transfer/code/scGPT/scgpt/model/generation_model.py

import os
import math
from typing import Mapping, Optional, Tuple, Any, Union

import torch
from torch import nn, Tensor
import torch.distributed as dist
import torch.nn.functional as F
from torch.nn import TransformerEncoder, TransformerEncoderLayer
from torch.distributions import Bernoulli
from torch.utils.data import dataset
from tqdm import trange

from .model import (
    ExprDecoder,
    MVCDecoder,
    ContinuousValueEncoder,
    FastTransformerEncoderWrapper,
    FlashTransformerEncoderLayer,
)
from ..utils import map_raw_id_to_vocab_id
from .. import logger


class TransformerGenerator(nn.Module):
    def __init__(
        self,
        ntoken: int,
        d_model: int,
        nhead: int,
        d_hid: int,
        nlayers: int,
        nlayers_cls: int,
        n_cls: int,
        vocab: Any,
        dropout: float = 0.5,
        pad_token: str = "<pad>",
        pad_value: int = 0,
        pert_pad_id: int = 2,
        do_mvc: bool = False,
        domain_spec_batchnorm: Union[bool, str] = False,
        n_input_bins: Optional[int] = 0,
        cell_emb_style: str = "cls",
        mvc_decoder_style: str = "inner product",
        decoder_activation: Optional[str] = None,
        decoder_adaptive_bias: bool = False,
        ecs_threshold: float = 0.3,
        explicit_zero_prob: bool = False,
        use_fast_transformer: bool = False,
        fast_transformer_backend: str = "flash",
        pre_norm: bool = False,
    ):
        super().__init__()
        self.model_type = "Transformer"
        self.d_model = d_model
        self.pad_token_id = vocab[pad_token]
        self.pad_value = pad_value
        self.pert_pad_id = pert_pad_id
        self.ecs_threshold = ecs_threshold
        self.domain_spec_batchnorm = domain_spec_batchnorm
        self.n_input_bins = n_input_bins
        self.cell_emb_style = cell_emb_style
        self.explicit_zero_prob = explicit_zero_prob
        self.norm_scheme = "pre" if pre_norm else "post"
        if cell_emb_style not in ["cls", "avg-pool", "w-pool"]:
            raise ValueError(f"Unknown cell_emb_style: {cell_emb_style}")
        if use_fast_transformer:
            try:
                from flash_attn.flash_attention import FlashMHA
            except ImportError:
                import warnings

                warnings.warn(
                    "flash-attn is not installed, using pytorch transformer instead. "
                    "Set use_fast_transformer=False to avoid this warning. "
                    "Installing flash-attn is highly recommended."
                )
                use_fast_transformer = False
        self.use_fast_transformer = use_fast_transformer

        self.encoder = GeneEncoder(ntoken, d_model, padding_idx=vocab[pad_token])
        self.value_encoder = ContinuousValueEncoder(d_model, dropout)
        self.pert_encoder = nn.Embedding(3, d_model, padding_idx=pert_pad_id)

        # print("Using simple batchnorm instead of domain specific batchnorm")
        # self.bn = nn.BatchNorm1d(d_model, eps=6.1e-5)

        if use_fast_transformer:
            if fast_transformer_backend == "linear":
                self.transformer_encoder = FastTransformerEncoderWrapper(
                    d_model, nhead, d_hid, nlayers, dropout
                )
            elif fast_transformer_backend == "flash":
                encoder_layers = FlashTransformerEncoderLayer(
                    d_model,
                    nhead,
                    d_hid,
                    dropout,
                    batch_first=True,
                    norm_scheme=self.norm_scheme,
                )
                self.transformer_encoder = TransformerEncoder(encoder_layers, nlayers)
        else:
            encoder_layers = TransformerEncoderLayer(
                d_model, nhead, d_hid, dropout, batch_first=True
            )
            self.transformer_encoder = TransformerEncoder(encoder_layers, nlayers)

