app.py

import numpy as np
import gradio as gr
import torch
from transformers import Dinov2Config, Dinov2Model, Dinov2ForImageClassification, AutoImageProcessor
import torch.nn as nn
import os
import json
from huggingface_hub import hf_hub_download

# DEFINE MODEL NAME
model_name = "dinov2-large-2024_01_24-with_data_aug_batch-size32_epochs93_freeze"
checkpoint_name = "lombardata/" + model_name

# IMPORT CLASSIFICATION MODEL
def create_head(num_features , number_classes ,dropout_prob=0.5 ,activation_func =nn.ReLU):
    features_lst = [num_features , num_features//2 , num_features//4]
    layers = []
    for in_f ,out_f in zip(features_lst[:-1] , features_lst[1:]):
        layers.append(nn.Linear(in_f , out_f))
        layers.append(activation_func())
        layers.append(nn.BatchNorm1d(out_f))
        if dropout_prob !=0 : layers.append(nn.Dropout(dropout_prob))
    layers.append(nn.Linear(features_lst[-1] , number_classes))
    return nn.Sequential(*layers)
from transformers import Dinov2Config, Dinov2Model

class NewheadDinov2ForImageClassification(Dinov2ForImageClassification):
    def __init__(self, config: Dinov2Config) -> None:
        super().__init__(config)

        self.num_labels = config.num_labels
        self.dinov2 = Dinov2Model(config)

        # Classifier head
        self.classifier = create_head(config.hidden_size * 2, config.num_labels)

model = NewheadDinov2ForImageClassification.from_pretrained(checkpoint_name)

# IMPORT MODEL CONFIG PARAMETERS
config_path = hf_hub_download(repo_id=checkpoint_name, filename="config.json")
# Opening JSON file
config_file = open(config_path)
# returns JSON object as 
config = json.load(config_file)
# import parameters
id2label = config["id2label"]
label2id = config["label2id"]
image_size = config["image_size"]
classes_names = list(label2id.keys())

# PREDICTIONS
def sigmoid(_outputs):
    return 1.0 / (1.0 + np.exp(-_outputs))
    
def predict(input_image):
    image_processor = AutoImageProcessor.from_pretrained(checkpoint_name)
    # predict
    inputs = image_processor(input_image, return_tensors="pt")
    inputs = inputs    
    with torch.no_grad():
        model_outputs = model(**inputs)
    outputs = model_outputs["logits"][0]
    scores = sigmoid(outputs)
    result = {}
    i = 0
    for score in scores:
        label = classes_names[i]
        result[label] = float(score)
        i += 1
    result = {key: result[key] for key in result if result[key] > 0.5}
    return result
    
# Define style
title = "DinoVd'eau image classification"
model_link = "https://huggingface.co/" + checkpoint_name
description = f"This application showcases the capability of artificial intelligence-based systems to identify objects within underwater images. To utilize it, you can either upload your own image or select one of the provided examples for analysis.\nFor predictions, we use this [open-source model]({model_link})"

gr.Interface(
    fn=predict,
    inputs=gr.Image(shape=(512, 512)),
    outputs="label",
    title=title, 
    description=description,
    examples=["GOPR0106.JPG", 
              "session_2021_08_30_Mayotte_10_image_00066.jpg", 
              "session_2018_11_17_kite_Le_Morne_Manawa_G0065777.JPG", 
              "session_2023_06_28_caplahoussaye_plancha_body_v1B_00_GP1_3_1327.jpeg"]).launch()