pytorch-transformer-ts/switch/module.py

from typing import List, Optional, Union, Callable

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.modules.transformer import _get_activation_fn, _get_clones
from gluonts.core.component import validated
from gluonts.time_feature import get_lags_for_frequency
from gluonts.torch.distributions import DistributionOutput, StudentTOutput
from gluonts.torch.modules.feature import FeatureEmbedder
from gluonts.torch.modules.scaler import MeanScaler, NOPScaler


class SwitchFeedForward(nn.Module):
    """
    ## Routing among multiple FFNs
    """

    def __init__(
        self,
        *,
        capacity_factor: float,
        drop_tokens: bool,
        is_scale_prob: bool,
        n_experts: int,
        expert: nn.Module,
        d_model: int,
        dim_feedforward: int,
    ):
        """
        * `capacity_factor` is the capacity of each expert as a factor relative to ideally balanced load
        * `drop_tokens` specifies whether to drop tokens if more tokens are routed to an expert than the capacity
        * `is_scale_prob` specifies whether to multiply the input to the FFN by the routing probability
        * `n_experts` is the number of experts
        * `expert` is the expert layer, a [FFN module](../feed_forward.html)
        * `d_model` is the number of features in a token embedding
        """
        super().__init__()

        self.capacity_factor = capacity_factor
        self.is_scale_prob = is_scale_prob
        self.n_experts = n_experts
        self.drop_tokens = drop_tokens
        self.dim_feedforward = dim_feedforward

        # make copies of the FFNs
        self.experts = _get_clones(expert, n_experts)
        # Routing layer and softmax
        self.switch = nn.Linear(d_model, n_experts)
        self.softmax = nn.Softmax(dim=-1)

    def forward(self, x: torch.Tensor):
        """
        * `x` is the input to the switching module with shape `[batch_size, seq_len, d_model]`
        """

        # Capture the shape to change shapes later
        batch_size, seq_len, d_model = x.shape
        # Flatten the sequence and batch dimensions
        x = x.view(-1, d_model)

        # Get routing probabilities for each of the tokens.
        # $$p_i(x) = \frac{e^{h(x)_i}}{\sum^N_j e^{h(x)_j}}$$
        # where $N$ is the number of experts `n_experts` and
        # $h(\cdot)$ is the linear transformation of token embeddings.
        route_prob = self.softmax(self.switch(x))

        # Get the maximum routing probabilities and the routes.
        # We route to the expert with highest probability
        route_prob_max, routes = torch.max(route_prob, dim=-1)

        # Get indexes of tokens going to each expert
        indexes_list = [
            torch.eq(routes, i).nonzero(as_tuple=True)[0] for i in range(self.n_experts)
        ]

        # Initialize an empty tensor to store outputs
        final_output = x.new_zeros((batch_size, seq_len, self.dim_feedforward))

        # Capacity of each expert.
        # $$\mathrm{expert\;capacity} =
        # \frac{\mathrm{tokens\;per\;batch}}{\mathrm{number\;of\;experts}}
        # \times \mathrm{capacity\;factor}$$
        capacity = int(self.capacity_factor * len(x) / self.n_experts)
        # Number of tokens routed to each expert.
        counts = x.new_tensor([len(indexes_list[i]) for i in range(self.n_experts)])

        # Initialize an empty list of dropped tokens
        dropped = []
        # Only drop tokens if `drop_tokens` is `True`.
        if self.drop_tokens:
            # Drop tokens in each of the experts
            for i in range(self.n_experts):
                # Ignore if the expert is not over capacity
                if len(indexes_list[i]) <= capacity:
                    continue
                # Shuffle indexes before dropping
                indexes_list[i] = indexes_list[i][torch.randperm(len(indexes_list[i]))]
                # Collect the tokens over capacity as dropped tokens
                dropped.append(indexes_list[i][capacity:])
                # Keep only the tokens upto the capacity of the expert
                indexes_list[i] = indexes_list[i][:capacity]

        # Get outputs of the expert FFNs
        expert_output = [
            self.experts[i](x[indexes_list[i], :]) for i in range(self.n_experts)
        ]

        # Assign to final output
        for i in range(self.n_experts):
            final_output[indexes_list[i], :] = expert_output[i]

        # Pass through the dropped tokens
        if dropped:
            dropped = torch.cat(dropped)
            final_output[dropped, :] = x[dropped, :]

        if self.is_scale_prob:
            # Multiply by the expert outputs by the probabilities $y = p_i(x) E_i(x)$
            final_output = final_output * route_prob_max.view(-1, 1)
        else:
            # Don't scale the values but multiply by $\frac{p}{\hat{p}} = 1$ so that the gradients flow
            # (this is something we experimented with).
            final_output = final_output * (
                route_prob_max / route_prob_max.detach()
            ).view(-1, 1)

