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
|
#!/usr/bin/env python3
import argparse
import json
import os
import torch
from datasets import Dataset, Image # Use PIL?
from functools import partial
from glob import glob
from peft import LoraConfig
from transformers import (
AutoProcessor,
AutoModelForCausalLM,
AutoModelForImageTextToText,
TrainingArguments,
BitsAndBytesConfig,
)
from trl import SFTTrainer
def make_dataset(base="./baseline"): # TODO: Make actual hf dataset
prompt = "You are a professional English-German translator and also a renowned photography critic.\n\nWrite a detailed caption for this image in a single sentence. Translate the caption into German. The output needs to be JSON, the keys being 'English' and 'German' for the respective captions. Only output the JSON, nothing else." + "<start_of_image>"
user_prompts = []
images = []
assistant_replies = []
for filename in glob(f"{base}/*.jsonl"):
with open(filename, "r") as f:
data = json.loads(f.read())
image_path = f"../d/Images/{os.path.basename(filename).removesuffix(".jsonl")}.jpg"
user_prompts.append(prompt)
assistant_replies.append(json.dumps({
"English": data["English"],
"German": data["Translation"],
}, ensure_ascii=False, indent=0))
images.append(image_path)
return Dataset.from_dict({"image": images, "user": user_prompts, "assistant": assistant_replies}).cast_column("image", Image())
def add_chat_text(example, processor):
messages = [
{"role": "user", "content": example["user"]},
{"role": "assistant", "content": example["assistant"]},
]
example["text"] = processor.tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=False,
)
return example
def collate(batch, processor): # FIXME: Support batch_size > 1
images = [i["image"] for i in batch]
texts = [i["text"] for i in batch]
out = processor(
text=texts,
images=images,
padding=True,
truncation=True,
max_length=512,
return_tensors="pt",
)
image_token_id = [
processor.tokenizer.convert_tokens_to_ids(
processor.tokenizer.special_tokens_map["boi_token"]
)
]
out["labels"] = out["input_ids"].clone()
out["labels"][labels == processor.tokenizer.pad_token_id] = -100
out["labels"][labels == image_token_id] = -100
out["labels"][labels == 262144] = -100
return out
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--model", default="google/gemma-3-4b-it")
parser.add_argument("--batch-size", default=1)
parser.add_argument("--gradient-accumulation", default=4)
parser.add_argument("--learning-rate", default=1e-4)
parser.add_argument("--epochs", default=1)
parser.add_argument("--warmup-ratio", default=0.03)
parser.add_argument("--scheduler-type", default="constant")
parser.add_argument("--logging-steps", default=10)
parser.add_argument("--lora-alpha", default=32)
parser.add_argument("--lora-dropout", default=0.05)
parser.add_argument("--lora-r", default=16)
args = parser.parse_args()
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.bfloat16,
)
processor = AutoProcessor.from_pretrained(args.model, use_fast=True)
#model = AutoModelForCausalLM.from_pretrained(
model = AutoModelForImageTextToText.from_pretrained(
args.model,
quantization_config=bnb_config,
device_map="auto",
low_cpu_mem_usage=True,
)
peft_config = LoraConfig(
lora_alpha=args.lora_alpha,
lora_dropout=args.lora_dropout,
r=args.lora_r,
task_type="CAUSAL_LM",
bias="none",
target_modules="all-linear",
modules_to_save=["lm_head", "embed_tokens"],
)
dev_ds = make_dataset("./baseline/files_dev").map(partial(add_chat_text, processor=processor))
train_ds = make_dataset("./baseline/files_train").map(partial(add_chat_text, processor=processor))
args = TrainingArguments(
output_dir="gemma3-mm-sft-lora",
per_device_train_batch_size=args.batch_size,
gradient_accumulation_steps=args.gradient_accumulation,
num_train_epochs=args.epochs,
learning_rate=args.learning_rate,
warmup_ratio=args.warmup_ratio,
lr_scheduler_type=args.scheduler_type,
fp16=False,
bf16=True,
gradient_checkpointing=True,
gradient_checkpointing_kwargs={"use_reentrant": False},
optim="adamw_torch_8bit", # Alternative from BnB: paged_adamw_8bit
remove_unused_columns=False,
logging_steps=args.logging_steps,
save_strategy="epochs",
)
trainer = SFTTrainer(
model=model,
train_dataset=train_ds,
eval_dataset=dev_ds,
data_collator=partial(collate, processor=processor),
args=args,
peft_config=peft_config,
)
trainer.train()
trainer.save_model()
if __name__ == "__main__":
main()
|
