Spaces:
Running
on
Zero
Running
on
Zero
Create train.py
Browse files- blip3o/train/train.py +1025 -0
blip3o/train/train.py
ADDED
@@ -0,0 +1,1025 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import io
|
3 |
+
import copy
|
4 |
+
from dataclasses import dataclass, field
|
5 |
+
import json
|
6 |
+
import logging
|
7 |
+
import pathlib
|
8 |
+
from typing import Dict, Optional, Sequence, List
|
9 |
+
import time
|
10 |
+
import torch, gc
|
11 |
+
import glob
|
12 |
+
import transformers
|
13 |
+
import tokenizers
|
14 |
+
import random
|
15 |
+
from blip3o.constants import IGNORE_INDEX, DEFAULT_IMAGE_TOKEN, IMAGE_TOKEN_IDX
|
16 |
+
from torch.utils.data import Dataset
|
17 |
+
from blip3o.train.blip3o_trainer import blip3oTrainer
|
18 |
+
from blip3o import conversation as conversation_lib
|
19 |
+
from blip3o.model import *
|
20 |
+
from blip3o.mm_utils import tokenizer_image_token
|
21 |
+
from PIL import Image, ImageFile
|
22 |
+
from datasets import load_dataset, concatenate_datasets
|
23 |
+
from pathlib import Path
|
24 |
+
from datasets.utils.logging import set_verbosity_info
|
25 |
+
from transformers import logging as tf_logging
|
26 |
+
import torchvision.transforms as T
|
27 |
+
from torchvision.transforms.functional import InterpolationMode
|
28 |
+
from transformers import AutoProcessor
|
29 |
+
|
30 |
+
ImageFile.LOAD_TRUNCATED_IMAGES = True
|
31 |
+
transform_und_images = T.Compose([T.Resize(448, interpolation=InterpolationMode.BICUBIC, antialias=True), T.CenterCrop(448)])
|
32 |
+
|
33 |
+
set_verbosity_info()
|
34 |
+
tf_logging.set_verbosity_info()
|
35 |
+
|
36 |
+
local_rank = None
|
37 |
+
|
38 |
+
|
39 |
+
|
40 |
+
|
41 |
+
def rank0_print(*args):
|
42 |
+
if local_rank == 0:
|
43 |
+
print(*args)
|
44 |
+
|
45 |
+
|
46 |
+
from packaging import version
|
47 |
+
|
48 |
+
IS_TOKENIZER_GREATER_THAN_0_14 = version.parse(tokenizers.__version__) >= version.parse("0.14")
|
49 |
+
|
50 |
+
|
51 |
+
@dataclass
|
52 |
+
class ModelArguments:
|
53 |
+
model_name_or_path: Optional[str] = field(default="facebook/opt-125m")
|
54 |
+
version: Optional[str] = field(default="v0")
|
55 |
+
freeze_backbone: bool = field(default=True)
|
56 |
+
tune_mm_mlp_adapter: bool = field(default=False)
|
57 |
+
vision_tower: Optional[str] = field(default=None)
|
58 |
+
gen_vision_tower: Optional[str] = field(default=None)
|
59 |
+
mm_vision_select_layer: Optional[int] = field(default=-1) # default to the last layer
|
60 |
+
pretrain_mm_mlp_adapter: Optional[str] = field(default=None)
|
61 |
+
pretrain_gen_mlp_adapter: Optional[str] = field(default=None)
|
62 |
+
vision_tower_pretrained: Optional[str] = field(default=None)
|
63 |
+
mm_projector_type: Optional[str] = field(default="linear")
|
64 |
+
gen_projector_type: Optional[str] = field(default="linear")
|
65 |
+
mm_use_im_start_end: bool = field(default=False)
|
66 |
+
mm_use_im_patch_token: bool = field(default=True)
|
67 |
+
mm_patch_merge_type: Optional[str] = field(default="flat")
|
68 |
+
mm_vision_select_feature: Optional[str] = field(default="patch")
|
69 |
+
n_query: Optional[int] = field(default=729) # clip 576, siglip 729
|
70 |
+
n_und_query: Optional[int] = field(default=729) # clip 576, siglip 729
|
71 |
+
gen_pooling: Optional[str] = field(default="all") # options are: pool2d_3, pool2d_9, seq_3, seq_9, seq_27
|
72 |
+
|
73 |
+
|
74 |
+
@dataclass
|
75 |
+
class DataArguments:
|
76 |
+
data_path: str = field(default=None, metadata={"help": "Path to the training data."})
|
77 |
+
lazy_preprocess: bool = False
|
78 |
+
is_multimodal: bool = False
|
79 |
+
image_folder: Optional[str] = field(default=None)
|
80 |
+
shortcaption_image_folder: Optional[str] = field(default=None)
|
81 |
+
data_type: Optional[str] = field(default="mix")
|
82 |
+
image_aspect_ratio: str = "square"
|
83 |
+
|
84 |
+
|
85 |
+
@dataclass
|
86 |
+
class TrainingArguments(transformers.TrainingArguments):
|
87 |
+
cache_dir: Optional[str] = field(default=None)
|
88 |
+
optim: str = field(default="adamw_torch")
|
89 |
+
remove_unused_columns: bool = field(default=False)
|
90 |
+
freeze_mm_mlp_adapter: bool = field(default=False)
|
91 |
+
mpt_attn_impl: Optional[str] = field(default="triton")
|
92 |
+
model_max_length: int = field(
|
93 |
+
default=512,
|
94 |
+
metadata={"help": "Maximum sequence length. Sequences will be right padded (and possibly truncated)."},
|
95 |
+
)
|
96 |
+
double_quant: bool = field(
|
97 |
+
default=True,
|
98 |
+
metadata={"help": "Compress the quantization statistics through double quantization."},
|
99 |
+
)
|
100 |
+
quant_type: str = field(
|
101 |
+
default="nf4",
|
102 |
+
metadata={"help": "Quantization data type to use. Should be one of `fp4` or `nf4`."},
|
103 |
+
)
|
104 |
+
bits: int = field(default=16, metadata={"help": "How many bits to use."})
|
105 |
+
lora_enable: bool = False
|
106 |
+
lora_r: int = 64
|
107 |
+
lora_alpha: int = 16
|
108 |
+
lora_dropout: float = 0.05
|
109 |
+
lora_weight_path: str = ""
|
110 |
+
lora_bias: str = "none"
|
111 |
+
mm_projector_lr: Optional[float] = None
|
112 |
+
group_by_modality_length: bool = field(default=False)
|
113 |
+
|
114 |
+
|
115 |
+
def maybe_zero_3(param, ignore_status=False, name=None):
|
116 |
+
from deepspeed import zero
|
117 |
+
from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus
|
118 |
+
|
119 |
+
if hasattr(param, "ds_id"):
|
120 |
+
if param.ds_status == ZeroParamStatus.NOT_AVAILABLE:
|
121 |
+
if not ignore_status:
|
122 |
+
logging.warning(f"{name}: param.ds_status != ZeroParamStatus.NOT_AVAILABLE: {param.ds_status}")
|
123 |
+
with zero.GatheredParameters([param]):
|
124 |
+
param = param.data.detach().cpu().clone()
|
125 |
+
else:
|
126 |
+
param = param.detach().cpu().clone()
|
127 |
+
return param
|
128 |
+
|
129 |
+
|
130 |
+
# Borrowed from peft.utils.get_peft_model_state_dict
|
131 |
+
def get_peft_state_maybe_zero_3(named_params, bias):
|
132 |
+
if bias == "none":
|
133 |
+
to_return = {k: t for k, t in named_params if "lora_" in k}
