eval.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Calculate mAP for YOLO model on some annotation dataset
"""
import os, argparse, time
import numpy as np
from PIL import Image
import operator
from collections import OrderedDict
import matplotlib.pyplot as plt
from tqdm import tqdm

from tensorflow.keras.models import load_model
import tensorflow.keras.backend as K
import tensorflow as tf
import MNN
import onnxruntime

from yolo3.postprocess_np import yolo3_postprocess_np
from yolo2.postprocess_np import yolo2_postprocess_np
from common.data_utils import preprocess_image
from common.utils import get_dataset, get_classes, get_anchors, get_colors, draw_boxes, optimize_tf_gpu, get_custom_objects

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

optimize_tf_gpu(tf, K)


def annotation_parse(annotation_lines, class_names):
    '''
    parse annotation lines to get image dict and ground truth class dict

    image dict would be like:
    annotation_records = {
        '/path/to/000001.jpg': {'100,120,200,235':'dog', '85,63,156,128':'car', ...},
        ...
    }

    ground truth class dict would be like:
    classes_records = {
        'car': [
                ['000001.jpg','100,120,200,235'],
                ['000002.jpg','85,63,156,128'],
                ...
               ],
        ...
    }
    '''
    annotation_records = OrderedDict()
    classes_records = OrderedDict({class_name: [] for class_name in class_names})

    for line in annotation_lines:
        box_records = {}
        image_name = line.split(' ')[0]
        boxes = line.split(' ')[1:]
        for box in boxes:
            #strip box coordinate and class
            class_name = class_names[int(box.split(',')[-1])]
            coordinate = ','.join(box.split(',')[:-1])
            box_records[coordinate] = class_name
            #append or add ground truth class item
            record = [os.path.basename(image_name), coordinate]
            if class_name in classes_records:
                classes_records[class_name].append(record)
            else:
                classes_records[class_name] = list([record])
        annotation_records[image_name] = box_records

    return annotation_records, classes_records


def transform_gt_record(gt_records, class_names):
    '''
    Transform the Ground Truth records of a image to prediction format, in
    order to show & compare in result pic.

    Ground Truth records is a dict with format:
        {'100,120,200,235':'dog', '85,63,156,128':'car', ...}

    Prediction format:
        (boxes, classes, scores)
    '''
    if gt_records is None or len(gt_records) == 0:
        return [], [], []

    gt_boxes = []
    gt_classes = []
    gt_scores = []
    for (coordinate, class_name) in gt_records.items():
        gt_box = [int(x) for x in coordinate.split(',')]
        gt_class = class_names.index(class_name)

        gt_boxes.append(gt_box)
        gt_classes.append(gt_class)
        gt_scores.append(1.0)

    return np.array(gt_boxes), np.array(gt_classes), np.array(gt_scores)



def yolo_predict_tflite(interpreter, image, anchors, num_classes, conf_threshold):
    input_details = interpreter.get_input_details()
    output_details = interpreter.get_output_details()

    # check the type of the input tensor
    #if input_details[0]['dtype'] == np.float32:
        #floating_model = True

    height = input_details[0]['shape'][1]
    width = input_details[0]['shape'][2]
    model_image_size = (height, width)

    image_data = preprocess_image(image, model_image_size)
    image_shape = image.size

    interpreter.set_tensor(input_details[0]['index'], image_data)
    interpreter.invoke()

    prediction = []
    for output_detail in output_details:
        output_data = interpreter.get_tensor(output_detail['index'])
        prediction.append(output_data)

    if len(anchors) == 5:
        # YOLOv2 use 5 anchors and have only 1 prediction
        assert len(prediction) == 1, 'invalid YOLOv2 prediction number.'
        pred_boxes, pred_classes, pred_scores = yolo2_postprocess_np(prediction[0], image_shape, anchors, num_classes, model_image_size, max_boxes=100, confidence=conf_threshold)
    else:
        pred_boxes, pred_classes, pred_scores = yolo3_postprocess_np(prediction, image_shape, anchors, num_classes, model_image_size, max_boxes=100, confidence=conf_threshold)

    return pred_boxes, pred_classes, pred_scores


def yolo_predict_mnn(interpreter, session, image, anchors, num_classes, conf_threshold):
    # assume only 1 input tensor for image
    input_tensor = interpreter.getSessionInput(session)
    # get input shape
    input_shape = input_tensor.getShape()
    if input_tensor.getDimensionType() == MNN.Tensor_DimensionType_Tensorflow:
        batch, height, width, channel = input_shape
    elif input_tensor.getDimensionType() == MNN.Tensor_DimensionType_Caffe:
        batch, channel, height, width = input_shape
    else:
        # should be MNN.Tensor_DimensionType_Caffe_C4, unsupported now
        raise ValueError('unsupported input tensor dimension type')

    model_image_size = (height, width)

    # prepare input image
    image_data = preprocess_image(image, model_image_size)
    image_shape = image.size

    # use a temp tensor to copy data
    tmp_input = MNN.Tensor(input_shape, input_tensor.getDataType(),\
                    image_data, input_tensor.getDimensionType())

    input_tensor.copyFrom(tmp_input)
    interpreter.runSession(session)

    def get_tensor_list(output_tensors):
        # transform the output tensor dict to ordered tensor list, for further postprocess
        #
        # output tensor list should be like (for YOLOv3):
        # [
        #  (name, tensor) for (13, 13, 3, num_classes+5),
        #  (name, tensor) for (26, 26, 3, num_classes+5),
        #  (name, tensor) for (52, 52, 3, num_classes+5)
        # ]
        output_list = []

        for (output_tensor_name, output_tensor) in output_tensors.items():
            tensor_shape = output_tensor.getShape()
            dim_type = output_tensor.getDimensionType()
            tensor_height, tensor_width = tensor_shape[2:4] if dim_type == MNN.Tensor_DimensionType_Caffe else tensor_shape[1:3]