        # self.decoder = nn.Linear(d_model, 1)
        self.decoder = AffineExprDecoder(
            d_model,
            explicit_zero_prob=explicit_zero_prob,
            activation=decoder_activation,
            adaptive_bias=decoder_adaptive_bias,
        )
        self.cls_decoder = ClsDecoder(d_model, n_cls, nlayers=nlayers_cls)
        if do_mvc:
            self.mvc_decoder = MVCDecoder(
                d_model,
                arch_style=mvc_decoder_style,
                explicit_zero_prob=explicit_zero_prob,
            )

        self.init_weights()

    def init_weights(self) -> None:
        initrange = 0.1
        self.encoder.embedding.weight.data.uniform_(-initrange, initrange)

    def _encode(
        self,
        src: Tensor,
        values: Tensor,
        input_pert_flags,
        src_key_padding_mask: Tensor,
    ) -> Tensor:
        src = self.encoder(src)  # (batch, seq_len, embsize)
        self.cur_gene_token_embs = src
        values = self.value_encoder(values)  # (batch, seq_len, embsize)
        perts = self.pert_encoder(input_pert_flags)  # (batch, seq_len, embsize)
        total_embs = src + values + perts

        # total_embs = self.bn(total_embs.permute(0, 2, 1)).permute(0, 2, 1)
        output = self.transformer_encoder(
            total_embs, src_key_padding_mask=src_key_padding_mask
        )
        return output  # (batch, seq_len, embsize)

    def _get_cell_emb_from_layer(
        self, layer_output: Tensor, weights: Tensor = None
    ) -> Tensor:
        """
        Args:
            layer_output(:obj:`Tensor`): shape (batch, seq_len, embsize)
            weights(:obj:`Tensor`): shape (batch, seq_len), optional and only used
                when :attr:`self.cell_emb_style` is "w-pool".

        Returns:
            :obj:`Tensor`: shape (batch, embsize)
        """
        if self.cell_emb_style == "cls":
            cell_emb = layer_output[:, 0, :]  # (batch, embsize)
        elif self.cell_emb_style == "avg-pool":
            cell_emb = torch.mean(layer_output, dim=1)
        elif self.cell_emb_style == "w-pool":
            if weights is None:
                raise ValueError("weights is required when cell_emb_style is w-pool")
            if weights.dim() != 2:
                raise ValueError("weights should be 2D")
            cell_emb = torch.sum(layer_output * weights.unsqueeze(2), dim=1)
            cell_emb = F.normalize(cell_emb, p=2, dim=1)  # (batch, embsize)

        return cell_emb

    def forward(
        self,
        src: Tensor,
        values: Tensor,
        input_pert_flags: Tensor,
        src_key_padding_mask: Tensor,
        CLS: bool = False,
        CCE: bool = False,
        MVC: bool = False,
        ECS: bool = False,
        do_sample: bool = False,
    ) -> Mapping[str, Tensor]:
        """
        Args:
            src (:obj:`Tensor`): token ids, shape [batch_size, seq_len]
            values (:obj:`Tensor`): token values, shape [batch_size, seq_len]
            src_key_padding_mask (:obj:`Tensor`): mask for src, shape [batch_size,
                seq_len]
            CLS (:obj:`bool`): if True, return the celltype classification objective
                (CLS) output
            CCE (:obj:`bool`): if True, return the contrastive cell embedding objective
                (CCE) output
            MVC (:obj:`bool`): if True, return the masked value prediction for cell
                embedding MVC output
            ECS (:obj:`bool`): if True, return the elastic cell similarity objective
                (ECS) output.

        Returns:
            dict of output Tensors.
        """
        if self.explicit_zero_prob and not do_sample and not self.training:
            do_sample = True
            logger.warning("Auto set do_sample to True when model is in eval mode.")