|
134 |
+
elif bias == "all":
|
135 |
+
to_return = {k: t for k, t in named_params if "lora_" in k or "bias" in k}
|
136 |
+
elif bias == "lora_only":
|
137 |
+
to_return = {}
|
138 |
+
maybe_lora_bias = {}
|
139 |
+
lora_bias_names = set()
|
140 |
+
for k, t in named_params:
|
141 |
+
if "lora_" in k:
|
142 |
+
to_return[k] = t
|
143 |
+
bias_name = k.split("lora_")[0] + "bias"
|
144 |
+
lora_bias_names.add(bias_name)
|
145 |
+
elif "bias" in k:
|
146 |
+
maybe_lora_bias[k] = t
|
147 |
+
for k, t in maybe_lora_bias:
|
148 |
+
if bias_name in lora_bias_names:
|
149 |
+
to_return[bias_name] = t
|
150 |
+
else:
|
151 |
+
raise NotImplementedError
|
152 |
+
to_return = {k: maybe_zero_3(v, ignore_status=True) for k, v in to_return.items()}
|
153 |
+
return to_return
|
154 |
+
|
155 |
+
|
156 |
+
def get_peft_state_non_lora_maybe_zero_3(named_params, require_grad_only=True):
|
157 |
+
to_return = {k: t for k, t in named_params if "lora_" not in k}
|
158 |
+
if require_grad_only:
|
159 |
+
to_return = {k: t for k, t in to_return.items() if t.requires_grad}
|
160 |
+
to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()}
|
161 |
+
return to_return
|
162 |
+
|
163 |
+
|
164 |
+
def get_mm_adapter_state_maybe_zero_3(named_params, keys_to_match):
|
165 |
+
to_return = {k: t for k, t in named_params if any(key_match in k for key_match in keys_to_match)}
|
166 |
+
to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()}
|
167 |
+
return to_return
|
168 |
+
|
169 |
+
|
170 |
+
def get_vision_tower_state_maybe_zero_3(named_params, keys_to_match=[""]):
|
171 |
+
to_return = {k: t for k, t in named_params if any(key_match in k for key_match in keys_to_match)}
|
172 |
+
to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()}
|
173 |
+
return to_return
|
174 |
+
|
175 |
+
|
176 |
+
def find_all_linear_names(model):
|
177 |
+
cls = torch.nn.Linear
|
178 |
+
lora_module_names = set()
|
179 |
+
multimodal_keywords = ["mm_projector", "vision_tower", "vision_resampler"]
|
180 |
+
for name, module in model.named_modules():
|
181 |
+
if any(mm_keyword in name for mm_keyword in multimodal_keywords):
|
182 |
+
continue
|
183 |
+
if isinstance(module, cls):
|
184 |
+
names = name.split(".")
|
185 |
+
lora_module_names.add(names[0] if len(names) == 1 else names[-1])
|
186 |
+
|
187 |
+
if "lm_head" in lora_module_names: # needed for 16-bit
|
188 |
+
lora_module_names.remove("lm_head")
|
189 |
+
return list(lora_module_names)
|
190 |
+
|
191 |
+
|
192 |
+
def safe_save_model_for_hf_trainer(trainer: transformers.Trainer, output_dir: str, vision_tower: str):
|
193 |
+
"""Collects the state dict and dump to disk."""
|
194 |
+
|
195 |
+
# if getattr(trainer.args, "tune_vision_model", False):
|
196 |
+
|
197 |
+
if trainer.deepspeed:
|
198 |
+
torch.cuda.synchronize()
|
199 |
+
|
200 |
+
|
201 |
+
# Only save Adapter
|
202 |
+
keys_to_match = ["mm_projector"]
|
203 |
+
if getattr(trainer.args, "use_im_start_end", False):
|
204 |
+
keys_to_match.extend(["embed_tokens", "embed_in"])
|
205 |
+
|
206 |
+
weight_to_save = get_mm_adapter_state_maybe_zero_3(trainer.model.named_parameters(), keys_to_match)
|
207 |
+
trainer.model.config.save_pretrained(output_dir)
|
208 |
+
|
209 |
+
current_folder = output_dir.split("/")[-1]
|
210 |
+
parent_folder = os.path.dirname(output_dir)
|
211 |
+
if trainer.args.local_rank == 0 or trainer.args.local_rank == -1:
|
212 |
+
if current_folder.startswith("checkpoint-"):
|
213 |
+
mm_projector_folder = os.path.join(parent_folder, "mm_projector")
|
214 |
+
os.makedirs(mm_projector_folder, exist_ok=True)
|
215 |
+
torch.save(
|
216 |
+
weight_to_save,
|
217 |
+
os.path.join(mm_projector_folder, f"{current_folder}.bin"),
|
218 |
+
)
|
219 |
+
else:
|
220 |
+
torch.save(weight_to_save, os.path.join(output_dir, f"mm_projector.bin"))
|
221 |
+
|
222 |
+
keys_to_match = ["gen_projector"]
|
223 |
+
if getattr(trainer.args, "use_im_start_end", False):
|
224 |
+
keys_to_match.extend(["embed_tokens", "embed_in"])
|
225 |
+
|
226 |
+
weight_to_save = get_mm_adapter_state_maybe_zero_3(trainer.model.named_parameters(), keys_to_match)
|
227 |
+
trainer.model.config.save_pretrained(output_dir)
|
228 |
+
|
229 |
+
current_folder = output_dir.split("/")[-1]
|
230 |
+
parent_folder = os.path.dirname(output_dir)
|
231 |
+
if trainer.args.local_rank == 0 or trainer.args.local_rank == -1:
|
232 |
+
if current_folder.startswith("checkpoint-"):
|
233 |
+
mm_projector_folder = os.path.join(parent_folder, "gen_projector")
|
234 |
+
os.makedirs(mm_projector_folder, exist_ok=True)
|
235 |
+
torch.save(
|
236 |
+
weight_to_save,
|
237 |
+
os.path.join(mm_projector_folder, f"{current_folder}.bin"),
|
238 |
+
)
|
239 |
+
else:
|
240 |
+
torch.save(weight_to_save, os.path.join(output_dir, f"gen_projector.bin"))
|
241 |
+
|
242 |
+
if trainer.deepspeed:
|
243 |
+
torch.cuda.synchronize()
|
244 |
+
trainer.save_model(output_dir)
|
245 |
+
return
|
246 |
+
|
247 |
+
state_dict = trainer.model.state_dict()
|
248 |
+
if trainer.args.should_save:
|
249 |
+
cpu_state_dict = {key: value.cpu() for key, value in state_dict.items()}
|
250 |
+
del state_dict
|
251 |
+
trainer._save(output_dir, state_dict=cpu_state_dict) # noqa
|
252 |
+
|
253 |
+
|
254 |
+
def smart_tokenizer_and_embedding_resize(
|
255 |
+
special_tokens_dict: Dict,
|
256 |
+
tokenizer: transformers.PreTrainedTokenizer,
|
257 |
+
model: transformers.PreTrainedModel,
|
258 |
+
):
|
259 |
+
|
260 |
+
|
261 |
+
num_new_tokens = tokenizer.add_special_tokens(special_tokens_dict)
|
262 |
+
model.resize_token_embeddings(len(tokenizer))
|
263 |
+
|
264 |
+
if num_new_tokens > 0:
|
265 |
+
input_embeddings = model.get_input_embeddings().weight.data
|
266 |
+
input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)
|
267 |
+
input_embeddings[-num_new_tokens:] = input_embeddings_avg
|
268 |
+
|
269 |
+
|
270 |
+
def _tokenize_fn(strings: Sequence[str], tokenizer: transformers.PreTrainedTokenizer) -> Dict:
|
271 |
+
"""Tokenize a list of strings."""
