Skip to content

feat: U-net style protein transformer #20

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
chungongyu wants to merge 1 commit into
base: main
Choose a base branch
from
Open

feat: U-net style protein transformer #20

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
322 changes: 322 additions & 0 deletions src/proxyz/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,322 @@
"""U-Net style protein language model.

Three-stage architecture:
-> AA encoder (residue granularity)
-> token transformmer (token granularity)
-> AA decoder (residue granularity)
=> predict next amino acid
"""

import random
from typing import List, Optional, Tuple

import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import LlamaConfig, LlamaDecoderLayer
from transformers import PreTrainedModel, PretrainedConfig
from transformers.modeling_outputs import CausalLMOutputWithPast

AA_VOCAB = "ACDEFGHIKLMNPQRSTVWY"
AA_TO_IDX = {aa: i for i, aa in enumerate(AA_VOCAB)}
IDX_TO_AA = {i: aa for i, aa in enumerate(AA_VOCAB)}
AA_VOCAB_SIZE = 20


# ======================================================================
# Config
# ======================================================================
class ProXYZConfig(PretrainedConfig):
model_type = "proxyz"

def __init__(
self,
vocab_size: int = 30000,
hidden_size: int = 2048,
intermediate_size: int = 5632,
num_attention_heads: int = 16,
num_key_value_heads: int = 4,
aa_vocab_size: int = AA_VOCAB_SIZE,
aa_encoder_layers: int = 4,
bpe_encoder_layers: int = 12,
aa_decoder_layers: int = 4,
max_position_embeddings: int = 4096,
rms_norm_eps: float = 1e-6,
rope_theta: float = 10000.0,
use_skip_connection: bool = True,
**kwargs,
):
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.num_attention_heads = num_attention_heads
self.num_key_value_heads = num_key_value_heads
self.aa_vocab_size = aa_vocab_size
self.aa_encoder_layers = aa_encoder_layers
self.bpe_encoder_layers = bpe_encoder_layers
self.aa_decoder_layers = aa_decoder_layers
self.max_position_embeddings = max_position_embeddings
self.rms_norm_eps = rms_norm_eps
self.rope_theta = rope_theta
self.use_skip_connection = use_skip_connection
super().__init__(**kwargs)


# ======================================================================
# Down / Up sampling
# ======================================================================
class _Downsample(nn.Module):
"""Pool amino-acid representations → BPE-token representations."""

def forward(self, aa_hidden: torch.Tensor, alignments: torch.Tensor) -> torch.Tensor:
B, M, _ = alignments.shape
D = aa_hidden.shape[-1]
device = aa_hidden.device

starts = alignments[:, :, 0] # (B, M)
ends = alignments[:, :, 1] # (B, M)
max_len = aa_hidden.shape[1]

pos = torch.arange(max_len, device=device).unsqueeze(0).unsqueeze(0)
mask = (pos >= starts.unsqueeze(-1)) & (pos < ends.unsqueeze(-1))

counts = mask.sum(dim=-1, keepdim=True).clamp(min=1) # (B, M, 1)
pooled = (aa_hidden.unsqueeze(1) * mask.unsqueeze(-1).float()).sum(dim=2) / counts
return pooled


class _Upsample(nn.Module):
"""Expand BPE representations → amino-acid granularity with shift+1 for causality."""

def forward(self, bpe_hidden: torch.Tensor, alignments: torch.Tensor, target_len: int) -> torch.Tensor:
B, M, D = bpe_hidden.shape
device = bpe_hidden.device

starts = alignments[:, :, 0] # (B, M)

# Map each AA position → its BPE token index, then shift -1 for causality
pos = torch.arange(target_len, device=device).unsqueeze(0).unsqueeze(0)
bpe_idx = (pos >= starts.unsqueeze(-1)).long().sum(dim=1) - 1 # (B, N)

gather_idx = bpe_idx.clamp(min=0).unsqueeze(-1).expand(-1, -1, D)
result = torch.gather(bpe_hidden, 1, gather_idx)

valid = (bpe_idx >= 0).unsqueeze(-1).float()
return result * valid


# ======================================================================
# Model
# ======================================================================
class ProXYZForCausalLM(PreTrainedModel):
"""U-Net style pLM: AA encoder -> token transformer -> AA decoder => predict next AA."""

