File size: 12,207 Bytes
858826c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 | import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional, Dict, Any
import models.vocabulary as vocab # For event type IDs
class XPostEncoder(nn.Module):
""" Encodes: <AuthorWalletEmbedding>, <PostTextEmbedding>, <MediaEmbedding> """
def __init__(self, d_model: int, dtype: torch.dtype):
super().__init__()
# Input: Wallet (d_model) + Text (d_model) + Media (d_model)
self.mlp = nn.Sequential(
nn.Linear(d_model * 3, d_model * 2),
nn.GELU(),
nn.LayerNorm(d_model * 2),
nn.Linear(d_model * 2, d_model)
).to(dtype)
def forward(self, author_emb: torch.Tensor, text_emb: torch.Tensor, media_emb: torch.Tensor) -> torch.Tensor:
combined = torch.cat([author_emb, text_emb, media_emb], dim=-1)
return self.mlp(combined)
class XRetweetEncoder(nn.Module):
""" Encodes: <RetweeterWalletEmbedding>, <OriginalAuthorWalletEmbedding>, <OriginalPostTextEmbedding>, <OriginalPostMediaEmbedding> """
def __init__(self, d_model: int, dtype: torch.dtype):
super().__init__()
# Input: Retweeter (d_model) + Original Author (d_model) + Original Text (d_model) + Original Media (d_model)
self.mlp = nn.Sequential(
nn.Linear(d_model * 4, d_model * 2),
nn.GELU(),
nn.LayerNorm(d_model * 2),
nn.Linear(d_model * 2, d_model)
).to(dtype)
def forward(self,
retweeter_emb: torch.Tensor,
orig_author_emb: torch.Tensor,
orig_text_emb: torch.Tensor,
orig_media_emb: torch.Tensor) -> torch.Tensor:
combined = torch.cat([retweeter_emb, orig_author_emb, orig_text_emb, orig_media_emb], dim=-1)
return self.mlp(combined)
class XReplyEncoder(nn.Module):
""" Encodes: <AuthorWalletEmbedding>, <PostTextEmbedding>, <MediaEmbedding>, <MainTweetEmbedding> """
def __init__(self, d_model: int, dtype: torch.dtype):
super().__init__()
# Input: Author (d_model) + Reply Text (d_model) + Reply Media (d_model) + Main Tweet Text (d_model)
self.mlp = nn.Sequential(
nn.Linear(d_model * 4, d_model * 2),
nn.GELU(),
nn.LayerNorm(d_model * 2),
nn.Linear(d_model * 2, d_model)
).to(dtype)
def forward(self,
author_emb: torch.Tensor,
text_emb: torch.Tensor,
media_emb: torch.Tensor,
main_tweet_emb: torch.Tensor) -> torch.Tensor:
combined = torch.cat([author_emb, text_emb, media_emb, main_tweet_emb], dim=-1)
return self.mlp(combined)
class XQuoteTweetEncoder(nn.Module):
""" Encodes: <QuoterWalletEmbedding>, <QuoterTextEmbedding>, <OriginalAuthorWalletEmbedding>, <OriginalPostTextEmbedding>, <OriginalPostMediaEmbedding> """
def __init__(self, d_model: int, dtype: torch.dtype):
super().__init__()
# Input: Quoter Wallet (d_model) + Quoter Text (d_model) + Orig Author (d_model) + Orig Text (d_model) + Orig Media (d_model)
self.mlp = nn.Sequential(
nn.Linear(d_model * 5, d_model * 2),
nn.GELU(),
nn.LayerNorm(d_model * 2),
nn.Linear(d_model * 2, d_model)
).to(dtype)
def forward(self,
quoter_wallet_emb: torch.Tensor,
quoter_text_emb: torch.Tensor,
orig_author_emb: torch.Tensor,