|
272 |
+
tokenized_list = [
|
273 |
+
tokenizer(
|
274 |
+
text,
|
275 |
+
return_tensors="pt",
|
276 |
+
padding="longest",
|
277 |
+
max_length=tokenizer.model_max_length,
|
278 |
+
truncation=True,
|
279 |
+
)
|
280 |
+
for text in strings
|
281 |
+
]
|
282 |
+
input_ids = labels = [tokenized.input_ids[0] for tokenized in tokenized_list]
|
283 |
+
input_ids_lens = labels_lens = [tokenized.input_ids.ne(tokenizer.pad_token_id).sum().item() for tokenized in tokenized_list]
|
284 |
+
return dict(
|
285 |
+
input_ids=input_ids,
|
286 |
+
labels=labels,
|
287 |
+
input_ids_lens=input_ids_lens,
|
288 |
+
labels_lens=labels_lens,
|
289 |
+
)
|
290 |
+
|
291 |
+
|
292 |
+
def _mask_targets(target, tokenized_lens, speakers):
|
293 |
+
# cur_idx = 0
|
294 |
+
cur_idx = tokenized_lens[0]
|
295 |
+
tokenized_lens = tokenized_lens[1:]
|
296 |
+
target[:cur_idx] = IGNORE_INDEX
|
297 |
+
for tokenized_len, speaker in zip(tokenized_lens, speakers):
|
298 |
+
if speaker == "human":
|
299 |
+
target[cur_idx + 2 : cur_idx + tokenized_len] = IGNORE_INDEX
|
300 |
+
cur_idx += tokenized_len
|
301 |
+
|
302 |
+
|
303 |
+
def _add_speaker_and_signal(header, source, get_conversation=True):
|
304 |
+
"""Add speaker and start/end signal on each round."""
|
305 |
+
BEGIN_SIGNAL = "### "
|
306 |
+
END_SIGNAL = "\n"
|
307 |
+
conversation = header
|
308 |
+
for sentence in source:
|
309 |
+
from_str = sentence["from"]
|
310 |
+
if from_str.lower() == "human":
|
311 |
+
from_str = conversation_lib.default_conversation.roles[0]
|
312 |
+
elif from_str.lower() == "gpt":
|
313 |
+
from_str = conversation_lib.default_conversation.roles[1]
|
314 |
+
else:
|
315 |
+
from_str = "unknown"
|
316 |
+
sentence["value"] = BEGIN_SIGNAL + from_str + ": " + sentence["value"] + END_SIGNAL
|
317 |
+
if get_conversation:
|
318 |
+
conversation += sentence["value"]
|
319 |
+
conversation += BEGIN_SIGNAL
|
320 |
+
return conversation
|
321 |
+
|
322 |
+
|
323 |
+
|
324 |
+
def preprocess_multimodal(sources: Sequence[str], data_args: DataArguments) -> Dict:
|
325 |
+
is_multimodal = data_args.is_multimodal
|
326 |
+
if not is_multimodal:
|
327 |
+
return sources
|
328 |
+
und_placeholder = "<|vision_start|>" + "<|image_pad|>" * data_args.n_und_query + "<|vision_end|>"
|
329 |
+
gen_placeholder = ""
|
330 |
+
# "[IMG]" + "<image>" * data_args.n_query + "[/IMG]"
|
331 |
+
inst_type = None
|
332 |
+
for source in sources: # [instance]
|
333 |
+
for sentence in source:
|
334 |
+
if sentence["from"] == "human" and "<image>" in sentence["value"]:
|
335 |
+
sentence["value"] = sentence["value"].replace(DEFAULT_IMAGE_TOKEN, und_placeholder).strip()
|
336 |
+
inst_type = "und"
|
337 |
+
elif sentence["from"] == "gpt" and "<image>" in sentence["value"]:
|
338 |
+
sentence["value"] = sentence["value"].replace(DEFAULT_IMAGE_TOKEN, gen_placeholder).strip()
|
339 |
+
inst_type = "gen"
|
340 |
+
return sources, inst_type
|
341 |
+
|
342 |
+
|
343 |
+
|
344 |
+
|
345 |
+
|
346 |
+
def preprocess_qwen(sources, tokenizer: transformers.PreTrainedTokenizer, has_image: bool = False, max_len=2048, system_message: str = "You are a helpful assistant.") -> Dict:
|
347 |
+
roles = {"human": "user", "gpt": "assistant"}
|
348 |
+
|
349 |
+
tokenizer = copy.deepcopy(tokenizer)
|
350 |
+
chat_template = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}"
|
351 |
+
tokenizer.chat_template = chat_template
|
352 |
+
|
353 |
+
# Apply prompt templates
|
354 |
+
input_ids, targets = [], []
|
355 |
+
for i, source in enumerate(sources):
|
356 |
+
if roles[source[0]["from"]] != roles["human"]:
|
357 |
+
source = source[1:]
|
358 |
+
|
359 |
+
input_id, target = [], []
|
360 |
+
|
361 |
+
# New version, use apply chat template
|
362 |
+
# Build system message for each sentence
|
363 |
+
input_id += tokenizer.apply_chat_template([{"role" : "system", "content" : system_message}])
|
364 |
+
target += [IGNORE_INDEX] * len(input_id)
|
365 |
+
|
366 |
+
for conv in source:
|
367 |
+
try:
|
368 |
+
role = conv["role"]
|
369 |
+
content = conv["content"]
|
370 |
+
except:
|
371 |
+
role = conv["from"]
|
372 |
+
content = conv["value"]
|
373 |
+
|
374 |
+
role = roles.get(role, role)
|
375 |
+
|
376 |
+
conv = [{"role" : role, "content" : content}]
|
377 |
+
encode_id = tokenizer.apply_chat_template(conv)
|
378 |
+
input_id += encode_id
|
379 |
+
if role in ["user", "system"]:
|
380 |
+
target += [IGNORE_INDEX] * len(encode_id)
|
381 |
+
else:
|
382 |
+
target += encode_id
|
383 |
+
|
384 |
+
|
385 |
+
|
386 |
+
assert len(input_id) == len(target), f"{len(input_id)} != {len(target)}"
|
387 |
+
|
388 |
+
input_ids.append(input_id)
|
389 |
+
targets.append(target)
|
390 |
+
input_ids = torch.tensor(input_ids, dtype=torch.long)
|
391 |
+
targets = torch.tensor(targets, dtype=torch.long)
|
392 |
+
|
393 |
+
return dict(
|
394 |
+
input_ids=input_ids, # tensor(bs x seq_len)
|
395 |
+
labels=targets, # tensor(bs x seq_len)
|
396 |
+
)
|
397 |
+
|
398 |
+
|
399 |
+
|
400 |
+
|
401 |
+
def preprocess_llama3(
|
402 |
+
sources,