            if len(anchors) == 6:
                # Tiny YOLOv3
                if tensor_height == height//32:
                    output_list.insert(0, (output_tensor_name, output_tensor))
                elif tensor_height == height//16:
                    output_list.insert(1, (output_tensor_name, output_tensor))
                else:
                    raise ValueError('invalid tensor shape')
            elif len(anchors) == 9:
                # YOLOv3
                if tensor_height == height//32:
                    output_list.insert(0, (output_tensor_name, output_tensor))
                elif tensor_height == height//16:
                    output_list.insert(1, (output_tensor_name, output_tensor))
                elif tensor_height == height//8:
                    output_list.insert(2, (output_tensor_name, output_tensor))
                else:
                    raise ValueError('invalid tensor shape')
            elif len(anchors) == 5:
                # YOLOv2 use 5 anchors and have only 1 prediction
                assert len(output_tensors) == 1, 'YOLOv2 model should have only 1 output tensor.'
                output_list.insert(0, (output_tensor_name, output_tensor))
            else:
                raise ValueError('invalid anchor number')

        return output_list

    output_tensors = interpreter.getSessionOutputAll(session)
    output_tensor_list = get_tensor_list(output_tensors)

    prediction = []
    for (output_tensor_name, output_tensor) in output_tensor_list:
        output_shape = output_tensor.getShape()

        assert output_tensor.getDataType() == MNN.Halide_Type_Float

        # copy output tensor to host, for further postprocess
        tmp_output = MNN.Tensor(output_shape, output_tensor.getDataType(),\
                    np.zeros(output_shape, dtype=float), output_tensor.getDimensionType())

        output_tensor.copyToHostTensor(tmp_output)
        #tmp_output.printTensorData()

        output_data = np.array(tmp_output.getData(), dtype=float).reshape(output_shape)
        # our postprocess code based on TF channel last format, so if the output format
        # doesn't match, we need to transpose
        if output_tensor.getDimensionType() == MNN.Tensor_DimensionType_Caffe:
            output_data = output_data.transpose((0,2,3,1))
        elif output_tensor.getDimensionType() == MNN.Tensor_DimensionType_Caffe_C4:
            raise ValueError('unsupported output tensor dimension type')

        prediction.append(output_data)

    if len(anchors) == 5:
        # YOLOv2 use 5 anchors and have only 1 prediction
        assert len(prediction) == 1, 'invalid YOLOv2 prediction number.'
        pred_boxes, pred_classes, pred_scores = yolo2_postprocess_np(prediction[0], image_shape, anchors, num_classes, model_image_size, max_boxes=100, confidence=conf_threshold)
    else:
        pred_boxes, pred_classes, pred_scores = yolo3_postprocess_np(prediction, image_shape, anchors, num_classes, model_image_size, max_boxes=100, confidence=conf_threshold)

    return pred_boxes, pred_classes, pred_scores


def yolo_predict_pb(model, image, anchors, num_classes, model_image_size, conf_threshold):
    # NOTE: TF 1.x frozen pb graph need to specify input/output tensor name
    # so we need to hardcode the input/output tensor names here to get them from model
    if len(anchors) == 6:
        output_tensor_names = ['graph/conv2d_1/BiasAdd:0', 'graph/conv2d_3/BiasAdd:0']
    elif len(anchors) == 9:
        output_tensor_names = ['graph/conv2d_3/BiasAdd:0', 'graph/conv2d_8/BiasAdd:0', 'graph/conv2d_13/BiasAdd:0']
    elif len(anchors) == 5:
        # YOLOv2 use 5 anchors and have only 1 prediction
        output_tensor_names = ['graph/predict_conv/BiasAdd:0']
    else:
        raise ValueError('invalid anchor number')

    # assume only 1 input tensor for image
    input_tensor_name = 'graph/image_input:0'

    # get input/output tensors
    image_input = model.get_tensor_by_name(input_tensor_name)
    output_tensors = [model.get_tensor_by_name(output_tensor_name) for output_tensor_name in output_tensor_names]

    batch, height, width, channel = image_input.shape
    model_image_size = (int(height), int(width))

    # prepare input image
    image_data = preprocess_image(image, model_image_size)
    image_shape = image.size

    with tf.Session(graph=model) as sess:
        prediction = sess.run(output_tensors, feed_dict={
            image_input: image_data
        })

    if len(anchors) == 5:
        # YOLOv2 use 5 anchors and have only 1 prediction
        assert len(prediction) == 1, 'invalid YOLOv2 prediction number.'
        pred_boxes, pred_classes, pred_scores = yolo2_postprocess_np(prediction[0], image_shape, anchors, num_classes, model_image_size, max_boxes=100, confidence=conf_threshold)
    else:
        pred_boxes, pred_classes, pred_scores = yolo3_postprocess_np(prediction, image_shape, anchors, num_classes, model_image_size, max_boxes=100, confidence=conf_threshold)

    return pred_boxes, pred_classes, pred_scores


def yolo_predict_onnx(model, image, anchors, num_classes, conf_threshold):
    input_tensors = []
    for i, input_tensor in enumerate(model.get_inputs()):
        input_tensors.append(input_tensor)

    # assume only 1 input tensor for image
    assert len(input_tensors) == 1, 'invalid input tensor number.'

    batch, height, width, channel = input_tensors[0].shape
    model_image_size = (height, width)

    # prepare input image
    image_data = preprocess_image(image, model_image_size)
    image_shape = image.size

    feed = {input_tensors[0].name: image_data}
    prediction = model.run(None, feed)

    if len(anchors) == 5:
        # YOLOv2 use 5 anchors and have only 1 prediction
        assert len(prediction) == 1, 'invalid YOLOv2 prediction number.'
        pred_boxes, pred_classes, pred_scores = yolo2_postprocess_np(prediction[0], image_shape, anchors, num_classes, model_image_size, max_boxes=100, confidence=conf_threshold)
    else:
        pred_boxes, pred_classes, pred_scores = yolo3_postprocess_np(prediction, image_shape, anchors, num_classes, model_image_size, max_boxes=100, confidence=conf_threshold)

    return pred_boxes, pred_classes, pred_scores


def yolo_predict_keras(model, image, anchors, num_classes, model_image_size, conf_threshold):
    image_data = preprocess_image(image, model_image_size)
    image_shape = image.size

