
    v_o_w = np.ceil((input_shape.width - k_w + 1.0) / float(s_w))
    if (output_shape.height == v_o_h) and (output_shape.width == v_o_w):
        return 'VALID'
    return None

def plot_weight_matrix(Z, outname, save=True):
    num = Z.shape[0]
    fig = plt.figure(1, (80, 80))
    fig.subplots_adjust(left=0.05, right=0.95)
    grid = AxesGrid(fig, (1, 4, 2),  # similar to subplot(142)
                    nrows_ncols=(int(np.ceil(num / 10.)), 10),
                    axes_pad=0.04,
                    share_all=True,
                    label_mode="L",
                    )

    for i in range(num):
        im = grid[i].imshow(Z[i, :, :, :].mean(
            axis=0), cmap='gray')
    for i in range(grid.ngrids):
        grid[i].axis('off')

    for cax in grid.cbar_axes:
        cax.toggle_label(False)
    if save:
        fig.savefig(outname, bbox_inches='tight')
        fig.clear()

def __init__(self, h, x0=None, **kwargs):
        assert type(h) is list, 'h must be a list'
        assert len(h) in [2, 3], "TreeMesh is only in 2D or 3D."

        if '_levels' in kwargs.keys():
            self._levels = kwargs.pop('_levels')

        BaseTensorMesh.__init__(self, h, x0, **kwargs)

        if self._levels is None:
            self._levels = int(np.log2(len(self.h[0])))

        # self._levels = levels
        self._levelBits = int(np.ceil(np.sqrt(self._levels)))+1

        self.__dirty__ = True  #: The numbering is dirty!

        if '_cells' in kwargs.keys():
            self._cells = kwargs.pop('_cells')
        else:
            self._cells.add(0)

def _optim(self, xys):
        idx = np.arange(len(xys))
        self.batch_size = np.ceil(len(xys) / self.nbatches)
        batch_idx = np.arange(self.batch_size, len(xys), self.batch_size)

        for self.epoch in range(1, self.max_epochs + 1):
            # shuffle training examples
            self._pre_epoch()
            shuffle(idx)

            # store epoch for callback
            self.epoch_start = timeit.default_timer()

            # process mini-batches
            for batch in np.split(idx, batch_idx):
                # select indices for current batch
                bxys = [xys[z] for z in batch]
                self._process_batch(bxys)

            # check callback function, if false return
            for f in self.post_epoch:
                if not f(self):
                    break

def dec_round(num, dprec=4, rnd='down', rto_zero=False):
    """
    Round up/down numeric ``num`` at specified decimal ``dprec``.

    Parameters
    ----------
    num: float
    dprec: int
        Decimal position for truncation.
    rnd: str (default: 'down')
        Set as 'up' or 'down' to return a rounded-up or rounded-down value.
    rto_zero: bool (default: False)
        Use a *round-towards-zero* method, e.g., ``floor(-3.5) == -3``.

    Returns
    ----------
    float (default: rounded-up)
    """
    dprec = 10**dprec
    if rnd == 'up' or (rnd == 'down' and rto_zero and num < 0.):
        return np.ceil(num*dprec)/dprec
    elif rnd == 'down' or (rnd == 'up' and rto_zero and num < 0.):
        return np.floor(num*dprec)/dprec
    return np.round(num, dprec)

def update(self, es, **kwargs):
        if es.countiter < 2:
            self.initialize(es)
            self.fit = es.fit.fit
        else:
            ft1, ft2 = self.fit[int(self.index_to_compare)], self.fit[int(np.ceil(self.index_to_compare))]
            ftt1, ftt2 = es.fit.fit[(es.popsize - 1) // 2], es.fit.fit[int(np.ceil((es.popsize - 1) / 2))]
            pt2 = self.index_to_compare - int(self.index_to_compare)
            # ptt2 = (es.popsize - 1) / 2 - (es.popsize - 1) // 2  # not in use
            s = 0
            if 1 < 3:
                s += pt2 * sum(es.fit.fit <= self.fit[int(np.ceil(self.index_to_compare))])
                s += (1 - pt2) * sum(es.fit.fit < self.fit[int(self.index_to_compare)])
                s -= es.popsize / 2.
                s *= 2. / es.popsize  # the range was popsize, is 2
            self.s = (1 - self.c) * self.s + self.c * s
            es.sigma *= exp(self.s / self.damp)
        # es.more_to_write.append(10**(self.s))

