Fine-tuning computer vision models

Multi-Modal Models with Hugging Face

James Chapman

Curriculum Manager, DataCamp

Purpose of fine-tuning vision models

New classes, e.g., binary classification of real or AI-generated images
New imaging domains, e.g., X-rays

General image used for pretraining of a seal

Example of an AI generated image

¹ https://image-net.org/index.php

Fine-tuning vision models

Example of an AI generated image

Adjust model output to the new predictions
Prepare dataset for training
Configure training options
Train!

Model updates

from datasets import load_dataset
dataset = load_dataset("ideepankarsharma2003/Midjourney_v6_Classification_small_shuf
fled")['train']

dataset = dataset.train_test_split(test_size=0.2, seed=42)


labels = dataset["train"].features["label"].names

label2id, id2label = dict(), dict()
for i, label in enumerate(labels):
    label2id[label] = str(i)
    id2label[str(i)] = label

Model updates

from transformers import AutoModelForImageClassification
checkpoint = "google/mobilenet_v2_1.0_224"
model = AutoModelForImageClassification.from_pretrained(
    checkpoint,
    num_labels=len(labels),

    id2label=id2label, 
    label2id=label2id,

    ignore_mismatched_sizes=True

)

Dataset preparation

from transformers import AutoImageProcessor
image_processor = AutoImageProcessor.from_pretrained(checkpoint)


from torchvision.transforms import Compose, Normalize, ToTensor

normalize = Normalize(mean=image_processor.image_mean, std=image_processor.image_std)

transform = Compose([ToTensor(), normalize])


def transforms(examples):
    examples["pixel_values"] = [transform(img.convert("RGB")) for img in examples["image"]]
    del examples["image"]
    return examples


dataset = dataset.with_transform(transforms)

Plotting transformed data

import matplotlib.pyplot as plt
plt.imshow(dataset["train"][0]["pixel_values"].permute(1, 2, 0))
plt.show()

Transformed image from the new dataset

Training

from transformers import TrainingArguments

training_args = TrainingArguments(
    output_dir="dataset_finetune",

    learning_rate=6e-5,

    gradient_accumulation_steps=4,

    num_train_epochs=3,

    push_to_hub=False
)

from transformers import Trainer,
    DefaultDataCollator

data_collator = DefaultDataCollator()

trainer = Trainer(

    model=model,

    args=training_args,

    train_dataset=dataset["train"],
    eval_dataset=dataset["test"],

    processing_class=image_processor,

    data_collator=data_collator

)

Evaluation

predictions = trainer.predict(dataset["test"])
predictions.metrics["test_accuracy"]

0.455

trainer.train()

{..., 'eval_accuracy': 0.93, ...}

Let's practice!

Multi-Modal Models with Hugging Face