    prediction = model.predict([image_data])
    if len(anchors) == 5:
        # YOLOv2 use 5 anchors
        pred_boxes, pred_classes, pred_scores = yolo2_postprocess_np(prediction, image_shape, anchors, num_classes, model_image_size, max_boxes=100, confidence=conf_threshold)
    else:
        pred_boxes, pred_classes, pred_scores = yolo3_postprocess_np(prediction, image_shape, anchors, num_classes, model_image_size, max_boxes=100, confidence=conf_threshold)

    return pred_boxes, pred_classes, pred_scores


def get_prediction_class_records(model, model_format, annotation_records, anchors, class_names, model_image_size, conf_threshold, save_result):
    '''
    Do the predict with YOLO model on annotation images to get predict class dict

    predict class dict would contain image_name, coordinary and score, and
    sorted by score:
    pred_classes_records = {
        'car': [
                ['000001.jpg','94,115,203,232',0.98],
                ['000002.jpg','82,64,154,128',0.93],
                ...
               ],
        ...
    }
    '''
    if model_format == 'MNN':
        #MNN inference engine need create session
        session = model.createSession()

    # create txt file to save prediction result, with
    # save format as annotation file but adding score, like:
    #
    # path/to/img1.jpg 50,100,150,200,0,0.86 30,50,200,120,3,0.95
    #
    os.makedirs('result', exist_ok=True)
    result_file = open(os.path.join('result','detection_result.txt'), 'w')

    pred_classes_records = OrderedDict()
    pbar = tqdm(total=len(annotation_records), desc='Eval model')
    for (image_name, gt_records) in annotation_records.items():
        image = Image.open(image_name)
        if image.mode != 'RGB':
            image = image.convert('RGB')
        image_array = np.array(image, dtype='uint8')

        # support of tflite model
        if model_format == 'TFLITE':
            pred_boxes, pred_classes, pred_scores = yolo_predict_tflite(model, image, anchors, len(class_names), conf_threshold)
        # support of MNN model
        elif model_format == 'MNN':
            pred_boxes, pred_classes, pred_scores = yolo_predict_mnn(model, session, image, anchors, len(class_names), conf_threshold)
        # support of TF 1.x frozen pb model
        elif model_format == 'PB':
            pred_boxes, pred_classes, pred_scores = yolo_predict_pb(model, image, anchors, len(class_names), model_image_size, conf_threshold)
        # support of ONNX model
        elif model_format == 'ONNX':
            pred_boxes, pred_classes, pred_scores = yolo_predict_onnx(model, image, anchors, len(class_names), conf_threshold)
        # normal keras h5 model
        elif model_format == 'H5':
            pred_boxes, pred_classes, pred_scores = yolo_predict_keras(model, image, anchors, len(class_names), model_image_size, conf_threshold)
        else:
            raise ValueError('invalid model format')

        #print('Found {} boxes for {}'.format(len(pred_boxes), image_name))
        pbar.update(1)

        # save prediction result to txt
        result_file.write(image_name)
        for box, cls, score in zip(pred_boxes, pred_classes, pred_scores):
            xmin, ymin, xmax, ymax = box
            box_annotation = " %d,%d,%d,%d,%d,%f" % (
                xmin, ymin, xmax, ymax, cls, score)
            result_file.write(box_annotation)
        result_file.write('\n')
        result_file.flush()

        if save_result:

            gt_boxes, gt_classes, gt_scores = transform_gt_record(gt_records, class_names)

            result_dir=os.path.join('result','detection')
            os.makedirs(result_dir, exist_ok=True)
            colors = get_colors(class_names)
            image_array = draw_boxes(image_array, gt_boxes, gt_classes, gt_scores, class_names, colors=None, show_score=False)
            image_array = draw_boxes(image_array, pred_boxes, pred_classes, pred_scores, class_names, colors)
            image = Image.fromarray(image_array)
            # here we handle the RGBA image
            if(len(image.split()) == 4):
                r, g, b, a = image.split()
                image = Image.merge("RGB", (r, g, b))
            image.save(os.path.join(result_dir, image_name.split(os.path.sep)[-1]))

        # Nothing detected
        if pred_boxes is None or len(pred_boxes) == 0:
            continue

        for box, cls, score in zip(pred_boxes, pred_classes, pred_scores):
            pred_class_name = class_names[cls]
            xmin, ymin, xmax, ymax = box
            coordinate = "{},{},{},{}".format(xmin, ymin, xmax, ymax)

            #append or add predict class item
            record = [os.path.basename(image_name), coordinate, score]
            if pred_class_name in pred_classes_records:
                pred_classes_records[pred_class_name].append(record)
            else:
                pred_classes_records[pred_class_name] = list([record])

    # sort pred_classes_records for each class according to score
    for pred_class_list in pred_classes_records.values():
        pred_class_list.sort(key=lambda ele: ele[2], reverse=True)

    pbar.close()
    result_file.close()
    return pred_classes_records


def box_iou(pred_box, gt_box):
    '''
    Calculate iou for predict box and ground truth box
    Param
         pred_box: predict box coordinate
                   (xmin,ymin,xmax,ymax) format
         gt_box: ground truth box coordinate
                 (xmin,ymin,xmax,ymax) format
    Return
         iou value
    '''
    # get intersection box
    inter_box = [max(pred_box[0], gt_box[0]), max(pred_box[1], gt_box[1]), min(pred_box[2], gt_box[2]), min(pred_box[3], gt_box[3])]
    inter_w = max(0.0, inter_box[2] - inter_box[0] + 1)
    inter_h = max(0.0, inter_box[3] - inter_box[1] + 1)

    # compute overlap (IoU) = area of intersection / area of union
    pred_area = (pred_box[2] - pred_box[0] + 1) * (pred_box[3] - pred_box[1] + 1)
    gt_area = (gt_box[2] - gt_box[0] + 1) * (gt_box[3] - gt_box[1] + 1)
    inter_area = inter_w * inter_h
    union_area = pred_area + gt_area - inter_area
    return 0 if union_area == 0 else float(inter_area) / float(union_area)


def match_gt_box(pred_record, gt_records, iou_threshold=0.5):
    '''
    Search gt_records list and try to find a matching box for the predict box

