Python tensorflow.cast() Examples

def _create_autosummary_var(name, value_expr):
    assert not _autosummary_finalized
    v = tf.cast(value_expr, tf.float32)
    if v.shape.ndims is 0:
        v = [v, np.float32(1.0)]
    elif v.shape.ndims is 1:
        v = [tf.reduce_sum(v), tf.cast(tf.shape(v)[0], tf.float32)]
    else:
        v = [tf.reduce_sum(v), tf.reduce_prod(tf.cast(tf.shape(v), tf.float32))]
    v = tf.cond(tf.is_finite(v[0]), lambda: tf.stack(v), lambda: tf.zeros(2))
    with tf.control_dependencies(None):
        var = tf.Variable(tf.zeros(2)) # [numerator, denominator]
    update_op = tf.cond(tf.is_variable_initialized(var), lambda: tf.assign_add(var, v), lambda: tf.assign(var, v))
    if name in _autosummary_vars:
        _autosummary_vars[name].append(var)
    else:
        _autosummary_vars[name] = [var]
    return update_op

#----------------------------------------------------------------------------
# Call filewriter.add_summary() with all summaries in the default graph,
# automatically finalizing and merging them on the first call. 

def minibatch_stddev_layer(x, group_size=4):
    with tf.variable_scope('MinibatchStddev'):
        group_size = tf.minimum(group_size, tf.shape(x)[0])     # Minibatch must be divisible by (or smaller than) group_size.
        s = x.shape                                             # [NCHW]  Input shape.
        y = tf.reshape(x, [group_size, -1, s[1], s[2], s[3]])   # [GMCHW] Split minibatch into M groups of size G.
        y = tf.cast(y, tf.float32)                              # [GMCHW] Cast to FP32.
        y -= tf.reduce_mean(y, axis=0, keep_dims=True)           # [GMCHW] Subtract mean over group.
        y = tf.reduce_mean(tf.square(y), axis=0)                # [MCHW]  Calc variance over group.
        y = tf.sqrt(y + 1e-8)                                   # [MCHW]  Calc stddev over group.
        y = tf.reduce_mean(y, axis=[1,2,3], keep_dims=True)      # [M111]  Take average over fmaps and pixels.
        y = tf.cast(y, x.dtype)                                 # [M111]  Cast back to original data type.
        y = tf.tile(y, [group_size, 1, s[2], s[3]])             # [N1HW]  Replicate over group and pixels.
        return tf.concat([x, y], axis=1)                        # [NCHW]  Append as new fmap.

#----------------------------------------------------------------------------
# Generator network used in the paper. 

def fprop(self, x, y, **kwargs):
        kwargs.update(self.kwargs)
        if self.attack is not None:
            x = x, self.attack(x)
        else:
            x = x,

        # Catching RuntimeError: Variable -= value not supported by tf.eager.
        try:
            y -= self.smoothing * (y - 1. / tf.cast(y.shape[-1], y.dtype))
        except RuntimeError:
            y.assign_sub(self.smoothing * (y - 1. / tf.cast(y.shape[-1],
                                                            y.dtype)))

        logits = [self.model.get_logits(x, **kwargs) for x in x]
        loss = sum(
            tf.reduce_mean(softmax_cross_entropy_with_logits(labels=y,
                                                             logits=logit))
            for logit in logits)
        return loss 

def fprop(self, x, y, **kwargs):
        if self.attack is not None:
            x = x, self.attack(x)
        else:
            x = x,

        # Catching RuntimeError: Variable -= value not supported by tf.eager.
        try:
            y -= self.smoothing * (y - 1. / tf.cast(y.shape[-1], tf.float32))
        except RuntimeError:
            y.assign_sub(self.smoothing * (y - 1. / tf.cast(y.shape[-1],
                                                            tf.float32)))

        logits = [self.model.get_logits(x, **kwargs) for x in x]
        loss = sum(
            softmax_cross_entropy_with_logits(labels=y,
                                              logits=logit)
            for logit in logits)
        warnings.warn("LossCrossEntropy is deprecated, switch to "
                      "CrossEntropy. LossCrossEntropy may be removed on "
                      "or after 2019-03-06.")
        return loss 

def from_float32_to_uint8(
        tensor,
        tensor_key='tensor',
        min_key='min',
        max_key='max'):
    """

