app.py

#!/usr/bin/env python
import os
import io
import numpy as np
import pandas as pd
import tensorflow as tf
from tensorflow.keras import layers, regularizers
from sklearn.preprocessing import MultiLabelBinarizer
from sklearn.model_selection import train_test_split
from google.cloud import storage
from huggingface_hub import hf_hub_download, notebook_login, login
from PIL import Image
import gradio as gr
import collections

login()

# ======================
# CONSTANTS & CONFIGURATION
# ======================

SCIN_GCP_PROJECT = 'dx-scin-public'
SCIN_GCS_BUCKET_NAME = 'dx-scin-public-data'
SCIN_GCS_CASES_CSV = 'dataset/scin_cases.csv'
SCIN_GCS_LABELS_CSV = 'dataset/scin_labels.csv'

SCIN_HF_MODEL_NAME = 'google/derm-foundation'
SCIN_HF_EMBEDDING_FILE = 'scin_dataset_precomputed_embeddings.npz'

# The 10 conditions we want to predict
CONDITIONS_TO_PREDICT = [
    'Eczema',
    'Allergic Contact Dermatitis',
    'Insect Bite',
    'Urticaria',
    'Psoriasis',
    'Folliculitis',
    'Irritant Contact Dermatitis',
    'Tinea',
    'Herpes Zoster',
    'Drug Rash'
]

# ======================
# HELPER FUNCTIONS FOR DATA LOADING
# ======================

def initialize_df_with_metadata(bucket, csv_path):
    csv_bytes = bucket.blob(csv_path).download_as_string()
    df = pd.read_csv(io.BytesIO(csv_bytes), dtype={'case_id': str})
    df['case_id'] = df['case_id'].astype(str)
    return df

def augment_metadata_with_labels(df, bucket, csv_path):
    csv_bytes = bucket.blob(csv_path).download_as_string()
    labels_df = pd.read_csv(io.BytesIO(csv_bytes), dtype={'case_id': str})
    labels_df['case_id'] = labels_df['case_id'].astype(str)
    merged_df = pd.merge(df, labels_df, on='case_id')
    return merged_df

def load_embeddings_from_file(repo_id, object_name):
    file_path = hf_hub_download(repo_id=repo_id, filename=object_name, local_dir='./')
    embeddings = {}
    with open(file_path, 'rb') as f:
        npz_file = np.load(f, allow_pickle=True)
        for key, value in npz_file.items():
            embeddings[key] = value
    return embeddings

# ======================
# DATA PREPARATION FUNCTION
# ======================

def prepare_data(df, embeddings):
    MINIMUM_CONFIDENCE = 0  # Adjust this if needed.
    X = []
    y = []
    poor_image_quality_counter = 0
    missing_embedding_counter = 0
    not_in_condition_counter = 0
    condition_confidence_low_counter = 0

    for row in df.itertuples():
        # Check if the image is marked as having sufficient quality.
        if getattr(row, 'dermatologist_gradable_for_skin_condition_1', None) != 'DEFAULT_YES_IMAGE_QUALITY_SUFFICIENT':
            poor_image_quality_counter += 1
            continue

        # Parse the labels and confidences.
        try:
            labels = eval(getattr(row, 'dermatologist_skin_condition_on_label_name', '[]'))
            confidences = eval(getattr(row, 'dermatologist_skin_condition_confidence', '[]'))
        except Exception as e:
            continue

        row_labels = []
        for label, conf in zip(labels, confidences):
            if label not in CONDITIONS_TO_PREDICT:
                not_in_condition_counter += 1
                continue
            if conf < MINIMUM_CONFIDENCE:
                condition_confidence_low_counter += 1
                continue
            row_labels.append(label)

        # For each image associated with this case, add its embedding and labels.
        for image_path in [getattr(row, 'image_1_path', None),
                           getattr(row, 'image_2_path', None),
                           getattr(row, 'image_3_path', None)]:
            if pd.isna(image_path) or image_path is None:
                continue
            if image_path not in embeddings:
                missing_embedding_counter += 1
                continue
            X.append(embeddings[image_path])
            y.append(row_labels)
    
    print(f'Poor image quality count: {poor_image_quality_counter}')
    print(f'Missing embedding count: {missing_embedding_counter}')
    print(f'Condition not in list count: {not_in_condition_counter}')
    print(f'Excluded due to low confidence count: {condition_confidence_low_counter}')
    return X, y

# ======================
# MODEL BUILDING FUNCTION
# ======================

def build_model(input_dim, output_dim, weight_decay=1e-4):
    inputs = tf.keras.Input(shape=(input_dim,))
    hidden = layers.Dense(256, activation="relu",
                          kernel_regularizer=regularizers.l2(weight_decay),
                          bias_regularizer=regularizers.l2(weight_decay))(inputs)
    hidden = layers.Dropout(0.1)(hidden)
    hidden = layers.Dense(128, activation="relu",
                          kernel_regularizer=regularizers.l2(weight_decay),
                          bias_regularizer=regularizers.l2(weight_decay))(hidden)
    hidden = layers.Dropout(0.1)(hidden)
    output = layers.Dense(output_dim, activation="sigmoid")(hidden)
    model = tf.keras.Model(inputs, output)
    model.compile(loss="binary_crossentropy",
                  optimizer=tf.keras.optimizers.Adam(learning_rate=1e-4))
    return model