    Param
         pred_record: with format ['image_file', 'xmin,ymin,xmax,ymax', score]
         gt_records: record list with format
                     [
                      ['image_file', 'xmin,ymin,xmax,ymax', 'usage'],
                      ['image_file', 'xmin,ymin,xmax,ymax', 'usage'],
                      ...
                     ]
         iou_threshold:

         pred_record and gt_records should be from same annotation image file

    Return
         matching gt_record index. -1 when there's no matching gt
    '''
    max_iou = 0.0
    max_index = -1
    #get predict box coordinate
    pred_box = [float(x) for x in pred_record[1].split(',')]

    for i, gt_record in enumerate(gt_records):
        #get ground truth box coordinate
        gt_box = [float(x) for x in gt_record[1].split(',')]
        iou = box_iou(pred_box, gt_box)

        # if the ground truth has been assigned to other
        # prediction, we couldn't reuse it
        if iou > max_iou and gt_record[2] == 'unused' and pred_record[0] == gt_record[0]:
            max_iou = iou
            max_index = i

    # drop the prediction if couldn't match iou threshold
    if max_iou < iou_threshold:
        max_index = -1

    return max_index

def voc_ap(rec, prec):
    """
    --- Official matlab code VOC2012---
    mrec=[0 ; rec ; 1];
    mpre=[0 ; prec ; 0];
    for i=numel(mpre)-1:-1:1
        mpre(i)=max(mpre(i),mpre(i+1));
    end
    i=find(mrec(2:end)~=mrec(1:end-1))+1;
    ap=sum((mrec(i)-mrec(i-1)).*mpre(i));
    """
    rec.insert(0, 0.0)  # insert 0.0 at begining of list
    rec.append(1.0)  # insert 1.0 at end of list
    mrec = rec[:]
    prec.insert(0, 0.0)  # insert 0.0 at begining of list
    prec.append(0.0)  # insert 0.0 at end of list
    mpre = prec[:]
    """
     This part makes the precision monotonically decreasing
      (goes from the end to the beginning)
    """
    # matlab indexes start in 1 but python in 0, so I have to do:
    #   range(start=(len(mpre) - 2), end=0, step=-1)
    # also the python function range excludes the end, resulting in:
    #   range(start=(len(mpre) - 2), end=-1, step=-1)
    for i in range(len(mpre) - 2, -1, -1):
        mpre[i] = max(mpre[i], mpre[i + 1])
    """
     This part creates a list of indexes where the recall changes
    """
    # matlab: i=find(mrec(2:end)~=mrec(1:end-1))+1;
    i_list = []
    for i in range(1, len(mrec)):
        if mrec[i] != mrec[i - 1]:
            i_list.append(i)  # if it was matlab would be i + 1
    """
     The Average Precision (AP) is the area under the curve
      (numerical integration)
    """
    # matlab: ap=sum((mrec(i)-mrec(i-1)).*mpre(i));
    ap = 0.0
    for i in i_list:
        ap += ((mrec[i] - mrec[i - 1]) * mpre[i])
    return ap, mrec, mpre

'''
def voc_ap(rec, prec, use_07_metric=False):
    if use_07_metric:
        # 11 point metric
        ap = 0.
        for t in np.arange(0., 1.1, 0.1):
            if np.sum(rec >= t) == 0:
                p = 0
            else:
                p = np.max(prec[rec >= t])
            ap = ap + p / 11.
    else:
        mrec = np.concatenate(([0.], rec, [1.]))
        mpre = np.concatenate(([0.], prec, [0.]))
        for i in range(mpre.size - 1, 0, -1):
            mpre[i - 1] = np.maximum(mpre[i - 1], mpre[i])
        i = np.where(mrec[1:] != mrec[:-1])[0]
        ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1])
    return ap, mrec, mpre
'''


def get_rec_prec(true_positive, false_positive, gt_records):
    '''
    Calculate precision/recall based on true_positive, false_positive
    result.
    '''
    cumsum = 0
    for idx, val in enumerate(false_positive):
        false_positive[idx] += cumsum
        cumsum += val

    cumsum = 0
    for idx, val in enumerate(true_positive):
        true_positive[idx] += cumsum
        cumsum += val

    rec = true_positive[:]
    for idx, val in enumerate(true_positive):
        rec[idx] = (float(true_positive[idx]) / len(gt_records)) if len(gt_records) != 0 else 0

    prec = true_positive[:]
    for idx, val in enumerate(true_positive):
        prec[idx] = float(true_positive[idx]) / (false_positive[idx] + true_positive[idx])

    return rec, prec


def draw_rec_prec(rec, prec, mrec, mprec, class_name, ap):
    """
     Draw plot
    """
    plt.plot(rec, prec, '-o')
    # add a new penultimate point to the list (mrec[-2], 0.0)
    # since the last line segment (and respective area) do not affect the AP value
    area_under_curve_x = mrec[:-1] + [mrec[-2]] + [mrec[-1]]
    area_under_curve_y = mprec[:-1] + [0.0] + [mprec[-1]]
    plt.fill_between(area_under_curve_x, 0, area_under_curve_y, alpha=0.2, edgecolor='r')
    # set window title
    fig = plt.gcf() # gcf - get current figure
    fig.canvas.set_window_title('AP ' + class_name)
    # set plot title
    plt.title('class: ' + class_name + ' AP = {}%'.format(ap*100))
    #plt.suptitle('This is a somewhat long figure title', fontsize=16)
    # set axis titles
    plt.xlabel('Recall')
    plt.ylabel('Precision')
    # optional - set axes
    axes = plt.gca() # gca - get current axes
    axes.set_xlim([0.0,1.0])
    axes.set_ylim([0.0,1.05]) # .05 to give some extra space
    # Alternative option -> wait for button to be pressed
    #while not plt.waitforbuttonpress(): pass # wait for key display
    # Alternative option -> normal display
    #plt.show()
    # save the plot
    rec_prec_plot_path = os.path.join('result','classes')
    os.makedirs(rec_prec_plot_path, exist_ok=True)
    fig.savefig(os.path.join(rec_prec_plot_path, class_name + ".png"))
    plt.cla() # clear axes for next plot