        # binning input gene values
        if self.n_input_bins > 0:
            from ..preprocess import binning

            processed_values = torch.stack(
                [binning(row, n_bins=self.n_input_bins) for row in values], dim=0
            ).to(values.device)
        else:
            processed_values = values

        transformer_output = self._encode(
            src, processed_values, input_pert_flags, src_key_padding_mask
        )
        output = {}
        mlm_output = self.decoder(transformer_output, values)
        if self.explicit_zero_prob and do_sample:
            bernoulli = Bernoulli(probs=mlm_output["zero_probs"])
            output["mlm_output"] = bernoulli.sample() * mlm_output["pred"]
        else:
            output["mlm_output"] = mlm_output["pred"]  # (batch, seq_len)
        if self.explicit_zero_prob:
            output["mlm_zero_probs"] = mlm_output["zero_probs"]

        cell_emb = self._get_cell_emb_from_layer(transformer_output, values)
        if CLS:
            output["cls_output"] = self.cls_decoder(cell_emb)  # (batch, n_cls)
        if MVC:
            mvc_output = self.mvc_decoder(
                cell_emb,
                self.cur_gene_token_embs,
            )  # (batch, seq_len)
            if self.explicit_zero_prob and do_sample:
                bernoulli = Bernoulli(probs=mvc_output["zero_probs"])
                output["mvc_output"] = bernoulli.sample() * mvc_output["pred"]
            else:
                output["mvc_output"] = mvc_output["pred"]  # (batch, seq_len)
            if self.explicit_zero_prob:
                output["mvc_zero_probs"] = mvc_output["zero_probs"]
        if ECS:
            # Here using customized cosine similarity instead of F.cosine_similarity
            # to avoid the pytorch issue of similarity larger than 1.0, pytorch # 78064
            # normalize the embedding
            cell_emb_normed = F.normalize(cell_emb, p=2, dim=1)
            cos_sim = torch.mm(cell_emb_normed, cell_emb_normed.t())  # (batch, batch)

            # mask out diagnal elements
            mask = torch.eye(cos_sim.size(0)).bool().to(cos_sim.device)
            cos_sim = cos_sim.masked_fill(mask, 0.0)
            # only optimize positive similarities
            cos_sim = F.relu(cos_sim)

            output["loss_ecs"] = torch.mean(1 - (cos_sim - self.ecs_threshold) ** 2)

        return output

    def encode_batch(
        self,
        src: Tensor,
        values: Tensor,
        src_key_padding_mask: Tensor,
        batch_size: int,
        output_to_cpu: bool = True,
    ) -> Tensor:
        """
        Args:
            src: Tensor, shape [N, seq_len]
            values: Tensor, shape [N, seq_len]
            src_key_padding_mask: Tensor, shape [N, seq_len]

        Returns:
            output Tensor of shape [N, seq_len, embsize]
        """
        outputs = []
        N = src.size(0)
        device = next(self.parameters()).device
        for i in trange(0, N, batch_size):
            output = self._encode(
                src[i : i + batch_size].to(device),
                values[i : i + batch_size].to(device),
                src_key_padding_mask[i : i + batch_size].to(device),
            )
            if output_to_cpu:
                output = output.cpu()
            outputs.append(output)
        return torch.cat(outputs, dim=0)

    def pred_perturb(
        self,
        batch_data,
        include_zero_gene="batch-wise",
        gene_ids=None,
        amp=True,
    ) -> Tensor:
        """
        Args:
            batch_data: a dictionary of input data with keys.

