relogic/pretrainkit/models/relationalsemparse/modeling_relational_bart.py [418:632]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
class DecoderLayer(nn.Module):
    def __init__(self, config: BartConfig):
        super().__init__()
        self.embed_dim = config.d_model
        self.output_attentions = config.output_attentions
        self.self_attn = SelfAttention(
            embed_dim=self.embed_dim, num_heads=config.decoder_attention_heads, dropout=config.attention_dropout,
        )
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout
        self.normalize_before = config.normalize_before

        self.self_attn_layer_norm = LayerNorm(self.embed_dim)
        self.encoder_attn = SelfAttention(
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            encoder_decoder_attention=True,
        )
        self.encoder_attn_layer_norm = LayerNorm(self.embed_dim)
        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = LayerNorm(self.embed_dim)

    def forward(
        self,
        x,
        encoder_hidden_states,
        encoder_attn_mask=None,
        layer_state=None,
        causal_mask=None,
        decoder_padding_mask=None,
    ):
        residual = x

        if layer_state is None:
            layer_state = {}
        if self.normalize_before:
            x = self.self_attn_layer_norm(x)
        # Self Attention

        x, self_attn_weights = self.self_attn(
            query=x,
            key=x,
            layer_state=layer_state,  # adds keys to layer state
            key_padding_mask=decoder_padding_mask,
            attn_mask=causal_mask,
            need_weights=self.output_attentions,
        )
        x = F.dropout(x, p=self.dropout, training=self.training)
        x = residual + x
        if not self.normalize_before:
            x = self.self_attn_layer_norm(x)

        # Cross attention
        residual = x
        assert self.encoder_attn.cache_key != self.self_attn.cache_key
        if self.normalize_before:
            x = self.encoder_attn_layer_norm(x)
        x, _ = self.encoder_attn(
            query=x,
            key=encoder_hidden_states,
            key_padding_mask=encoder_attn_mask,
            layer_state=layer_state,  # mutates layer state
        )
        x = F.dropout(x, p=self.dropout, training=self.training)
        x = residual + x
        if not self.normalize_before:
            x = self.encoder_attn_layer_norm(x)

        # Fully Connected
        residual = x
        if self.normalize_before:
            x = self.final_layer_norm(x)
        x = self.activation_fn(self.fc1(x))
        x = F.dropout(x, p=self.activation_dropout, training=self.training)
        x = self.fc2(x)
        x = F.dropout(x, p=self.dropout, training=self.training)
        x = residual + x
        if not self.normalize_before:
            x = self.final_layer_norm(x)
        return (
            x,
            self_attn_weights,
            layer_state,
        )  # just self_attn weights for now, following t5, layer_state = cache for decoding


class BartDecoder(nn.Module):
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer
    is a :class:`DecoderLayer`.
    Args:
        config: BartConfig
        embed_tokens (torch.nn.Embedding): output embedding
    """

    def __init__(self, config: BartConfig, embed_tokens: nn.Embedding):
        super().__init__()
        self.output_attentions = config.output_attentions
        self.output_hidden_states = config.output_hidden_states
        self.dropout = config.dropout
        self.layerdrop = config.decoder_layerdrop
        self.padding_idx = embed_tokens.padding_idx
        # print("I am in Decoder, padding_idx ", self.padding_idx)
        # self.padding_idx = 0
        self.max_target_positions = config.max_position_embeddings
        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
        self.embed_tokens = embed_tokens
        if config.static_position_embeddings:
            self.embed_positions = SinusoidalPositionalEmbedding(
                config.max_position_embeddings, config.d_model, config.pad_token_id
            )
        else:
            self.embed_positions = LearnedPositionalEmbedding(
                config.max_position_embeddings, config.d_model, self.padding_idx,
            )
        self.layers = nn.ModuleList(
            [DecoderLayer(config) for _ in range(config.decoder_layers)]
        )  # type: List[DecoderLayer]
        self.layernorm_embedding = LayerNorm(config.d_model) if config.normalize_embedding else nn.Identity()
        self.layer_norm = LayerNorm(config.d_model) if config.add_final_layer_norm else None

    def forward(
        self,
        input_ids,
        input_embed,
        encoder_hidden_states,
        encoder_padding_mask,
        decoder_padding_mask,
        decoder_causal_mask,
        decoder_cached_states=None,
        use_cache=False,
        **unused
    ):
        """
        Includes several features from "Jointly Learning to Align and
        Translate with Transformer Models" (Garg et al., EMNLP 2019).