def adjust_axes(r, t, fig, axes):
    """
     Plot - adjust axes
    """
    # get text width for re-scaling
    bb = t.get_window_extent(renderer=r)
    text_width_inches = bb.width / fig.dpi
    # get axis width in inches
    current_fig_width = fig.get_figwidth()
    new_fig_width = current_fig_width + text_width_inches
    propotion = new_fig_width / current_fig_width
    # get axis limit
    x_lim = axes.get_xlim()
    axes.set_xlim([x_lim[0], x_lim[1]*propotion])


def draw_plot_func(dictionary, n_classes, window_title, plot_title, x_label, output_path, to_show, plot_color, true_p_bar):
    """
     Draw plot using Matplotlib
    """
    # sort the dictionary by decreasing value, into a list of tuples
    sorted_dic_by_value = sorted(dictionary.items(), key=operator.itemgetter(1))
    # unpacking the list of tuples into two lists
    sorted_keys, sorted_values = zip(*sorted_dic_by_value)
    #
    if true_p_bar != "":
        """
         Special case to draw in (green=true predictions) & (red=false predictions)
        """
        fp_sorted = []
        tp_sorted = []
        for key in sorted_keys:
            fp_sorted.append(dictionary[key] - true_p_bar[key])
            tp_sorted.append(true_p_bar[key])
        plt.barh(range(n_classes), fp_sorted, align='center', color='crimson', label='False Predictions')
        plt.barh(range(n_classes), tp_sorted, align='center', color='forestgreen', label='True Predictions', left=fp_sorted)
        # add legend
        plt.legend(loc='lower right')
        """
         Write number on side of bar
        """
        fig = plt.gcf() # gcf - get current figure
        axes = plt.gca()
        r = fig.canvas.get_renderer()
        for i, val in enumerate(sorted_values):
            fp_val = fp_sorted[i]
            tp_val = tp_sorted[i]
            fp_str_val = " " + str(fp_val)
            tp_str_val = fp_str_val + " " + str(tp_val)
            # trick to paint multicolor with offset:
            #   first paint everything and then repaint the first number
            t = plt.text(val, i, tp_str_val, color='forestgreen', va='center', fontweight='bold')
            plt.text(val, i, fp_str_val, color='crimson', va='center', fontweight='bold')
            if i == (len(sorted_values)-1): # largest bar
                adjust_axes(r, t, fig, axes)
    else:
      plt.barh(range(n_classes), sorted_values, color=plot_color)
      """
       Write number on side of bar
      """
      fig = plt.gcf() # gcf - get current figure
      axes = plt.gca()
      r = fig.canvas.get_renderer()
      for i, val in enumerate(sorted_values):
          str_val = " " + str(val) # add a space before
          if val < 1.0:
              str_val = " {0:.2f}".format(val)
          t = plt.text(val, i, str_val, color=plot_color, va='center', fontweight='bold')
          # re-set axes to show number inside the figure
          if i == (len(sorted_values)-1): # largest bar
              adjust_axes(r, t, fig, axes)
    # set window title
    fig.canvas.set_window_title(window_title)
    # write classes in y axis
    tick_font_size = 12
    plt.yticks(range(n_classes), sorted_keys, fontsize=tick_font_size)
    """
     Re-scale height accordingly
    """
    init_height = fig.get_figheight()
    # comput the matrix height in points and inches
    dpi = fig.dpi
    height_pt = n_classes * (tick_font_size * 1.4) # 1.4 (some spacing)
    height_in = height_pt / dpi
    # compute the required figure height
    top_margin = 0.15    # in percentage of the figure height
    bottom_margin = 0.05 # in percentage of the figure height
    figure_height = height_in / (1 - top_margin - bottom_margin)
    # set new height
    if figure_height > init_height:
        fig.set_figheight(figure_height)

    # set plot title
    plt.title(plot_title, fontsize=14)
    # set axis titles
    # plt.xlabel('classes')
    plt.xlabel(x_label, fontsize='large')
    # adjust size of window
    fig.tight_layout()
    # save the plot
    fig.savefig(output_path)
    # show image
    if to_show:
        plt.show()
    # close the plot
    plt.close()


def calc_AP(gt_records, pred_records, class_name, iou_threshold, show_result):
    '''
    Calculate AP value for one class records

    Param
         gt_records: ground truth records list for one class, with format:
                     [
                      ['image_file', 'xmin,ymin,xmax,ymax'],
                      ['image_file', 'xmin,ymin,xmax,ymax'],
                      ...
                     ]
         pred_record: predict records for one class, with format:
                     [
                      ['image_file', 'xmin,ymin,xmax,ymax', score],
                      ['image_file', 'xmin,ymin,xmax,ymax', score],
                      ...
                     ]
    Return
         AP value for the class
    '''
    # append usage flag in gt_records for matching gt search
    gt_records = [gt_record + ['unused'] for gt_record in gt_records]

    # init true_positive and false_positive list
    nd = len(pred_records)  # number of predict data
    true_positive = [0] * nd
    false_positive = [0] * nd
    true_positive_count = 0
    # assign predictions to ground truth objects
    for idx, pred_record in enumerate(pred_records):
        # filter out gt record from same image
        image_gt_records = [ gt_record for gt_record in gt_records if gt_record[0] == pred_record[0]]

        i = match_gt_box(pred_record, image_gt_records, iou_threshold=iou_threshold)
        if i != -1:
            # find a valid gt obj to assign, set
            # true_positive list and mark image_gt_records.
            #
            # trick: gt_records will also be marked
            # as 'used', since image_gt_records is a
            # reference list
            image_gt_records[i][2] = 'used'
            true_positive[idx] = 1
            true_positive_count += 1
        else:
            false_positive[idx] = 1

