【Python】Understanding Convolutional Neural Networks for

    !pip install tensorflow==2.3.0
    !pip install opencv-python numpy matplotlib pillow
    
      
    
    代码解读

    import cv2
    import numpy as np
    import tensorflow as tf
    from google.colab.patches import cv2_imshow
    
    def load_model():
    model = tf.saved_model.load('ssd_mobilenet_v2_fpnlite_320x320_coco17_tpu-8')
    return model
    
    def inference(image):
    h, w = image.shape[:2]
    
    # Resize to the input size of the model
    resize_ratio = min(320 / w, 320 / h)
    resized_image = cv2.resize(image, None, fx=resize_ratio, fy=resize_ratio, interpolation=cv2.INTER_AREA)
    
    # Normalize and add a batch dimension
    input_tensor = tf.convert_to_tensor(resized_image)[tf.newaxis,...] * (2./255) - 1.0
    
    detections = detect_fn(input_tensor)
    num_detections = int(detections.pop('num_detections'))
    detections = {key: value[0, :num_detections].numpy()
                   for key, value in detections.items()}
    detections['num_detections'] = num_detections
    
    # Rescale bboxes to original image size
    boxes = detections['detection_boxes']
    scale_x, scale_y = w / resized_image.shape[1], h / resized_image.shape[0]
    boxes[:, [0, 2]] *= scale_x
    boxes[:, [1, 3]] *= scale_y
    
    scores = detections['detection_scores']
    classes = detections['detection_classes'].astype(np.int32)
    bboxes = boxes.astype(np.int32)
    
    return scores, classes, bboxes
    
    
    if __name__ == '__main__':
    detect_fn = load_model().signatures['serving_default']
    
    image = cv2.imread(img_path)
    scores, classes, bboxes = inference(image)
    print("Number of objects detected:", len(scores))
    
    for i in range(len(scores)):
        score = float(scores[i])
        bbox = tuple(bboxes[i])
        class_id = int(classes[i])
    
        if score > 0.5:
            label = CLASSES[class_id] + f"({score:.2f})"
    
            cv2.rectangle(image, bbox[0:2], bbox[2:], (0,255,0), thickness=2)
            cv2.putText(image, label, (bbox[0]+10,bbox[1]+30), cv2.FONT_HERSHEY_SIMPLEX, 1, (0,255,0), 2, cv2.LINE_AA)
    
    cv2_imshow(image)
    
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
    
    代码解读

    train_datagen = tf.keras.preprocessing.image.ImageDataGenerator(
    rotation_range=45,
    width_shift_range=.15, 
    height_shift_range=.15, 
    zoom_range=0.5,
    shear_range=.15,
    horizontal_flip=True, 
    rescale=1./255.,
    )
    
    val_datagen = tf.keras.preprocessing.image.ImageDataGenerator(rescale=1./255.)
    
    train_ds = train_datagen.flow_from_directory(
    TRAIN_DIR, target_size=(IMAGE_SIZE, IMAGE_SIZE), batch_size=BATCH_SIZE)
    
    val_ds = val_datagen.flow_from_directory(
    VAL_DIR, target_size=(IMAGE_SIZE, IMAGE_SIZE), batch_size=BATCH_SIZE)
    
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
    
    代码解读

    model = ssd_mobiledet_cpu_320x320(num_classes=NUM_CLASSES+1, pretrained_backbone='MobileDetCPU', freeze_batchnorm=True)
    
    base_lr = 0.004
    steps_per_epoch = train_ds.samples // BATCH_SIZE
    total_epochs = EPOCHS
    
    loss = {
    "cls_out": tf.losses.SparseCategoricalCrossentropy(from_logits=True),
    "box_out": smooth_l1_loss(),
    }
    
    model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=base_lr),
              steps_per_execution=steps_per_epoch,
              loss=loss)
    
    checkpoint_filepath = os.path.join(CHECKPOINT_PATH, "{epoch}.h5")
    callbacks = [
    tf.keras.callbacks.ModelCheckpoint(filepath=checkpoint_filepath, save_weights_only=True, verbose=1),
    tf.keras.callbacks.TensorBoard(log_dir=LOG_DIR),
    ]
    
    history = model.fit(train_ds, epochs=EPOCHS, validation_data=val_ds, callbacks=callbacks)
    
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
    
    代码解读

    latest = tf.train.latest_checkpoint(CHECKPOINT_PATH)
    print("Loading", latest)
    model.load_weights(latest)
    
    _, _, APs, _ = evaluate(model, val_ds)
    mean_ap = np.nanmean(APs[:-1])
    print("mAP:", mean_ap)
    
      
      
      
      
      
      
    
    代码解读