orig_text_emb: torch.Tensor,
orig_media_emb: torch.Tensor) -> torch.Tensor:
combined = torch.cat([quoter_wallet_emb, quoter_text_emb, orig_author_emb, orig_text_emb, orig_media_emb], dim=-1)
return self.mlp(combined)
class PumpReplyEncoder(nn.Module):
""" Encodes: <UserWalletEmbedding>, <ReplyTextEmbedding> """
def __init__(self, d_model: int, dtype: torch.dtype):
super().__init__()
# Input: User Wallet (d_model) + Reply Text (d_model)
self.mlp = nn.Sequential(
nn.Linear(d_model * 2, d_model * 2),
nn.GELU(),
nn.LayerNorm(d_model * 2),
nn.Linear(d_model * 2, d_model)
).to(dtype)
def forward(self, user_emb: torch.Tensor, text_emb: torch.Tensor) -> torch.Tensor:
combined = torch.cat([user_emb, text_emb], dim=-1)
return self.mlp(combined)
# --- NEW: Encoders for other text-based events ---
class DexProfileUpdatedEncoder(nn.Module):
""" Encodes: <4_flags_projection>, <website_emb>, <twitter_emb>, <telegram_emb>, <description_emb> """
def __init__(self, d_model: int, dtype: torch.dtype):
super().__init__()
# Input: flags_proj (d_model) + 4x text_embeds (d_model)
self.mlp = nn.Sequential(
nn.Linear(d_model * 4, d_model * 2), # Corrected from 5 to 4, flags are separate
nn.GELU(),
nn.LayerNorm(d_model * 2),
nn.Linear(d_model * 2, d_model)
).to(dtype)
def forward(self, website_emb: torch.Tensor, twitter_emb: torch.Tensor, telegram_emb: torch.Tensor, description_emb: torch.Tensor) -> torch.Tensor:
combined = torch.cat([website_emb, twitter_emb, telegram_emb, description_emb], dim=-1)
return self.mlp(combined)
class GlobalTrendingEncoder(nn.Module):
""" Encodes: <hashtag_emb> """
def __init__(self, d_model: int, dtype: torch.dtype):
super().__init__()
# Input: hashtag_emb (d_model)
self.mlp = nn.Sequential(
nn.Linear(d_model, d_model),
nn.GELU(),
nn.Linear(d_model, d_model)
).to(dtype)
def forward(self, hashtag_emb: torch.Tensor) -> torch.Tensor:
return self.mlp(hashtag_emb)
class SocialEncoder(nn.Module):
"""
A single module to house all social event encoders.
This simplifies instantiation in the main Oracle model.
"""
def __init__(self, d_model: int, dtype: torch.dtype):
super().__init__()
self.x_post_encoder = XPostEncoder(d_model, dtype)
self.x_retweet_encoder = XRetweetEncoder(d_model, dtype)
self.x_reply_encoder = XReplyEncoder(d_model, dtype)
self.x_quote_tweet_encoder = XQuoteTweetEncoder(d_model, dtype)
self.pump_reply_encoder = PumpReplyEncoder(d_model, dtype)
# --- NEW: Add the other text-based encoders ---
self.dex_profile_encoder = DexProfileUpdatedEncoder(d_model, dtype)
self.global_trending_encoder = GlobalTrendingEncoder(d_model, dtype)
# Store for convenience
self.d_model = d_model
self.dtype = dtype
def forward(self, batch: Dict[str, Any], gathered_embeds: Dict[str, torch.Tensor]) -> torch.Tensor:
"""
REFACTORED: Processes all text-based events for the entire batch in a vectorized way.
This replaces the inefficient loops in the main Oracle model.