    # compute precision/recall
    rec, prec = get_rec_prec(true_positive, false_positive, gt_records)
    ap, mrec, mprec = voc_ap(rec, prec)
    if show_result:
        draw_rec_prec(rec, prec, mrec, mprec, class_name, ap)

    return ap, true_positive_count


def plot_Pascal_AP_result(count_images, count_true_positives, num_classes,
                          gt_counter_per_class, pred_counter_per_class,
                          precision_dict, recall_dict, mPrec, mRec,
                          APs, mAP, iou_threshold):
    '''
     Plot the total number of occurences of each class in the ground-truth
    '''
    window_title = "Ground-Truth Info"
    plot_title = "Ground-Truth\n" + "(" + str(count_images) + " files and " + str(num_classes) + " classes)"
    x_label = "Number of objects per class"
    output_path = os.path.join('result','Ground-Truth_Info.png')
    draw_plot_func(gt_counter_per_class, num_classes, window_title, plot_title, x_label, output_path, to_show=False, plot_color='forestgreen', true_p_bar='')

    '''
     Plot the total number of occurences of each class in the "predicted" folder
    '''
    window_title = "Predicted Objects Info"
    # Plot title
    plot_title = "Predicted Objects\n" + "(" + str(count_images) + " files and "
    count_non_zero_values_in_dictionary = sum(int(x) > 0 for x in list(pred_counter_per_class.values()))
    plot_title += str(count_non_zero_values_in_dictionary) + " detected classes)"
    # end Plot title
    x_label = "Number of objects per class"
    output_path = os.path.join('result','Predicted_Objects_Info.png')
    draw_plot_func(pred_counter_per_class, len(pred_counter_per_class), window_title, plot_title, x_label, output_path, to_show=False, plot_color='forestgreen', true_p_bar=count_true_positives)

    '''
     Draw mAP plot (Show AP's of all classes in decreasing order)
    '''
    window_title = "mAP"
    plot_title = "mAP@IoU={0}: {1:.2f}%".format(iou_threshold, mAP)
    x_label = "Average Precision"
    output_path = os.path.join('result','mAP.png')
    draw_plot_func(APs, num_classes, window_title, plot_title, x_label, output_path, to_show=False, plot_color='royalblue', true_p_bar='')

    '''
     Draw Precision plot (Show Precision of all classes in decreasing order)
    '''
    window_title = "Precision"
    plot_title = "mPrec@IoU={0}: {1:.2f}%".format(iou_threshold, mPrec)
    x_label = "Precision rate"
    output_path = os.path.join('result','Precision.png')
    draw_plot_func(precision_dict, len(precision_dict), window_title, plot_title, x_label, output_path, to_show=False, plot_color='royalblue', true_p_bar='')

    '''
     Draw Recall plot (Show Recall of all classes in decreasing order)
    '''
    window_title = "Recall"
    plot_title = "mRec@IoU={0}: {1:.2f}%".format(iou_threshold, mRec)
    x_label = "Recall rate"
    output_path = os.path.join('result','Recall.png')
    draw_plot_func(recall_dict, len(recall_dict), window_title, plot_title, x_label, output_path, to_show=False, plot_color='royalblue', true_p_bar='')


def get_mean_metric(metric_records, gt_classes_records):
    '''
    Calculate mean metric, but only count classes which have ground truth object

    Param
        metric_records: metric dict like:
            metric_records = {
                'aeroplane': 0.79,
                'bicycle': 0.79,
                    ...
                'tvmonitor': 0.71,
            }
        gt_classes_records: ground truth class dict like:
            gt_classes_records = {
                'car': [
                    ['000001.jpg','100,120,200,235'],
                    ['000002.jpg','85,63,156,128'],
                    ...
                    ],
                ...
            }
    Return
         mean_metric: float value of mean metric
    '''
    mean_metric = 0.0
    count = 0
    for (class_name, metric) in metric_records.items():
        if (class_name in gt_classes_records) and (len(gt_classes_records[class_name]) != 0):
            mean_metric += metric
            count += 1
    mean_metric = (mean_metric/count)*100 if count != 0 else 0.0
    return mean_metric


def compute_mAP_PascalVOC(annotation_records, gt_classes_records, pred_classes_records, class_names, iou_threshold, show_result=True):
    '''
    Compute PascalVOC style mAP
    '''
    APs = {}
    count_true_positives = {class_name: 0 for class_name in list(gt_classes_records.keys())}
    #get AP value for each of the ground truth classes
    for _, class_name in enumerate(class_names):
        #if there's no gt obj for a class, record 0
        if class_name not in gt_classes_records:
            APs[class_name] = 0.
            continue
        gt_records = gt_classes_records[class_name]
        #if we didn't detect any obj for a class, record 0
        if class_name not in pred_classes_records:
            APs[class_name] = 0.
            continue
        pred_records = pred_classes_records[class_name]
        ap, true_positive_count = calc_AP(gt_records, pred_records, class_name, iou_threshold, show_result)
        APs[class_name] = ap
        count_true_positives[class_name] = true_positive_count

    #sort AP result by value, in descending order
    APs = OrderedDict(sorted(APs.items(), key=operator.itemgetter(1), reverse=True))

    #get mAP percentage value
    #mAP = np.mean(list(APs.values()))*100
    mAP = get_mean_metric(APs, gt_classes_records)

    #get GroundTruth count per class
    gt_counter_per_class = {}
    for (class_name, info_list) in gt_classes_records.items():
        gt_counter_per_class[class_name] = len(info_list)

    #get Precision count per class
    pred_counter_per_class = {class_name: 0 for class_name in list(gt_classes_records.keys())}
    for (class_name, info_list) in pred_classes_records.items():
        pred_counter_per_class[class_name] = len(info_list)


    #get the precision & recall
    precision_dict = {}
    recall_dict = {}
    for (class_name, gt_count) in gt_counter_per_class.items():
        if (class_name not in pred_counter_per_class) or (class_name not in count_true_positives) or pred_counter_per_class[class_name] == 0:
            precision_dict[class_name] = 0.
        else:
            precision_dict[class_name] = float(count_true_positives[class_name]) / pred_counter_per_class[class_name]

        if class_name not in count_true_positives or gt_count == 0:
            recall_dict[class_name] = 0.
        else:
            recall_dict[class_name] = float(count_true_positives[class_name]) / gt_count