|
403 |
+
tokenizer: transformers.PreTrainedTokenizer,
|
404 |
+
has_image: bool = False,
|
405 |
+
max_len=2048,
|
406 |
+
system_message: str = "You are a helpful language and vision assistant. You are able to understand the visual content that the user provides, and assist the user with a variety of tasks using natural language.",
|
407 |
+
) -> Dict:
|
408 |
+
# roles = {"human": "<|start_header_id|>user<|end_header_id|>", "gpt": "<|start_header_id|>assistant<|end_header_id|>"}
|
409 |
+
roles = {"human": "user", "gpt": "assistant"}
|
410 |
+
|
411 |
+
# Add image tokens to tokenizer as a special tokens
|
412 |
+
# Use a deepcopy of tokenizer so that we don't modify on the tokenizer
|
413 |
+
tokenizer = copy.deepcopy(tokenizer)
|
414 |
+
# When there is actually an image, we add the image tokens as a special token
|
415 |
+
if has_image:
|
416 |
+
tokenizer.add_tokens(["<image>"], special_tokens=True)
|
417 |
+
image_token_index = tokenizer.convert_tokens_to_ids("<image>")
|
418 |
+
bos_token_id = tokenizer.convert_tokens_to_ids("<|begin_of_text|>")
|
419 |
+
start_header_id = tokenizer.convert_tokens_to_ids("<|start_header_id|>")
|
420 |
+
end_header_id = tokenizer.convert_tokens_to_ids("<|end_header_id|>")
|
421 |
+
eot_id = tokenizer.convert_tokens_to_ids("<|eot_id|>")
|
422 |
+
|
423 |
+
unmask_tokens = ["<|begin_of_text|>", "<|start_header_id|>", "<|end_header_id|>", "<|eot_id|>", "\n\n"]
|
424 |
+
unmask_tokens_idx = [tokenizer.convert_tokens_to_ids(tok) for tok in unmask_tokens]
|
425 |
+
|
426 |
+
# After update, calling tokenizer of llama3 will
|
427 |
+
# auto add bos id for the tokens. ヽ(`⌒´)ノ
|
428 |
+
def safe_tokenizer_llama3(text):
|
429 |
+
input_ids = tokenizer(text).input_ids
|
430 |
+
if input_ids[0] == bos_token_id:
|
431 |
+
input_ids = input_ids[1:]
|
432 |
+
return input_ids
|
433 |
+
|
434 |
+
nl_tokens = tokenizer.convert_tokens_to_ids("\n\n")
|
435 |
+
# Apply prompt templates
|
436 |
+
input_ids, targets = [], []
|
437 |
+
for i, source in enumerate(sources):
|
438 |
+
if roles[source[0]["from"]] != roles["human"]:
|
439 |
+
source = source[1:]
|
440 |
+
|
441 |
+
input_id, target = [], []
|
442 |
+
|
443 |
+
# New version, use apply chat template
|
444 |
+
# Build system message for each sentence
|
445 |
+
input_id += tokenizer.apply_chat_template([{"role" : "system", "content" : system_message}])
|
446 |
+
target += [IGNORE_INDEX] * len(input_id)
|
447 |
+
|
448 |
+
for conv in source:
|
449 |
+
try:
|
450 |
+
role = conv["role"]
|
451 |
+
content = conv["content"]
|
452 |
+
except:
|
453 |
+
role = conv["from"]
|
454 |
+
content = conv["value"]
|
455 |
+
|
456 |
+
role = roles.get(role, role)
|
457 |
+
|
458 |
+
conv = [{"role" : role, "content" : content}]
|
459 |
+
# First is bos token we don't need here
|
460 |
+
encode_id = tokenizer.apply_chat_template(conv)[1:]
|
461 |
+
input_id += encode_id
|
462 |
+
if role in ["user", "system"]:
|
463 |
+
target += [IGNORE_INDEX] * len(encode_id)
|
464 |
+
else:
|
465 |
+
target += encode_id
|
466 |
+
|
467 |
+
|
468 |
+
|
469 |
+
assert len(input_id) == len(target), f"{len(input_id)} != {len(target)}"
|
470 |
+
for idx, encode_id in enumerate(input_id):
|
471 |
+
if encode_id in unmask_tokens_idx:
|
472 |
+
target[idx] = encode_id
|
473 |
+
if encode_id == image_token_index:
|
474 |
+
input_id[idx] = IMAGE_TOKEN_INDEX
|
475 |
+
input_ids.append(input_id)
|
476 |
+
targets.append(target)
|
477 |
+
input_ids = torch.tensor(input_ids, dtype=torch.long)
|
478 |
+
targets = torch.tensor(targets, dtype=torch.long)
|
479 |
+
|
480 |
+
return dict(
|
481 |
+
input_ids=input_ids, # tensor(bs x seq_len)
|
482 |
+
labels=targets, # tensor(bs x seq_len)
|
483 |
+
)
|
484 |
+
|
485 |
+
|
486 |
+
|
487 |
+
def preprocess_plain(
|
488 |
+
sources: Sequence[str],
|
489 |
+
tokenizer: transformers.PreTrainedTokenizer,
|
490 |
+
) -> Dict:
|
491 |
+
# add end signal and concatenate together
|
492 |
+
conversations = []
|
493 |
+
for source in sources:
|
494 |
+
assert len(source) == 2
|
495 |
+
# assert DEFAULT_IMAGE_TOKEN in source[0]['value'] or DEFAULT_IMAGE_TOKEN in source[1]['value']
|
496 |
+
conversation = source[0]["value"] + source[1]["value"] + conversation_lib.default_conversation.sep
|
497 |
+
conversations.append(conversation)
|
498 |
+
# tokenize conversations
|
499 |
+
input_ids = [tokenizer_image_token(prompt, tokenizer, return_tensors="pt") for prompt in conversations]
|
500 |
+
targets = copy.deepcopy(input_ids)
|
501 |
+
for target, source in zip(targets, sources):
|
502 |
+
tokenized_len = len(tokenizer_image_token(source[0]["value"], tokenizer))
|
503 |
+
target[:tokenized_len] = IGNORE_INDEX
|
504 |
+
|
505 |
+
return dict(input_ids=input_ids, labels=targets)
|
506 |
+
|
507 |
+
|
508 |
+
def preprocess(
|
509 |
+
sources: Sequence[str],
|
510 |
+
tokenizer: transformers.PreTrainedTokenizer,
|
511 |
+
has_image: bool = False,
|
512 |
+
) -> Dict:
|
513 |
+
"""
|
514 |
+
Given a list of sources, each is a conversation list. This transform:
