multi-images training
xinyanghuang7
parent
2adb3d15a0
commit
b45247a32d
|
|
@ -1,244 +0,0 @@
|
||||||
import os
|
|
||||||
import platform
|
|
||||||
import argparse
|
|
||||||
import time
|
|
||||||
import math
|
|
||||||
import warnings
|
|
||||||
import json
|
|
||||||
|
|
||||||
import pandas as pd
|
|
||||||
import torch
|
|
||||||
import torch.nn.functional as F
|
|
||||||
import torch.distributed as dist
|
|
||||||
from contextlib import nullcontext
|
|
||||||
|
|
||||||
from torch import optim
|
|
||||||
from torch.nn.parallel import DistributedDataParallel
|
|
||||||
from torch.utils.data import DataLoader, DistributedSampler
|
|
||||||
from transformers import AutoTokenizer, AutoModel
|
|
||||||
from model.model import Transformer
|
|
||||||
from model.LMConfig import LMConfig
|
|
||||||
from model.dataset import SFTDataset_multi
|
|
||||||
from model.vision_utils import get_vision_model, get_img_embedding
|
|
||||||
|
|
||||||
warnings.filterwarnings('ignore')
|
|
||||||
|
|
||||||
|
|
||||||
def count_parameters(model):
|
|
||||||
return sum(p.numel() for p in model.parameters() if p.requires_grad)
|
|
||||||
|
|
||||||
|
|
||||||
def Logger(content):
|
|
||||||
if not ddp or dist.get_rank() == 0:
|
|
||||||
print(content)
|
|
||||||
|
|
||||||
|
|
||||||
def get_lr(it, all):
|
|
||||||
warmup_iters = args.warmup_iters
|
|
||||||
lr_decay_iters = all
|
|
||||||
min_lr = args.learning_rate / 10
|
|
||||||
|
|
||||||
if it < warmup_iters:
|
|
||||||
return args.learning_rate * it / warmup_iters
|
|
||||||
if it > lr_decay_iters:
|
|
||||||
return min_lr
|
|
||||||
decay_ratio = (it - warmup_iters) / (lr_decay_iters - warmup_iters)
|
|
||||||
assert 0 <= decay_ratio <= 1
|
|
||||||
coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))
|
|
||||||
return min_lr + coeff * (args.learning_rate - min_lr)
|
|
||||||
|
|
||||||
|
|
||||||
def train_epoch(epoch, wandb):
|
|
||||||
start_time = time.time()
|
|
||||||
for step, (X, Y, loss_mask, image_process) in enumerate(train_loader):
|
|
||||||
X = X.to(args.device)
|
|
||||||
Y = Y.to(args.device)
|
|
||||||
loss_mask = loss_mask.to(args.device)
|
|
||||||
image_process = image_process.to(args.device)
|
|
||||||
image_encoders = get_img_embedding(image_process, vision_model) # torch.Size([32, 4, 50, 768])
|
|
||||||
lr = get_lr(epoch * iter_per_epoch + step, args.epochs * iter_per_epoch)
|
|
||||||
for param_group in optimizer.param_groups:
|
|
||||||
param_group['lr'] = lr
|
|
||||||
|
|
||||||
with ctx:
|
|
||||||
logits = model(X, Y, image_encoders=image_encoders).logits
|
|
||||||
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), Y.view(-1), ignore_index=0, reduction='none')
|
|
||||||
loss_mask = loss_mask.view(-1)
|
|
||||||
loss = torch.sum(loss * loss_mask) / loss_mask.sum()
|
|
||||||
|
|
||||||
scaler.scale(loss).backward()
|
|
||||||
|
|
||||||
if (step + 1) % args.accumulation_steps == 0:
|
|
||||||
scaler.unscale_(optimizer)
|
|
||||||
torch.nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
|
|
||||||
|
|
||||||
scaler.step(optimizer)
|
|
||||||
scaler.update()
|
|
||||||
|
|
||||||
optimizer.zero_grad(set_to_none=True)
|
|
||||||
|
|
||||||
if step % args.log_interval == 0:
|
|
||||||
spend_time = time.time() - start_time
|
|
||||||
Logger(
|
|
||||||
'Epoch:[{}/{}]({}/{}) loss:{:.3f} lr:{:.7f} epoch_Time:{}min:'.format(
|
|
||||||
epoch+1,
|
|
||||||
args.epochs,
|
|
||||||
step,
|
|
||||||
iter_per_epoch,
|
|
||||||
loss.item(),
|
|
||||||
optimizer.param_groups[-1]['lr'],
|
|
||||||
spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60))
|
|
||||||
|
|
||||||
if (wandb is not None) and (not ddp or dist.get_rank() == 0):
|
|
||||||
wandb.log({"loss": loss,
|
|
||||||