    #get mPrec, mRec
    #mPrec = np.mean(list(precision_dict.values()))*100
    #mRec = np.mean(list(recall_dict.values()))*100
    mPrec = get_mean_metric(precision_dict, gt_classes_records)
    mRec = get_mean_metric(recall_dict, gt_classes_records)


    if show_result:
        plot_Pascal_AP_result(len(annotation_records), count_true_positives, len(gt_classes_records),
                                  gt_counter_per_class, pred_counter_per_class,
                                  precision_dict, recall_dict, mPrec, mRec,
                                  APs, mAP, iou_threshold)
        #show result
        print('\nPascal VOC AP evaluation')
        for (class_name, AP) in APs.items():
            print('%s: AP %.4f, precision %.4f, recall %.4f' % (class_name, AP, precision_dict[class_name], recall_dict[class_name]))
        print('mAP@IoU=%.2f result: %f' % (iou_threshold, mAP))
        print('mPrec@IoU=%.2f result: %f' % (iou_threshold, mPrec))
        print('mRec@IoU=%.2f result: %f' % (iou_threshold, mRec))

    #return mAP percentage value
    return mAP, APs



def compute_AP_COCO(annotation_records, gt_classes_records, pred_classes_records, class_names, show_result=True):
    '''
    Compute MSCOCO AP list on AP 0.5:0.05:0.95
    '''
    iou_threshold_list = np.arange(0.50, 1.00, 0.05)
    APs = {}
    pbar = tqdm(total=len(iou_threshold_list), desc='Eval COCO')
    for iou_threshold in iou_threshold_list:
        iou_threshold = round(iou_threshold, 2)
        mAP, _ = compute_mAP_PascalVOC(annotation_records, gt_classes_records, pred_classes_records, class_names, iou_threshold, show_result=False)
        APs[iou_threshold] = round(mAP, 6)
        pbar.update(1)

    pbar.close()

    #sort AP result by value, in descending order
    APs = OrderedDict(sorted(APs.items(), key=operator.itemgetter(1), reverse=True))

    #get overall AP percentage value
    AP = np.mean(list(APs.values()))

    if show_result:
        '''
         Draw MS COCO AP plot
        '''
        os.makedirs('result', exist_ok=True)
        window_title = "MSCOCO AP on different IOU"
        plot_title = "COCO AP = {0:.2f}%".format(AP)
        x_label = "Average Precision"
        output_path = os.path.join('result','COCO_AP.png')
        draw_plot_func(APs, len(APs), window_title, plot_title, x_label, output_path, to_show=False, plot_color='royalblue', true_p_bar='')

        print('\nMS COCO AP evaluation')
        for (iou_threshold, AP_value) in APs.items():
            print('IOU %.2f: AP %f' % (iou_threshold, AP_value))
        print('total AP: %f' % (AP))

    #return AP percentage value
    return AP, APs


def compute_AP_COCO_Scale(annotation_records, scale_gt_classes_records, pred_classes_records, class_names):
    '''
    Compute MSCOCO AP on different scale object: small, medium, large
    '''
    scale_APs = {}
    for scale_key in ['small','medium','large']:
        gt_classes_records = scale_gt_classes_records[scale_key]
        scale_AP, _ = compute_AP_COCO(annotation_records, gt_classes_records, pred_classes_records, class_names, show_result=False)
        scale_APs[scale_key] = round(scale_AP, 4)

    #get overall AP percentage value
    scale_mAP = np.mean(list(scale_APs.values()))

    '''
     Draw Scale AP plot
    '''
    os.makedirs('result', exist_ok=True)
    window_title = "MSCOCO AP on different scale"
    plot_title = "scale mAP = {0:.2f}%".format(scale_mAP)
    x_label = "Average Precision"
    output_path = os.path.join('result','COCO_scale_AP.png')
    draw_plot_func(scale_APs, len(scale_APs), window_title, plot_title, x_label, output_path, to_show=False, plot_color='royalblue', true_p_bar='')

    '''
     Draw Scale Object Sum plot
    '''
    for scale_key in ['small','medium','large']:
        gt_classes_records = scale_gt_classes_records[scale_key]
        gt_classes_sum = {}

        for _, class_name in enumerate(class_names):
            # summarize the gt object number for every class on different scale
            gt_classes_sum[class_name] = np.sum(len(gt_classes_records[class_name])) if class_name in gt_classes_records else 0

        total_sum = np.sum(list(gt_classes_sum.values()))

        window_title = "{} object number".format(scale_key)
        plot_title = "total {} object number = {}".format(scale_key, total_sum)
        x_label = "Object Number"
        output_path = os.path.join('result','{}_object_number.png'.format(scale_key))
        draw_plot_func(gt_classes_sum, len(gt_classes_sum), window_title, plot_title, x_label, output_path, to_show=False, plot_color='royalblue', true_p_bar='')

    print('\nMS COCO AP evaluation on different scale')
    for (scale, AP_value) in scale_APs.items():
        print('%s scale: AP %f' % (scale, AP_value))
    print('total AP: %f' % (scale_mAP))


def add_gt_record(gt_records, gt_record, class_name):
    # append or add ground truth class item
    if class_name in gt_records:
        gt_records[class_name].append(gt_record)
    else:
        gt_records[class_name] = list([gt_record])

    return gt_records


def get_scale_gt_dict(gt_classes_records, class_names):
    '''
    Get ground truth class dict on different object scales, according to MS COCO metrics definition:
        small objects: area < 32^2
        medium objects: 32^2 < area < 96^2
        large objects: area > 96^2

    input gt_classes_records would be like:
    gt_classes_records = {
        'car': [
                ['000001.jpg','100,120,200,235'],
                ['000002.jpg','85,63,156,128'],
                ...
               ],
        ...
    }
    return a record dict with following format, for AP/AR eval on different scale:
        scale_gt_classes_records = {
            'small': {
                'car': [
                        ['000001.jpg','100,120,200,235'],
                        ['000002.jpg','85,63,156,128'],
                        ...
                       ],
                ...
            },

            'medium': {
                'car': [
                        ['000003.jpg','100,120,200,235'],
                        ['000004.jpg','85,63,156,128'],
                        ...
                       ],
                ...
            },