"""
device = gathered_embeds['wallet'].device
B, L, D = gathered_embeds['wallet'].shape
final_embeds = torch.zeros(B, L, D, device=device, dtype=self.dtype)
textual_event_indices = batch['textual_event_indices']
textual_event_data = batch.get('textual_event_data', [])
precomputed_lookup = gathered_embeds['precomputed']
zero_emb = torch.zeros(self.d_model, device=device, dtype=self.dtype)
# --- Create masks for each event type ---
event_type_ids = batch['event_type_ids']
event_masks = {
'XPost': (event_type_ids == vocab.EVENT_TO_ID.get('XPost', -1)),
'XReply': (event_type_ids == vocab.EVENT_TO_ID.get('XReply', -1)),
'XRetweet': (event_type_ids == vocab.EVENT_TO_ID.get('XRetweet', -1)),
'XQuoteTweet': (event_type_ids == vocab.EVENT_TO_ID.get('XQuoteTweet', -1)),
'PumpReply': (event_type_ids == vocab.EVENT_TO_ID.get('PumpReply', -1)),
'DexProfile_Updated': (event_type_ids == vocab.EVENT_TO_ID.get('DexProfile_Updated', -1)),
'TikTok_Trending_Hashtag': (event_type_ids == vocab.EVENT_TO_ID.get('TikTok_Trending_Hashtag', -1)),
'XTrending_Hashtag': (event_type_ids == vocab.EVENT_TO_ID.get('XTrending_Hashtag', -1)),
}
# --- Gather all necessary pre-computed embeddings in one go ---
# Flatten indices for efficient lookup, then reshape
flat_indices = textual_event_indices.flatten()
# Create a default event structure for padding indices (idx=0)
default_event = {'event_type': 'PAD'}
# Use 1-based index from collator, so textual_event_data[idx-1]
raw_events_flat = [textual_event_data[idx-1] if idx > 0 else default_event for idx in flat_indices.tolist()]
# Helper to gather embeddings for a specific key
def gather_precomputed(key: str) -> torch.Tensor:
indices = torch.tensor([e.get(key, 0) for e in raw_events_flat], device=device, dtype=torch.long)
return F.embedding(indices, precomputed_lookup).view(B, L, -1)
# --- Process each event type ---
# XPost
if event_masks['XPost'].any():
text_emb = gather_precomputed('text_emb_idx')
media_emb = gather_precomputed('media_emb_idx')
post_embeds = self.x_post_encoder(gathered_embeds['wallet'], text_emb, media_emb)
final_embeds += post_embeds * event_masks['XPost'].unsqueeze(-1)
# XReply
if event_masks['XReply'].any():
text_emb = gather_precomputed('text_emb_idx')
media_emb = gather_precomputed('media_emb_idx')
main_tweet_emb = gather_precomputed('main_tweet_text_emb_idx')
reply_embeds = self.x_reply_encoder(gathered_embeds['wallet'], text_emb, media_emb, main_tweet_emb)
final_embeds += reply_embeds * event_masks['XReply'].unsqueeze(-1)
# XRetweet
if event_masks['XRetweet'].any():
orig_text_emb = gather_precomputed('original_post_text_emb_idx')
orig_media_emb = gather_precomputed('original_post_media_emb_idx')
retweet_embeds = self.x_retweet_encoder(gathered_embeds['wallet'], gathered_embeds['original_author'], orig_text_emb, orig_media_emb)
final_embeds += retweet_embeds * event_masks['XRetweet'].unsqueeze(-1)
# XQuoteTweet
if event_masks['XQuoteTweet'].any():
quoter_text_emb = gather_precomputed('quoter_text_emb_idx')
orig_text_emb = gather_precomputed('original_post_text_emb_idx')
orig_media_emb = gather_precomputed('original_post_media_emb_idx')
quote_embeds = self.x_quote_tweet_encoder(gathered_embeds['wallet'], quoter_text_emb, gathered_embeds['original_author'], orig_text_emb, orig_media_emb)
final_embeds += quote_embeds * event_masks['XQuoteTweet'].unsqueeze(-1)
# PumpReply
if event_masks['PumpReply'].any():
text_emb = gather_precomputed('reply_text_emb_idx')
pump_reply_embeds = self.pump_reply_encoder(gathered_embeds['wallet'], text_emb)
final_embeds += pump_reply_embeds * event_masks['PumpReply'].unsqueeze(-1)
# DexProfile_Updated
if event_masks['DexProfile_Updated'].any():
website_emb = gather_precomputed('website_emb_idx')
twitter_emb = gather_precomputed('twitter_link_emb_idx')
telegram_emb = gather_precomputed('telegram_link_emb_idx')
description_emb = gather_precomputed('description_emb_idx')
profile_embeds = self.dex_profile_encoder(website_emb, twitter_emb, telegram_emb, description_emb)
# Note: The flags are handled separately in the main model now, so we just add the text embeds
final_embeds += profile_embeds * event_masks['DexProfile_Updated'].unsqueeze(-1)
# Global Trending Hashtags
trending_mask = event_masks['TikTok_Trending_Hashtag'] | event_masks['XTrending_Hashtag']
if trending_mask.any():
hashtag_emb = gather_precomputed('hashtag_name_emb_idx')
trending_embeds = self.global_trending_encoder(hashtag_emb)
final_embeds += trending_embeds * trending_mask.unsqueeze(-1)
return final_embeds |