File size: 12,008 Bytes
acbfbc3 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 | from transformers import (
AutoProcessor,
AutoTokenizer,
)
from typing import Dict, Any, Union, List, Optional, Callable, Type
from trl.data_utils import maybe_apply_chat_template
import torch
from bioreason.dna_modules.dna_module import DNABaseModule
from model.blip2_stage2 import Blip2Stage2
class Blip2DNAModule(DNABaseModule):
"""
DNA module implementation for BLIP2-based models.
This module provides the interface between BLIP2 models and the GRPO training
infrastructure, handling model loading, processing setup, and reward functions.
"""
def __init__(self):
"""Initialize the Blip2DNAModule."""
super().__init__()
def get_dnallm_key(self) -> str:
"""
Get the key identifier for this DNA-LLM implementation.
Returns:
String identifier for this module type
"""
return "blip2"
def get_model_class(self, model_id: str, model_init_kwargs: Dict[str, Any]) -> Type:
"""
Return the appropriate model class based on model ID.
Args:
model_id: Identifier for the model
model_init_kwargs: Initialization arguments for the model
Returns:
The model class to instantiate
Raises:
ValueError: If the model is not supported
"""
if "blip2" in model_id.lower() or "stage2" in model_id.lower():
model_cls = Blip2Stage2
else:
raise ValueError(f"Unsupported model: {model_id}")
return model_cls
def post_model_init(self, model: Any, processing_class: Any) -> None:
"""
Perform any post-initialization setup on the model.
Args:
model: The initialized model
processing_class: The processor for the model
"""
# BLIP2 models might need specific post-init setup
if hasattr(model, 'blip2') and hasattr(model.blip2, 'llm_tokenizer'):
# Ensure the tokenizer is properly configured
if not hasattr(model.blip2.llm_tokenizer, 'pad_token') or model.blip2.llm_tokenizer.pad_token is None:
model.blip2.llm_tokenizer.pad_token = model.blip2.llm_tokenizer.eos_token
def get_processing_class(self) -> Type:
"""
Get the processing class to use with this BLIP2 model.
Returns:
The processing class
"""
return Blip2Processor
def get_dnallm_modules_keywords(self) -> List[str]:
"""
Get keywords to identify DNA-specific modules in the model.
Used to exclude DNA modules from LoRA adaptation during training.
Returns:
List of keywords that identify DNA modules
"""
return ["plm", "qformer", "opt_proj"]
def get_custom_multimodal_keywords(self) -> List[str]:
"""
Get keywords for multimodal inputs that should be passed to the model.
Returns:
List of input keywords for multimodal processing
"""
return ["prot_batch", "prompt_batch"]
def get_non_generate_params(self) -> List[str]:
"""
Get parameter names that should be excluded from generation.
Returns:
List of parameter names to exclude from generation calls
"""
return ["prot_batch"]
def get_custom_processing_keywords(self) -> List[tuple]:
"""
Get custom processing keywords for the processor.
Returns:
List of (component, parameter) tuples for custom processing
"""
return [("plm_tokenizer", "max_length"), ("llm_tokenizer", "max_length")]
def prepare_prompt(
self, processing_class: Any, inputs: List[Dict[str, Union[torch.Tensor, Any]]]
) -> List[str]:
"""
Prepare prompts from input examples.
Args:
processing_class: The processor to use
inputs: List of input examples
Returns:
List of prepared prompts
"""
prompts_text = []
for example in inputs:
if "prompt" in example:
# Extract text content from conversational format
if isinstance(example["prompt"], list) and len(example["prompt"]) > 0:
user_content = example["prompt"][0].get("content", "")
if isinstance(user_content, list):
# Extract text from multimodal content
text_parts = [item.get("text", "") for item in user_content if item.get("type") == "text"]
prompt_text = " ".join(text_parts)
else:
prompt_text = str(user_content)
else:
prompt_text = str(example["prompt"])
else:
prompt_text = ""
prompts_text.append(prompt_text)
return prompts_text
def prepare_model_inputs(
self,
processing_class: Any,
model: Any,
prompts_text: List[str],
batch_dna_sequences: List[List[str]],
return_tensors: str = "pt",
padding: bool = True,
padding_side: str = "left",
add_special_tokens: bool = False,
) -> Dict[str, Any]:
"""
Prepare inputs for the BLIP2 model.