            'large': {
                'car': [
                        ['000005.jpg','100,120,200,235'],
                        ['000006.jpg','85,63,156,128'],
                        ...
                       ],
                ...
            }
        }
    '''
    scale_gt_classes_records = {}
    small_gt_records = {}
    medium_gt_records = {}
    large_gt_records = {}

    for _, class_name in enumerate(class_names):
        gt_records = gt_classes_records[class_name]

        for (image_file, box) in gt_records:
            # get box area based on coordinate
            box_coord = [int(p) for p in box.split(',')]
            box_area = (box_coord[2] - box_coord[0]) * (box_coord[3] - box_coord[1])

            # add to corresponding gt records dict according to area size
            if box_area <= 32*32:
                small_gt_records = add_gt_record(small_gt_records, [image_file, box], class_name)
            elif box_area > 32*32 and box_area <= 96*96:
                medium_gt_records = add_gt_record(medium_gt_records, [image_file, box], class_name)
            elif box_area > 96*96:
                large_gt_records = add_gt_record(large_gt_records, [image_file, box], class_name)

    # form up scale_gt_classes_records
    scale_gt_classes_records['small'] = small_gt_records
    scale_gt_classes_records['medium'] = medium_gt_records
    scale_gt_classes_records['large'] = large_gt_records

    return scale_gt_classes_records



def eval_AP(model, model_format, annotation_lines, anchors, class_names, model_image_size, eval_type, iou_threshold, conf_threshold, save_result):
    '''
    Compute AP for detection model on annotation dataset
    '''
    annotation_records, gt_classes_records = annotation_parse(annotation_lines, class_names)
    pred_classes_records = get_prediction_class_records(model, model_format, annotation_records, anchors, class_names, model_image_size, conf_threshold, save_result)
    AP = 0.0

    if eval_type == 'VOC':
        AP, _ = compute_mAP_PascalVOC(annotation_records, gt_classes_records, pred_classes_records, class_names, iou_threshold)
    elif eval_type == 'COCO':
        AP, _ = compute_AP_COCO(annotation_records, gt_classes_records, pred_classes_records, class_names)
        # get AP for different scale: small, medium, large
        scale_gt_classes_records = get_scale_gt_dict(gt_classes_records, class_names)
        compute_AP_COCO_Scale(annotation_records, scale_gt_classes_records, pred_classes_records, class_names)
    else:
        raise ValueError('Unsupported evaluation type')

    return AP


#load TF 1.x frozen pb graph
def load_graph(model_path):
    # We parse the graph_def file
    with tf.gfile.GFile(model_path, "rb") as f:
        graph_def = tf.GraphDef()
        graph_def.ParseFromString(f.read())

    # We load the graph_def in the default graph
    with tf.Graph().as_default() as graph:
        tf.import_graph_def(
            graph_def,
            input_map=None,
            return_elements=None,
            name="graph",
            op_dict=None,
            producer_op_list=None
        )
    return graph


def load_eval_model(model_path):
    # support of tflite model
    if model_path.endswith('.tflite'):
        from tensorflow.lite.python import interpreter as interpreter_wrapper
        model = interpreter_wrapper.Interpreter(model_path=model_path)
        model.allocate_tensors()
        model_format = 'TFLITE'

    # support of MNN model
    elif model_path.endswith('.mnn'):
        model = MNN.Interpreter(model_path)
        model_format = 'MNN'

    # support of TF 1.x frozen pb model
    elif model_path.endswith('.pb'):
        model = load_graph(model_path)
        model_format = 'PB'

    # support of ONNX model
    elif model_path.endswith('.onnx'):
        model = onnxruntime.InferenceSession(model_path)
        model_format = 'ONNX'

    # normal keras h5 model
    elif model_path.endswith('.h5'):
        custom_object_dict = get_custom_objects()

        model = load_model(model_path, compile=False, custom_objects=custom_object_dict)
        model_format = 'H5'
        K.set_learning_phase(0)
    else:
        raise ValueError('invalid model file')

    return model, model_format


def main():
    # class YOLO defines the default value, so suppress any default here
    parser = argparse.ArgumentParser(argument_default=argparse.SUPPRESS, description='evaluate YOLO model (h5/pb/onnx/tflite/mnn) with test dataset')
    '''
    Command line options
    '''
    parser.add_argument(
        '--model_path', type=str, required=True,
        help='path to model file')

    parser.add_argument(
        '--anchors_path', type=str, required=True,
        help='path to anchor definitions')

    parser.add_argument(
        '--classes_path', type=str, required=False,
        help='path to class definitions, default configs/voc_classes.txt', default=os.path.join('configs' , 'voc_classes.txt'))

    parser.add_argument(
        '--annotation_file', type=str, required=True,
        help='test annotation txt file')

    parser.add_argument(
        '--eval_type', type=str,
        help='evaluation type (VOC/COCO), default=VOC', default='VOC')

    parser.add_argument(
        '--iou_threshold', type=float,
        help='IOU threshold for PascalVOC mAP, default=0.5', default=0.5)

    parser.add_argument(
        '--conf_threshold', type=float,
        help='confidence threshold for filtering box in postprocess, default=0.001', default=0.001)

    parser.add_argument(
        '--model_image_size', type=str,
        help='model image input size as <num>x<num>, default 416x416', default='416x416')

    parser.add_argument(
        '--save_result', default=False, action="store_true",
        help='Save the detection result image in result/detection dir'
    )

    args = parser.parse_args()

    # param parse
    anchors = get_anchors(args.anchors_path)
    class_names = get_classes(args.classes_path)
    height, width = args.model_image_size.split('x')
    model_image_size = (int(height), int(width))

    annotation_lines = get_dataset(args.annotation_file, shuffle=False)
    model, model_format = load_eval_model(args.model_path)

    start = time.time()
    eval_AP(model, model_format, annotation_lines, anchors, class_names, model_image_size, args.eval_type, args.iou_threshold, args.conf_threshold, args.save_result)
    end = time.time()
    print("Evaluation time cost: {:.6f}s".format(end - start))


if __name__ == '__main__':
    main()