def __init__(self, dict_args):
super(SequenceEncoder, self).__init__()
self.configuration = dict_args['encoder_configuration']
self.dropout_rate = dict_args['encoder_dropout_rate']
self.input_dim = dict_args['encoder_input_dim']
self.contextvector = None
self.inputvector = None
self.recurrent, self.attention = None, None
self.dropout_layer = nn.Dropout(p=self.dropout_rate)
#self.linear = nn.Linear(self.input_dim, self.input_dim)
if self.configuration == 'MP':
self.recurrent, self.attention = False, False
elif self.configuration == 'LSTM':
self.recurrent, self.attention = True, False
elif self.configuration == 'MPAttn':
self.recurrent, self.attention = False, True
elif self.configuration == 'LSTMAttn':
self.recurrent, self.attention = True, True
if self.recurrent:
self.hidden_dim = dict_args['encoder_rnn_hdim']
self.rnn_type = dict_args['encoder_rnn_type']
self.num_layers = dict_args['encoder_num_layers']
recurrent_layer_args = {
'input_dim': self.input_dim,
'rnn_hdim': self.hidden_dim,
'rnn_type': self.rnn_type,
'num_layers': self.num_layers,
'dropout_rate': self.dropout_rate
}
self.recurrent_layer = BiDirEncoder(recurrent_layer_args)
if self.attention:
self.projection_dim = dict_args['encoderattn_projection_dim']
if self.configuration == 'LSTMAttn':
self.context_dim = self.hidden_dim
elif self.configuration == 'MPAttn':
self.context_dim = self.input_dim
self.query_dim = dict_args['encoderattn_query_dim']
similarity_function_args = {
'sequence1_dim': self.context_dim,
'sequence2_dim': self.query_dim,
'projection_dim': self.projection_dim
}
self.similarity_function = similarity.LinearProjectionSimilarity(
similarity_function_args)
self.attention_function = UniDirAttention(
{'similarity_function': self.similarity_function})
def __init__(self, dict_args):
super(UniDirAttention, self).__init__()
self.similarity_function_name = dict_args['similarity_function']
#dict_args should contain the arguments for the similarity function as well
self.similarity_function = None
if self.similarity_function_name == 'DotProduct':
self.similarity_function = similarity.DotProductSimilarity(
dict_args)
elif self.similarity_function_name == 'WeightedInputsConcatenation':
self.similarity_function = similarity.LinearConcatenationSimilarity(
dict_args)
elif self.similarity_function_name == 'WeightedInputsDotConcatenation':
self.similarity_function = similarity.LinearConcatenationDotSimilarity(
dict_args)
elif self.similarity_function_name == 'WeightedSumProjection':
self.similarity_function = similarity.LinearProjectionSimilarity(
dict_args)
elif self.similarity_function_name == 'ProjectionSimilaritySharedWeights':
self.similarity_function = dict_args['similarity_function_pointer']
def __init__(self, dict_args):
super(BiDirAttention, self).__init__()
self.similarity_function_name = dict_args['similarity_function']
#attends a sequence(instead of word) over an other sequence
self.one_shot_attention = dict_args['one_shot_attention']
self.self_match_attention = dict_args['self_match_attention']
self.similarity_function = None
if self.similarity_function_name == 'DotProduct':
self.similarity_function = similarity.DotProductSimilarity(
dict_args)
elif self.similarity_function_name == 'WeightedInputsConcatenation':
self.similarity_function = similarity.LinearConcatenationSimilarity(
dict_args)
elif self.similarity_function_name == 'WeightedInputsDotConcatenation':
self.similarity_function = similarity.LinearConcatenationDotSimilarity(
dict_args)
elif self.similarity_function_name == 'WeightedSumProjection':
self.similarity_function = similarity.LinearProjectionSimilarity(
dict_args)
elif self.similarity_function_name == 'ProjectionSimilaritySharedWeights':
self.similarity_function = dict_args['similarity_function_pointer']
def __init__(self, dict_args):
super(VideoFrameEncoder, self).__init__()
self.configuration = dict_args['encoder_configuration']
self.dropout_rate = dict_args['encoder_dropout_rate']
self.channel_dim = dict_args['frame_channel_dim']
self.spatial_dim = dict_args['frame_spatial_dim']
self.projection_dim = dict_args['encoderattn_projection_dim']
self.use_linear = dict_args["encoder_linear"]
if self.use_linear:
self.channelred_dim = dict_args["frame_channelred_dim"]
self.linear = nn.Linear(self.channel_dim, self.channelred_dim)
self.channel_dim = self.channelred_dim
self.recurrent = None
self.spatialattention = None
self.temporalattention = None
self.dropout_layer = nn.Dropout(p=self.dropout_rate)
if self.configuration == 'LSTMTrackSpatial':
self.recurrent, self.spatialattention, self.temporalattention = True, True, False
elif self.configuration == 'LSTMTrackSpatialTemporal':
self.recurrent, self.spatialattention, self.temporalattention = True, True, True
elif self.configuration == 'SpatialTemporal':
self.recurrent, self.spatialattention, self.temporalattention = False, True, True
if self.recurrent:
self.hidden_dim = dict_args['encoder_rnn_hdim']
self.rnn_type = dict_args['encoder_rnn_type']
if self.rnn_type == 'LSTM':
self.trackrnn = nn.LSTMCell(self.channel_dim, self.hidden_dim)
elif self.rnn_type == 'GRU':
self.trackrnn = nn.GRUCell(self.channel_dim, self.hidden_dim)
elif self.rnn_type == 'RNN':
pass
if self.spatialattention:
self.query_dim = dict_args['encoderattn_query_dim']
spatial_similarity_function_args = {
'sequence1_dim' : self.channel_dim,
'sequence2_dim' : self.query_dim,
'projection_dim' : self.projection_dim
}
self.spatial_similarity_function = similarity.LinearProjectionSimilarity(spatial_similarity_function_args)
self.spatial_attention_function = UniDirAttention({'similarity_function': self.spatial_similarity_function})
if self.temporalattention:
self.query_dim = dict_args['encoderattn_query_dim']
self.context_dim = self.channel_dim
if self.recurrent: self.context_dim = self.hidden_dim
temporal_similarity_function_args = {
'sequence1_dim' : self.context_dim,
'sequence2_dim' : self.query_dim,
'projection_dim' : self.projection_dim
}
self.temporal_similarity_function = similarity.LinearProjectionSimilarity(temporal_similarity_function_args)
self.temporal_attention_function = UniDirAttention({'similarity_function': self.temporal_similarity_function})