config_class = ProXYZConfig
base_model_prefix = "proxyz"
supports_gradient_checkpointing = False

def __init__(self, config: ProXYZConfig):
super().__init__(config)

# --- Embeddings ---
self.aa_embedding = nn.Embedding(config.aa_vocab_size, config.hidden_size)
self.bpe_embedding = nn.Embedding(config.vocab_size, config.hidden_size)

# --- Shared LlamaDecoderLayer config ---
layer_cfg = LlamaConfig(
vocab_size=config.vocab_size,
hidden_size=config.hidden_size,
intermediate_size=config.intermediate_size,
num_hidden_layers=1,
num_attention_heads=config.num_attention_heads,
num_key_value_heads=config.num_key_value_heads,
max_position_embeddings=config.max_position_embeddings,
rms_norm_eps=config.rms_norm_eps,
rope_theta=config.rope_theta,
tie_word_embeddings=False,
)

# --- AA Encoder (residue granularity, causal) ---
self.aa_encoder = nn.ModuleList(
[LlamaDecoderLayer(layer_cfg, layer_idx=i) for i in range(config.aa_encoder_layers)]
)
self.aa_encoder_norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

# --- Down / Up ---
self.downsample = _Downsample()
self.upsample = _Upsample()

# --- BPE Encoder (token granularity, causal) ---
self.bpe_encoder = nn.ModuleList(
[LlamaDecoderLayer(layer_cfg, layer_idx=i) for i in range(config.bpe_encoder_layers)]
)
self.bpe_encoder_norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

# --- Skip projection (concat → hidden_size) ---
if config.use_skip_connection:
self.skip_proj = nn.Linear(config.hidden_size * 2, config.hidden_size, bias=False)

# --- AA Decoder (residue granularity, causal) ---
self.aa_decoder = nn.ModuleList(
[LlamaDecoderLayer(layer_cfg, layer_idx=i) for i in range(config.aa_decoder_layers)]
)
self.aa_decoder_norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

# --- LM Head → 20 amino acids ---
self.lm_head = nn.Linear(config.hidden_size, config.aa_vocab_size, bias=False)

# Initialize weights
self.post_init()

# ------------------------------------------------------------------
def _init_weights(self, module):
if isinstance(module, nn.Linear):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
elif isinstance(module, nn.RMSNorm):
nn.init.ones_(module.weight)

# ------------------------------------------------------------------
def forward(
self,
aa_input_ids: torch.Tensor,
bpe_input_ids: torch.Tensor,
alignments: torch.Tensor,
labels: Optional[torch.Tensor] = None,
aa_attention_mask: Optional[torch.Tensor] = None,
bpe_attention_mask: Optional[torch.Tensor] = None,
) -> CausalLMOutputWithPast:
"""
Args:
aa_input_ids: (B, N) amino-acid indices 0-19
bpe_input_ids: (B, M) BPE token ids (incl. BOS/EOS)
alignments: (B, M, 2) AA start/end per BPE token
labels: (B, N) next-AA targets, -100 = ignore
aa_attention_mask: (B, N) 1=real, 0=pad (None → all real)
bpe_attention_mask:(B, M) 1=real, 0=pad (None → all real)
"""
B, N = aa_input_ids.shape
M = bpe_input_ids.shape[1]
device = aa_input_ids.device

# 2-D masks: LlamaDecoderLayer internally builds causal + padding mask
if aa_attention_mask is None:
aa_attention_mask = aa_input_ids.new_ones(B, N)
if bpe_attention_mask is None:
bpe_attention_mask = bpe_input_ids.new_ones(B, M)

aa_pos = torch.arange(N, device=device).unsqueeze(0)
bpe_pos = torch.arange(M, device=device).unsqueeze(0)

# ====== AA Encoder ======
h = self.aa_embedding(aa_input_ids)
for layer in self.aa_encoder:
h = layer(h, attention_mask=aa_attention_mask, position_ids=aa_pos)[0]
aa_hidden = self.aa_encoder_norm(h) # (B, N, D)