"lr": optimizer.param_groups[-1]['lr'],
|
|
||||||
"epoch_Time": spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60})
|
|
||||||
|
|
||||||
if (step + 1) % args.save_interval == 0 and (not ddp or dist.get_rank() == 0):
|
|
||||||
model.eval()
|
|
||||||
moe_path = '_moe' if lm_config.use_moe else ''
|
|
||||||
ckp = f'{args.save_dir}/{lm_config.dim}{moe_path}_vlm_sft.pth'
|
|
||||||
|
|
||||||
if isinstance(model, torch.nn.parallel.DistributedDataParallel):
|
|
||||||
state_dict = model.module.state_dict()
|
|
||||||
else:
|
|
||||||
state_dict = model.state_dict()
|
|
||||||
|
|
||||||
torch.save(state_dict, ckp)
|
|
||||||
model.train()
|
|
||||||
|
|
||||||
if args.save_last and (not ddp or dist.get_rank() == 0):
|
|
||||||
model.eval()
|
|
||||||
moe_path = '_moe' if lm_config.use_moe else ''
|
|
||||||
ckp = f'{args.save_dir}/{lm_config.dim}{moe_path}_vlm_sft_multi.pth'
|
|
||||||
|
|
||||||
if isinstance(model, torch.nn.parallel.DistributedDataParallel):
|
|
||||||
state_dict = model.module.state_dict()
|
|
||||||
else:
|
|
||||||
state_dict = model.state_dict()
|
|
||||||
|
|
||||||
torch.save(state_dict, ckp)
|
|
||||||
model.train()
|
|
||||||
|
|
||||||
|
|
||||||
def init_model(lm_config):
|
|
||||||
tokenizer = AutoTokenizer.from_pretrained('./model/minimind_tokenizer')
|
|
||||||
moe_path = '_moe' if lm_config.use_moe else ''
|
|
||||||
# ckp = f'./out/{lm_config.dim}{moe_path}_vlm_pretrain.pth'
|
|
||||||
ckp = f'./out/{lm_config.dim}{moe_path}_vlm_sft.pth'
|
|
||||||
|
|
||||||
model = Transformer(lm_config)
|
|
||||||
state_dict = torch.load(ckp, map_location=args.device)
|
|
||||||
|
|
||||||
# 处理不需要的前缀
|
|
||||||
unwanted_prefix = '_orig_mod.'
|
|
||||||
for k, v in list(state_dict.items()):
|
|
||||||
if k.startswith(unwanted_prefix):
|
|
||||||
state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k)
|
|
||||||
|
|
||||||
# 加载到模型中
|
|
||||||
model.load_state_dict(state_dict, strict=False)
|
|
||||||
model = model.to(args.device)
|
|
||||||
|
|
||||||
print(f'模型可学习参数: {count_parameters(model) / 1e6} 百万 = {count_parameters(model) / 1e9} B (Billion)')
|
|
||||||
|
|
||||||
(vision_model, preprocess) = get_vision_model()
|
|
||||||
vision_model = vision_model.to(args.device)
|
|
||||||
return model, tokenizer, (vision_model, preprocess)
|
|
||||||
|
|
||||||
|
|
||||||
def init_distributed_mode():
|
|
||||||
if not ddp: return
|
|
||||||
global ddp_local_rank, DEVICE
|
|
||||||
|
|
||||||
dist.init_process_group(backend="nccl")
|
|
||||||
ddp_rank = int(os.environ["RANK"])
|
|
||||||
ddp_local_rank = int(os.environ["LOCAL_RANK"])
|
|
||||||
ddp_world_size = int(os.environ["WORLD_SIZE"])
|
|
||||||
DEVICE = f"cuda:{ddp_local_rank}"
|
|
||||||
torch.cuda.set_device(DEVICE)
|
|
||||||
|
|
||||||
|
|
||||||
if __name__ == "__main__":
|
|
||||||
parser = argparse.ArgumentParser(description="MiniMind-V-SFT")
|
|
||||||
parser.add_argument("--out_dir", type=str, default="out", help="Output directory")
|
|
||||||
parser.add_argument("--epochs", type=int, default=30, help="Number of epochs")
|
|
||||||
parser.add_argument("--batch_size", type=int, default=32, help="Batch size")
|
|
||||||
parser.add_argument("--learning_rate", type=float, default=1e-5, help="Learning rate")
|
|
||||||
parser.add_argument("--device", type=str, default="cuda:0" if torch.cuda.is_available() else "cpu",
|
|
||||||
help="Device to use")
|
|
||||||
parser.add_argument("--dtype", type=str, default="bfloat16", help="Data type")
|
|
||||||
parser.add_argument("--use_wandb", default=False, action="store_true", help="Use Weights & Biases")
|
|
||||||
parser.add_argument("--wandb_project", type=str, default="MiniMind-V", help="Weights & Biases project name")
|
|
||||||