        Returns:
            output Tensor of shape [N, seq_len]
        """
        self.eval()
        device = next(self.parameters()).device
        batch_data.to(device)
        batch_size = len(batch_data.pert)
        x: torch.Tensor = batch_data.x
        ori_gene_values = x[:, 0].view(batch_size, -1)  # (batch_size, n_genes)
        pert_flags = x[:, 1].long().view(batch_size, -1)

        if include_zero_gene in ["all", "batch-wise"]:
            assert gene_ids is not None
            if include_zero_gene == "all":
                input_gene_ids = torch.arange(ori_gene_values.size(1), device=device)
            else:  # batch-wise
                input_gene_ids = (
                    ori_gene_values.nonzero()[:, 1].flatten().unique().sort()[0]
                )
            input_values = ori_gene_values[:, input_gene_ids]
            input_pert_flags = pert_flags[:, input_gene_ids]

            mapped_input_gene_ids = map_raw_id_to_vocab_id(input_gene_ids, gene_ids)
            mapped_input_gene_ids = mapped_input_gene_ids.repeat(batch_size, 1)

            src_key_padding_mask = torch.zeros_like(
                input_values, dtype=torch.bool, device=device
            )
            with torch.cuda.amp.autocast(enabled=amp):
                output_dict = self(
                    mapped_input_gene_ids,
                    input_values,
                    input_pert_flags,
                    src_key_padding_mask=src_key_padding_mask,
                    CLS=False,
                    CCE=False,
                    MVC=False,
                    ECS=False,
                    do_sample=True,
                )
            output_values = output_dict["mlm_output"].float()
            pred_gene_values = torch.zeros_like(ori_gene_values)
            pred_gene_values[:, input_gene_ids] = output_values
        return pred_gene_values


def generate_square_subsequent_mask(sz: int) -> Tensor:
    """Generates an upper-triangular matrix of -inf, with zeros on diag."""
    return torch.triu(torch.ones(sz, sz) * float("-inf"), diagonal=1)


class GeneEncoder(nn.Module):
    def __init__(
        self,
        num_embeddings: int,
        embedding_dim: int,
        padding_idx: Optional[int] = None,
    ):
        super().__init__()
        self.embedding = nn.Embedding(
            num_embeddings, embedding_dim, padding_idx=padding_idx
        )
        self.enc_norm = nn.LayerNorm(embedding_dim)

    def forward(self, x: Tensor) -> Tensor:
        x = self.embedding(x)  # (batch, seq_len, embsize)
        x = self.enc_norm(x)
        return x


class AffineExprDecoder(nn.Module):
    def __init__(
        self,
        d_model: int,
        explicit_zero_prob: bool = False,
        activation: Optional[str] = None,
        tanh_coeff: bool = False,
        adaptive_bias: bool = False,
    ):
        """
        Predict the expression value of each gene in an affine like form of Ax + b.
        This decoder takes two ExprDecoder intrinsically to genrate the coefficient A and bias b.

        Args:
            d_model: The embedding dimension.
            explicit_zero_prob: If True, predict the probability of each gene being
                zero.
            activation: The activation function for the coefficient A and bias b.
            tanh_coeff: If True, use tanh activation for the coefficient A.
            adaptive_bias: If True, use a learnable bias for the bias b.
        """
        super().__init__()
        self.explicit_zero_prob = explicit_zero_prob
        self.tanh_coeff = tanh_coeff
        self.adaptive_bias = adaptive_bias
        self.coeff_decoder = ExprDecoder(d_model, explicit_zero_prob=explicit_zero_prob)
        self.bias_decoder = ExprDecoder(d_model, explicit_zero_prob=explicit_zero_prob)

        self.activation = activation
        if activation is not None:
            assert hasattr(nn, activation), f"Unknown activation: {activation}"
            self.activation = getattr(nn, activation)()

    def forward(self, x: Tensor, values: Tensor) -> Tensor:
        """
        Args:
            x: Tensor, shape [batch_size, seq_len, embsize]
            values: Tensor, shape [batch_size, seq_len]

        Returns:
            output Tensor of shape [batch_size, seq_len]
        """
        coeff = self.coeff_decoder(x)
        bias = self.bias_decoder(x)

        if self.activation is not None:
            coeff["pred"] = self.activation(coeff["pred"])
            bias["pred"] = self.activation(bias["pred"])