# ====== Downsample: AA → BPE ======
bpe_from_aa = self.downsample(aa_hidden, alignments) # (B, M, D)

# ====== BPE Encoder ======
h = self.bpe_embedding(bpe_input_ids) + bpe_from_aa
for layer in self.bpe_encoder:
h = layer(h, attention_mask=bpe_attention_mask, position_ids=bpe_pos)[0]
bpe_hidden = self.bpe_encoder_norm(h) # (B, M, D)

# ====== Upsample: BPE → AA (shift +1 for causality) ======
upsampled = self.upsample(bpe_hidden, alignments, N) # (B, N, D)

# ====== Skip connection ======
if self.config.use_skip_connection:
h = self.skip_proj(torch.cat([aa_hidden, upsampled], dim=-1))
else:
h = upsampled

# ====== AA Decoder ======
for layer in self.aa_decoder:
h = layer(h, attention_mask=aa_attention_mask, position_ids=aa_pos)[0]
aa_decoded = self.aa_decoder_norm(h) # (B, N, D)

# ====== LM Head ======
logits = self.lm_head(aa_decoded) # (B, N, 20)

# ====== Loss ======
loss = None
if labels is not None:
loss = F.cross_entropy(
logits.reshape(-1, logits.size(-1)),
labels.reshape(-1),
ignore_index=-100,
)

return CausalLMOutputWithPast(loss=loss, logits=logits)

# ------------------------------------------------------------------
@torch.no_grad()
def generate(
self,
tokenizer,
prompt_aa: str = "",
max_length: int = 500,
temperature: float = 1.0,
top_k: int = 1,
) -> str:
"""Autoregressively generate amino acids one at a time."""
device = next(self.parameters()).device

aa_seq: List[int] = [AA_TO_IDX[aa] for aa in prompt_aa if aa in AA_TO_IDX]
if not aa_seq:
aa_seq = [random.randint(0, AA_VOCAB_SIZE - 1)]

for _ in range(max_length - len(aa_seq)):
N = len(aa_seq)
aa_tensor = torch.tensor([aa_seq], dtype=torch.long, device=device)

# Build BPE tokens from current AA string
aa_str = "".join(IDX_TO_AA[a] for a in aa_seq)
bpe_wrapped = f"{tokenizer.bos_token}{aa_str}{tokenizer.eos_token}"
bpe_enc = tokenizer(bpe_wrapped, return_offsets=True)
bpe_ids = torch.tensor([bpe_enc["input_ids"]], dtype=torch.long, device=device)

# Build alignment (content only, skip BOS/EOS)
offsets = bpe_enc["offsets"]
M_bpe = len(bpe_enc["input_ids"])
align = [[0, 0]] * M_bpe
for j in range(1, M_bpe - 1):
s, e = offsets[j]
if e > 0 and s < len(aa_str):
align[j] = [max(s, 0), min(e, len(aa_str))]

alignments = torch.tensor([align], dtype=torch.long, device=device)

outputs = self.forward(aa_tensor, bpe_ids, alignments)
next_logits = outputs.logits[0, -1, :].clone()

if temperature != 1.0:
next_logits = next_logits / temperature
if top_k > 1:
topk_vals, _ = torch.topk(next_logits, top_k)
next_logits[next_logits < topk_vals[-1]] = float("-inf")

probs = F.softmax(next_logits, dim=-1)
next_aa = torch.multinomial(probs, 1).item()
aa_seq.append(next_aa)

return "".join(IDX_TO_AA.get(a, "") for a in aa_seq)


# ======================================================================
# Utility
# ======================================================================
def build_alignments(tokenizer, text: str) -> Tuple[torch.Tensor, int]:
"""Tokenize *text* (without BOS/EOS) and return (alignments, num_bpe_tokens)."""
enc = tokenizer(text, return_offsets=True, add_special_tokens=False)
offsets = enc["offsets"]
M = len(enc["input_ids"])
alignments = [[0, 0]] * M
for j, (s, e) in enumerate(offsets):
alignments[j] = [s, e]
return torch.tensor(alignments, dtype=torch.long), M
Loading