parser.add_argument("--num_workers", type=int, default=8, help="Number of workers for data loading")
|
|
||||||
parser.add_argument("--data_path", type=str, default="./dataset/sft_multi_images/output.json",
|
|
||||||
help="Path to training data")
|
|
||||||
parser.add_argument("--ddp", action="store_true", help="Use DistributedDataParallel")
|
|
||||||
parser.add_argument("--accumulation_steps", type=int, default=1, help="Gradient accumulation steps")
|
|
||||||
parser.add_argument("--grad_clip", type=float, default=1.0, help="Gradient clipping threshold")
|
|
||||||
parser.add_argument("--warmup_iters", type=int, default=0, help="Number of warmup iterations")
|
|
||||||
parser.add_argument("--log_interval", type=int, default=100, help="Logging interval")
|
|
||||||
parser.add_argument("--save_interval", type=int, default=100, help="Model saving interval")
|
|
||||||
parser.add_argument('--local_rank', type=int, default=-1, help='local rank for distributed training')
|
|
||||||
parser.add_argument('--save_last', type=bool, default=True, help='save last step model')
|
|
||||||
|
|
||||||
args = parser.parse_args()
|
|
||||||
|
|
||||||
lm_config = LMConfig()
|
|
||||||
max_seq_len = lm_config.max_seq_len
|
|
||||||
args.save_dir = os.path.join(args.out_dir)
|
|
||||||
os.makedirs(args.save_dir, exist_ok=True)
|
|
||||||
os.makedirs(args.out_dir, exist_ok=True)
|
|
||||||
tokens_per_iter = args.batch_size * max_seq_len
|
|
||||||
torch.manual_seed(1337)
|
|
||||||
device_type = "cuda" if "cuda" in args.device else "cpu"
|
|
||||||
|
|
||||||
args.wandb_run_name = f"MiniMind-V-SFT-Epoch-{args.epochs}-BatchSize-{args.batch_size}-LearningRate-{args.learning_rate}"
|
|
||||||
|
|
||||||
ctx = nullcontext() if device_type == "cpu" else torch.cuda.amp.autocast()
|
|
||||||
ddp = int(os.environ.get("RANK", -1)) != -1 # is this a ddp run?
|
|
||||||
ddp_local_rank, DEVICE = 0, "cuda:0"
|
|
||||||
if ddp:
|
|
||||||
init_distributed_mode()
|
|
||||||
args.device = torch.device(DEVICE)
|
|
||||||
|
|
||||||
if args.use_wandb and (not ddp or ddp_local_rank == 0):
|
|
||||||
import wandb
|
|
||||||
|
|
||||||
wandb.init(project=args.wandb_project, name=args.wandb_run_name)
|
|
||||||
else:
|
|
||||||
wandb = None
|
|
||||||
|
|
||||||
model, tokenizer, (vision_model, preprocess) = init_model(lm_config)
|
|
||||||
|
|
||||||
use_version = 0
|
|
||||||
train_ds = SFTDataset_multi(args.data_path, tokenizer, vision_model=(vision_model, preprocess),
|
|
||||||
image_special_token=lm_config.image_special_token,
|
|
||||||
max_length=max_seq_len)
|
|
||||||
train_sampler = DistributedSampler(train_ds) if ddp else None
|
|
||||||
train_loader = DataLoader(
|
|
||||||
train_ds,
|
|
||||||
batch_size=args.batch_size,
|
|
||||||
pin_memory=True,
|
|
||||||
drop_last=False,
|
|
||||||
shuffle=False,
|
|
||||||
num_workers=args.num_workers,
|
|
||||||
sampler=train_sampler
|
|
||||||
)
|
|
||||||
|
|
||||||
scaler = torch.cuda.amp.GradScaler(enabled=(args.dtype in ['float16', 'bfloat16']))
|
|
||||||
optimizer = optim.Adam(model.parameters(), lr=args.learning_rate)
|
|
||||||
|
|
||||||
if False and not lm_config.use_moe and platform.system() != 'Windows' and float(
|
|
||||||
torch.__version__.split('.')[0]) >= 2:
|
|
||||||
Logger("compiling the model... (takes a ~minute)")
|
|
||||||
unoptimized_model = model
|
|
||||||
model = torch.compile(model)
|
|
||||||
|
|
||||||
if ddp:
|
|
||||||
model._ddp_params_and_buffers_to_ignore = {"pos_cis"}
|
|
||||||
model = DistributedDataParallel(model, device_ids=[ddp_local_rank])
|
|
||||||
|
|
||||||
iter_per_epoch = len(train_loader)
|
|
||||||
for epoch in range(args.epochs):
|
|
||||||
train_epoch(epoch, wandb)
|
|
||||||
|
|
@ -1,167 +0,0 @@
|
||||||
import os
|
|
||||||
import random
|