        #es.more_to_write.append(10**((2 / es.popsize) * (sum(es.fit.fit < self.fit[int(self.index_to_compare)]) - (es.popsize + 1) / 2)))
        # # es.more_to_write.append(10**(self.index_to_compare - sum(self.fit <= es.fit.fit[es.popsize // 2])))
        # # es.more_to_write.append(10**(np.sign(self.fit[int(self.index_to_compare)] - es.fit.fit[es.popsize // 2])))
        self.fit = es.fit.fit

def __init__(self, env, n, max_path_length, scope=None):
        if scope is None:
            # initialize random scope
            scope = str(uuid.uuid4())

        envs_per_worker = int(np.ceil(n * 1.0 / singleton_pool.n_parallel))
        alloc_env_ids = []
        rest_alloc = n
        start_id = 0
        for _ in range(singleton_pool.n_parallel):
            n_allocs = min(envs_per_worker, rest_alloc)
            alloc_env_ids.append(list(range(start_id, start_id + n_allocs)))
            start_id += n_allocs
            rest_alloc = max(0, rest_alloc - envs_per_worker)

        singleton_pool.run_each(worker_init_envs, [(alloc, scope, env) for alloc in alloc_env_ids])

        self._alloc_env_ids = alloc_env_ids
        self._action_space = env.action_space
        self._observation_space = env.observation_space
        self._num_envs = n
        self.scope = scope
        self.ts = np.zeros(n, dtype='int')
        self.max_path_length = max_path_length

def view_samples(self, show=True):
        """Displays the samples."""
        if not self.samples:
            return  # Nothing to show...
        plt.figure("Sample views")
        num = len(self.samples)
        rows = math.floor(num ** .5)
        cols = math.ceil(num / rows)
        for idx, img in enumerate(self.samples):
            plt.subplot(rows, cols, idx+1)
            plt.imshow(img, interpolation='nearest')
        if show:
            plt.show()


# EXPERIMENT: Try breaking out each output encoder by type instead of
# concatenating them all together.  Each type of sensors would then get its own
# HTM.  Maybe keep the derivatives with their source?
#

def make_train_test_split(prms):
    '''
    # I will just make one split and consider the last 5% of the iamges as the val images. 
    # Randomly sampling in this data is a bad idea, because many images appear together as 
    # pairs. Selecting from the end will maximize the chances of using unique and different
    # imahes in the train and test splits. 
    '''
    # Read the source pairs. 
    fid    = open(prms['paths']['pairList']['raw'],'r')
    lines  = fid.readlines()
    fid.close()
    numIm, numPairs = int(lines[0].split()[0]), int(lines[0].split()[1])
    lines = lines[1:]

    #Make train and val splits
    N = len(lines)
    trainNum   = int(np.ceil(0.95 * N))
    trainLines = lines[0:trainNum]
    testLines  = lines[trainNum:]
    _write_pairs(prms['paths']['pairList']['train'], trainLines, numIm)
    _write_pairs(prms['paths']['pairList']['test'] , testLines, numIm)

##
# Get the list of tar files for downloading the image data

def gen_samples(self, z0=None, n=32, batch_size=32, use_transform=True):
        assert n % batch_size == 0

        samples = []

        if z0 is None:
            z0 = np_rng.uniform(-1., 1., size=(n, self.nz))
        else:
            n = len(z0)
            batch_size = max(n, 64)
        n_batches = int(np.ceil(n/float(batch_size)))
        for i in range(n_batches):
            zmb = floatX(z0[batch_size * i:min(n, batch_size * (i + 1)), :])
            xmb = self._gen(zmb)
            samples.append(xmb)

        samples = np.concatenate(samples, axis=0)
        if use_transform:
            samples = self.inverse_transform(samples, npx=self.npx, nc=self.nc)
            samples = (samples * 255).astype(np.uint8)
        return samples

def __init__(self, opt_engine, topK=16, grid_size=None, nps=320, model_name='tmp'):
        QWidget.__init__(self)
        self.topK = topK
        if grid_size is None:
            self.n_grid = int(np.ceil(np.sqrt(self.topK)))
            self.grid_size = (self.n_grid, self.n_grid) # (width, height)
        else:
            self.grid_size = grid_size
        self.select_id = 0
        self.ims = None
        self.vis_results = None
        self.width = int(np.round(nps/ (4 * float(self.grid_size[1])))) * 4
        self.winWidth = self.width * self.grid_size[0]
        self.winHeight = self.width * self.grid_size[1]

        self.setFixedSize(self.winWidth, self.winHeight)
        self.opt_engine = opt_engine
        self.frame_id = -1
        self.sr = save_result.SaveResult(model_name=model_name)