        # if self.tanh_coeff:
        #     coeff["pred"] = 1 + torch.tanh(coeff["pred"])

        if self.adaptive_bias:
            # bias["pred"] = bias["pred"] * values.mean(dim=1, keepdim=True)
            non_zero_value_mean = values.sum(dim=1, keepdim=True) / (values != 0).sum(
                dim=1, keepdim=True
            )
            bias["pred"] = bias["pred"] * non_zero_value_mean

        if self.explicit_zero_prob:
            return {
                "pred": coeff["pred"] * values + bias["pred"],
                "zero_probs": coeff["zero_probs"],
            }

        return dict(pred=coeff["pred"] * values + bias["pred"])


class TokenEmbedding(nn.Module):
    def __init__(
        self,
        num_embeddings: int,
        embedding_dim: int,
        padding_idx: Optional[int] = None,
        zero_out_idx: Optional[int] = None,
    ):
        """
        Generic token embedding module.

        Args:
            num_embeddings: The number of tokens.
            embedding_dim: The embedding dimension.
            padding_idx: The index of the padding token.
            zero_out_idx: Indicate if any idx embedding should be zero vector.
        """
        super().__init__()
        self.embedding = nn.Embedding(
            num_embeddings, embedding_dim, padding_idx=padding_idx
        )
        self.enc_norm = nn.LayerNorm(embedding_dim)

        self.zero_out_idx = zero_out_idx
        if zero_out_idx is not None:
            self._fill_idx_with_zero(zero_out_idx)
            zero_vector = self(zero_out_idx)
            assert torch.all(zero_vector == 0.0)
            assert not zero_vector.requires_grad

    def _fill_idx_with_zero(self, idx) -> None:
        with torch.no_grad():
            self.embedding.weight[idx].fill_(0)

    def forward(self, x: Tensor) -> Tensor:
        x = self.embedding(x)  # (batch, seq_len, embsize)
        x = self.enc_norm(x)
        return x


class PositionalEncoding(nn.Module):
    def __init__(self, d_model: int, dropout: float = 0.1, max_len: int = 5000):
        super().__init__()
        self.dropout = nn.Dropout(p=dropout)

        position = torch.arange(max_len).unsqueeze(1)
        div_term = torch.exp(
            torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model)
        )
        pe = torch.zeros(max_len, 1, d_model)
        pe[:, 0, 0::2] = torch.sin(position * div_term)
        pe[:, 0, 1::2] = torch.cos(position * div_term)
        self.register_buffer("pe", pe)

    def forward(self, x: Tensor) -> Tensor:
        """
        Args:
            x: Tensor, shape [seq_len, batch_size, embedding_dim]
        """
        x = x + self.pe[: x.size(0)]
        return self.dropout(x)


class Similarity(nn.Module):
    """
    Dot product or cosine similarity
    """

    def __init__(self, temp):
        super().__init__()
        self.temp = temp
        self.cos = nn.CosineSimilarity(dim=-1)

    def forward(self, x, y):
        return self.cos(x, y) / self.temp


class ClsDecoder(nn.Module):
    """
    Decoder for classification task.
    """

    def __init__(
        self,
        d_model: int,
        n_cls: int,
        nlayers: int = 3,
        activation: callable = nn.ReLU,
    ):
        super().__init__()
        # module list
        self._decoder = nn.ModuleList()
        for i in range(nlayers - 1):
            self._decoder.append(nn.Linear(d_model, d_model))
            self._decoder.append(activation())
            self._decoder.append(nn.LayerNorm(d_model))
        self.out_layer = nn.Linear(d_model, n_cls)

    def forward(self, x: Tensor) -> Tensor:
        """
        Args:
            x: Tensor, shape [batch_size, embsize]
        """
        for layer in self._decoder:
            x = layer(x)
        return self.out_layer(x)