File size: 19,973 Bytes
cbff41a |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 |
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
import os
import re
import torch
import random
import pandas as pd
import numpy as np
from tqdm import tqdm
from torch.utils.data import DataLoader
from transformers import AutoTokenizer, HfArgumentParser
from utils.utils import seed_everything
from typing import Optional
from dataclasses import dataclass, field
from utils.data_collator import MyDataCollatorForWPathVLM
from datasets import load_dataset, concatenate_datasets, load_from_disk
from model.my_model import WPathVLM
from model.my_model_vision import WPathVLM as WPathVLM_Vision
from peft import LoraConfig, get_peft_model
from utils.formatting_funcs import wsi_formatting_des_test, wsi_formatting_qa_open_test, wsi_formatting_qa_close_test,wsi_formatting_qa_open, wsi_formatting_qa_close
from utils.eval_utils import calculate_prf_score, compute_bleu_scores, split_sentence, compute_cider_scores, compute_spice_scores
from transformers import BatchEncoding
device = 'cuda'
@dataclass
class ScriptArguments:
"""
The name of the Casual LM model we wish to fine-tune with SFTTrainer.
"""
# System config
load_in_8bit: Optional[bool] = field(default=False, metadata={"help": "Load the model in 8 bits precision"})
load_in_4bit: Optional[bool] = field(default=False, metadata={"help": "Load the model in 4 bits precision"})
trust_remote_code: Optional[bool] = field(default=False, metadata={"help": "Enable `trust_remote_code`"})
token: Optional[bool] = field(default=True, metadata={"help": "Use HF auth token to access the model"})
seed: Optional[int] = field(default=42, metadata={"help": "Random seed"})
# Model
llm_name: Optional[str] = field(default="meta-llama/Meta-Llama-3-8B", metadata={"help": "The model name"})
max_seq_length: Optional[int] = field(default=512, metadata={"help": "Input sequence length"})
llm_requires_grad: Optional[bool] = field(default=False, metadata={"help": "True or /output/checkpoint-1400"})
vision_adaptor: Optional[bool] = field(default=False, metadata={"help": "True or False, only for longnet and qformer"})
hierachical_token: Optional[bool] = field(default=True, metadata={"help": "True or False"})
hierachical_adaptor: Optional[bool] = field(default=True, metadata={"help": "True or False, only for longnet and qformer"})
# Data
select_data_num: Optional[int] = field(default=-1, metadata={"help": "the number of training data, -1 mean use all data"})
dataset_name_list: Optional[str] = field(default="CNX-PathLLM/TCGA-WSI-Text", metadata={"help": "Dataset names separated by comma"})
dataset_text_field: Optional[str] = field(default="text", metadata={"help": "The text field of the dataset"})
data_local_dir: Optional[str] = field(default=None, metadata={"help": "Local directory to load data from"})
eval_fold_index: Optional[int] = field(default=9, metadata={"help": "The test fold index"})
data_cache_dir: Optional[str] = field(default="~/.cache", metadata={"help": "Cache directory for dataset and model"})
results_save_path: Optional[str] = field(default="output.csv", metadata={"help": "the save path for prediction results"})
fea_root: Optional[str] = field(default="/bask/homes/a/asiw9691/PathVLM/WSI_Dataset/Conch/", metadata={"help": "the root path for WSI feature"})
gmm_root: Optional[str] = field(default="/data_local/pxb/CNX-PathLLM/GMM_PT", metadata={"help": "the root path for WSI feature"})
# WSI hyperparameters
n_heads: Optional[str] = field(default='32,16,8', metadata={"help": "Number of attention heads for WSI aggregation"})
n_level: Optional[int] = field(default=3, metadata={"help": "Number of hierarchical levels for WSI embedding"})
embed_dim: Optional[int] = field(default=512, metadata={"help": "Embedding dimension of each patch"})
agg_strategy: Optional[str] = field(default='abmil', metadata={"help": "the strategy for WSI aggregation, sample, kmeans, gmm, abmil"})
# Evaluation
batch_size: Optional[int] = field(default=4, metadata={"help": "Batch size"})
ckpt_path: Optional[str] = field(default=None, metadata={"help": "Checkpoint path"})
shuffle: Optional[bool] = field(default=False, metadata={"help": "shuffle eval_dataloader or not"})
eval_sample_size: Optional[int] = field(default=-1, metadata={"help": "-1 indicate evaluating on all"})
#lora
use_peft: Optional[bool] = field(default=False, metadata={"help": "Wether to use PEFT or not to train adapters"})
peft_lora_r: Optional[int] = field(default=64, metadata={"help": "the r parameter of the LoRA adapters, 4 to 64"})
peft_lora_alpha: Optional[int] = field(default=16, metadata={"help": "the alpha parameter of the LoRA adapters, 16 to 128"})
def tokenize(element):
"""Tokenize the input text."""
outputs = tokenizer(
element,
add_special_tokens=True,
truncation=True,
padding=False,
max_length=script_args.max_seq_length,
return_overflowing_tokens=False,
return_length=False
)
return {"input_ids": outputs["input_ids"], "attention_mask": outputs["attention_mask"]}
def move_to(obj, device="cuda"):
if torch.is_tensor(obj):
return obj.to(device)
elif isinstance(obj, dict) or isinstance(obj,BatchEncoding):
return {k: move_to(v, device) for k, v in obj.items()}
elif isinstance(obj, list):
return [move_to(v, device) for v in obj]
elif isinstance(obj, tuple):
return tuple(move_to(v, device) for v in obj)
else:
return obj
def tokenize_instruct(element):
"""Tokenize the input text."""
outputs = tokenizer(
element,
add_special_tokens=True,
truncation=True,
padding=False,
max_length=script_args.max_seq_length,
return_overflowing_tokens=False,
return_length=False
)
return {"input_ids_instruct": outputs["input_ids"], "attention_mask_instruct": outputs["attention_mask"]}
def evaluate_model(model, eval_dataloader, script_args, mode='open'):
"""
Evaluate the model on the provided data loader and save results.
Parameters:
model: The model to evaluate.
eval_dataloader: DataLoader for evaluation data.
device: The device (CPU or GPU) on which to perform computations.
script_args: Arguments containing settings for evaluation, e.g., eval_sample_size and results_save_path.
"""
slide_id_list, qes_list, ans_list, res_list = [], [], [], []
total_batches = len(eval_dataloader)
N = script_args.eval_sample_size if script_args.eval_sample_size != -1 else total_batches
for i, batch in enumerate(tqdm(eval_dataloader, total=N, desc="Evaluating")):
if i >= N:
break
# Move inputs to the device
# input_ids = batch['input_ids'].to(device)
# attention_masks = batch['attention_mask'].to(device)
# input_ids_instruct = batch['input_ids_instruct'].to(device)
# attention_mask_instruct = batch['attention_mask_instruct'].to(device)
# fea0, fea1, fea2 = batch['fea0'].to(device), batch['fea1'].to(device), batch['fea2'].to(device)
# cor0, cor1, cor2 = batch['cor0'].to(device), batch['cor1'].to(device), batch['cor2'].to(device)
# mask0, mask1, mask2 = batch['mask0'].to(device), batch['mask1'].to(device), batch['mask2'].to(device)
# input_ids_instruct = batch["input_ids_instruct"].to(device)
# attention_mask_instruct = batch["attention_mask_instruct"].to(device)
questions = batch[script_args.agg_strategy.split(",")[0]]['questions']
answers = batch[script_args.agg_strategy.split(",")[0]]['answers']
slide_ids = batch[script_args.agg_strategy.split(",")[0]]['slide_ids']
batch = move_to(obj=batch)
# Model inference
res = model.generate(
**batch
)
# Collect results
slide_id_list.extend(slide_ids)
qes_list.extend(questions)
ans_list.extend(answers)
res_list.extend(res)
# Save results in a dictionary
results = {
"slide_id": slide_id_list,
"question": qes_list,
"answer": ans_list,
"prediction": res_list,
}
# Evaluate using the specified metrics function
if mode == 'open':
metrics_open_ended(res_list, ans_list, script_args)
else:
metrics_close_ended(res_list, ans_list, script_args)
# Save results to a CSV file
df_results = pd.DataFrame(results)
slide_metadata_1 = pd.read_csv("./dataset_csv/gtex_slide_url_info.csv")[['slide_id', 'slide_url']]
slide_metadata_2 = pd.read_csv("./dataset_csv/tcga_slide_url_info.csv")[['slide_id', 'slide_url']]
slide_metadata = pd.concat([slide_metadata_1, slide_metadata_2], axis=0)
df_results = df_results.merge(slide_metadata, on='slide_id', how='left')
filename, ext = os.path.splitext(script_args.results_save_path)
if mode == 'open':
save_path = f"{filename}_open{ext}"
else:
save_path = f"{filename}_close{ext}"
df_results.to_csv(save_path, index=False)
print(f"Results saved to {save_path}")
def metrics_open_ended(open_candidate, open_reference, script_args):
open_ques_pre = []
open_ques_rec = []
open_ques_f1 = []
open_bleu_score = []
# calculate f1 score for open ended problem
for i in range(len(open_reference)):
precision, recall, f1_score = calculate_prf_score(open_candidate[i], open_reference[i])
open_ques_pre.append(precision)
open_ques_rec.append(recall)
open_ques_f1.append(f1_score)
open_bleu_score = compute_bleu_scores(open_candidate, open_reference, avg=True)
open_cider_score = compute_cider_scores(open_candidate, open_reference)
# open_spice_score = compute_spice_scores(open_candidate, open_reference)
open_ques_pre = np.mean(open_ques_pre)
open_ques_rec = np.mean(open_ques_rec)
open_ques_f1 = np.mean(open_ques_f1)
print("open question macro f1_score: {}\n".format(open_ques_f1))
print("open question macro precision: {}\n".format(open_ques_pre))
print("open question macro recall: {}\n".format(open_ques_rec))
print("open question bleu score: {}\n".format(open_bleu_score))
print("open question cider score: {}\n".format(open_cider_score))
# print("open question spice score: {}\n".format(open_spice_score))
output_path = script_args.ckpt_path.replace('.bin', '.txt')
with open(output_path, 'a') as file:
file.write("open question macro f1_score: {}\n".format(open_ques_f1))
file.write("open question macro precision: {}\n".format(open_ques_pre))
file.write("open question macro recall: {}\n".format(open_ques_rec))
file.write("open question bleu score: {}\n".format(open_bleu_score))
file.write("open question cider score: {}\n".format(open_cider_score))
# file.write("open question spice score: {}\n".format(open_spice_score))
print("Results have been written to {}".format(output_path))
def metrics_close_ended(close_candidate, close_reference, script_args):
correct_predictions = 0
total_predictions = len(close_candidate)
for answer, result in zip(close_candidate, close_reference):
answer = answer.strip().lower()
result = str(result).strip().lower()
# 判断题的匹配逻辑(“yes”或“no”)
if answer in ["yes", "no"]:
if answer in result:
correct_predictions += 1
else:
answer_choice = re.search(r'\b([a-d])\.\s?', answer)
if answer_choice:
answer_choice = answer_choice.group(1)
result_choice = re.search(r'\b([a-d])\.\s?', result)
if result_choice:
result_choice = result_choice.group(1)
else:
result_choice = re.search(r'\b([1-4])\.\s?', result)
choice_map = {'1': 'a', '2': 'b', '3': 'c', '4': 'd'}
if result_choice:
result_choice = choice_map[result_choice.group(1)]
if answer_choice == result_choice:
correct_predictions += 1
print("close question accuracy: {}\n".format(correct_predictions/total_predictions))
output_path = script_args.ckpt_path.replace('.bin', '.txt')
with open(output_path, 'a') as file:
file.write("close ended accuracy: {}\n".format(correct_predictions/total_predictions))
print("Results have been written to {}".format(output_path))
# Parse arguments and set seed
parser = HfArgumentParser(ScriptArguments)
script_args = parser.parse_args_into_dataclasses()[0]
seed_everything(script_args.seed)
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(script_args.llm_name)
# tokenizer.pad_token = tokenizer.eos_token
tokenizer.pad_token = "<|finetune_right_pad_id|>"
tokenizer.padding_side = 'left'
tokenizer.truncation_side = 'left'
# Add new tokens
if script_args.hierachical_token:
new_tokens = ['<|Question|>', '<|Prompt|>', '<|Answer|>', '<|Image|>', '<|High|>', '<|`Mid`|>', '<|Low|>']
else:
new_tokens = ['<|Question|>', '<|Prompt|>', '<|Answer|>', '<|Image|>']
# num_added_toks = tokenizer.add_tokens(new_tokens)
num_added_toks = tokenizer.add_special_tokens({"additional_special_tokens": new_tokens})
new_tokens_ids = tokenizer.convert_tokens_to_ids(new_tokens)
print("New tokens IDs:", new_tokens_ids)
# Determine data split
split_text = f"test[:{script_args.select_data_num}]" if script_args.select_data_num > 0 else "test"
# if script_args.data_local_dir is None:
open_dataset = []
close_dataset = []
for dataset_name in script_args.dataset_name_list.split(","):
# columns_to_remove = ['slide_id'] # keep slide id
columns_to_remove = []
one_dataset = load_dataset(dataset_name, split=split_text, cache_dir=script_args.data_cache_dir)
# one_dataset.cleanup_cache_files()
if 'project' in one_dataset.column_names:
columns_to_remove.append('project')
elif 'site' in one_dataset.column_names:
columns_to_remove.append('site')
if 'QA' in dataset_name: # for QA instruction dataset
# columns_to_remove += ['question']
if 'Open' in dataset_name: # for OpenQA instruction dataset
print("********MAPPED********")
one_dataset = one_dataset.map(wsi_formatting_qa_open_test, fn_kwargs={'tokenizer': tokenizer},
num_proc=20, remove_columns=columns_to_remove)
open_dataset.append(one_dataset)
else: # for CloseQA instruction dataset
one_dataset = one_dataset.map(wsi_formatting_qa_close_test, fn_kwargs={'tokenizer': tokenizer, 'prompt_tag': False},
num_proc=20, remove_columns=columns_to_remove)
close_dataset.append(one_dataset)
else:
columns_to_remove += ['description']
one_dataset = one_dataset.map(wsi_formatting_des_test, fn_kwargs={'tokenizer': tokenizer},
num_proc=20, remove_columns=columns_to_remove)
open_dataset.append(one_dataset)
if open_dataset!=[]:
open_dataset = concatenate_datasets(open_dataset)
if close_dataset!=[]:
close_dataset = concatenate_datasets(close_dataset)
# Load model
print(open_dataset)
print(close_dataset)
print()
print()
print(open_dataset.column_names)
if script_args.vision_adaptor:
model = WPathVLM_Vision(script_args.llm_requires_grad,
script_args.load_in_8bit,
script_args.load_in_4bit,
script_args.llm_name,
script_args.trust_remote_code, # False
script_args.token, # True
tokenizer,
new_tokens_ids[3:],
n_heads = script_args.n_heads,
n_level = script_args.n_level,
embed_dim = script_args.embed_dim,
agg_strategy = script_args.agg_strategy,
hierachical_token = script_args.hierachical_token,
hierachical_adaptor=script_args.hierachical_adaptor,
data_cache_dir = script_args.data_cache_dir,
)
else:
model = WPathVLM(script_args.llm_requires_grad,
script_args.load_in_8bit,
script_args.load_in_4bit,
script_args.llm_name,
script_args.trust_remote_code, # False
script_args.token, # True
tokenizer,
new_tokens_ids[3:],
n_heads = script_args.n_heads,
n_level = script_args.n_level,
embed_dim = script_args.embed_dim,
agg_strategy = script_args.agg_strategy,
hierachical_token = script_args.hierachical_token,
hierachical_adaptor=script_args.hierachical_adaptor,
data_cache_dir = script_args.data_cache_dir,
)
if script_args.use_peft:
peft_config = LoraConfig(
r=script_args.peft_lora_r, # Use a moderate rank
lora_alpha=script_args.peft_lora_alpha, # Scaling factor
bias="none", # No bias adaptation
task_type="CAUSAL_LM", # For causal language modeling tasks
)
model.llm = get_peft_model(model.llm, peft_config)
model.llm.print_trainable_parameters()
else:
peft_config = None
model.load_state_dict(torch.load(script_args.ckpt_path, map_location=device))
# model = model.to(torch.bfloat16)
model.to(device)
# Prepare data loader
data_collator = MyDataCollatorForWPathVLM(tokenizer=tokenizer,
fea_root=script_args.fea_root,
gmm_root=script_args.gmm_root,
fea_dim=script_args.embed_dim,
n_level=script_args.n_level,
n_heads=list(map(int, script_args.n_heads.split(','))),
agg_strategy=script_args.agg_strategy,
test=True)
# if script_args.adaptor == 'qformer':
# remove_columns = []
# else:
dataloader_params = {"batch_size": script_args.batch_size, "collate_fn": data_collator, "shuffle": script_args.shuffle}
if open_dataset!=[]:
tokenized_open_dataset = open_dataset.map(tokenize, batched=False, remove_columns=['text'], num_proc=4,
batch_size=script_args.batch_size,input_columns=['text'])
tokenized_open_dataset = tokenized_open_dataset.map(tokenize_instruct, batched=False, remove_columns=['text_input'], num_proc=4,
batch_size=script_args.batch_size,input_columns=['text_input'])
open_dataloader = DataLoader(tokenized_open_dataset, **dataloader_params)
print("### Start evaluating open-ended!")
evaluate_model(model, open_dataloader, script_args, mode='open')
if close_dataset!=[]:
tokenized_close_dataset = close_dataset.map(tokenize, batched=False, remove_columns=['text'], num_proc=4,
batch_size=script_args.batch_size,input_columns=['text'])
tokenized_close_dataset = tokenized_close_dataset.map(tokenize_instruct, batched=False, remove_columns=['text_input'], num_proc=4,
batch_size=script_args.batch_size,input_columns=['text_input'])
close_dataloader = DataLoader(tokenized_close_dataset, **dataloader_params)
print("### Start evaluating close-ended!")
evaluate_model(model, close_dataloader, script_args, mode='close') |