conformer_u2

the transformer model

Module Contents

Classes

ConformerU2

A transformer model. User is able to modify the attributes as needed.

ConformerU2Encoder

TransformerEncoder is a stack of N encoder layers

ConformerU2Decoder

TransformerDecoder is a stack of N decoder layers
class conformer_u2.ConformerU2(d_model=512, nhead=8, cnn_module_kernel=15, num_encoder_layers=6, num_decoder_layers=6, dim_feedforward=2048, dropout=0.1, activation='gelu', unidirectional=False, look_ahead=0, custom_encoder=None, custom_decoder=None, static_chunk_size: int = 0, use_dynamic_chunk: bool = False, use_dynamic_left_chunk: bool = False)

Bases: athena.layers.transformer.Transformer

A transformer model. User is able to modify the attributes as needed.

Parameters

  • d_model – the number of expected features in the encoder/decoder inputs (default=512).

  • nhead – the number of heads in the multiheadattention models (default=8).

  • num_encoder_layers – the number of sub-encoder-layers in the encoder (default=6).

  • num_decoder_layers – the number of sub-decoder-layers in the decoder (default=6).

  • dim_feedforward – the dimension of the feedforward network model (default=2048).

  • dropout – the dropout value (default=0.1).

  • activation – the activation function of encoder/decoder intermediate layer, relu or gelu (default=relu).

  • custom_encoder – custom encoder (default=None).

  • custom_decoder – custom decoder (default=None).

Examples

>>> transformer_model = Transformer(nhead=16, num_encoder_layers=12)
>>> src = tf.random.normal((10, 32, 512))
>>> tgt = tf.random.normal((20, 32, 512))
>>> out = transformer_model(src, tgt)
call(src, tgt, src_mask=None, tgt_mask=None, memory_mask=None, return_encoder_output=False, return_attention_weights=False, training=None)

Take in and process masked source/target sequences.

Parameters

  • src – the sequence to the encoder (required).

  • tgt – the sequence to the decoder (required).

  • src_mask – the additive mask for the src sequence (optional).

  • tgt_mask – the additive mask for the tgt sequence (optional).

  • memory_mask – the additive mask for the encoder output (optional).

  • src_key_padding_mask – the ByteTensor mask for src keys per batch (optional).

  • tgt_key_padding_mask – the ByteTensor mask for tgt keys per batch (optional).

  • memory_key_padding_mask – the ByteTensor mask for memory keys per batch (optional).

Shape:

  • src: \((N, S, E)\).

  • tgt: \((N, T, E)\).

  • src_mask: \((N, S)\).

  • tgt_mask: \((N, T)\).

  • memory_mask: \((N, S)\).

Note: [src/tgt/memory]_mask should be a ByteTensor where True values are positions that should be masked with float(‘-inf’) and False values will be unchanged. This mask ensures that no information will be taken from position i if it is masked, and has a separate mask for each sequence in a batch.

  • output: \((N, T, E)\).

Note: Due to the multi-head attention architecture in the transformer model, the output sequence length of a transformer is same as the input sequence (i.e. target) length of the decode.

where S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number

Examples

>>> output = transformer_model(src, tgt, src_mask=src_mask, tgt_mask=tgt_mask)
class conformer_u2.ConformerU2Encoder(encoder_layers, static_chunk_size: int = 0, use_dynamic_chunk: bool = False, use_dynamic_left_chunk: bool = False)

Bases: tensorflow.keras.layers.Layer

TransformerEncoder is a stack of N encoder layers

Parameters

  • encoder_layer – an instance of the TransformerEncoderLayer() class (required).

  • num_layers – the number of sub-encoder-layers in the encoder (required).

  • norm – the layer normalization component (optional).

Examples

>>> encoder_layer = [ConformerU2EncoderLayer(d_model=512, nhead=8)
>>>                    for _ in range(num_layers)]
>>> transformer_encoder = ConformerU2Encoder(encoder_layer)
>>> src = tf.random.gamma(10, 32, 512)
>>> out = transformer_encoder(src)
call(src, src_mask=None, mask_pad=None, training=None)

Pass the input through the endocder layers in turn.