|
515 |
+
1. Add signal '### ' at the beginning each sentence, with end signal '\n';
|
516 |
+
2. Concatenate conversations together;
|
517 |
+
3. Tokenize the concatenated conversation;
|
518 |
+
4. Make a deepcopy as the target. Mask human words with IGNORE_INDEX.
|
519 |
+
"""
|
520 |
+
if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.PLAIN:
|
521 |
+
return preprocess_plain(sources, tokenizer)
|
522 |
+
if conversation_lib.default_conversation.version == "llama3":
|
523 |
+
return preprocess_llama3(sources, tokenizer, has_image=has_image)
|
524 |
+
if conversation_lib.default_conversation.version == "qwen":
|
525 |
+
return preprocess_qwen(sources, tokenizer, has_image=has_image)
|
526 |
+
# add end signal and concatenate together
|
527 |
+
conversations = []
|
528 |
+
for source in sources:
|
529 |
+
header = f"{conversation_lib.default_conversation.system}\n\n"
|
530 |
+
conversation = _add_speaker_and_signal(header, source)
|
531 |
+
conversations.append(conversation)
|
532 |
+
|
533 |
+
# tokenize conversations
|
534 |
+
def get_tokenize_len(prompts):
|
535 |
+
return [len(tokenizer_image_token(prompt, tokenizer)) for prompt in prompts]
|
536 |
+
|
537 |
+
if has_image:
|
538 |
+
input_ids = [tokenizer_image_token(prompt, tokenizer, return_tensors="pt") for prompt in conversations]
|
539 |
+
else:
|
540 |
+
conversations_tokenized = _tokenize_fn(conversations, tokenizer)
|
541 |
+
input_ids = conversations_tokenized["input_ids"]
|
542 |
+
|
543 |
+
targets = copy.deepcopy(input_ids)
|
544 |
+
for target, source in zip(targets, sources):
|
545 |
+
if has_image:
|
546 |
+
tokenized_lens = get_tokenize_len([header] + [s["value"] for s in source])
|
547 |
+
else:
|
548 |
+
tokenized_lens = _tokenize_fn([header] + [s["value"] for s in source], tokenizer)["input_ids_lens"]
|
549 |
+
speakers = [sentence["from"] for sentence in source]
|
550 |
+
_mask_targets(target, tokenized_lens, speakers)
|
551 |
+
|
552 |
+
return dict(input_ids=input_ids, labels=targets)
|
553 |
+
|
554 |
+
|
555 |
+
|
556 |
+
class LazySupervisedMixDataset(Dataset):
|
557 |
+
"""Dataset for supervised fine-tuning."""
|
558 |
+
|
559 |
+
def __init__(
|
560 |
+
self,
|
561 |
+
data_path: str,
|
562 |
+
tokenizer: transformers.PreTrainedTokenizer,
|
563 |
+
data_args: DataArguments,
|
564 |
+
):
|
565 |
+
super(LazySupervisedMixDataset, self).__init__()
|
566 |
+
|
567 |
+
self.data_args = data_args
|
568 |
+
list_data_dict = []
|
569 |
+
|
570 |
+
|
571 |
+
###################################### text to image #######################################
|
572 |
+
data_files = glob.glob(os.path.join(self.data_args.image_folder, "*.tar"))
|
573 |
+
## text to image
|
574 |
+
train_dataset = load_dataset("webdataset", data_files=data_files, split="train", num_proc=128)
|
575 |
+
train_dataset = train_dataset.rename_column("jpg", "image")
|
576 |
+
train_dataset = train_dataset.add_column('type', len(train_dataset) * ['T2I'])
|
577 |
+
train_dataset = train_dataset.add_column('image_path', len(train_dataset) * [None])
|
578 |
+
train_dataset = train_dataset.remove_columns([col for col in train_dataset.column_names if not col in (
|
579 |
+
["image", "txt", "type", "image_path"])])
|
580 |
+
print(f"finish loading image {len(train_dataset)}")
|
581 |
+
list_data_dict.append(train_dataset)
|
582 |
+
|
583 |
+
|
584 |
+
if len(list_data_dict) > 1:
|
585 |
+
list_data_dict = concatenate_datasets(list_data_dict)
|
586 |
+
else:
|
587 |
+
list_data_dict = list_data_dict[0]
|
588 |
+
list_data_dict = list_data_dict.shuffle(seed=42)
|
589 |
+
|
590 |
+
rank0_print(f"Totoal number of training instance: {len(list_data_dict)}")
|
591 |
+
self.tokenizer = tokenizer
|
592 |
+
self.list_data_dict = list_data_dict
|
593 |
+
|
594 |
+
def __len__(self):
|
595 |
+
return len(self.list_data_dict)
|
596 |
+
|
597 |
+
@property
|
598 |
+
def lengths(self):
|
599 |
+
length_list = []
|
600 |
+
for sample in self.list_data_dict:
|
601 |
+
img_tokens = 128 if "image" in sample else 0
|
602 |
+
length_list.append(sum(len(conv["value"].split()) for conv in sample["conversations"]) + img_tokens)
|
603 |
+
return length_list
|
604 |
+
|
605 |
+
@property
|
606 |
+
def modality_lengths(self):
|
607 |
+
length_list = []
|
608 |
+
for sample in self.list_data_dict:
|
609 |
+
cur_len = sum(len(conv["value"].split()) for conv in sample["conversations"])
|
610 |
+
cur_len = cur_len if "image" in sample else -cur_len
|
611 |
+
length_list.append(cur_len)
|
612 |
+
return length_list
|
613 |
+
|
614 |
+
def __getitem__(self, i) -> Dict[str, torch.Tensor]:
|
615 |
+
|
616 |
+
while True:
|
617 |
+
sources = self.list_data_dict[i]
|
618 |
+
|
619 |
+
if sources["type"] == "T2I" or sources["type"] == "journeyDB_T2I":
|
620 |
+
sources["conversations"] = [
|
621 |
+
{"from": "human", "value": f"Please generate image based on the following caption: {sources['txt']}"},
|
622 |
+
{"from": "gpt", "value": "<image>"},
|
623 |
+
]
|
624 |
+
|
625 |
+
|
626 |
+
elif sources["type"] == "I2I" or sources["type"] == "journeyDB_I2I":
|
627 |
+
sources["conversations"] = [
|
628 |
+
{
|
629 |
+
"from": "human",
|
630 |
+
"value": f"<image>\nPlease reconstruct the given image.",
|
631 |
+
},
|
632 |
+
{"from": "gpt", "value": ""},
|
633 |
+
]
|
634 |
+
|
635 |
+
else:
|
636 |
+
raise ValueError("Unknown source type. Please check the 'type' in 'sources'.")
