README.md

---
base_model: HuggingFaceM4/Idefics3-8B-Llama3
library_name: peft
license: apache-2.0
tags:
- generated_from_trainer
model-index:
- name: idefics3-llama-gui-dense-descriptions
  results: []
datasets:
- Agent-Eval-Refine/GUI-Dense-Descriptions
language:
- en
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

# idefics3-llama-gui-dense-descriptions

This model is a fine-tuned version of [HuggingFaceM4/Idefics3-8B-Llama3](https://huggingface.co/HuggingFaceM4/Idefics3-8B-Llama3) on https://huggingface.co/datasets/Agent-Eval-Refine/GUI-Dense-Descriptions dataset

## Finetuning script

```python
# !pip install git+https://github.com/andimarafioti/transformers.git@e1b7c0a05ab65e4ddb62a407fe12f8ec13a916f0"
# !pip install accelerate datasets peft bitsandbytes
# !pip install flash-attn --no-build-isolation

import pandas as pd
import torch
from peft import LoraConfig, prepare_model_for_kbit_training, get_peft_model
from transformers import (
    AutoProcessor,
    BitsAndBytesConfig,
    Idefics3ForConditionalGeneration,
)
import os
from PIL import Image
from datasets import load_dataset
from transformers import TrainingArguments, Trainer
from huggingface_hub import notebook_login

notebook_login()

gui_dense_desc_dataset = load_dataset("Agent-Eval-Refine/GUI-Dense-Descriptions")
train_ds = gui_dense_desc_dataset["train"]

# os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
# os.environ["CUDA_VISIBLE_DEVICES"] = "2"

USE_LORA = False
USE_QLORA = True
model_id = "HuggingFaceM4/Idefics3-8B-Llama3"

processor = AutoProcessor.from_pretrained(model_id)

if USE_QLORA or USE_LORA:
    lora_config = LoraConfig(
        r=8,
        lora_alpha=8,
        lora_dropout=0.1,
        target_modules=[
            "down_proj",
            "o_proj",
            "k_proj",
            "q_proj",
            "gate_proj",
            "up_proj",
            "v_proj",
        ],
        use_dora=False if USE_QLORA else True,
        init_lora_weights="gaussian",
    )
    lora_config.inference_mode = False
    if USE_QLORA:
        bnb_config = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_use_double_quant=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype=torch.bfloat16,
        )

    model = Idefics3ForConditionalGeneration.from_pretrained(
        model_id,
        quantization_config=bnb_config if USE_QLORA else None,
        _attn_implementation="flash_attention_2",
        device_map="auto",
        torch_dtype=torch.bfloat16,
    )
    model.add_adapter(lora_config)
    model.enable_adapters()
    model = prepare_model_for_kbit_training(model)
    model = get_peft_model(model, lora_config)
    print(model.get_nb_trainable_parameters())
else:
    model = Idefics3ForConditionalGeneration.from_pretrained(
        model_id,
        torch_dtype=torch.bfloat16,
        _attn_implementation="flash_attention_2",
        device_map="auto",
    )

    # if you'd like to only fine-tune LLM
    for param in model.model.vision_model.parameters():
        param.requires_grad = False

image_token_id = processor.tokenizer.additional_special_tokens_ids[
    processor.tokenizer.additional_special_tokens.index("<image>")
]


def collate_fn(examples):
    texts = []
    images = []
    for example in examples:
        image = example["image"]
        image_description = example["text"]
        messages = [
            {
                "role": "user",
                "content": [
                    {"type": "image"},
                    {
                        "type": "text",
                        "text": "Provide a detailed description of the image.",
                    },
                ],
            },
            {
                "role": "assistant",
                "content": [{"type": "text", "text": image_description}],
            },
        ]
        text = processor.apply_chat_template(messages, add_generation_prompt=False)
        texts.append(text.strip())
        images.append([image])

        batch = processor(text=texts, images=images, return_tensors="pt", padding=True)
        labels = batch["input_ids"].clone()
        labels[labels == processor.tokenizer.pad_token_id] = -100
        labels[labels == image_token_id] = -100
        batch["labels"] = labels

    return batch

training_args = TrainingArguments(
    num_train_epochs=1,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    warmup_steps=50,
    learning_rate=1e-4,
    weight_decay=0.01,
    logging_steps=5,
    save_strategy="steps",
    save_steps=250,
    save_total_limit=1,
    optim="adamw_torch",
    bf16=True,
    output_dir="./idefics3-llama-gui-dense-descriptions",
    hub_model_id="idefics3-llama-gui-dense-descriptions",
    remove_unused_columns=False,
)

trainer = Trainer(
    model=model,
    args=training_args,
    data_collator=collate_fn,
    train_dataset=train_ds,
)

trainer.train()

trainer.push_to_hub()

```

Training took approx. 40 min. on 2xH100 (80 Gb each) devices.

## Intended usage

```python
from peft import PeftModel
from transformers import AutoProcessor, Idefics3ForConditionalGeneration
from transformers.image_utils import load_image
import torch

adapter_path = "Maverick17/idefics3-llama-gui-dense-descriptions"
base_model_id = "HuggingFaceM4/Idefics3-8B-Llama3"

# Load Model base model
model = Idefics3ForConditionalGeneration.from_pretrained(
    base_model_id,
    _attn_implementation="flash_attention_2",
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

# Merge LoRA and base model
peft_model = PeftModel.from_pretrained(model, adapter_path)
merged_model = peft_model.merge_and_unload()

processor = AutoProcessor.from_pretrained(base_model_id)

image = load_image("path/to/ui/image.png")

# Create inputs
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {
                "type": "text",
                "text": "Provide a detailed description of the image.",
            },
        ],
    },
]

prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(text=prompt, images=[image], return_tensors="pt")
inputs = {k: v.to("cuda") for k, v in inputs.items()}

generation_args = {
    "max_new_tokens": 1024,
    "repetition_penalty": 1,
}

generation_args["do_sample"] = False
generation_args.update(inputs)

# Generate
generated_ids = model.generate(**generation_args)

generated_texts = processor.batch_decode(
    generated_ids[:, generation_args["input_ids"].size(1) :], skip_special_tokens=True
)

print(generated_texts[0].strip())
```

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 2
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 8
- total_train_batch_size: 16
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 50
- num_epochs: 1

### Framework versions

- PEFT 0.13.0
- Transformers 4.44.0.dev0
- Pytorch 2.4.1+cu121
- Datasets 3.0.1
- Tokenizers 0.19.1