def rasta_plp_extractor(x, sr, plp_order=0, do_rasta=True):
    spec = log_power_spectrum_extractor(x, int(sr*0.02), int(sr*0.01), 'hamming', False)
    bark_filters = int(np.ceil(freq2bark(sr//2)))
    wts = get_fft_bark_mat(sr, int(sr*0.02), bark_filters)
    bark_spec = np.matmul(wts, spec)
    if do_rasta:
        bark_spec = np.where(bark_spec == 0.0, np.finfo(float).eps, bark_spec)
        log_bark_spec = np.log(bark_spec)
        rasta_log_bark_spec = rasta_filt(log_bark_spec)
        bark_spec = np.exp(rasta_log_bark_spec)
    post_spec = postaud(bark_spec, sr/2.)
    if plp_order > 0:
        lpcas = do_lpc(post_spec, plp_order)
    else:
        lpcas = post_spec
    return lpcas

def _wav_to_framed_samples(wav_audio, hparams):
  """Transforms the contents of a wav file into a series of framed samples."""
  y = audio_io.wav_data_to_samples(wav_audio, hparams.sample_rate)

  hl = hparams.spec_hop_length
  n_frames = int(np.ceil(y.shape[0] / hl))
  frames = np.zeros((n_frames, hl), dtype=np.float32)

  # Fill in everything but the last frame which may not be the full length
  cutoff = (n_frames - 1) * hl
  frames[:n_frames - 1, :] = np.reshape(y[:cutoff], (n_frames - 1, hl))
  # Fill the last frame
  remain_len = len(y[cutoff:])
  frames[n_frames - 1, :remain_len] = y[cutoff:]

  return frames

def comp_ola_deconv(fs_gpu, ys_gpu, L_gpu, alpha, beta):
    """
    Computes the division in Fourier space needed for direct deconvolution
    """

    sfft = fs_gpu.shape
    block_size = (16,16,1)   
    grid_size = (int(np.ceil(np.float32(sfft[0]*sfft[1])/block_size[0])),
                 int(np.ceil(np.float32(sfft[2])/block_size[1])))

    mod = cu.module_from_buffer(cubin)
    comp_ola_deconv_Kernel = mod.get_function("comp_ola_deconv_Kernel")

    z_gpu = cua.zeros(sfft, np.complex64)

    comp_ola_deconv_Kernel(z_gpu.gpudata,
                           np.int32(sfft[0]), np.int32(sfft[1]), np.int32(sfft[2]),
                           fs_gpu.gpudata, ys_gpu.gpudata, L_gpu.gpudata,
                           np.float32(alpha), np.float32(beta),
                           block=block_size, grid=grid_size)

    return z_gpu

def crop_gpu2cpu(x_gpu, sz, offset=(0,0)):

    sfft = x_gpu.shape

    block_size = (16, 16 ,1)
    grid_size = (int(np.ceil(np.float32(sfft[1])/block_size[1])),
                 int(np.ceil(np.float32(sfft[0])/block_size[0])))

    if x_gpu.dtype == np.float32:
        mod = cu.module_from_buffer(cubin)
        cropKernel = mod.get_function("crop_Kernel")

    elif x_gpu.dtype == np.complex64:
        mod = cu.module_from_buffer(cubin)
        cropKernel = mod.get_function("crop_ComplexKernel")

    x_cropped_gpu = cua.empty(tuple((int(sz[0]),int(sz[1]))), np.float32)

    cropKernel(x_cropped_gpu.gpudata,   np.int32(sz[0]),       np.int32(sz[1]),
                       x_gpu.gpudata, np.int32(sfft[0]),     np.int32(sfft[1]),
                                    np.int32(offset[0]), np.int32(offset[1]),
                                    block=block_size   , grid=grid_size)

    return x_cropped_gpu

def comp_ola_sdeconv(gx_gpu, gy_gpu, xx_gpu, xy_gpu, Ftpy_gpu, f_gpu, L_gpu, alpha, beta, gamma=0):
    """
    Computes the division in Fourier space needed for sparse deconvolution
    """

    sfft = xx_gpu.shape
    block_size = (16,16,1)   
    grid_size = (int(np.ceil(np.float32(sfft[0]*sfft[1])/block_size[0])),
                 int(np.ceil(np.float32(sfft[2])/block_size[1])))

    mod = cu.module_from_buffer(cubin)
    comp_ola_sdeconv_Kernel = mod.get_function("comp_ola_sdeconv_Kernel")

    z_gpu = cua.zeros(sfft, np.complex64)