        # Change the shape of the final output back to `[batch_size, seq_len, d_ff]`
        final_output = final_output.view(batch_size, seq_len, -1)

        # Return
        #
        # * the final output
        # * number of tokens routed to each expert
        # * sum of probabilities for each expert
        # * number of tokens dropped.
        # * routing probabilities of the selected experts
        #
        # These are used for the load balancing loss and logging
        return final_output, counts, route_prob.sum(0), len(dropped), route_prob_max


class TransformerEncoderLayer(nn.Module):

    __constants__ = ["batch_first", "norm_first"]

    def __init__(
        self,
        d_model: int,
        nhead: int,
        capacity_factor: int,
        drop_tokens: bool,
        is_scale_prob: bool,
        n_experts: int = 1,
        dim_feedforward: int = 2048,
        dropout: float = 0.1,
        activation: Union[str, Callable[[torch.Tensor], torch.Tensor]] = F.relu,
        layer_norm_eps: float = 1e-5,
        batch_first: bool = True,
        norm_first: bool = False,
        device=None,
        dtype=None,
    ) -> None:
        factory_kwargs = {"device": device, "dtype": dtype}
        super(TransformerEncoderLayer, self).__init__()
        self.self_attn = nn.MultiheadAttention(
            d_model, nhead, dropout=dropout, batch_first=batch_first, **factory_kwargs
        )
        # Implementation of Feedforward model
        linear = nn.Linear(d_model, dim_feedforward, **factory_kwargs)
        self.linear1 = SwitchFeedForward(
            capacity_factor,
            drop_tokens,
            is_scale_prob,
            n_experts,
            expert=linear,
            d_model=d_model,
            dim_feedforward=dim_feedforward,
        )
        self.dropout = nn.Dropout(dropout)
        self.linear2 = nn.Linear(dim_feedforward, d_model, **factory_kwargs)

        self.norm_first = norm_first
        self.norm1 = nn.LayerNorm(d_model, eps=layer_norm_eps, **factory_kwargs)
        self.norm2 = nn.LayerNorm(d_model, eps=layer_norm_eps, **factory_kwargs)
        self.dropout1 = nn.Dropout(dropout)
        self.dropout2 = nn.Dropout(dropout)

        # Legacy string support for activation function.
        if isinstance(activation, str):
            self.activation = _get_activation_fn(activation)
        else:
            self.activation = activation

    def __setstate__(self, state):
        if "activation" not in state:
            state["activation"] = F.relu
        super(TransformerEncoderLayer, self).__setstate__(state)

    def forward(
        self,
        src: torch.Tensor,
        src_mask: Optional[torch.Tensor] = None,
        src_key_padding_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:

        x = src
        if self.norm_first:
            x = x + self._sa_block(self.norm1(x), src_mask, src_key_padding_mask)
            x = x + self._ff_block(self.norm2(x))
        else:
            x = self.norm1(x + self._sa_block(x, src_mask, src_key_padding_mask))
            x = self.norm2(x + self._ff_block(x))

        return x

    # self-attention block
    def _sa_block(
        self,
        x: torch.Tensor,
        attn_mask: Optional[torch.Tensor],
        key_padding_mask: Optional[torch.Tensor],
    ) -> torch.Tensor:
        x = self.self_attn(
            x,
            x,
            x,
            attn_mask=attn_mask,
            key_padding_mask=key_padding_mask,
            need_weights=False,
        )[0]
        return self.dropout1(x)

    # feed forward block
    def _ff_block(self, x: torch.Tensor) -> torch.Tensor:
        x = self.linear2(self.dropout(self.activation(self.linear1(x))))
        return self.dropout2(x)


class SwitchTransformerModel(nn.Module):
    @validated()
    def __init__(
        self,
        freq: str,
        context_length: int,
        prediction_length: int,
        num_feat_dynamic_real: int,
        num_feat_static_real: int,
        num_feat_static_cat: int,
        cardinality: List[int],
        # switch transformer arguments
        nhead: int,
        num_encoder_layers: int,
        num_decoder_layers: int,
        dim_feedforward: int,
        activation: str = "gelu",
        dropout: float = 0.1,
        # univariate input
        input_size: int = 1,
        embedding_dimension: Optional[List[int]] = None,
        distr_output: DistributionOutput = StudentTOutput(),
        lags_seq: Optional[List[int]] = None,
        scaling: bool = True,
        num_parallel_samples: int = 100,
    ) -> None:
        super().__init__()