Args:
processing_class: The processor to use
model: The model to prepare inputs for
prompts_text: List of text prompts
batch_dna_sequences: List of lists of DNA sequences (treated as protein sequences)
return_tensors: Return format for tensors
padding: Whether to pad inputs
padding_side: Side to pad on
add_special_tokens: Whether to add special tokens
Returns:
Processed inputs for the model
"""
# Get the BLIP2 model from the wrapper
blip2_model = model.blip2 if hasattr(model, 'blip2') else model
# Prepare protein batch (using DNA sequences as protein sequences)
# Flatten all DNA sequences to treat them as individual protein sequences
all_sequences = []
for sequences in batch_dna_sequences:
all_sequences.extend(sequences)
if all_sequences:
prot_batch = blip2_model.plm_tokenizer(
all_sequences,
padding=padding,
truncation=True,
max_length=512, # Default protein sequence length
return_tensors=return_tensors,
)
else:
# Empty batch handling
prot_batch = {
'input_ids': torch.empty(0, 1, dtype=torch.long),
'attention_mask': torch.empty(0, 1, dtype=torch.long)
}
# Prepare prompt batch
prompt_batch = blip2_model.llm_tokenizer(
prompts_text,
padding=padding,
truncation=True,
max_length=256, # Default prompt length
return_tensors=return_tensors,
)
return {
"prot_batch": prot_batch,
"prompt_batch": prompt_batch,
"input_ids": prompt_batch["input_ids"], # For compatibility
"attention_mask": prompt_batch["attention_mask"], # For compatibility
}
def is_embeds_input(self) -> bool:
"""
Whether the model uses embeddings as input (instead of token IDs).
Returns:
Boolean indicating if the model takes embedding inputs
"""
return True # BLIP2 uses embeddings internally
@staticmethod
def get_question_template() -> str:
"""
Get the template for formatting questions.
Returns:
String template for questions
"""
return "{Question}"
@staticmethod
def format_reward_rec(completions: List[Dict[str, Any]], **kwargs) -> List[float]:
"""
Check if the BLIP2 model output matches a specific format.
Args:
completions: List of model completions
**kwargs: Additional arguments
Returns:
List of reward scores (1.0 for match, 0.0 for no match)
"""
import re
import os
from datetime import datetime
# Pattern to match the expected output format
pattern = r"<think>.*?</think>\s*<answer>.*?\{.*\[\d+,\s*\d+,\s*\d+,\s*\d+\].*\}.*?</answer>"
completion_contents = [completion[0]["content"] for completion in completions]
matches = [
re.search(pattern, content, re.DOTALL) is not None
for content in completion_contents
]
# Log format results if in debug mode
current_time = datetime.now().strftime("%d-%H-%M-%S-%f")
if os.getenv("DEBUG_MODE") == "true":
log_path = os.getenv("LOG_PATH")
with open(
log_path.replace(".txt", "_format.txt"), "a", encoding="utf-8"
) as f:
f.write(f"------------- {current_time} Format reward -------------\n")
for content, match in zip(completion_contents, matches):
f.write(f"Content: {content}\n")
f.write(f"Has format: {bool(match)}\n")
return [1.0 if match else 0.0 for match in matches]
@staticmethod
def select_reward_func(func: str, task_type: str) -> Callable:
"""
Select the appropriate reward function based on function name and task type.
Args:
func: The type of reward function ('accuracy', 'format', etc.)
task_type: The type of task ('rec', etc.)
Returns:
The reward function to use
Raises:
ValueError: If the function or task type is not supported
"""
if func == "accuracy":
match task_type:
case "rec":
return Blip2DNAModule.iou_reward
case _:
raise ValueError(f"Unsupported reward function: {func}")
elif func == "format":
match task_type:
case "rec":
return Blip2DNAModule.format_reward_rec
case _:
raise ValueError(f"Unsupported reward function: {func}")
else:
raise ValueError(f"Unsupported reward function: {func}")
@staticmethod
def iou_reward(completions: List[Dict[str, Any]], **kwargs) -> List[float]:
"""
Placeholder IoU reward function.
Args:
completions: List of model completions
**kwargs: Additional arguments
Returns:
List of reward scores
"""
# Placeholder implementation
return [1.0] * len(completions)
class Blip2Processor:
"""
Simple processor wrapper for BLIP2 models to maintain compatibility
with the GRPO trainer interface.
"""
def __init__(self, plm_tokenizer=None, llm_tokenizer=None):
self.plm_tokenizer = plm_tokenizer
self.llm_tokenizer = llm_tokenizer
# Set compatibility attributes
if llm_tokenizer:
self.eos_token_id = llm_tokenizer.eos_token_id
self.pad_token_id = llm_tokenizer.pad_token_id
def __call__(self, *args, **kwargs):
"""
Process inputs for BLIP2 model.
This is a simplified version that delegates to the appropriate tokenizer.
"""
# For compatibility, return a simple tokenization result
if self.llm_tokenizer:
return self.llm_tokenizer(*args, **kwargs)
else:
# Fallback behavior
return {"input_ids": torch.tensor([[1]]), "attention_mask": torch.tensor([[1]])}
def batch_decode(self, *args, **kwargs):
"""Decode token sequences."""
if self.llm_tokenizer:
return self.llm_tokenizer.batch_decode(*args, **kwargs)
else:
return [""] |