app.py

import gradio as gr
import PIL.Image
import pandas as pd
import numpy as np
import boto3
from io import BytesIO, StringIO
from fastai.vision.all import *

def get_x(r): return r['Image Path']
def get_y(r): return r['Survived']

def ProjectReportSplitter(df):
    valid_pct = 0.2
    unique_reports = df['Project Report'].unique()
    valid_reports = np.random.choice(unique_reports, int(len(unique_reports) * valid_pct), replace=False)
    valid_idx = df.index[df['Project Report'].isin(valid_reports)].tolist()
    train_idx = df.index[~df.index.isin(valid_idx)].tolist()
    return train_idx, valid_idx

# Use a function to resolve path
def get_x_transformed(r): return open_image_from_s3(get_x(r))

dblock = DataBlock(
    blocks=(ImageBlock(cls=PILImage), CategoryBlock),
    splitter=ProjectReportSplitter,
    get_x=get_x_transformed,
    get_y=get_y,
    item_tfms=Resize(460, method='pad', pad_mode='zeros'),
    batch_tfms=aug_transforms(mult=2, do_flip=True, max_rotate=20, max_zoom=1.1, max_warp=0.2)
)

# Load your model
learn = load_learner("templateClassifierDATAhalfEPOCHoneVISION.pkl")

# Print the vocabulary of the model
print("Model Vocabulary:", learn.dls.vocab)

labels = learn.dls.vocab
def predict(img):
    img = PILImage.create(img)
    pred,pred_idx,probs = learn.predict(img)
    return {labels[i]: float(probs[i]) for i in range(len(labels))}

# def predict(img):
#     img = PILImage.create(img)
#     pred, pred_idx, probs = learn.predict(img)
#     results = {labels[i]: float(probs[i]) for i in range(len(labels))}
#     # Adjust results to highlight when 'Survived' meets the 75% threshold
#     if results['Survived'] >= 0.75:
#         results['Survived'] = 1.0  # Indicating high confidence of survival
#     else:
#         results['Survived'] = 0.0  # Indicating it did not meet the threshold
#     return results
# def predict(img):
#     img = PILImage.create(img)
#     pred, pred_idx, probs = learn.predict(img)
#     results = {labels[i]: float(probs[i]) for i in range(len(labels))}
#     # Adjusting to display survival status based on the threshold
#     survival_status = 'Survived' if results['Survived'] >= 0.75 else 'Not Survived'
#     results['Survival Status'] = survival_status
#     return results


# Gradio interface setup
title = "Photo Culling AI"
description = "Upload your photo to check if it survives culling."
article = "This interface uses a model trained to predict whether a photo is relevant for a project report."
gr.Interface(fn=predict, inputs=gr.Image(), outputs=gr.Label(num_top_classes=2), title=title, description=description, article=article).launch(share=True,show_error=True)




# import gradio as gr
# import PIL.Image
# import pandas as pd
# import boto3
# from io import BytesIO, StringIO
# from fastai.vision.all import *


# def get_x(r): return r['Image Path']
    
# def get_y(r): return r['Survived']

# def ProjectReportSplitter(df):
#     valid_pct = 0.2
#     unique_reports = df['Project Report'].unique()
#     valid_reports = np.random.choice(unique_reports, int(len(unique_reports) * valid_pct), replace=False)
#     valid_idx = df.index[df['Project Report'].isin(valid_reports)].tolist()
#     train_idx = df.index[~df.index.isin(valid_idx)].tolist()
#     return train_idx, valid_idx


# # Use a function to resolve path
# def get_x_transformed(r): return open_image_from_s3(get_x(r))

# dblock = DataBlock(
#     blocks=(ImageBlock(cls=PILImage), CategoryBlock),
#     splitter=ProjectReportSplitter,
#     get_x=get_x_transformed,
#     get_y=get_y,
#     item_tfms=Resize(460, method='pad', pad_mode='zeros'),
#     batch_tfms=aug_transforms(mult=2, do_flip=True, max_rotate=20, max_zoom=1.1, max_warp=0.2)
# )

# # Load your model
# learn = load_learner("templateClassifierDATAhalfEPOCHoneVISION.pkl")

# # Print the vocabulary of the model
# print("Model Vocabulary:", learn.dls.vocab)


# # Update prediction function to directly read from S3
# def predict(img_path):
#     pred, pred_idx, probs = learn.predict(img_path)
#     return {learn.dls.vocab[i]: float(probs[i]) for i in range(len(probs))}

# # Gradio interface setup
# title = "Photo Culling AI"
# description = "Upload your photo to check if it survives culling."
# article = "This interface uses a model trained to predict whether a photo is relevant for a project report."
# gr.Interface(fn=predict, inputs=gr.Image(), outputs=gr.Label(num_top_classes=2), title=title, description=description, article=article).launch(share=True)