        self.input_size = input_size

        self.target_shape = distr_output.event_shape
        self.num_feat_dynamic_real = num_feat_dynamic_real
        self.num_feat_static_cat = num_feat_static_cat
        self.num_feat_static_real = num_feat_static_real
        self.embedding_dimension = (
            embedding_dimension
            if embedding_dimension is not None or cardinality is None
            else [min(50, (cat + 1) // 2) for cat in cardinality]
        )
        self.lags_seq = lags_seq or get_lags_for_frequency(freq_str=freq)
        self.num_parallel_samples = num_parallel_samples
        self.history_length = context_length + max(self.lags_seq)
        self.embedder = FeatureEmbedder(
            cardinalities=cardinality,
            embedding_dims=self.embedding_dimension,
        )
        if scaling:
            self.scaler = MeanScaler(dim=1, keepdim=True)
        else:
            self.scaler = NOPScaler(dim=1, keepdim=True)

        # total feature size
        d_model = self.input_size * len(self.lags_seq) + self._number_of_features

        self.context_length = context_length
        self.prediction_length = prediction_length
        self.distr_output = distr_output
        self.param_proj = distr_output.get_args_proj(d_model)

        # transformer enc-decoder and mask initializer
        self.transformer = nn.Transformer(
            d_model=d_model,
            nhead=nhead,
            num_encoder_layers=num_encoder_layers,
            num_decoder_layers=num_decoder_layers,
            dim_feedforward=dim_feedforward,
            dropout=dropout,
            activation=activation,
            batch_first=True,
        )

        # causal decoder tgt mask
        self.register_buffer(
            "tgt_mask",
            self.transformer.generate_square_subsequent_mask(prediction_length),
        )

    @property
    def _number_of_features(self) -> int:
        return (
            sum(self.embedding_dimension)
            + self.num_feat_dynamic_real
            + self.num_feat_static_real
            + 1  # the log(scale)
        )

    @property
    def _past_length(self) -> int:
        return self.context_length + max(self.lags_seq)

    def get_lagged_subsequences(
        self, sequence: torch.Tensor, subsequences_length: int, shift: int = 0
    ) -> torch.Tensor:
        """
        Returns lagged subsequences of a given sequence.
        Parameters
        ----------
        sequence : Tensor
            the sequence from which lagged subsequences should be extracted.
            Shape: (N, T, C).
        subsequences_length : int
            length of the subsequences to be extracted.
        shift: int
            shift the lags by this amount back.
        Returns
        --------
        lagged : Tensor
            a tensor of shape (N, S, C, I), where S = subsequences_length and
            I = len(indices), containing lagged subsequences. Specifically,
            lagged[i, j, :, k] = sequence[i, -indices[k]-S+j, :].
        """
        sequence_length = sequence.shape[1]
        indices = [lag - shift for lag in self.lags_seq]

        assert max(indices) + subsequences_length <= sequence_length, (
            f"lags cannot go further than history length, found lag {max(indices)} "
            f"while history length is only {sequence_length}"
        )

        lagged_values = []
        for lag_index in indices:
            begin_index = -lag_index - subsequences_length
            end_index = -lag_index if lag_index > 0 else None
            lagged_values.append(sequence[:, begin_index:end_index, ...])
        return torch.stack(lagged_values, dim=-1)

    def _check_shapes(
        self,
        prior_input: torch.Tensor,
        inputs: torch.Tensor,
        features: Optional[torch.Tensor],
    ) -> None:
        assert len(prior_input.shape) == len(inputs.shape)
        assert (
            len(prior_input.shape) == 2 and self.input_size == 1
        ) or prior_input.shape[2] == self.input_size
        assert (len(inputs.shape) == 2 and self.input_size == 1) or inputs.shape[
            -1
        ] == self.input_size
        assert (
            features is None or features.shape[2] == self._number_of_features
        ), f"{features.shape[2]}, expected {self._number_of_features}"

    def create_network_inputs(
        self,
        feat_static_cat: torch.Tensor,
        feat_static_real: torch.Tensor,
        past_time_feat: torch.Tensor,
        past_target: torch.Tensor,
        past_observed_values: torch.Tensor,
        future_time_feat: Optional[torch.Tensor] = None,
        future_target: Optional[torch.Tensor] = None,
    ):
        # time feature
        time_feat = (
            torch.cat(
                (
                    past_time_feat[:, self._past_length - self.context_length :, ...],
                    future_time_feat,
                ),
                dim=1,
            )
            if future_target is not None
            else past_time_feat[:, self._past_length - self.context_length :, ...]
        )