        Args:
            input_ids (LongTensor): previous decoder outputs of shape
                `(batch, tgt_len)`, for teacher forcing
            encoder_hidden_states: output from the encoder, used for
                encoder-side attention
            encoder_padding_mask: for ignoring pad tokens
            decoder_cached_states (dict or None): dictionary used for storing state during generation

        Returns:
            tuple:
                - the decoder's features of shape `(batch, tgt_len, embed_dim)`
                - hidden states
                - attentions
        """
        # check attention mask and invert
        if encoder_padding_mask is not None:
            encoder_padding_mask = invert_mask(encoder_padding_mask)

        # embed positions
        positions, p_idx = self.embed_positions(input_ids, use_cache=use_cache)
        # print("I am in decoder, use_cache", use_cache, p_idx)

        if use_cache:
            input_ids = input_ids[:, -1:]
            input_embed = input_embed[:, -1:]
            positions = positions[:, -1:]  # happens after we embed them
            # assert input_ids.ne(self.padding_idx).any()

        # x = self.embed_tokens(input_ids) * self.embed_scale
        x = input_embed * self.embed_scale
        x += positions
        x = self.layernorm_embedding(x)
        x = F.dropout(x, p=self.dropout, training=self.training)

        # Convert to Bart output format: (seq_len, BS, model_dim) -> (BS, seq_len, model_dim)
        x = x.transpose(0, 1)
        encoder_hidden_states = encoder_hidden_states.transpose(0, 1)

        # decoder layers
        all_hidden_states = ()
        all_self_attns = ()
        next_decoder_cache = []
        for idx, decoder_layer in enumerate(self.layers):
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            if self.output_hidden_states:
                all_hidden_states += (x,)
            dropout_probability = random.uniform(0, 1)
            if self.training and (dropout_probability < self.layerdrop):
                continue

            layer_state = decoder_cached_states[idx] if decoder_cached_states is not None else None

            x, layer_self_attn, layer_past = decoder_layer(
                x,
                encoder_hidden_states,
                encoder_attn_mask=encoder_padding_mask,
                decoder_padding_mask=decoder_padding_mask,
                layer_state=layer_state,
                causal_mask=decoder_causal_mask,
            )

            if use_cache:
                next_decoder_cache.append(layer_past.copy())

            if self.layer_norm and (idx == len(self.layers) - 1):  # last layer of mbart
                x = self.layer_norm(x)
            if self.output_attentions:
                all_self_attns += (layer_self_attn,)

        # Convert to standard output format: (seq_len, BS, model_dim) -> (BS, seq_len, model_dim)
        all_hidden_states = [hidden_state.transpose(0, 1) for hidden_state in all_hidden_states]
        x = x.transpose(0, 1)
        encoder_hidden_states = encoder_hidden_states.transpose(0, 1)

        if use_cache:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



relogic/pretrainkit/models/semparse/modeling_bart_copy.py [352:566]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
class DecoderLayer(nn.Module):
    def __init__(self, config: BartConfig):
        super().__init__()
        self.embed_dim = config.d_model
        self.output_attentions = config.output_attentions
        self.self_attn = SelfAttention(
            embed_dim=self.embed_dim, num_heads=config.decoder_attention_heads, dropout=config.attention_dropout,
        )
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout
        self.normalize_before = config.normalize_before

        self.self_attn_layer_norm = LayerNorm(self.embed_dim)
        self.encoder_attn = SelfAttention(
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            encoder_decoder_attention=True,
        )
        self.encoder_attn_layer_norm = LayerNorm(self.embed_dim)
        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = LayerNorm(self.embed_dim)

    def forward(
        self,
        x,
        encoder_hidden_states,
        encoder_attn_mask=None,
        layer_state=None,
        causal_mask=None,
        decoder_padding_mask=None,
    ):
        residual = x

        if layer_state is None:
            layer_state = {}
        if self.normalize_before:
            x = self.self_attn_layer_norm(x)
        # Self Attention

        x, self_attn_weights = self.self_attn(
            query=x,
            key=x,
            layer_state=layer_state,  # adds keys to layer state
            key_padding_mask=decoder_padding_mask,
            attn_mask=causal_mask,
            need_weights=self.output_attentions,
        )
        x = F.dropout(x, p=self.dropout, training=self.training)
        x = residual + x
        if not self.normalize_before:
            x = self.self_attn_layer_norm(x)

        # Cross attention
        residual = x
        assert self.encoder_attn.cache_key != self.self_attn.cache_key
        if self.normalize_before:
            x = self.encoder_attn_layer_norm(x)
        x, _ = self.encoder_attn(
            query=x,
            key=encoder_hidden_states,
            key_padding_mask=encoder_attn_mask,
            layer_state=layer_state,  # mutates layer state
        )
        x = F.dropout(x, p=self.dropout, training=self.training)
        x = residual + x
        if not self.normalize_before:
            x = self.encoder_attn_layer_norm(x)