|
||||||
import time
|
|
||||||
|
|
||||||
import numpy as np
|
|
||||||
import torch
|
|
||||||
import warnings
|
|
||||||
|
|
||||||
from PIL import Image
|
|
||||||
from transformers import AutoTokenizer, AutoModelForCausalLM
|
|
||||||
from model.model import Transformer
|
|
||||||
from model.LMConfig import LMConfig
|
|
||||||
from model.vision_utils import get_vision_model, get_img_process, get_img_embedding
|
|
||||||
|
|
||||||
warnings.filterwarnings('ignore')
|
|
||||||
|
|
||||||
|
|
||||||
def count_parameters(model):
|
|
||||||
return sum(p.numel() for p in model.parameters() if p.requires_grad)
|
|
||||||
|
|
||||||
|
|
||||||
def init_model(lm_config):
|
|
||||||
tokenizer = AutoTokenizer.from_pretrained('./model/minimind_tokenizer')
|
|
||||||
model_from = 1 # 1从权重，2用transformers
|
|
||||||
|
|
||||||
if model_from == 1:
|
|
||||||
moe_path = '_moe' if lm_config.use_moe else ''
|
|
||||||
ckp = f'./out/{lm_config.dim}{moe_path}_vlm_sft_multi.pth'
|
|
||||||
|
|
||||||
model = Transformer(lm_config)
|
|
||||||
state_dict = torch.load(ckp, map_location=device)
|
|
||||||
|
|
||||||
# 处理不需要的前缀
|
|
||||||
unwanted_prefix = '_orig_mod.'
|
|
||||||
for k, v in list(state_dict.items()):
|
|
||||||
if k.startswith(unwanted_prefix):
|
|
||||||
state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k)
|
|
||||||
|
|
||||||
# 加载到模型中
|
|
||||||
model.load_state_dict(state_dict, strict=False)
|
|
||||||
else:
|
|
||||||
model = AutoModelForCausalLM.from_pretrained('minimind-v-v1-small',
|
|
||||||
trust_remote_code=True)
|
|
||||||
|
|
||||||
model = model.to(device)
|
|
||||||
|
|
||||||
print(f'模型参数: {count_parameters(model) / 1e6} 百万 = {count_parameters(model) / 1e9} B (Billion)')
|
|
||||||
|
|
||||||
(vision_model, preprocess) = get_vision_model()
|
|
||||||
vision_model = vision_model.to(device)
|
|
||||||
return model, tokenizer, (vision_model, preprocess)
|
|
||||||
|
|
||||||
|
|
||||||
def setup_seed(seed):
|
|
||||||
random.seed(seed) # 设置 Python 的随机种子
|
|
||||||
np.random.seed(seed) # 设置 NumPy 的随机种子
|
|
||||||
torch.manual_seed(seed) # 设置 PyTorch 的随机种子
|
|
||||||
torch.cuda.manual_seed(seed) # 为当前 GPU 设置随机种子（如果有）
|
|
||||||
torch.cuda.manual_seed_all(seed) # 为所有 GPU 设置随机种子（如果有）
|
|
||||||
torch.backends.cudnn.deterministic = True # 确保每次返回的卷积算法是确定的
|
|
||||||
torch.backends.cudnn.benchmark = False # 关闭 cuDNN 的自动调优，避免不确定性
|
|
||||||
|
|
||||||
|
|
||||||
if __name__ == "__main__":
|
|
||||||
# -----------------------------------------------------------------------------
|
|
||||||
out_dir = 'out'
|
|
||||||
start = ""
|
|
||||||
temperature = 0.5
|
|
||||||
top_k = 8
|
|
||||||
setup_seed(1337)
|
|
||||||
# device = 'cpu'
|
|
||||||
device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
|
|
||||||
dtype = 'bfloat16'
|
|
||||||
max_seq_len = 1 * 1024
|
|
||||||
lm_config = LMConfig()
|
|
||||||
lm_config.max_seq_len = max_seq_len
|
|
||||||
# -----------------------------------------------------------------------------
|
|
||||||
|
|
||||||
model, tokenizer, (vision_model, preprocess) = init_model(lm_config)
|
|
||||||
|
|
||||||
model = model.eval()
|
|
||||||
# 推送到huggingface
|
|
||||||
# model.push_to_hub("minimind")
|
|
||||||
# tokenizer.push_to_hub("minimind")
|
|
||||||
|
|
||||||
stream = True
|
|
||||||
|
|
||||||
i = 0
|
|
||||||
# 获取图像文件列表
|
|
||||||
# image_dir = './dataset/eval_multi_images/' # 对于图像转换数据集
|
|
||||||
image_dir = './dataset/eval_multi_images/bird/' # 对于鸟类数据集
|
|
||||||
image_files = sorted(os.listdir(image_dir))
|
|
||||||
|
|
||||||
# 对于图像转换数据集
|
|
||||||
# prompt = "context: Source Image: <image> Target Image: <image> Instruction: What is the correct image edit instruction that can transfrom the source image to target image?"