|
637 |
+
|
638 |
+
if "image" in sources:
|
639 |
+
|
640 |
+
def img_process(images, processor, image_aspect_ratio):
|
641 |
+
if image_aspect_ratio == "pad":
|
642 |
+
|
643 |
+
def expand2square(pil_img, background_color):
|
644 |
+
width, height = pil_img.size
|
645 |
+
if width == height:
|
646 |
+
return pil_img
|
647 |
+
elif width > height:
|
648 |
+
result = Image.new(pil_img.mode, (width, width), background_color)
|
649 |
+
result.paste(pil_img, (0, (width - height) // 2))
|
650 |
+
return result
|
651 |
+
else:
|
652 |
+
result = Image.new(pil_img.mode, (height, height), background_color)
|
653 |
+
result.paste(pil_img, ((height - width) // 2, 0))
|
654 |
+
return result
|
655 |
+
|
656 |
+
images = [expand2square(img, tuple(int(x * 255) for x in processor.image_mean)) for img in images]
|
657 |
+
images = processor.preprocess(images, return_tensors="pt")["pixel_values"]
|
658 |
+
else:
|
659 |
+
images = processor.preprocess(images, return_tensors="pt")["pixel_values"]
|
660 |
+
return images
|
661 |
+
|
662 |
+
if sources["type"] == "T2I" or sources["type"] == "I2I":
|
663 |
+
image_files = self.list_data_dict[i]["image"]
|
664 |
+
else:
|
665 |
+
image_files = self.list_data_dict[i]["image_path"]
|
666 |
+
|
667 |
+
if not isinstance(image_files, list):
|
668 |
+
image_files = [image_files]
|
669 |
+
|
670 |
+
images = []
|
671 |
+
|
672 |
+
def read_bin_as_bytesio(bin_file_path):
|
673 |
+
with open(bin_file_path, "rb") as f:
|
674 |
+
return io.BytesIO(f.read())
|
675 |
+
|
676 |
+
for img in image_files:
|
677 |
+
try:
|
678 |
+
if sources["type"] == "T2I" or sources["type"] == "I2I":
|
679 |
+
img = img.convert("RGB")
|
680 |
+
elif sources["type"] == "journeyDB_T2I" or sources["type"] == "journeyDB_I2I":
|
681 |
+
if sources["type"] == "journeyDB_T2I" or sources["type"] == "journeyDB_I2I":
|
682 |
+
image_path = os.path.join('/fsx/sfr/data/jiuhai/hub/datasets--JourneyDB--JourneyDB/snapshots/e191aa61ca37e5e4418707ade4df5deb5c6d5d8f/data/train/imgs', img)
|
683 |
+
else:
|
684 |
+
raise ValueError("Unknown source type. Please check the 'type' in 'sources'.")
|
685 |
+
img = Image.open(image_path).convert("RGB")
|
686 |
+
images.append(img)
|
687 |
+
except Exception as e:
|
688 |
+
print(f"Error opening image {img}: {e}")
|
689 |
+
images = None
|
690 |
+
break # Skip to the next image if there's an error
|
691 |
+
|
692 |
+
if not images is None:
|
693 |
+
try:
|
694 |
+
temp = img_process(
|
695 |
+
images,
|
696 |
+
self.data_args.gen_image_processor,
|
697 |
+
self.data_args.image_aspect_ratio,
|
698 |
+
)
|
699 |
+
except Exception as e:
|
700 |
+
print(f"Error wrong number of channels: {e}")
|
701 |
+
images = None
|
702 |
+
|
703 |
+
|
704 |
+
# If no valid images were found, randomly pick another item
|
705 |
+
if images is None:
|
706 |
+
print(sources)
|
707 |
+
print(f"warning false image!!!!!!")
|
708 |
+
i = random.randint(0, len(self.list_data_dict) - 1)
|
709 |
+
continue
|
710 |
+
|
711 |
+
|
712 |
+
sources, inst_type = preprocess_multimodal(copy.deepcopy([sources["conversations"]]), self.data_args)
|
713 |
+
else:
|
714 |
+
sources = copy.deepcopy([sources["conversations"]])
|
715 |
+
data_dict = preprocess(sources, self.tokenizer, has_image=("image" in self.list_data_dict[i]))
|
716 |
+
if isinstance(i, int):
|
717 |
+
data_dict = dict(input_ids=data_dict["input_ids"][0], labels=data_dict["labels"][0])
|
718 |
+
|
719 |
+
# image exist in the data
|
720 |
+
if "image" in self.list_data_dict[i]:
|
721 |
+
if inst_type == "gen":
|
722 |
+
data_dict["gen_image"] = img_process(
|
723 |
+
images,
|
724 |
+
self.data_args.gen_image_processor,
|
725 |
+
self.data_args.image_aspect_ratio,
|
726 |
+
)
|
727 |
+
|
728 |
+
elif inst_type == "und":
|
729 |
+
|
730 |
+
resized_images = [transform_und_images(img) for img in images]
|
731 |
+
|
732 |
+
image_inputs = self.data_args.image_processor(resized_images, return_tensors="pt")
|
733 |
+
|
734 |
+
data_dict["und_image"] = image_inputs.pixel_values
|
735 |
+
data_dict["grid_thw"] = image_inputs.image_grid_thw
|
736 |
+
data_dict["gen_image"] = img_process(
|
737 |
+
resized_images,
|
738 |
+
self.data_args.gen_image_processor,
|
739 |
+
self.data_args.image_aspect_ratio,
|
740 |
+
)
|
741 |
+
|
742 |
+
elif self.data_args.is_multimodal:
|
743 |
+
crop_size = self.data_args.image_processor.crop_size
|
744 |
+
data_dict["image"] = torch.zeros(3, crop_size["height"], crop_size["width"])
|
745 |
+
|
746 |
+
data_dict["ids"] = self.list_data_dict[i]["id"] if "id" in self.list_data_dict[i] else "unk"
|
747 |
+
return data_dict
|
748 |
+
|
749 |
+
|
750 |
+
@dataclass
|
751 |
+
class DataCollatorForSupervisedDataset(object):
|
752 |
+
"""Collate examples for supervised fine-tuning."""
