RESM.py

# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

from typing import Union
import torch
import torch.nn as nn

import esm
from esm.multihead_attention import MultiheadAttention
from esm.modules import ContactPredictionHead, ESM1bLayerNorm, RobertaLMHead, ESM1LayerNorm
from data import Alphabet_RNA
import math

def gelu(x):
    """Implementation of the gelu activation function.

    For information: OpenAI GPT's gelu is slightly different
    (and gives slightly different results):
    0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
    """
    return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))

class TransformerLayer(nn.Module):
    """Transformer layer block."""

    def __init__(
        self,
        embed_dim,
        ffn_embed_dim,
        attention_heads,
        add_bias_kv=True,
        use_esm1b_layer_norm=False,
        use_rotary_embeddings: bool = False,
        checkpoint = False
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.ffn_embed_dim = ffn_embed_dim
        self.attention_heads = attention_heads
        self.use_rotary_embeddings = use_rotary_embeddings
        self._init_submodules(add_bias_kv, use_esm1b_layer_norm)
        self.checkpoint = checkpoint

    def _init_submodules(self, add_bias_kv, use_esm1b_layer_norm):
        BertLayerNorm = ESM1bLayerNorm if use_esm1b_layer_norm else ESM1LayerNorm

        self.self_attn = MultiheadAttention(
            self.embed_dim,
            self.attention_heads,
            add_bias_kv=add_bias_kv,
            add_zero_attn=False,
            use_rotary_embeddings=self.use_rotary_embeddings,
            
        )
        self.self_attn_layer_norm = BertLayerNorm(self.embed_dim)

        self.fc1 = nn.Linear(self.embed_dim, self.ffn_embed_dim)
        self.fc2 = nn.Linear(self.ffn_embed_dim, self.embed_dim)

        self.final_layer_norm = BertLayerNorm(self.embed_dim)

    def forward(
        self, x, self_attn_mask=None, self_attn_padding_mask=None, need_head_weights=False
    ):
        residual = x
        x = self.self_attn_layer_norm(x)
        if self.checkpoint:
            x, attn = self.self_attn(
            query=x,
            key=x,
            value=x,
            key_padding_mask=self_attn_padding_mask,
            need_weights=True,
            need_head_weights=need_head_weights,
            attn_mask=self_attn_mask,
        )
            x = residual + x

            residual = x
            x = self.final_layer_norm(x)
            x = gelu(self.fc1(x))
            x = self.fc2(x)
            x = residual + x
        
        else:
            x, attn = self.self_attn(
                query=x,
                key=x,
                value=x,
                key_padding_mask=self_attn_padding_mask,
                need_weights=True,
                need_head_weights=need_head_weights,
                attn_mask=self_attn_mask,
            )
            x = residual + x

            residual = x
            x = self.final_layer_norm(x)
            x = gelu(self.fc1(x))
            x = self.fc2(x)
            x = residual + x

        return x, attn

class RESM(nn.Module):
    def __init__(
        self,
        alphabet: Alphabet_RNA,
        num_layers: int = 33,
        embed_dim: int = 1280,
        attention_heads: int = 20,
        token_dropout: bool = True,
    ):
        super().__init__()
        self.num_layers = num_layers
        self.embed_dim = embed_dim
        self.attention_heads = attention_heads
        self.alphabet = alphabet
        self.alphabet_size = len(alphabet)
        self.padding_idx = alphabet.padding_idx
        self.mask_idx = alphabet.mask_idx
        self.cls_idx = alphabet.cls_idx
        self.eos_idx = alphabet.eos_idx
        self.prepend_bos = alphabet.prepend_bos
        self.append_eos = alphabet.append_eos
        self.token_dropout = token_dropout

        self._init_submodules()

    def _init_submodules(self):
        self.embed_scale = 1
        self.embed_tokens = nn.Embedding(
            self.alphabet_size,
            self.embed_dim,
            padding_idx=self.padding_idx,
        )

        self.layers = nn.ModuleList(
            [
                TransformerLayer(
                    self.embed_dim,
                    4 * self.embed_dim,
                    self.attention_heads,
                    add_bias_kv=False,
                    use_esm1b_layer_norm=True,
                    use_rotary_embeddings=True,
                )
                for _ in range(self.num_layers)
            ]
        )

        self.contact_head = ContactPredictionHead(
            self.num_layers * self.attention_heads,
            self.prepend_bos,
            self.append_eos,
            eos_idx=self.eos_idx,
        )
        self.emb_layer_norm_after = ESM1bLayerNorm(self.embed_dim)

        self.lm_head = RobertaLMHead(
            embed_dim=self.embed_dim,
            output_dim=self.alphabet_size,
            weight=self.embed_tokens.weight,
        )

    def forward(self, tokens, repr_layers=[], need_head_weights=False, return_contacts=False):
        if return_contacts:
            need_head_weights = True

        assert tokens.ndim == 2
        padding_mask = tokens.eq(self.padding_idx)  # B, T

        x = self.embed_scale * self.embed_tokens(tokens)

        if self.token_dropout:
            x.masked_fill_((tokens == self.mask_idx).unsqueeze(-1), 0.0)
            # x: B x T x C
            mask_ratio_train = 0.15 * 0.8
            src_lengths = (~padding_mask).sum(-1)
            mask_ratio_observed = (tokens == self.mask_idx).sum(-1).to(x.dtype) / src_lengths
            x = x * (1 - mask_ratio_train) / (1 - mask_ratio_observed)[:, None, None]

        if padding_mask is not None:
            x = x * (1 - padding_mask.unsqueeze(-1).type_as(x))

        repr_layers = set(repr_layers)
        hidden_representations = {}
        if 0 in repr_layers:
            hidden_representations[0] = x

        if need_head_weights:
            attn_weights = []

        # (B, T, E) => (T, B, E)
        x = x.transpose(0, 1)

        if not padding_mask.any():
            padding_mask = None

        for layer_idx, layer in enumerate(self.layers):
            x, attn = layer(
                x,
                self_attn_padding_mask=padding_mask,
                need_head_weights=need_head_weights,
            )
            if (layer_idx + 1) in repr_layers:
                hidden_representations[layer_idx + 1] = x.transpose(0, 1)
            if need_head_weights:
                # (H, B, T, T) => (B, H, T, T)
                attn_weights.append(attn.transpose(1, 0))

        x = self.emb_layer_norm_after(x)
        x = x.transpose(0, 1)  # (T, B, E) => (B, T, E)

        # last hidden representation should have layer norm applied
        if (layer_idx + 1) in repr_layers:
            hidden_representations[layer_idx + 1] = x
        x = self.lm_head(x)

        result = {"logits": x, "representations": hidden_representations}
        if need_head_weights:
            # attentions: B x L x H x T x T
            attentions = torch.stack(attn_weights, 1)
            if padding_mask is not None:
                attention_mask = 1 - padding_mask.type_as(attentions)
                attention_mask = attention_mask.unsqueeze(1) * attention_mask.unsqueeze(2)
                attentions = attentions * attention_mask[:, None, None, :, :]
            result["attentions"] = attentions
            if return_contacts:
                contacts = self.contact_head(tokens, attentions)
                result["contacts"] = contacts

        return result

    def predict_contacts(self, tokens):
        return self(tokens, return_contacts=True)["contacts"]