def __init__(self, d_model, d_ff, n_heads, kernel_size,
dropout, dropout_att, dropout_layer,
layer_norm_eps, ffn_activation, param_init,
ffn_bottleneck_dim=0):
super(ConformerEncoderBlock, self).__init__()
self.n_heads = n_heads
self.fc_factor = 0.5
# first half position-wise feed-forward
self.norm1 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.feed_forward1 = FFN(d_model, d_ff, dropout, ffn_activation, param_init,
ffn_bottleneck_dim)
# conv module
self.norm2 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.conv = ConformerConvBlock(d_model, kernel_size, param_init)
# self-attention
self.norm3 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.self_attn = RelMHA(kdim=d_model,
qdim=d_model,
adim=d_model,
odim=d_model,
n_heads=n_heads,
dropout=dropout_att,
param_init=param_init)
# second half position-wise feed-forward
self.norm4 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.feed_forward2 = FFN(d_model, d_ff, dropout, ffn_activation, param_init,
ffn_bottleneck_dim)
self.dropout = nn.Dropout(dropout)
self.dropout_layer = dropout_layer
def __init__(self,
d_model,
d_ff,
n_heads,
kernel_size,
dropout,
dropout_att,
dropout_layer,
layer_norm_eps,
ffn_activation,
param_init,
pe_type,
clamp_len,
ffn_bottleneck_dim,
unidirectional,
normalization='layer_norm'):
super(ConformerEncoderBlock, self).__init__()
self.n_heads = n_heads
self.fc_factor = 0.5
# first half position-wise feed-forward
self.norm1 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.feed_forward_macaron = FFN(d_model, d_ff, dropout, ffn_activation,
param_init, ffn_bottleneck_dim)
# self-attention
self.norm2 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.self_attn = RelMHA(kdim=d_model,
qdim=d_model,
adim=d_model,
odim=d_model,
n_heads=n_heads,
dropout=dropout_att,
param_init=param_init,
xl_like=pe_type == 'relative_xl',
clamp_len=clamp_len)
# conv module
self.norm3 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.conv = ConformerConvBlock(d_model,
kernel_size,
param_init,
normalization,
causal=unidirectional)
self.conv_context = kernel_size
# second half position-wise feed-forward
self.norm4 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.feed_forward = FFN(d_model, d_ff, dropout, ffn_activation,
param_init, ffn_bottleneck_dim)
self.norm5 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.dropout = nn.Dropout(dropout)
self.dropout_layer = dropout_layer # probability to skip
logger.info('Stochastic depth prob: %.3f' % dropout_layer)
self.reset_visualization()
def __init__(self, d_model, d_ff, n_heads, kernel_size,
dropout, dropout_att, dropout_layer,
layer_norm_eps, ffn_activation, param_init,
pe_type, clamp_len, ffn_bottleneck_dim, unidirectional,
normalization='batch_norm'):
super(ConformerEncoderBlock_v2, self).__init__()
self.n_heads = n_heads
self.fc_factor = 0.5
# first half position-wise feed-forward
self.norm1 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.feed_forward_macaron = FFN(d_model, d_ff, dropout, ffn_activation, param_init,
ffn_bottleneck_dim)
# conv module
self.norm2 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.conv = ConformerConvBlock(d_model, kernel_size, param_init, normalization,
causal=unidirectional)
self.conv_context = kernel_size
# self-attention
self.norm3 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.self_attn = MHA(kdim=d_model,
qdim=d_model,
adim=d_model,
odim=d_model,
n_heads=n_heads,
dropout=dropout_att,
param_init=param_init)
# second half position-wise feed-forward
self.norm4 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.feed_forward = FFN(d_model, d_ff, dropout, ffn_activation, param_init,
ffn_bottleneck_dim)
self.norm5 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.dropout = nn.Dropout(dropout)
self.dropout_layer = dropout_layer
self.reset_visualization()