|
753 |
+
|
754 |
+
tokenizer: transformers.PreTrainedTokenizer
|
755 |
+
|
756 |
+
def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]:
|
757 |
+
input_ids, labels, ids = tuple([instance[key] for instance in instances] for key in ("input_ids", "labels", "ids"))
|
758 |
+
multi_input_ids = []
|
759 |
+
multi_labels = []
|
760 |
+
i_s_pos = []
|
761 |
+
for input_id, label in zip(input_ids, labels):
|
762 |
+
input_id = input_id[: self.tokenizer.model_max_length - 65]
|
763 |
+
label = label[: self.tokenizer.model_max_length - 65]
|
764 |
+
i_s_pos.append(input_id.shape[0]+1)
|
765 |
+
img_id = torch.full((65,), IMAGE_TOKEN_IDX, dtype=input_id.dtype, device=input_id.device)
|
766 |
+
img_id[0] = 151665
|
767 |
+
input_id = torch.cat([input_id, img_id])
|
768 |
+
img_label = torch.full((65,), IMAGE_TOKEN_IDX, dtype=label.dtype, device=label.device)
|
769 |
+
img_label[0] = 151665
|
770 |
+
label = torch.cat([label, img_label])
|
771 |
+
multi_input_ids.append(input_id)
|
772 |
+
multi_labels.append(label)
|
773 |
+
|
774 |
+
input_ids = multi_input_ids
|
775 |
+
labels = multi_labels
|
776 |
+
|
777 |
+
input_ids = torch.nn.utils.rnn.pad_sequence(input_ids, batch_first=True, padding_value=self.tokenizer.pad_token_id)
|
778 |
+
labels = torch.nn.utils.rnn.pad_sequence(labels, batch_first=True, padding_value=IGNORE_INDEX)
|
779 |
+
if input_ids.shape[1] > self.tokenizer.model_max_length:
|
780 |
+
print(f"Warning input with length {input_ids.shape[1]} is longer than max length {self.tokenizer.model_max_length}")
|
781 |
+
input_ids = input_ids[:, : self.tokenizer.model_max_length]
|
782 |
+
labels = labels[:, : self.tokenizer.model_max_length]
|
783 |
+
batch = dict(
|
784 |
+
input_ids=input_ids,
|
785 |
+
labels=labels,
|
786 |
+
attention_mask=input_ids.ne(self.tokenizer.pad_token_id),
|
787 |
+
)
|
788 |
+
|
789 |
+
batch_gen_images = []
|
790 |
+
batch_und_images = []
|
791 |
+
batch_grid_thw = []
|
792 |
+
|
793 |
+
for instance in instances:
|
794 |
+
if "gen_image" in instance:
|
795 |
+
batch_gen_images.append(instance["gen_image"])
|
796 |
+
|
797 |
+
|
798 |
+
if len(batch_gen_images) > 0:
|
799 |
+
if all(x is not None and y.shape == batch_gen_images[0][0].shape for x in batch_gen_images for y in x):
|
800 |
+
batch["gen_image"] = torch.cat([images for images in batch_gen_images], dim=0)
|
801 |
+
else:
|
802 |
+
batch["gen_image"] = batch_gen_images
|
803 |
+
else:
|
804 |
+
batch["gen_image"] = None
|
805 |
+
|
806 |
+
|
807 |
+
for instance in instances:
|
808 |
+
if "und_image" in instance:
|
809 |
+
batch_und_images.append(instance["und_image"].unsqueeze(0)) ## 1*1024*1176
|
810 |
+
batch_grid_thw.append(instance["grid_thw"]) ## 1*3
|
811 |
+
|
812 |
+
|
813 |
+
# print(f"batch_und_images {batch_und_images}")
|
814 |
+
if len(batch_und_images) > 0:
|
815 |
+
batch["und_image"] = torch.cat([images for images in batch_und_images], dim=0)
|
816 |
+
batch["grid_thw"] = torch.cat([images for images in batch_grid_thw], dim=0)
|
817 |
+
else:
|
818 |
+
batch["und_image"] = None
|
819 |
+
batch["grid_thw"] = None
|
820 |
+
|
821 |
+
batch["ids"] = ids
|
822 |
+
|
823 |
+
batch["i_s_pos"] = i_s_pos
|
824 |
+
|
825 |
+
return batch
|
826 |
+
|
827 |
+
|
828 |
+
def make_supervised_data_module(tokenizer: transformers.PreTrainedTokenizer, data_args) -> Dict:
|
829 |
+
|
830 |
+
if data_args.data_type == "mix":
|
831 |
+
train_dataset = LazySupervisedMixDataset(tokenizer=tokenizer, data_path=data_args.data_path, data_args=data_args)
|
832 |
+
else:
|
833 |
+
raise ValueError("Unknown data type. Please check the Dataloader type.")
|
834 |
+
|
835 |
+
data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer)
|
836 |
+
return dict(train_dataset=train_dataset, eval_dataset=None, data_collator=data_collator)
|
837 |
+
|
838 |
+
|
839 |
+
def unlock_vit(training_args, model_args, vision_tower):
|
840 |
+
for n, p in vision_tower.named_parameters():
|
841 |
+
p.requires_grad = True
|
842 |
+
|
843 |
+
|
844 |
+
def train(attn_implementation=None):
|
845 |
+
global local_rank
|
846 |
+
|
847 |
+
parser = transformers.HfArgumentParser((ModelArguments, DataArguments, TrainingArguments))
|
848 |
+
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
|
849 |
+
print(model_args, data_args, training_args)
|
850 |
+
local_rank = training_args.local_rank
|
851 |
+
compute_dtype = torch.float16 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32)
|
852 |
+
|
853 |
+
bnb_model_from_pretrained_args = {}
|
854 |
+
if training_args.bits in [4, 8]:
|
855 |
+
from transformers import BitsAndBytesConfig
|
856 |
+
|
857 |
+
bnb_model_from_pretrained_args.update(
|
858 |
+
dict(
|
859 |
+
device_map={"": training_args.device},
|
860 |
+
load_in_4bit=training_args.bits == 4,
|
861 |
+
load_in_8bit=training_args.bits == 8,
|
862 |
+
quantization_config=BitsAndBytesConfig(
|
863 |
+
load_in_4bit=training_args.bits == 4,
|
864 |
+
load_in_8bit=training_args.bits == 8,
|
865 |
+
llm_int8_skip_modules=["mm_projector"],
|
866 |
+
llm_int8_threshold=6.0,
|
867 |
+
llm_int8_has_fp16_weight=False,
|
868 |
+
bnb_4bit_compute_dtype=compute_dtype,
|
869 |
+
bnb_4bit_use_double_quant=training_args.double_quant,
|
870 |
+
bnb_4bit_quant_type=training_args.quant_type, # {'fp4', 'nf4'}
|
871 |
+
),
|
872 |
+
)
|
873 |
+
)
|
874 |
+
|
875 |
+
if model_args.vision_tower is not None:
|
876 |
+
model = blip3oLlamaForCausalLM.from_pretrained(
|
877 |
+
model_args.model_name_or_path,
|
878 |
+
cache_dir=training_args.cache_dir,
|
879 |
+
attn_implementation=attn_implementation,
|
880 |
+
torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
|
881 |
+
**bnb_model_from_pretrained_args,
|
882 |
+
)
|
883 |
+
else:
|
884 |
+
if "Qwen" in model_args.model_name_or_path or "qwen" in model_args.model_name_or_path :
|
885 |
+
model = blip3oQwenForCausalLM.from_pretrained(
|
886 |
+
model_args.model_name_or_path,
|
887 |
+
cache_dir=training_args.cache_dir,
|
888 |
+
attn_implementation=attn_implementation,
|
889 |
+
torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
|
890 |
+
**bnb_model_from_pretrained_args,
|
891 |
+
)
|
892 |
+
else:
|
893 |
+
model = transformers.LlamaForCausalLM.from_pretrained(
|
894 |
+
model_args.model_name_or_path,
|
895 |
+
cache_dir=training_args.cache_dir,
|
896 |
+
attn_implementation=attn_implementation,
|
897 |
+
torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
|
898 |
+
**bnb_model_from_pretrained_args,
|
899 |
+
)
|
900 |
+
model.config.use_cache = False
|
901 |
+
|
902 |
+
if model_args.freeze_backbone:
|
903 |
+
for (n, p) in model.get_model().named_parameters():
|
904 |
+
p.requires_grad = False
|
905 |
+
for (n, p) in model.visual.named_parameters():
|
906 |
+
p.requires_grad = False
|
907 |
+
for (n, p) in model.lm_head.named_parameters():
|
908 |
+
p.requires_grad = False
|
909 |
+
|
910 |
+
if training_args.gradient_checkpointing:
|
911 |
+
if hasattr(model, "enable_input_require_grads"):
|
912 |
+
model.enable_input_require_grads()
|
913 |
+
else:
|
914 |
+
|
915 |
+
def make_inputs_require_grad(module, input, output):
|
916 |
+
output.requires_grad_(True)
|
917 |
+
|
918 |
+
model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
|
919 |
+
if "Qwen" in model_args.model_name_or_path or "qwen" in model_args.model_name_or_path:
|
920 |
+
tokenizer = AutoProcessor.from_pretrained(model_args.model_name_or_path).tokenizer
|
921 |
+
tokenizer.model_max_length = training_args.model_max_length
|
922 |
+
else:
|
923 |
+
tokenizer = transformers.AutoTokenizer.from_pretrained(
|
924 |
+
model_args.model_name_or_path,
|
925 |
+
cache_dir=training_args.cache_dir,
|
926 |
+
model_max_length=training_args.model_max_length,
|
927 |
+
padding_side="right",
|
928 |
+
use_fast=False,
|
929 |
+
)
|
930 |
+
# tokenizer.pad_token = tokenizer.unk_token
|
931 |
+
if tokenizer.pad_token is None:
|
932 |
+
smart_tokenizer_and_embedding_resize(
|
933 |
+
special_tokens_dict=dict(
|
934 |
+
pad_token="<pad>",
|
935 |
+
additional_special_tokens=["[IMG]", "[/IMG]", "<image>"],
|
936 |
+
),
|
937 |
+
tokenizer=tokenizer,
|
938 |
+
model=model,
|
939 |
+
)
|
940 |
+
elif not "<image>" in tokenizer.get_added_vocab():
|
941 |
+
smart_tokenizer_and_embedding_resize(
|
942 |
+
special_tokens_dict=dict(additional_special_tokens=["[IMG]", "[/IMG]", "<image>"]),
|
943 |
+
tokenizer=tokenizer,
|
944 |
+
model=model,
|
945 |
+
)
|
946 |
+
if model_args.version in conversation_lib.conv_templates:
|
947 |
+
conversation_lib.default_conversation = conversation_lib.conv_templates[model_args.version]
|
948 |
+
else:
|
949 |
+
conversation_lib.default_conversation = conversation_lib.conv_templates["llama3"]
|
950 |
+
rank0_print(f"Using conversation format: {conversation_lib.default_conversation.version}")
|
951 |
+
|
952 |
+
|
953 |
+
|
954 |
+
# if model_args.vision_tower is not None:
|
955 |
+
model.get_model().initialize_vision_modules(model_args=model_args, fsdp=training_args.fsdp)
|
956 |
+
|
957 |
+
## generation vision tower
|
958 |
+
gen_vision_tower = model.get_gen_vision_tower()
|
959 |
+
gen_vision_tower.to(
|
960 |
+
dtype=torch.bfloat16 if training_args.bf16 else torch.float16,
|
961 |
+
device=training_args.device,
|
962 |
+
)
|
963 |
+
gen_vision_tower.requires_grad_(False)
|
964 |
+
|
965 |
+
data_args.gen_image_processor = gen_vision_tower.image_processor
|
966 |
+
data_args.image_processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct").image_processor
|
967 |
+
|
968 |
+
data_args.is_multimodal = True
|
969 |
+
data_args.n_query = model_args.n_query
|
970 |
+
data_args.n_und_query = model_args.n_und_query
|
971 |
+
|
972 |
+
model.config.image_aspect_ratio = data_args.image_aspect_ratio
|
973 |
+
model.config.tokenizer_padding_side = tokenizer.padding_side
|
974 |
+
model.config.tokenizer_model_max_length = tokenizer.model_max_length
|
975 |
+
|
976 |
+
model.config.tune_mm_mlp_adapter = training_args.tune_mm_mlp_adapter = model_args.tune_mm_mlp_adapter
|
977 |
+
|
978 |
+
model.config.freeze_mm_mlp_adapter = training_args.freeze_mm_mlp_adapter
|
979 |
+
|
980 |
+
# Calculate total parameters and trainable parameters
|
981 |
+
total_params = sum(p.numel() for p in model.get_model().parameters())
|
982 |
+
trainable_params = sum(p.numel() for p in model.get_model().parameters() if p.requires_grad)
|
983 |
+
|
984 |
+
print(f"Total parameters: {total_params}")
|
985 |
+
print(f"Trainable parameters: {trainable_params}")
|
986 |
+
|
987 |
+
|
988 |
+
model.config.mm_use_im_start_end = data_args.mm_use_im_start_end = model_args.mm_use_im_start_end
|
989 |
+
model.config.mm_projector_lr = training_args.mm_projector_lr
|
990 |
+
training_args.use_im_start_end = model_args.mm_use_im_start_end
|
991 |
+
model.config.mm_use_im_patch_token = model_args.mm_use_im_patch_token
|
992 |
+
model.initialize_vision_tokenizer(model_args, tokenizer=tokenizer)
|
993 |
+
model.config.pad_token_id = tokenizer.pad_token_id
|
994 |
+
|
995 |
+
data_module = make_supervised_data_module(tokenizer=tokenizer, data_args=data_args)
|
996 |
+
|
997 |
+
trainer = blip3oTrainer(
|
998 |
+
model=model,
|
999 |
+
tokenizer=tokenizer,
|
1000 |
+
args=training_args,
|
1001 |
+
**data_module,
|
1002 |
+
)
|
1003 |
+
from tabulate import tabulate
|
1004 |
+
|
1005 |
+
if trainer.is_world_process_zero():
|
1006 |
+
stat = []
|
1007 |
+
for i, (n, p) in enumerate(trainer.model.named_parameters()):
|
1008 |
+
stat.append([i, n, p.shape, p.requires_grad])
|
1009 |
+
print(tabulate(stat, headers=["idx", "name", "shape", "trainable"]))
|
1010 |
+
if list(pathlib.Path(training_args.output_dir).glob("checkpoint-*")):
|
1011 |
+
trainer.train(resume_from_checkpoint=True)
|
1012 |
+
else:
|
1013 |
+
trainer.train()
|
1014 |
+
trainer.save_state()
|
1015 |
+
|
1016 |
+
model.config.use_cache = True
|
1017 |
+
safe_save_model_for_hf_trainer(
|
1018 |
+
trainer=trainer,
|
1019 |
+
output_dir=training_args.output_dir,
|
1020 |
+
vision_tower=model_args.vision_tower,
|
1021 |
+
)
|
1022 |
+
|
1023 |
+
|
1024 |
+
if __name__ == "__main__":
|
1025 |
+
train()
|