mamkit.modules package#

Submodules#

mamkit.modules.rnn module#

class mamkit.modules.rnn.LSTMStack(input_size, lstm_weigths, return_hidden=True)#

Bases: Module

forward(x)#

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool#

mamkit.modules.transformer module#

class mamkit.modules.transformer.CustomEncoder(d_model, ffn_hidden, n_head, n_layers, drop_prob)#

Bases: Module

Encoder Class

forward(embedding, text_mask, audio_mask)#
Parameters:

  • embedding – input tensor

  • text_mask – mask for text sequence

  • audio_mask – mask for audio sequence

training: bool#
class mamkit.modules.transformer.CustomEncoderLayer(d_model, ffn_hidden, n_head, drop_prob)#

Bases: Module

Encoder Layer Class

forward(x, text_mask, audio_mask)#
Parameters:

  • x – input tensor

  • text_mask – mask for text sequence

  • audio_mask – mask for audio sequence

training: bool#
class mamkit.modules.transformer.CustomMultiHeadAttention(d_model, n_head)#

Bases: Module

Multi Head Attention Class for Transformer

concat(tensor)#

inverse function of self.split(tensor : torch.Tensor) :param tensor: [batch_size, head, length, d_tensor]

forward(q, k, v, text_mask, audio_mask)#
Parameters:

  • q – query (decoder)

  • k – key (encoder)

  • v – value (encoder)

  • text_mask – mask for text sequence

  • audio_mask – mask for audio sequence

split(tensor)#

split tensor by number of head

Parameters:

tensor – [batch_size, length, d_model]

training: bool#
class mamkit.modules.transformer.CustomScaleDotProductAttention#

Bases: Module

compute scale dot product attention

Query : given sentence that we focused on (decoder) Key : every sentence to check relationship with Qeury(encoder) Value : every sentence same with Key (encoder)

forward(q, k, v, text_mask, audio_mask, e=1e-12)#
Parameters:

  • q – query (decoder)

  • k – key (encoder)

  • v – value (encoder)

  • text_mask – mask for text sequence

  • audio_mask – mask for audio sequence

  • e – epsilon value for masking

training: bool#
class mamkit.modules.transformer.LayerNorm(d_model, eps=1e-12)#

Bases: Module

Layer Normalization Class

forward(x)#
Parameters:

x – input tensor

training: bool#
class mamkit.modules.transformer.MulTA_CrossAttentionBlock(embedding_dim, d_ffn, num_heads=4, dropout_prob=0.1)#

Bases: Module

Class for the cross modal attention block

forward(elem_a, elem_b, attn_mask)#

Forward pass of the model :param elem_a: elements of the modality A :param elem_b: elements of the modality B :param attn_mask: attention mask to use

training: bool#
class mamkit.modules.transformer.PositionalEncoding(d_model, dual_modality=False, dropout=0.1, max_len=5000)#

Bases: Module

Positional Encoding for Transformer

forward(x, is_first=True)#
Parameters:

  • x – input tensor (bs, sqlen, emb)

  • is_first – True if the first modality, False if the second modality

training: bool#
class mamkit.modules.transformer.PositionwiseFeedForward(d_model, hidden, drop_prob=0.1)#

Bases: Module

Position-wise Feed Forward Layer

forward(x)#
Parameters:

x – input tensor

training: bool#

Module contents#