|
|
||||||
|
|
||||||
# 对于鸟类数据集
|
|
||||||
prompt = "<image>\n<image>\nName all the differences between these two birds."
|
|
||||||
|
|
||||||
# 遍历图像文件
|
|
||||||
image_encoders = []
|
|
||||||
for i, image_file in enumerate(image_files):
|
|
||||||
image_path = os.path.join(image_dir, image_file)
|
|
||||||
|
|
||||||
# 打开图像并转换为RGB格式
|
|
||||||
image = Image.open(image_path).convert('RGB')
|
|
||||||
image_process = get_img_process(image, preprocess).to(vision_model.device)
|
|
||||||
|
|
||||||
# 对图像进行编码
|
|
||||||
image_encoder = get_img_embedding(image_process, vision_model).unsqueeze(0)
|
|
||||||
image_encoders.append(image_encoder)
|
|
||||||
print(f'[Image]: {image_file}')
|
|
||||||
image_encoders = torch.cat(image_encoders, dim=0).unsqueeze(0)
|
|
||||||
|
|
||||||
# 准备问题和消息
|
|
||||||
prompt_ = prompt.replace("\n", "")
|
|
||||||
print(f'[Q]: {prompt_}')
|
|
||||||
|
|
||||||
messages = [{"role": "user", "content": prompt}]
|
|
||||||
|
|
||||||
new_prompt = tokenizer.apply_chat_template(
|
|
||||||
messages,
|
|
||||||
tokenize=False,
|
|
||||||
add_generation_prompt=True
|
|
||||||
)[-(max_seq_len - 1):]
|
|
||||||
|
|
||||||
x = tokenizer(new_prompt).data['input_ids']
|
|
||||||
x = (torch.tensor(x, dtype=torch.long, device=device)[None, ...])
|
|
||||||
|
|
||||||
with torch.no_grad():
|
|
||||||
res_y = model.generate(x, tokenizer.eos_token_id, max_new_tokens=max_seq_len, temperature=temperature,
|
|
||||||
top_k=top_k, stream=stream, image_encoders=image_encoders)
|
|
||||||
|
|
||||||
print('[A]: ', end='')
|
|
||||||
try:
|
|
||||||
y = next(res_y)
|
|
||||||
except StopIteration:
|
|
||||||
print("No answer")
|
|
||||||
exit()
|
|
||||||
|
|
||||||
history_idx = 0
|
|
||||||
while y is not None:
|
|
||||||
answer = tokenizer.decode(y[0].tolist())
|
|
||||||
if answer and answer[-1] == '<EFBFBD>':
|
|
||||||
try:
|
|
||||||
y = next(res_y)
|
|
||||||
except StopIteration:
|
|
||||||
break
|
|
||||||
continue
|
|
||||||
if not len(answer):
|
|
||||||
try:
|
|
||||||
y = next(res_y)
|
|
||||||
except StopIteration:
|
|
||||||
break
|
|
||||||
continue
|
|
||||||
|
|
||||||
print(answer[history_idx:], end='', flush=True)
|
|
||||||
try:
|
|
||||||
y = next(res_y)
|
|
||||||
except StopIteration:
|
|
||||||
break
|
|
||||||
history_idx = len(answer)
|
|
||||||
|
|
||||||
if not stream:
|
|
||||||
break
|
|
||||||
|
|
||||||
print('\n')
|
|
||||||
Loading…
Reference in New Issue