predict_genetic_ucid.py

# predict_genetic_ucid.py
# -*- coding: utf-8 -*-
import os, sys, argparse
import numpy as np
import cv2
import tensorflow as tf
sys.path.append(r"D:\Imhistory\ImageHistory\utils")
from Network import *

# ---- GA weights of UCID dataset (SE-Dense, SE-Inception, SE-ResNet50, ViT) ----
GA_WEIGHTS = {
    75:  [0.20061, 0.28446, 0.14702, 0.36790],
    85:  [0.23021, 0.40494, 0.04245, 0.32239],
    95:  [0.06745, 0.33683, 0.18399, 0.41174],
}

# # ---- GA weights of BOSSBASE (SE-Dense, SE-Inception, SE-ResNet50, ViT) ----
# GA_WEIGHTS = {
#     75:  [0.13952, 0.46086, 0.01697, 0.38264],
#     85:  [0.16546, 0.22289, 0.26274, 0.34891],
#     95:  [0.09371, 0.45073, 0.00427, 0.45129],
# }

# # ---- GA weights of IMAGENET-Mini10K (SE-Dense, SE-Inception, SE-ResNet50, ViT) ----
# GA_WEIGHTS = {
#     75:  [0.34034, 0.22858, 0.03842, 0.39266],
#     85:  [0.29683, 0.46990, 0.11377, 0.11949],
#     95:  [ 0.00265, 0.49270, 0.39497, 0.10968],
# }

IMHIST_CLASSES = ["Adjustment", "Denoising", "Highpass", "Lowpass", "Uniform"]

def load_image(path, size=(512,512)):
    img = cv2.imread(path, 1)
    if img is None:
        raise FileNotFoundError(f"Image not found: {path}")
    # img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    # img = cv2.resize(img, size, interpolation=cv2.INTER_AREA)
   
    img = img.astype(np.float32) / img.max()   
    return img

def build_and_load(model_selection, weights_path, input_shape, optimizer, loss, metric, height, width, dropbool, tf_bool, transformer_type, suffix):
    model = get_Network(tf_bool, len(IMHIST_CLASSES), input_shape, loss, optimizer, metric,
                        height, width, dropbool, model_selection, transformer_type)
    model.load_weights(weights_path)
    model.compile(loss=loss, optimizer=optimizer, metrics=[metric])
    # isim çakışmasın diye
    for i, layer in enumerate(model.layers):
        try:
            layer._name = f"{layer.name}_{suffix}_{i}"
        except Exception:
            pass
    return model

def softmax_ensemble(weights, probs_stack):
    """
    weights: (M,), probs_stack: (M, C)
    dönen: (C,) (ensemble olasılıkları)
    """
    w = np.array(weights, dtype=np.float64)
    w = w / np.sum(w)
    ens = np.tensordot(w, probs_stack, axes=1)  # (C,)
    return ens

def main(args):
    # ---- Parametreler (eğitim yapılmıyor, derleme için lazım) ----
    num_classes = len(IMHIST_CLASSES)
    height, width, n_channels = 512, 512, 3
    input_shape = (height, width, n_channels)
    loss = "categorical_crossentropy"
    metric = ["categorical_accuracy"]
    dropbool = 1
    tf_bool = 2           # 1: pretrained, 2: None  (senin koddaki gibi)
    transformer_type = 2  # 1-b16, 2-b32, 3-l16, 4-l32

    # Optimizer sadece compile için
    from tensorflow.keras.optimizers import SGD
    optimizer = SGD(learning_rate=0.0002, decay=1e-6, momentum=0.9, nesterov=True)

    # ---- Model yolları (senin yapına göre) ----
    # Not: Aşağıdaki yollar QF=75 klasörlerinden; 85/95 için de aynı mimariye ait en iyi ağırlıkların yolunu ver.
    w_dense = r"D:\Imhistory\ImageHistory\Ensemble_Models\UCID\Ucid_75_Imhistory_weights1_SEDensenet.hdf5"
    w_incp  = r"D:\Imhistory\ImageHistory\Ensemble_Models\UCID\Ucid_75_Imhistory_weights1_SEInception.hdf5"
    w_res   = r"D:\Imhistory\ImageHistory\Ensemble_Models\UCID\Ucid_75_Imhistory_weights1_SEResnet.hdf5"
    w_vit   = r"D:\Imhistory\ImageHistory\Ensemble_Models\UCID\Ucid_75_Imhistory_weights1_Transformer.hdf5"

    # ---- Görüntü ----
    img = load_image(args.image, size=(width, height))
    x = np.expand_dims(img, axis=0)  # (1, H, W, 3)

    # ---- Modelleri kur/yükle ----
    # model_selection: 3: SE-DenseNet121, 4: SE-InceptionV3, 2: SE-ResNet50, 1: ViT
    m_dense = build_and_load(3, w_dense, input_shape, optimizer, loss, metric, height, width, dropbool, tf_bool, transformer_type, "sedense121")
    m_incp  = build_and_load(4, w_incp,  input_shape, optimizer, loss, metric, height, width, dropbool, tf_bool, transformer_type, "seinceptionv3")
    m_res   = build_and_load(2, w_res,   input_shape, optimizer, loss, metric, height, width, dropbool, tf_bool, transformer_type, "seresnet50")
    m_vit   = build_and_load(1, w_vit,   input_shape, optimizer, loss, metric, height, width, dropbool, tf_bool, transformer_type, "vitb32")

    models = [m_dense, m_incp, m_res, m_vit]

    # ---- Tek görüntü tahminleri ----
    probs_stack = []
    for m in models:
        p = m.predict(x, verbose=0)[0]  # (C,)
        probs_stack.append(p)
    probs_stack = np.stack(probs_stack, axis=0)  # (M, C)

    # ---- GA ağırlıkları ----
    if args.qf not in GA_WEIGHTS:
        raise ValueError(f"Unsupported QF={args.qf}. Choose from 75, 85, 95.")
    weights = GA_WEIGHTS[args.qf]

    ens_probs = softmax_ensemble(weights, probs_stack)  # (C,)
    pred_idx = int(np.argmax(ens_probs))
    pred_lab = IMHIST_CLASSES[pred_idx]

    # ---- Çıktı yazdır ----
    print("\n== Single-Image Prediction ==")
    print(f"Image: {args.image}")
    print(f"QF: {args.qf}  |  GA Weights: {np.round(weights, 5).tolist()}")
    print("Class probabilities:")
    for i, c in enumerate(IMHIST_CLASSES):
        print(f"  {c:10s}: {ens_probs[i]:.4f}")
    print(f"\nPredicted: {pred_lab} (idx={pred_idx})\n")

# if __name__ == "__main__":
#     ap = argparse.ArgumentParser()
#     ap.add_argument("--image", type=str, required=True, help=r"D:\Imhistory\ImageHistory\Datasets\UCID\Test\QF=75\Adjustment\3_uniformT75_adj1.jpg)")
#     ap.add_argument("--qf", type=int, default=75, choices=[75,85,95], help="JPEG quality factor used in pipeline")
#     args = ap.parse_args()
#     main(args)

# Run cmd 
# python predict_single_ucid.py --image "D:\Imhistory\ImageHistory\Datasets\UCID\Test\QF=75\Adjustment\3_uniformT75_adj1.jpg" --qf 75


if __name__ == "__main__":
    # Spyder'da doğrudan çalıştırmak için argümanları burada tanımla
    class Args:
        image = r"D:\Imhistory\ImageHistory\Datasets\UCID\Test\QF=75\Uniform\3_uniformT75.jpg"
        qf = 75
    args = Args()

    main(args)