    comp_ola_sdeconv_Kernel(z_gpu.gpudata,
                            np.int32(sfft[0]), np.int32(sfft[1]), np.int32(sfft[2]),
                            gx_gpu.gpudata, gy_gpu.gpudata,
                            xx_gpu.gpudata, xy_gpu.gpudata, 
                            Ftpy_gpu.gpudata, f_gpu.gpudata, L_gpu.gpudata,
                            np.float32(alpha), np.float32(beta),
                            np.float32(gamma),
                            block=block_size, grid=grid_size)

    return z_gpu

def impad_gpu(y_gpu, sf):

  sf = np.array(sf)
  shape = (np.array(y_gpu.shape) + sf).astype(np.uint32)
  dtype = y_gpu.dtype
  block_size = (16,16,1)
  grid_size = (int(np.ceil(float(shape[1])/block_size[0])),
               int(np.ceil(float(shape[0])/block_size[1])))

  preproc = _generate_preproc(dtype, shape)
  mod = SourceModule(preproc + kernel_code, keep=True)

  padded_gpu = cua.empty((int(shape[0]), int(shape[1])), dtype)
  impad_fun = mod.get_function("impad")

  upper_left = np.uint32(np.floor(sf / 2.))
  original_size = np.uint32(np.array(y_gpu.shape))

  impad_fun(padded_gpu.gpudata, y_gpu.gpudata,
            upper_left[1], upper_left[0],
            original_size[0], original_size[1],
            block=block_size, grid=grid_size)

  return padded_gpu

def laplace_stack_gpu(y_gpu, mode='valid'):
  """
  This funtion computes the Laplacian of each slice of a stack of images
  """
  shape = np.array(y_gpu.shape).astype(np.uint32)
  dtype = y_gpu.dtype
  block_size = (6,int(np.floor(512./6./float(shape[0]))),int(shape[0]))
  grid_size = (int(np.ceil(float(shape[1])/block_size[0])),
               int(np.ceil(float(shape[0])/block_size[1])))
  shared_size = int((2+block_size[0])*(2+block_size[1])*(2+block_size[2])
                    *dtype.itemsize)

  preproc = _generate_preproc(dtype, (shape[1],shape[2]))
  mod = SourceModule(preproc + kernel_code, keep=True)

  laplace_fun_gpu = mod.get_function("laplace_stack_same")
  laplace_gpu = cua.empty((y_gpu.shape[0], y_gpu.shape[1], y_gpu.shape[2]),
                          y_gpu.dtype)

  laplace_fun_gpu(laplace_gpu.gpudata, y_gpu.gpudata,
                  block=block_size, grid=grid_size, shared=shared_size)

  return laplace_gpu

def morph(roi):
    ratio = min(28. / np.size(roi, 0), 28. / np.size(roi, 1))
    roi = cv2.resize(roi, None, fx=ratio, fy=ratio,
                     interpolation=cv2.INTER_NEAREST)
    dx = 28 - np.size(roi, 1)
    dy = 28 - np.size(roi, 0)
    px = ((int(dx / 2.)), int(np.ceil(dx / 2.)))
    py = ((int(dy / 2.)), int(np.ceil(dy / 2.)))
    squared = np.pad(roi, (py, px), 'constant', constant_values=0)
    return squared

def computePad(dims,depth):
    y1=y2=x1=x2=0; 
    y,x = [numpy.ceil(dims[i]/float(2**depth)) * (2**depth) for i in range(-2,0)]
    x = float(x); y = float(y);
    y1 = int(numpy.floor((y - dims[-2])/2)); y2 = int(numpy.ceil((y - dims[-2])/2))
    x1 = int(numpy.floor((x - dims[-1])/2)); x2 = int(numpy.ceil((x - dims[-1])/2))
    return y1,y2,x1,x2

def view_waveforms_clusters(data, halo, threshold, templates, amps_lim, n_curves=200, save=False):

    nb_templates = templates.shape[1]
    n_panels     = numpy.ceil(numpy.sqrt(nb_templates))
    mask         = numpy.where(halo > -1)[0]
    clust_idx    = numpy.unique(halo[mask])
    fig          = pylab.figure()    
    square       = True
    center       = len(data[0] - 1)//2
    for count, i in enumerate(xrange(nb_templates)):
        if square:
            pylab.subplot(n_panels, n_panels, count + 1)
            if (numpy.mod(count, n_panels) != 0):
                pylab.setp(pylab.gca(), yticks=[])
            if (count < n_panels*(n_panels - 1)):
                pylab.setp(pylab.gca(), xticks=[])

        subcurves = numpy.where(halo == clust_idx[count])[0]
        for k in numpy.random.permutation(subcurves)[:n_curves]:
            pylab.plot(data[k], '0.5')