Parameters

  • src – the sequnce to the encoder (required).

  • mask – the mask for the src sequence (optional).

forward_chunk(xs: tensorflow.Tensor, offset: int, required_cache_size: int, subsampling_cache: Optional[tensorflow.Tensor] = None, elayers_output_cache: Optional[List[tensorflow.Tensor]] = None, conformer_cnn_cache: Optional[List[tensorflow.Tensor]] = None) → Tuple[tensorflow.Tensor, tensorflow.Tensor, List[tensorflow.Tensor], List[tensorflow.Tensor]]

Forward just one chunk

Parameters

  • xs (torch.Tensor) – chunk input

  • offset (int) – current offset in encoder output time stamp

  • required_cache_size (int) – cache size required for next chunk compuation >=0: actual cache size <0: means all history cache is required

  • subsampling_cache (Optional[torch.Tensor]) – subsampling cache

  • elayers_output_cache (Optional[List[torch.Tensor]]) – transformer/conformer encoder layers output cache

  • conformer_cnn_cache (Optional[List[torch.Tensor]]) – conformer cnn cache

Returns

output of current input xs torch.Tensor: subsampling cache required for next chunk computation List[torch.Tensor]: encoder layers output cache required for next

chunk computation

List[torch.Tensor]: conformer cnn cache

Return type

torch.Tensor

forward_chunk_by_chunk(xs: tensorflow.Tensor, decoding_chunk_size: int, num_decoding_left_chunks: int = -1) → Tuple[tensorflow.Tensor, tensorflow.Tensor]
Forward input chunk by chunk with chunk_size like a streaming

fashion

Here we should pay special attention to computation cache in the streaming style forward chunk by chunk. Three things should be taken into account for computation in the current network:

  1. transformer/conformer encoder layers output cache

  2. convolution in conformer

  3. convolution in subsampling

However, we don’t implement subsampling cache for:

  1. We can control subsampling module to output the right result by overlapping input instead of cache left context, even though it wastes some computation, but subsampling only takes a very small fraction of computation in the whole model.

  2. Typically, there are several covolution layers with subsampling in subsampling module, it is tricky and complicated to do cache with different convolution layers with different subsampling rate.

  3. Currently, nn.Sequential is used to stack all the convolution layers in subsampling, we need to rewrite it to make it work with cache, which is not prefered.

Parameters

  • xs (torch.Tensor) – (1, max_len, dim)

  • chunk_size (int) – decoding chunk size

set_unidirectional(uni=False)

whether to apply trianglar masks to make transformer unidirectional

class conformer_u2.ConformerU2Decoder(decoder_layers)

Bases: tensorflow.keras.layers.Layer

TransformerDecoder is a stack of N decoder layers

Parameters

  • decoder_layer – an instance of the TransformerDecoderLayer() class (required).

  • num_layers – the number of sub-decoder-layers in the decoder (required).

  • norm – the layer normalization component (optional).

Examples

>>> decoder_layer = [ConformerU2DecoderLayer(d_model=512, nhead=8)
>>>                     for _ in range(num_layers)]
>>> transformer_decoder = ConformerU2Decoder(decoder_layer)
>>> memory = tf.random.gamma(10, 32, 512)
>>> tgt = tf.random.gamma(20, 32, 512)
>>> out = transformer_decoder(tgt, memory)
call(tgt, memory, tgt_mask=None, memory_mask=None, return_attention_weights=False, training=None)

Pass the inputs (and mask) through the decoder layer in turn.

Parameters

  • tgt – the sequence to the decoder (required).

  • memory – the sequnce from the last layer of the encoder (required).

  • tgt_mask – the mask for the tgt sequence (optional).

  • memory_mask – the mask for the memory sequence (optional).