# ======================
# MAIN FUNCTION & GRADIO INTERFACE
# ======================

def main():
    # Connect to the Google Cloud Storage bucket.
    storage_client = storage.Client(SCIN_GCP_PROJECT)
    bucket = storage_client.bucket(SCIN_GCS_BUCKET_NAME)
    
    # Load SCIN dataset CSVs and merge them.
    df_cases = initialize_df_with_metadata(bucket, SCIN_GCS_CASES_CSV)
    df_full = augment_metadata_with_labels(df_cases, bucket, SCIN_GCS_LABELS_CSV)
    df_full.set_index('case_id', inplace=True)
    
    # Load precomputed embeddings from Hugging Face.
    print("Loading embeddings...")
    embeddings = load_embeddings_from_file(SCIN_HF_MODEL_NAME, SCIN_HF_EMBEDDING_FILE)
    
    # Prepare the training data.
    print("Preparing training data...")
    X, y = prepare_data(df_full, embeddings)
    X = np.array(X)
    # Convert the list of label lists to binary arrays.
    mlb = MultiLabelBinarizer(classes=CONDITIONS_TO_PREDICT)
    y_bin = mlb.fit_transform(y)
    
    # Split the data into train and test sets.
    X_train, X_test, y_train, y_test = train_test_split(X, y_bin, test_size=0.2, random_state=42)
    
    # Build the model.
    model = build_model(input_dim=6144, output_dim=len(CONDITIONS_TO_PREDICT))
    
    # If a saved model exists, load it; otherwise, train and save it.
    model_file = "model.h5"
    if os.path.exists(model_file):
        print("Loading existing model from", model_file)
        model = tf.keras.models.load_model(model_file)
    else:
        print("Training model... This may take a few minutes.")
        train_ds = tf.data.Dataset.from_tensor_slices((X_train, y_train)).batch(32)
        test_ds = tf.data.Dataset.from_tensor_slices((X_test, y_test)).batch(32)
        model.fit(train_ds, validation_data=test_ds, epochs=15)
        model.save(model_file)
    
    # Extract a list of case IDs for dropdown
    case_ids = list(df_full.index)

    def predict_case(case_id: str):
        """Fetch images and predictions for a given case ID."""
        if case_id not in df_full.index:
            return [], "Case ID not found!", "N/A", "N/A"

        row = df_full.loc[case_id]
        image_paths = [row.get('image_1_path'), row.get('image_2_path'), row.get('image_3_path')]
        images, predictions_text = [], []

        # Get Dermatologist's Labels
        dermatologist_conditions = row.get('dermatologist_skin_condition_on_label_name', "N/A")
        dermatologist_confidence = row.get('dermatologist_skin_condition_confidence', "N/A")

        if isinstance(dermatologist_conditions, str):
            try:
                dermatologist_conditions = eval(dermatologist_conditions)
                dermatologist_confidence = eval(dermatologist_confidence)
            except:
                pass

        # Process images & generate predictions
        for path in image_paths:
            if isinstance(path, str) and (path in embeddings):
                try:
                    img_bytes = bucket.blob(path).download_as_string()
                    img = Image.open(io.BytesIO(img_bytes)).convert("RGB")
                    images.append(img)
                except:
                    continue

                # Model Prediction
                emb = np.expand_dims(embeddings[path], axis=0)
                pred = model.predict(emb)[0]
                pred_dict = {cond: round(float(prob), 3) for cond, prob in zip(mlb.classes_, pred)}
                predictions_text.append(str(pred_dict))

        # Format the output
        predictions_text = "\n".join(predictions_text) if predictions_text else "No predictions available."
        dermatologist_conditions = str(dermatologist_conditions)
        dermatologist_confidence = str(dermatologist_confidence)

        return images, predictions_text, dermatologist_conditions, dermatologist_confidence

    # Create the Gradio Interface with a Dropdown
    iface = gr.Interface(
        fn=predict_case,
        inputs=gr.Dropdown(choices=case_ids, label="Select a Case ID"),
        outputs=[
            gr.Gallery(label="Case Images"),
            gr.Textbox(label="Model's Predictions"),
            gr.Textbox(label="Dermatologist's Skin Conditions"),
            gr.Textbox(label="Dermatologist's Confidence Ratings")
        ],
        title="Derm Foundation Skin Conditions Explorer",
        description="Select a Case ID from the dropdown to view images and predictions."
    )

    iface.launch()


if __name__ == "__main__":
    main()