        pylab.plot(templates[:, count], 'r')        
        pylab.plot(amps_lim[count][0]*templates[:, count], 'b', alpha=0.5)
        pylab.plot(amps_lim[count][1]*templates[:, count], 'b', alpha=0.5)

        xmin, xmax = pylab.xlim()
        pylab.plot([xmin, xmax], [-threshold, -threshold], 'k--')
        pylab.plot([xmin, xmax], [threshold, threshold], 'k--')
        #pylab.ylim(-1.5*threshold, 1.5*threshold)
        ymin, ymax = pylab.ylim()
        pylab.plot([center, center], [ymin, ymax], 'k--')
        pylab.title('Cluster %d' %i)

    if nb_templates > 0:
        pylab.tight_layout()
    if save:
        pylab.savefig(os.path.join(save[0], 'waveforms_%s' %save[1]))
        pylab.close()
    else:
        pylab.show()
    del fig

def draw_circles(image,cands,origin,spacing):
    #make empty matrix, which will be filled with the mask
    image_mask = np.zeros(image.shape, dtype=np.int16)

    #run over all the nodules in the lungs
    for ca in cands.values:
        #get middel x-,y-, and z-worldcoordinate of the nodule
        #radius = np.ceil(ca[4])/2     ## original:  replaced the ceil with a very minor increase of 1% ....
        radius = (ca[4])/2 + 0.51 * spacing[0]  # increasing by circa half of distance in z direction .... (trying to capture wider region/border for learning ... and adress the rough net .

        coord_x = ca[1]
        coord_y = ca[2]
        coord_z = ca[3]
        image_coord = np.array((coord_z,coord_y,coord_x))

        #determine voxel coordinate given the worldcoordinate
        image_coord = world_2_voxel(image_coord,origin,spacing)

        #determine the range of the nodule
        #noduleRange = seq(-radius, radius, RESIZE_SPACING[0])  # original, uniform spacing 
        noduleRange_z = seq(-radius, radius, spacing[0])
        noduleRange_y = seq(-radius, radius, spacing[1])
        noduleRange_x = seq(-radius, radius, spacing[2])

          #x = y = z = -2
        #create the mask
        for x in noduleRange_x:
            for y in noduleRange_y:
                for z in noduleRange_z:
                    coords = world_2_voxel(np.array((coord_z+z,coord_y+y,coord_x+x)),origin,spacing)
                    #if (np.linalg.norm(image_coord-coords) * RESIZE_SPACING[0]) < radius:  ### original (contrained to a uniofrm RESIZE)
                    if (np.linalg.norm((image_coord-coords) * spacing)) < radius:
                        image_mask[int(np.round(coords[0])),int(np.round(coords[1])),int(np.round(coords[2]))] = int(1)


    return image_mask

def draw_circles(image,cands,origin,spacing):
    #make empty matrix, which will be filled with the mask
    image_mask = np.zeros(image.shape, dtype=np.int16)

    #run over all the nodules in the lungs
    for ca in cands.values:
        #get middel x-,y-, and z-worldcoordinate of the nodule
        #radius = np.ceil(ca[4])/2     ## original:  replaced the ceil with a very minor increase of 1% ....
        radius = (ca[4])/2 + 0.51 * spacing[0]  # increasing by circa half of distance in z direction .... (trying to capture wider region/border for learning ... and adress the rough net .

        coord_x = ca[1]
        coord_y = ca[2]
        coord_z = ca[3]
        image_coord = np.array((coord_z,coord_y,coord_x))

        #determine voxel coordinate given the worldcoordinate
        image_coord = world_2_voxel(image_coord,origin,spacing)

        #determine the range of the nodule
        #noduleRange = seq(-radius, radius, RESIZE_SPACING[0])  # original, uniform spacing 
        noduleRange_z = seq(-radius, radius, spacing[0])
        noduleRange_y = seq(-radius, radius, spacing[1])
        noduleRange_x = seq(-radius, radius, spacing[2])

          #x = y = z = -2
        #create the mask
        for x in noduleRange_x:
            for y in noduleRange_y:
                for z in noduleRange_z:
                    coords = world_2_voxel(np.array((coord_z+z,coord_y+y,coord_x+x)),origin,spacing)
                    #if (np.linalg.norm(image_coord-coords) * RESIZE_SPACING[0]) < radius:  ### original (contrained to a uniofrm RESIZE)
                    if (np.linalg.norm((image_coord-coords) * spacing)) < radius:
                        image_mask[int(np.round(coords[0])),int(np.round(coords[1])),int(np.round(coords[2]))] = int(1)


    return image_mask