Python model_utils.FramePooling() Examples

def forward(self, reshaped_input):

        feature_size = self.feature_size
        cluster_size = self.cluster_size
        add_batch_norm = self.add_batch_norm
        max_frames = self.max_frames
        is_training = self.is_training

        cluster_weights = tf.get_variable("cluster_weights",
          [feature_size, cluster_size],
          initializer = tf.random_normal_initializer(stddev=1 / math.sqrt(feature_size)))
        
        tf.summary.histogram("cluster_weights", cluster_weights)
        activation = tf.matmul(reshaped_input, cluster_weights)
        
        if add_batch_norm:
          activation = slim.batch_norm(
              activation,
              center=True,
              scale=True,
              is_training=is_training,
              scope="cluster_bn")
        else:
          cluster_biases = tf.get_variable("cluster_biases",
            [cluster_size],
            initializer = tf.random_normal(stddev=1 / math.sqrt(feature_size)))
          tf.summary.histogram("cluster_biases", cluster_biases)
          activation += cluster_biases

        if activation == 'glu':
            space_ind = range(cluster_size/2)
            gate_ind = range(cluster_size/2,cluster_size)

            gates = tf.sigmoid(activation[:,gate_ind])
            activation = tf.multiply(activation[:,space_ind],gates)

        elif activation == 'relu':
            activation = tf.nn.relu6(activation)
        
        tf.summary.histogram("cluster_output", activation)

        activation = tf.reshape(activation, [-1, max_frames, cluster_size])

        avg_activation = utils.FramePooling(activation, 'average')
        avg_activation = tf.nn.l2_normalize(avg_activation,1)

        max_activation = utils.FramePooling(activation, 'max')
        max_activation = tf.nn.l2_normalize(max_activation,1)
        
        return tf.concat([avg_activation,max_activation],1) 

def create_model(self, model_input, vocab_size, num_frames,
                   l2_penalty=1e-4, **unused_params):
    """Creates a model which uses a logistic classifier over the average of the
    frame-level features.

    This class is intended to be an example for implementors of frame level
    models. If you want to train a model over averaged features it is more
    efficient to average them beforehand rather than on the fly.

    Args:
      model_input: A 'batch_size' x 'max_frames' x 'num_features' matrix of
                   input features.
      vocab_size: The number of classes in the dataset.
      num_frames: A vector of length 'batch' which indicates the number of
           frames for each video (before padding).

    Returns:
      A dictionary with a tensor containing the probability predictions of the
      model in the 'predictions' key. The dimensions of the tensor are
      'batch_size' x 'num_classes'.
    """
#    num_frames = tf.cast(tf.expand_dims(num_frames, 1), tf.float32)
#    feature_size = model_input.get_shape().as_list()[2]
#        
#    logging.info('model_input shape: {}'.format(
#            model_input.get_shape().as_list()))
#
#    denominators = tf.reshape(
#        tf.tile(num_frames, [1, feature_size]), [-1, feature_size])
#    avg_pooled = tf.reduce_sum(model_input, axis=[1]) / denominators
    
    avg_pooled = utils.FramePooling(model_input, 'average')
    
    logging.info( 'avg_pooled shape: {}'.format(
            avg_pooled.get_shape().as_list() )) 
    
    aggregated_model = getattr(video_level_models,
                               FLAGS.video_level_classifier_model)
    return aggregated_model().create_model(
        model_input=avg_pooled,
        vocab_size=vocab_size,
        num_mixtures=2,
        **unused_params) 

def create_model(self, model_input, vocab_size, num_frames,
                   l2_penalty=1e-4, **unused_params):
    """Creates a model which uses a logistic classifier over the average of the
    frame-level features.

    This class is intended to be an example for implementors of frame level
    models. If you want to train a model over averaged features it is more
    efficient to average them beforehand rather than on the fly.

    Args:
      model_input: A 'batch_size' x 'max_frames' x 'num_features' matrix of
                   input features.
      vocab_size: The number of classes in the dataset.
      num_frames: A vector of length 'batch' which indicates the number of
           frames for each video (before padding).

    Returns:
      A dictionary with a tensor containing the probability predictions of the
      model in the 'predictions' key. The dimensions of the tensor are
      'batch_size' x 'num_classes'.
    """
#    num_frames = tf.cast(tf.expand_dims(num_frames, 1), tf.float32)
#    feature_size = model_input.get_shape().as_list()[2]
#        
#    logging.info('model_input shape: {}'.format(
#            model_input.get_shape().as_list()))
#
#    denominators = tf.reshape(
#        tf.tile(num_frames, [1, feature_size]), [-1, feature_size])
#    avg_pooled = tf.reduce_sum(model_input, axis=[1]) / denominators
    
    avg_pooled = utils.FramePooling(model_input, 'average')
    
    logging.info( 'avg_pooled shape: {}'.format(
            avg_pooled.get_shape().as_list() )) 
    
    aggregated_model = getattr(video_level_models,
                               FLAGS.video_level_classifier_model)
    return aggregated_model().create_model(
        model_input=avg_pooled,
        vocab_size=vocab_size,
        num_mixtures=2,
        **unused_params)