        # Fully Connected
        residual = x
        if self.normalize_before:
            x = self.final_layer_norm(x)
        x = self.activation_fn(self.fc1(x))
        x = F.dropout(x, p=self.activation_dropout, training=self.training)
        x = self.fc2(x)
        x = F.dropout(x, p=self.dropout, training=self.training)
        x = residual + x
        if not self.normalize_before:
            x = self.final_layer_norm(x)
        return (
            x,
            self_attn_weights,
            layer_state,
        )  # just self_attn weights for now, following t5, layer_state = cache for decoding


class BartDecoder(nn.Module):
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer
    is a :class:`DecoderLayer`.
    Args:
        config: BartConfig
        embed_tokens (torch.nn.Embedding): output embedding
    """

    def __init__(self, config: BartConfig, embed_tokens: nn.Embedding):
        super().__init__()
        self.output_attentions = config.output_attentions
        self.output_hidden_states = config.output_hidden_states
        self.dropout = config.dropout
        self.layerdrop = config.decoder_layerdrop
        self.padding_idx = embed_tokens.padding_idx
        # print("I am in Decoder, padding_idx ", self.padding_idx)
        # self.padding_idx = 0
        self.max_target_positions = config.max_position_embeddings
        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
        self.embed_tokens = embed_tokens
        if config.static_position_embeddings:
            self.embed_positions = SinusoidalPositionalEmbedding(
                config.max_position_embeddings, config.d_model, config.pad_token_id
            )
        else:
            self.embed_positions = LearnedPositionalEmbedding(
                config.max_position_embeddings, config.d_model, self.padding_idx,
            )
        self.layers = nn.ModuleList(
            [DecoderLayer(config) for _ in range(config.decoder_layers)]
        )  # type: List[DecoderLayer]
        self.layernorm_embedding = LayerNorm(config.d_model) if config.normalize_embedding else nn.Identity()
        self.layer_norm = LayerNorm(config.d_model) if config.add_final_layer_norm else None

    def forward(
        self,
        input_ids,
        input_embed,
        encoder_hidden_states,
        encoder_padding_mask,
        decoder_padding_mask,
        decoder_causal_mask,
        decoder_cached_states=None,
        use_cache=False,
        **unused
    ):
        """
        Includes several features from "Jointly Learning to Align and
        Translate with Transformer Models" (Garg et al., EMNLP 2019).

        Args:
            input_ids (LongTensor): previous decoder outputs of shape
                `(batch, tgt_len)`, for teacher forcing
            encoder_hidden_states: output from the encoder, used for
                encoder-side attention
            encoder_padding_mask: for ignoring pad tokens
            decoder_cached_states (dict or None): dictionary used for storing state during generation

        Returns:
            tuple:
                - the decoder's features of shape `(batch, tgt_len, embed_dim)`
                - hidden states
                - attentions
        """
        # check attention mask and invert
        if encoder_padding_mask is not None:
            encoder_padding_mask = invert_mask(encoder_padding_mask)

        # embed positions
        positions, p_idx = self.embed_positions(input_ids, use_cache=use_cache)
        # print("I am in decoder, use_cache", use_cache, p_idx)

        if use_cache:
            input_ids = input_ids[:, -1:]
            input_embed = input_embed[:, -1:]
            positions = positions[:, -1:]  # happens after we embed them
            # assert input_ids.ne(self.padding_idx).any()

        # x = self.embed_tokens(input_ids) * self.embed_scale
        x = input_embed * self.embed_scale
        x += positions
        x = self.layernorm_embedding(x)
        x = F.dropout(x, p=self.dropout, training=self.training)

        # Convert to Bart output format: (seq_len, BS, model_dim) -> (BS, seq_len, model_dim)
        x = x.transpose(0, 1)
        encoder_hidden_states = encoder_hidden_states.transpose(0, 1)

        # decoder layers
        all_hidden_states = ()
        all_self_attns = ()
        next_decoder_cache = []
        for idx, decoder_layer in enumerate(self.layers):
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            if self.output_hidden_states:
                all_hidden_states += (x,)
            dropout_probability = random.uniform(0, 1)
            if self.training and (dropout_probability < self.layerdrop):
                continue

            layer_state = decoder_cached_states[idx] if decoder_cached_states is not None else None

            x, layer_self_attn, layer_past = decoder_layer(
                x,
                encoder_hidden_states,
                encoder_attn_mask=encoder_padding_mask,
                decoder_padding_mask=decoder_padding_mask,
                layer_state=layer_state,
                causal_mask=decoder_causal_mask,
            )

            if use_cache:
                next_decoder_cache.append(layer_past.copy())

            if self.layer_norm and (idx == len(self.layers) - 1):  # last layer of mbart
                x = self.layer_norm(x)
            if self.output_attentions:
                all_self_attns += (layer_self_attn,)

        # Convert to standard output format: (seq_len, BS, model_dim) -> (BS, seq_len, model_dim)
        all_hidden_states = [hidden_state.transpose(0, 1) for hidden_state in all_hidden_states]
        x = x.transpose(0, 1)
        encoder_hidden_states = encoder_hidden_states.transpose(0, 1)

        if use_cache:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -