app.py

# AUTOGENERATED! DO NOT EDIT! File to edit: car_damage_detector.ipynb.

# %% auto 0
__all__ = ['assets_path', 'models_path', 'examples_path', 'imagenet_labels', 'model', 'transform', 'catogories', 'title',
           'description', 'examples', 'learn_damaged_or_not', 'learn_damage_location', 'learn_damage_severity', 'intf',
           'get_imagenet_classes', 'create_model', 'car_or_not_inference', 'predict', 'main_predictor']

# %% car_damage_detector.ipynb 2
# imports
# import os
import timm
# import json
import torch
import gradio as gr
import pickle as pk
# from PIL import Image
import fastbook
fastbook.setup_book()

from fastbook import *
from fastai.vision.widgets import * 
# from collections import Counter, defaultdict

assets_path = 'assets/'
models_path = 'assets/models/'
examples_path = 'assets/examples/'

# %% car_damage_detector.ipynb 3
# Imagenet Class
def get_imagenet_classes():
    # read idx file
    imagenet_file = open(assets_path+"imagenet_class_index.txt", "r").read()
    # seperate elements and onvert string to list
    imagenet_labels_raw = imagenet_file.strip().split('\n')
    # keep first label
    imagenet_labels = [item.split(',')[0] for item in imagenet_labels_raw]
    return imagenet_labels

imagenet_labels = get_imagenet_classes()

# Create Model
def create_model(model_name='vgg16.tv_in1k'):
    # import required model
    model = timm.create_model(model_name, pretrained=True).eval()
    # transform data as required by the model
    transform = timm.data.create_transform(
        **timm.data.resolve_data_config(model.pretrained_cfg)
    )
    return model, transform

model, transform = create_model()

# Car or Not : Main Inferene Code
catogories = ('Is a Car', 'Not a Car')
def car_or_not_inference(input_image):

    # print ("Validating that this is a picture of a car...")

    # retain the top 'n' most occuring items \\ n=36
    top_n_cat_list  = ['sports_car', 'minivan', 'convertible', 'beach_wagon', 'limousine', 'pickup', 'car_wheel', 'grille', 'racer', 'minibus', 'jeep', 'moving_van', 'tow_truck', 'cab', 'police_van', 'snowplow', 'amphibian', 'trailer_truck', 'recreational_vehicle', 'ambulance', 'motor_scooter', 'cassette_player', 'fire_engine', 'car_mirror', 'mobile_home', 'crash_helmet', 'mouse', 'snowmobile', 'Model_T', 'passenger_car', 'solar_dish', 'garbage_truck', 'photocopier', 'mountain_tent', 'half_track', 'speedboat']

    # image = PILImage.create(input_image)
    # transform image as required for prediction
    image_tensor = transform(input_image)
    # predict on image
    output = model(image_tensor.unsqueeze(0))
    # get probabilites
    probabilities = torch.nn.functional.softmax(output[0], dim=0)
    # select top 5 probs
    _, indices = torch.topk(probabilities, 5)

    for idx in indices:
        pred_label = imagenet_labels[idx]
        if pred_label in top_n_cat_list:
            return 1.0 #dict(zip(catogories, [1.0, 0.0])) #"Validation complete - proceed to damage evaluation"

    return 0.0 #dict(zip(catogories, [0.0, 1.0]))#"Are you sure this is a picture of your car? Please take another picture (try a different angle or lighting) and try again."


# %% car_damage_detector.ipynb 5
title = "Car Care"
description = "A vision based car damage identifier."
examples = [examples_path+'lambo.jpg', examples_path+'dog.jpg', examples_path+'front_moderate.jpg']

learn_damaged_or_not = load_learner(models_path+'car_damaged_or_not.pkl')
learn_damage_location = load_learner(models_path+'car_damage_side.pkl')
learn_damage_severity = load_learner(models_path+'car_damage_severity.pkl')

def predict(img, learn):
    # img = PILImage.create(img)
    pred, idx, probs = learn.predict(img)
    return pred#, float(probs[idx])

def main_predictor(img,  progress=gr.Progress()):
    
    progress((0,4), desc="Starting Analysis...")    
    input_image = PILImage.create(img)
    car_or_not = car_or_not_inference(input_image)

    progress((1,4))
    if car_or_not:
        gr.Info("Car check completed.")
        damaged_or_not = predict(input_image, learn_damaged_or_not)

        progress((2,4))
        if damaged_or_not == 'damage':
            gr.Info("Damage check completed.")
            damaged_location = predict(input_image, learn_damage_location)
            progress((3,4))
            gr.Info("Damage Location identified.")
            damaged_severity = predict(input_image, learn_damage_severity)
            progress((4,4), desc="Analysis Complete")
            gr.Info("Damage Severity assessed.")
            # refer below sections for Location and Severity
            return f"""Results: \n Car Check: it's a Car \n Damage Check: Car is Damaged \n Location: {damaged_location} \n Severity: {damaged_severity}"""
        else:
            progress((4,4))
            return "Are you sure your car is damaged ?. \nMake sure you click a clear picture of the damaged portion. \nPlease resubmit the picture"
    else:
        progress((4,4))
        return "Are you sure this is a picture of your car? \nPlease take another picture (try a different angle or lighting) and try again."

# input_image = 'assets/examples/severe.jpg'
# main_predictor(input_image)

# %% car_damage_detector.ipynb 6
intf = gr.Interface(fn=main_predictor,inputs=gr.Image(),outputs=gr.Textbox(),title=title,description=description,examples=examples)
intf.launch(share=True)