    :param tensor:
    :param tensor_key:
    :param min_key:
    :param max_key:
    :returns:
    """
    tensor_min = tf.reduce_min(tensor)
    tensor_max = tf.reduce_max(tensor)
    return {
        tensor_key: tf.cast(
            (tensor - tensor_min) / (tensor_max - tensor_min + 1e-16)
            * 255.9999, dtype=tf.uint8),
        min_key: tensor_min,
        max_key: tensor_max
    } 

def time_stretch(
        spectrogram,
        factor=1.0,
        method=tf.image.ResizeMethod.BILINEAR):
    """ Time stretch a spectrogram preserving shape in tensorflow. Note that
    this is an approximation in the frequency domain.

    :param spectrogram: Input spectrogram to be time stretched as tensor.
    :param factor: (Optional) Time stretch factor, must be >0, default to 1.
    :param mehtod: (Optional) Interpolation method, default to BILINEAR.
    :returns: Time stretched spectrogram as tensor with same shape.
    """
    T = tf.shape(spectrogram)[0]
    T_ts = tf.cast(tf.cast(T, tf.float32) * factor, tf.int32)[0]
    F = tf.shape(spectrogram)[1]
    ts_spec = tf.image.resize_images(
        spectrogram,
        [T_ts, F],
        method=method,
        align_corners=True)
    return tf.image.resize_image_with_crop_or_pad(ts_spec, T, F) 

def pitch_shift(
        spectrogram,
        semitone_shift=0.0,
        method=tf.image.ResizeMethod.BILINEAR):
    """ Pitch shift a spectrogram preserving shape in tensorflow. Note that
    this is an approximation in the frequency domain.

    :param spectrogram: Input spectrogram to be pitch shifted as tensor.
    :param semitone_shift: (Optional) Pitch shift in semitone, default to 0.0.
    :param mehtod: (Optional) Interpolation method, default to BILINEAR.
    :returns: Pitch shifted spectrogram (same shape as spectrogram).
    """
    factor = 2 ** (semitone_shift / 12.)
    T = tf.shape(spectrogram)[0]
    F = tf.shape(spectrogram)[1]
    F_ps = tf.cast(tf.cast(F, tf.float32) * factor, tf.int32)[0]
    ps_spec = tf.image.resize_images(
        spectrogram,
        [T, F_ps],
        method=method,
        align_corners=True)
    paddings = [[0, 0], [0, tf.maximum(0, F - F_ps)], [0, 0]]
    return tf.pad(ps_spec[:, :F, :], paddings, 'CONSTANT') 

def loop_decode(self):
        # decoder_initial_state: Tuple Tensor (c,h) of size [batch_size x cell.state_size]
        # decoder_first_input: Tensor [batch_size x cell.state_size]

        # Loop the decoding process and collect results
        s,i = self.decoder_initial_state,  tf.cast(self.decoder_first_input,tf.float32)
        for step in range(self.seq_length):
            s, i = self.decode(s,i,step)

        # Return to start
        self.positions.append(self.first_city)

        # Stack visited indices
        self.positions=tf.stack(self.positions,axis=1)  # [Batch,seq_length+1]

        # Sum log_softmax over output steps
        self.log_softmax=tf.add_n(self.log_softmax)  # [Batch,seq_length]

        # Stack attending & pointing distribution
        self.attending=tf.stack(self.attending,axis=1) # [Batch,seq_length,seq_length]
        self.pointing=tf.stack(self.pointing,axis=1) # [Batch,seq_length,seq_length]
        
        # Return stacked lists of visited_indices and log_softmax for backprop
        return self.positions,self.log_softmax 

def distorted_inputs():
  """Construct distorted input for CIFAR training using the Reader ops.

  Returns:
    images: Images. 4D tensor of [batch_size, IMAGE_SIZE, IMAGE_SIZE, 3] size.
    labels: Labels. 1D tensor of [batch_size] size.

  Raises:
    ValueError: If no data_dir
  """
  if not FLAGS.data_dir:
    raise ValueError('Please supply a data_dir')
  data_dir = os.path.join(FLAGS.data_dir, 'cifar-10-batches-bin')
  images, labels = cifar10_input.distorted_inputs(data_dir=data_dir,
                                                  batch_size=FLAGS.batch_size)
  if FLAGS.use_fp16:
    images = tf.cast(images, tf.float16)
    labels = tf.cast(labels, tf.float16)
  return images, labels 

def inputs(eval_data):
  """Construct input for CIFAR evaluation using the Reader ops.

  Args:
    eval_data: bool, indicating if one should use the train or eval data set.

  Returns:
    images: Images. 4D tensor of [batch_size, IMAGE_SIZE, IMAGE_SIZE, 3] size.
    labels: Labels. 1D tensor of [batch_size] size.

  Raises:
    ValueError: If no data_dir
  """
  if not FLAGS.data_dir:
    raise ValueError('Please supply a data_dir')
  data_dir = os.path.join(FLAGS.data_dir, 'cifar-10-batches-bin')
  images, labels = cifar10_input.inputs(eval_data=eval_data,
                                        data_dir=data_dir,
                                        batch_size=FLAGS.batch_size)
  if FLAGS.use_fp16:
    images = tf.cast(images, tf.float16)
    labels = tf.cast(labels, tf.float16)
  return images, labels 

def loss(logits, labels):
  """Add L2Loss to all the trainable variables.

  Add summary for "Loss" and "Loss/avg".
  Args:
    logits: Logits from inference().
    labels: Labels from distorted_inputs or inputs(). 1-D tensor
            of shape [batch_size]

  Returns:
    Loss tensor of type float.
  """
  # Calculate the average cross entropy loss across the batch.
  labels = tf.cast(labels, tf.int64)
  cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
      labels=labels, logits=logits, name='cross_entropy_per_example')
  cross_entropy_mean = tf.reduce_mean(cross_entropy, name='cross_entropy')
  tf.add_to_collection('losses', cross_entropy_mean)

  # The total loss is defined as the cross entropy loss plus all of the weight
  # decay terms (L2 loss).
  return tf.add_n(tf.get_collection('losses'), name='total_loss') 

def char_predictions(self, chars_logit):
    """Returns confidence scores (softmax values) for predicted characters.

    Args:
      chars_logit: chars logits, a tensor with shape
        [batch_size x seq_length x num_char_classes]

    Returns:
      A tuple (ids, log_prob, scores), where:
        ids - predicted characters, a int32 tensor with shape
          [batch_size x seq_length];
        log_prob - a log probability of all characters, a float tensor with
          shape [batch_size, seq_length, num_char_classes];
        scores - corresponding confidence scores for characters, a float
        tensor
          with shape [batch_size x seq_length].
    """
    log_prob = utils.logits_to_log_prob(chars_logit)
    ids = tf.to_int32(tf.argmax(log_prob, dimension=2), name='predicted_chars')
    mask = tf.cast(
        slim.one_hot_encoding(ids, self._params.num_char_classes), tf.bool)
    all_scores = tf.nn.softmax(chars_logit)
    selected_scores = tf.boolean_mask(all_scores, mask, name='char_scores')
    scores = tf.reshape(selected_scores, shape=(-1, self._params.seq_length))
    return ids, log_prob, scores 

def compute_first_or_last(self, select, first=True):
    #perform first ot last operation on row select with probabilistic row selection
    answer = tf.zeros_like(select)
    running_sum = tf.zeros([self.batch_size, 1], self.data_type)
    for i in range(self.max_elements):
      if (first):
        current = tf.slice(select, [0, i], [self.batch_size, 1])
      else:
        current = tf.slice(select, [0, self.max_elements - 1 - i],
                           [self.batch_size, 1])
      curr_prob = current * (1 - running_sum)
      curr_prob = curr_prob * tf.cast(curr_prob >= 0.0, self.data_type)
      running_sum += curr_prob
      temp_ans = []
      curr_prob = tf.expand_dims(tf.reshape(curr_prob, [self.batch_size]), 0)
      for i_ans in range(self.max_elements):
        if (not (first) and i_ans == self.max_elements - 1 - i):
          temp_ans.append(curr_prob)
        elif (first and i_ans == i):
          temp_ans.append(curr_prob)
        else:
          temp_ans.append(tf.zeros_like(curr_prob))
      temp_ans = tf.transpose(tf.concat(axis=0, values=temp_ans))
      answer += temp_ans
    return answer 

def batch_of_random_bools(batch_size, n):
  """Return a batch of random "boolean" numbers.

  Args:
    batch_size:  Batch size dimension of returned tensor.
    n:  number of entries per batch.

  Returns:
    A [batch_size, n] tensor of "boolean" numbers, where each number is
    preresented as -1 or 1.
  """

  as_int = tf.random_uniform(
      [batch_size, n], minval=0, maxval=2, dtype=tf.int32)
  expanded_range = (as_int * 2) - 1
  return tf.cast(expanded_range, tf.float32) 

def one_hot_encoding(labels, num_classes, scope=None):
  """Transform numeric labels into onehot_labels.

  Args:
    labels: [batch_size] target labels.
    num_classes: total number of classes.
    scope: Optional scope for name_scope.
  Returns:
    one hot encoding of the labels.
  """
  with tf.name_scope(scope, 'OneHotEncoding', [labels]):
    batch_size = labels.get_shape()[0]
    indices = tf.expand_dims(tf.range(0, batch_size), 1)
    labels = tf.cast(tf.expand_dims(labels, 1), indices.dtype)
    concated = tf.concat(axis=1, values=[indices, labels])
    onehot_labels = tf.sparse_to_dense(
        concated, tf.stack([batch_size, num_classes]), 1.0, 0.0)
    onehot_labels.set_shape([batch_size, num_classes])
    return onehot_labels 

def _reshape_instance_masks(self, keys_to_tensors):
    """Reshape instance segmentation masks.

    The instance segmentation masks are reshaped to [num_instances, height,
    width] and cast to boolean type to save memory.

    Args:
      keys_to_tensors: a dictionary from keys to tensors.

    Returns:
      A 3-D boolean tensor of shape [num_instances, height, width].
    """
    masks = keys_to_tensors['image/segmentation/object']
    if isinstance(masks, tf.SparseTensor):
      masks = tf.sparse_tensor_to_dense(masks)
    height = keys_to_tensors['image/height']
    width = keys_to_tensors['image/width']
    to_shape = tf.cast(tf.stack([-1, height, width]), tf.int32)

    return tf.cast(tf.reshape(masks, to_shape), tf.bool) 

def _create_regression_weights(self, match):
    """Set regression weight for each anchor.

    Only positive anchors are set to contribute to the regression loss, so this
    method returns a weight of 1 for every positive anchor and 0 for every
    negative anchor.

    Args:
      match: a matcher.Match object that provides a matching between anchors
        and groundtruth boxes.

    Returns:
      reg_weights: a float32 tensor with shape [num_anchors] representing
        regression weights
    """
    reg_weights = tf.cast(match.matched_column_indicator(), tf.float32)
    return reg_weights 

def scale(keypoints, y_scale, x_scale, scope=None):
  """Scales keypoint coordinates in x and y dimensions.

  Args:
    keypoints: a tensor of shape [num_instances, num_keypoints, 2]
    y_scale: (float) scalar tensor
    x_scale: (float) scalar tensor
    scope: name scope.

  Returns:
    new_keypoints: a tensor of shape [num_instances, num_keypoints, 2]
  """
  with tf.name_scope(scope, 'Scale'):
    y_scale = tf.cast(y_scale, tf.float32)
    x_scale = tf.cast(x_scale, tf.float32)
    new_keypoints = keypoints * [[[y_scale, x_scale]]]
    return new_keypoints 

def setup_graph(self, input_audio_batch, target_phrase): 
        batch_size = input_audio_batch.shape[0]
        weird = (input_audio_batch.shape[1] - 1) // 320 
        logits_arg2 = np.tile(weird, batch_size)
        dense_arg1 = np.array(np.tile(target_phrase, (batch_size, 1)), dtype=np.int32)
        dense_arg2 = np.array(np.tile(target_phrase.shape[0], batch_size), dtype=np.int32)
        
        pass_in = np.clip(input_audio_batch, -2**15, 2**15-1)
        seq_len = np.tile(weird, batch_size).astype(np.int32)
        
        with tf.variable_scope('', reuse=tf.AUTO_REUSE):
            
            inputs = tf.placeholder(tf.float32, shape=pass_in.shape, name='a')
            len_batch = tf.placeholder(tf.float32, name='b')
            arg2_logits = tf.placeholder(tf.int32, shape=logits_arg2.shape, name='c')
            arg1_dense = tf.placeholder(tf.float32, shape=dense_arg1.shape, name='d')
            arg2_dense = tf.placeholder(tf.int32, shape=dense_arg2.shape, name='e')
            len_seq = tf.placeholder(tf.int32, shape=seq_len.shape, name='f')
            
            logits = get_logits(inputs, arg2_logits)
            target = ctc_label_dense_to_sparse(arg1_dense, arg2_dense, len_batch)
            ctcloss = tf.nn.ctc_loss(labels=tf.cast(target, tf.int32), inputs=logits, sequence_length=len_seq)
            decoded, _ = tf.nn.ctc_greedy_decoder(logits, arg2_logits, merge_repeated=True)
            
            sess = tf.Session()
            saver = tf.train.Saver(tf.global_variables())
            saver.restore(sess, "models/session_dump")
            
        func1 = lambda a, b, c, d, e, f: sess.run(ctcloss, 
            feed_dict={inputs: a, len_batch: b, arg2_logits: c, arg1_dense: d, arg2_dense: e, len_seq: f})
        func2 = lambda a, b, c, d, e, f: sess.run([ctcloss, decoded], 
            feed_dict={inputs: a, len_batch: b, arg2_logits: c, arg1_dense: d, arg2_dense: e, len_seq: f})
        return (func1, func2) 

def compute_mfcc(audio, **kwargs):
    """
    Compute the MFCC for a given audio waveform. This is
    identical to how DeepSpeech does it, but does it all in
    TensorFlow so that we can differentiate through it.
    """

    batch_size, size = audio.get_shape().as_list()
    audio = tf.cast(audio, tf.float32)

    # 1. Pre-emphasizer, a high-pass filter
    audio = tf.concat((audio[:, :1], audio[:, 1:] - 0.97*audio[:, :-1], np.zeros((batch_size,1000),dtype=np.float32)), 1)

    # 2. windowing into frames of 320 samples, overlapping
    windowed = tf.stack([audio[:, i:i+400] for i in range(0,size-320,160)],1)

    # 3. Take the FFT to convert to frequency space
    ffted = tf.spectral.rfft(windowed, [512])
    ffted = 1.0 / 512 * tf.square(tf.abs(ffted))

    # 4. Compute the Mel windowing of the FFT
    energy = tf.reduce_sum(ffted,axis=2)+1e-30
    filters = np.load("filterbanks.npy").T
    feat = tf.matmul(ffted, np.array([filters]*batch_size,dtype=np.float32))+1e-30

    # 5. Take the DCT again, because why not
    feat = tf.log(feat)
    feat = tf.spectral.dct(feat, type=2, norm='ortho')[:,:,:26]

    # 6. Amplify high frequencies for some reason
    _,nframes,ncoeff = feat.get_shape().as_list()
    n = np.arange(ncoeff)
    lift = 1 + (22/2.)*np.sin(np.pi*n/22)
    feat = lift*feat
    width = feat.get_shape().as_list()[1]

    # 7. And now stick the energy next to the features
    feat = tf.concat((tf.reshape(tf.log(energy),(-1,width,1)), feat[:, :, 1:]), axis=2)
    
    return feat 

def create_attention_mask_from_input_mask(from_tensor, to_mask):
    """Create 3D attention mask from a 2D tensor mask.

