Added documentation for ChemNet layers, Updated directory for Chembl25 loader

class Stem(tf.keras.layers.Layer):

  """

  Stem Layer as defined in https://arxiv.org/abs/1710.02238. The structure is

  significantly altered from the original Inception-ResNet architecture,

  (https://arxiv.org/abs/1602.07261) but the idea behind this layer is to

  downsample the image as a preprocessing step for the Inception-ResNet layers,

  and reduce computational complexity.

  """

  def __init__(self, num_filters, **kwargs):

    """

    self._layers = [self.conv_layer, self.activation_layer]

  def call(self, inputs):

    """Invoked when __call__ method of the layer is used."""

    conv1 = self.conv_layer(inputs)

    return self.activation_layer(conv1)

class InceptionResnetA(tf.keras.layers.Layer):

  """

  Variant A of the three InceptionResNet layers described in

  https://arxiv.org/abs/1710.02238. All variants use multiple

  convolutional blocks with varying kernel sizes and number of filters. This

  allows capturing patterns over different scales in the inputs. Residual

  connections are additionally used and have been shown previously to improve

  convergence and training in deep networks. A 1x1 convolution is used on the

  concatenated feature maps from the different convolutional blocks, to ensure

  shapes of inputs and feature maps are same for the residual connection.

  """

  def __init__(self, num_filters, input_dim, **kwargs):

    """

        [self.concat_layer, self.add_layer, self.activation_layer])

  def call(self, inputs):

    """Invoked when __call__ method of the layer is used."""

    conv1 = inputs

    for layer in self.conv_block1:

      conv1 = layer(conv1)

class InceptionResnetB(tf.keras.layers.Layer):

  """

  Variant B of the three InceptionResNet layers described in

  https://arxiv.org/abs/1710.02238. All variants use multiple

  convolutional blocks with varying kernel sizes and number of filters. This

  allows capturing patterns over different scales in the inputs. Residual

  connections are additionally used and have been shown previously to improve

  convergence and training in deep networks. A 1x1 convolution is used on the

  concatenated feature maps from the different convolutional blocks, to ensure

  shapes of inputs and feature maps are same for the residual connection.

  """

  def __init__(self, num_filters, input_dim, **kwargs):

    """

        [self.concat_layer, self.add_layer, self.activation_layer])

  def call(self, inputs):

    """Invoked when __call__ method of the layer is used."""

    conv1 = inputs

    for layer in self.conv_block1:

      conv1 = layer(conv1)

class InceptionResnetC(tf.keras.layers.Layer):

  """

  Variant C of the three InceptionResNet layers described in

  https://arxiv.org/abs/1710.02238. All variants use multiple

  convolutional blocks with varying kernel sizes and number of filters. This

  allows capturing patterns over different scales in the inputs. Residual

  connections are additionally used and have been shown previously to improve

  convergence and training in deep networks. A 1x1 convolution is used on the

  concatenated feature maps from the different convolutional blocks, to ensure

  shapes of inputs and feature maps are same for the residual connection.

  """

  def __init__(self, num_filters, input_dim, **kwargs):

    """

        [self.concat_layer, self.add_layer, self.activation_layer])

  def call(self, inputs):

    """Invoked when __call__ method of the layer is used."""

    conv1 = inputs

    for layer in self.conv_block1:

      conv1 = layer(conv1)

class ReductionA(tf.keras.layers.Layer):

  """

  Variant A of the two Reduction layers described in

  https://arxiv.org/abs/1710.02238. All variants use multiple convolutional

  blocks with varying kernel sizes and number of filters, to reduce the spatial

  extent of the image and reduce computational complexity for downstream layers.

  """

  def __init__(self, num_filters, **kwargs):

    """

        [self.max_pool1, self.concat_layer, self.activation_layer])

  def call(self, inputs):

    """Invoked when __call__ method of the layer is used."""

    maxpool1 = self.max_pool1(inputs)

    conv1 = inputs

    for layer in self.conv_block1:

class ReductionB(tf.keras.layers.Layer):

  """

  Variant B of the two Reduction layers described in

  https://arxiv.org/abs/1710.02238. All variants use multiple convolutional

  blocks with varying kernel sizes and number of filters, to reduce the spatial

  extent of the image and reduce computational complexity for downstream layers.

  """

  def __init__(self, num_filters, **kwargs):

    """

        [self.max_pool1, self.concat_layer, self.activation_layer])

  def call(self, inputs):

    """Invoked when __call__ method of the layer is used."""

    maxpool1 = self.max_pool1(inputs)

    conv1 = inputs

    for layer in self.conv_block1:

    save_dir = DEFAULT_DIR

  save_folder = os.path.join(save_dir, "chembl_25-featurized", str(featurizer))

  if featurizer == "smiles2img":

    img_spec = kwargs.get("img_spec", "engd")

    save_folder = os.path.join(save_folder, img_spec)

  if reload:

    if not os.path.exists(save_folder):

  elif featurizer == "smiles2img":

    img_size = kwargs.get("img_size", 80)

    img_mode = kwargs.get("img_mode", "engd")

    img_spec = kwargs.get("img_spec", "engd")

    res = kwargs.get("res", 0.5)

    featurizer = SmilesToImage(img_size=img_size, img_mode=img_mode, res=res)

    featurizer = SmilesToImage(img_size=img_size, img_spec=img_spec, res=res)

  else:

    raise ValueError(