        # target
        context = past_target[:, -self.context_length :]
        observed_context = past_observed_values[:, -self.context_length :]
        _, scale = self.scaler(context, observed_context)

        inputs = (
            torch.cat((past_target, future_target), dim=1) / scale
            if future_target is not None
            else past_target / scale
        )

        inputs_length = (
            self._past_length + self.prediction_length
            if future_target is not None
            else self._past_length
        )
        assert inputs.shape[1] == inputs_length

        subsequences_length = (
            self.context_length + self.prediction_length
            if future_target is not None
            else self.context_length
        )

        # embeddings
        embedded_cat = self.embedder(feat_static_cat)
        static_feat = torch.cat(
            (embedded_cat, feat_static_real, scale.log()),
            dim=1,
        )
        expanded_static_feat = static_feat.unsqueeze(1).expand(
            -1, time_feat.shape[1], -1
        )

        features = torch.cat((expanded_static_feat, time_feat), dim=-1)

        # self._check_shapes(prior_input, inputs, features)

        # sequence = torch.cat((prior_input, inputs), dim=1)
        lagged_sequence = self.get_lagged_subsequences(
            sequence=inputs,
            subsequences_length=subsequences_length,
        )

        lags_shape = lagged_sequence.shape
        reshaped_lagged_sequence = lagged_sequence.reshape(
            lags_shape[0], lags_shape[1], -1
        )

        transformer_inputs = torch.cat((reshaped_lagged_sequence, features), dim=-1)

        return transformer_inputs, scale, static_feat

    def output_params(self, transformer_inputs):
        enc_input = transformer_inputs[:, : self.context_length, ...]
        dec_input = transformer_inputs[:, self.context_length :, ...]

        enc_out = self.transformer.encoder(enc_input)
        dec_output = self.transformer.decoder(
            dec_input, enc_out, tgt_mask=self.tgt_mask
        )

        return self.param_proj(dec_output)

    @torch.jit.ignore
    def output_distribution(
        self, params, scale=None, trailing_n=None
    ) -> torch.distributions.Distribution:
        sliced_params = params
        if trailing_n is not None:
            sliced_params = [p[:, -trailing_n:] for p in params]
        return self.distr_output.distribution(sliced_params, scale=scale)

    # for prediction
    def forward(
        self,
        feat_static_cat: torch.Tensor,
        feat_static_real: torch.Tensor,
        past_time_feat: torch.Tensor,
        past_target: torch.Tensor,
        past_observed_values: torch.Tensor,
        future_time_feat: torch.Tensor,
        num_parallel_samples: Optional[int] = None,
    ) -> torch.Tensor:

        if num_parallel_samples is None:
            num_parallel_samples = self.num_parallel_samples

        encoder_inputs, scale, static_feat = self.create_network_inputs(
            feat_static_cat,
            feat_static_real,
            past_time_feat,
            past_target,
            past_observed_values,
        )

        enc_out = self.transformer.encoder(encoder_inputs)

        repeated_scale = scale.repeat_interleave(
            repeats=self.num_parallel_samples, dim=0
        )

        repeated_past_target = (
            past_target.repeat_interleave(repeats=self.num_parallel_samples, dim=0)
            / repeated_scale
        )

        expanded_static_feat = static_feat.unsqueeze(1).expand(
            -1, future_time_feat.shape[1], -1
        )
        features = torch.cat((expanded_static_feat, future_time_feat), dim=-1)
        repeated_features = features.repeat_interleave(
            repeats=self.num_parallel_samples, dim=0
        )

        repeated_enc_out = enc_out.repeat_interleave(
            repeats=self.num_parallel_samples, dim=0
        )

        future_samples = []

        # greedy decoding
        for k in range(self.prediction_length):
            # self._check_shapes(repeated_past_target, next_sample, next_features)
            # sequence = torch.cat((repeated_past_target, next_sample), dim=1)

            lagged_sequence = self.get_lagged_subsequences(
                sequence=repeated_past_target,
                subsequences_length=1 + k,
                shift=1,
            )

            lags_shape = lagged_sequence.shape
            reshaped_lagged_sequence = lagged_sequence.reshape(
                lags_shape[0], lags_shape[1], -1
            )

            decoder_input = torch.cat(
                (reshaped_lagged_sequence, repeated_features[:, : k + 1]), dim=-1
            )

            output = self.transformer.decoder(decoder_input, repeated_enc_out)

            params = self.param_proj(output[:, -1:])
            distr = self.output_distribution(params, scale=repeated_scale)
            next_sample = distr.sample()

            repeated_past_target = torch.cat(
                (repeated_past_target, next_sample / repeated_scale), dim=1
            )
            future_samples.append(next_sample)

        concat_future_samples = torch.cat(future_samples, dim=1)
        return concat_future_samples.reshape(
            (-1, self.num_parallel_samples, self.prediction_length) + self.target_shape,
        )