    Args:
      from_tensor: 2D or 3D Tensor of shape [batch_size, from_seq_length, ...].
      to_mask: int32 Tensor of shape [batch_size, to_seq_length].

    Returns:
      float Tensor of shape [batch_size, from_seq_length, to_seq_length].
    """
    from_shape = get_shape_list(from_tensor, expected_rank=[2, 3])
    batch_size = from_shape[0]
    from_seq_length = from_shape[1]

    to_shape = get_shape_list(to_mask, expected_rank=2)
    to_seq_length = to_shape[1]

    to_mask = tf.cast(
        tf.reshape(to_mask, [batch_size, 1, to_seq_length]), tf.float32)

    # We don't assume that `from_tensor` is a mask (although it could be). We
    # don't actually care if we attend *from* padding tokens (only *to* padding)
    # tokens so we create a tensor of all ones.
    #
    # `broadcast_ones` = [batch_size, from_seq_length, 1]
    broadcast_ones = tf.ones(
        shape=[batch_size, from_seq_length, 1], dtype=tf.float32)

    # Here we broadcast along two dimensions to create the mask.
    mask = broadcast_ones * to_mask

    return mask 

def register_gradients(self, loss, vars):
        assert not self._updates_applied

        # Validate arguments.
        if isinstance(vars, dict):
            vars = list(vars.values()) # allow passing in Network.trainables as vars
        assert isinstance(vars, list) and len(vars) >= 1
        assert all(is_tf_expression(expr) for expr in vars + [loss])
        if self._grad_shapes is None:
            self._grad_shapes = [shape_to_list(var.shape) for var in vars]
        assert len(vars) == len(self._grad_shapes)
        assert all(shape_to_list(var.shape) == var_shape for var, var_shape in zip(vars, self._grad_shapes))
        dev = loss.device
        assert all(var.device == dev for var in vars)

        # Register device and compute gradients.
        with tf.name_scope(self.id + '_grad'), tf.device(dev):
            if dev not in self._dev_opt:
                opt_name = self.scope.replace('/', '_') + '_opt%d' % len(self._dev_opt)
                self._dev_opt[dev] = self.optimizer_class(name=opt_name, learning_rate=self.learning_rate, **self.optimizer_kwargs)
                self._dev_grads[dev] = []
            loss = self.apply_loss_scaling(tf.cast(loss, tf.float32))
            grads = self._dev_opt[dev].compute_gradients(loss, vars, gate_gradients=tf.train.Optimizer.GATE_NONE) # disable gating to reduce memory usage
            grads = [(g, v) if g is not None else (tf.zeros_like(v), v) for g, v in grads] # replace disconnected gradients with zeros
            self._dev_grads[dev].append(grads)

    # Construct training op to update the registered variables based on their gradients. 

def fp32(*values):
    if len(values) == 1 and isinstance(values[0], tuple):
        values = values[0]
    values = tuple(tf.cast(v, tf.float32) for v in values)
    return values if len(values) >= 2 else values[0]

#----------------------------------------------------------------------------
# Extract (ws x ws) patches from a tensor 

def dense(x, fmaps, gain=np.sqrt(2), use_wscale=False):
    if len(x.shape) > 2:
        x = tf.reshape(x, [-1, np.prod([d.value for d in x.shape[1:]])])
    w = get_weight([x.shape[1].value, fmaps], gain=gain, use_wscale=use_wscale)
    w = tf.cast(w, x.dtype)
    return tf.matmul(x, w)

#----------------------------------------------------------------------------
# Convolutional layer. 

def apply_bias(x):
    b = tf.get_variable('bias', shape=[x.shape[1]], initializer=tf.initializers.zeros())
    b = tf.cast(b, x.dtype)
    if len(x.shape) == 2:
        return x + b
    else:
        return x + tf.reshape(b, [1, -1, 1, 1])

#----------------------------------------------------------------------------
# Leaky ReLU activation. Same as tf.nn.leaky_relu, but supports FP16. 

def upscale2d_conv2d(x, fmaps, kernel, gain=np.sqrt(2), use_wscale=False):
    assert kernel >= 1 and kernel % 2 == 1
    w = get_weight([kernel, kernel, fmaps, x.shape[1].value], gain=gain, use_wscale=use_wscale, fan_in=(kernel**2)*x.shape[1].value)
    w = tf.pad(w, [[1,1], [1,1], [0,0], [0,0]], mode='CONSTANT')
    w = tf.add_n([w[1:, 1:], w[:-1, 1:], w[1:, :-1], w[:-1, :-1]])
    w = tf.cast(w, x.dtype)
    os = [tf.shape(x)[0], fmaps, x.shape[2] * 2, x.shape[3] * 2]
    return tf.nn.conv2d_transpose(x, w, os, strides=[1,1,2,2], padding='SAME', data_format='NCHW')

#----------------------------------------------------------------------------
# Box filter downscaling layer. 

def conv2d_downscale2d(x, fmaps, kernel, gain=np.sqrt(2), use_wscale=False):
    assert kernel >= 1 and kernel % 2 == 1
    w = get_weight([kernel, kernel, x.shape[1].value, fmaps], gain=gain, use_wscale=use_wscale)
    w = tf.pad(w, [[1,1], [1,1], [0,0], [0,0]], mode='CONSTANT')
    w = tf.add_n([w[1:, 1:], w[:-1, 1:], w[1:, :-1], w[:-1, :-1]]) * 0.25
    w = tf.cast(w, x.dtype)
    return tf.nn.conv2d(x, w, strides=[1,1,2,2], padding='SAME', data_format='NCHW')

#----------------------------------------------------------------------------
# Pixelwise feature vector normalization. 

def process_reals(x, lod, mirror_augment, drange_data, drange_net):
    with tf.name_scope('ProcessReals'):
        with tf.name_scope('DynamicRange'):
            x = tf.cast(x, tf.float32)
            x = misc.adjust_dynamic_range(x, drange_data, drange_net)
        if mirror_augment:
            with tf.name_scope('MirrorAugment'):
                s = tf.shape(x)
                mask = tf.random_uniform([s[0], 1, 1, 1], 0.0, 1.0)
                mask = tf.tile(mask, [1, s[1], s[2], s[3]])
                x = tf.where(mask < 0.5, x, tf.reverse(x, axis=[3]))
        with tf.name_scope('FadeLOD'): # Smooth crossfade between consecutive levels-of-detail.
            s = tf.shape(x)
            y = tf.reshape(x, [-1, s[1], s[2]//2, 2, s[3]//2, 2])
            y = tf.reduce_mean(y, axis=[3, 5], keep_dims=True)
            y = tf.tile(y, [1, 1, 1, 2, 1, 2])
            y = tf.reshape(y, [-1, s[1], s[2], s[3]])
            x = tfutil.lerp(x, y, lod - tf.floor(lod))
        with tf.name_scope('UpscaleLOD'): # Upscale to match the expected input/output size of the networks.
            s = tf.shape(x)
            factor = tf.cast(2 ** tf.floor(lod), tf.int32)
            x = tf.reshape(x, [-1, s[1], s[2], 1, s[3], 1])
            x = tf.tile(x, [1, 1, 1, factor, 1, factor])
            x = tf.reshape(x, [-1, s[1], s[2] * factor, s[3] * factor])
        return x

#----------------------------------------------------------------------------
# Just-in-time processing of masks before feeding them to the networks. 

def get_minibatch_tf(self): # => images, labels
        with tf.name_scope('SyntheticDataset'):
            shrink = tf.cast(2.0 ** tf.cast(self._tf_lod_var, tf.float32), tf.int32)
            shape = [self.shape[0], self.shape[1] // shrink, self.shape[2] // shrink]
            images = self._generate_images(self._tf_minibatch_var, self._tf_lod_var, shape)
            labels = self._generate_labels(self._tf_minibatch_var)
            return images, labels 

def _top_1_accuracy(logits, labels):
    return tf.reduce_mean(
        tf.cast(tf.nn.in_top_k(logits, labels, 1), tf.float32))