bugs of integer division and subplot for one class

[YuCosine]

Thanks for your work! It's really handy, saving me a lot of time when learning YOLO in keras :)

I understand this program is for python 3. I run it with python 2.7 though, and a small modification will get it work.
In File "mean_average_precision/mean_average_precision/utils/bbox.py", function jaccard returns 0 when the coordinates of bounding boxes are all integers, because python 2.7 yields an integer for integer division. Modifying line 49 from
return inter / union
to
return inter.astype('float32') / union
works well.

Second, when calculating mAP for just one class, in File "mean_average_precision/mean_average_precision/detection_map.py", function plot raises an error

Traceback (most recent call last):
  File "mean_average_precision/detection_map.py", line 229, in plot
    for i, ax in enumerate(axes.flat):
AttributeError: 'AxesSubplot' object has no attribute 'flat'

because when grid == 1, fig, axes = plt.subplots(nrows=grid, ncols=grid) return axes as type of matplotlib.axes._subplots.AxesSubplot.
I modify it from

        grid = int(math.ceil(math.sqrt(self.n_class)))
        fig, axes = plt.subplots(nrows=grid, ncols=grid)
        mean_average_precision = []
        # TODO: data structure not optimal for this operation...
        for i, ax in enumerate(axes.flat):
            if i > self.n_class - 1:
                break
            precisions, recalls = self.compute_precision_recall_(i, interpolated)
            average_precision = self.compute_ap(precisions, recalls)
            self.plot_pr(ax, i, precisions, recalls, average_precision)
            mean_average_precision.append(average_precision)

to

        grid = int(math.ceil(math.sqrt(self.n_class)))
        if grid == 1:
            fig, ax = plt.subplots()
            mean_average_precision = []
            precisions, recalls = self.compute_precision_recall_(0, interpolated)
            average_precision = self.compute_ap(precisions, recalls)
            self.plot_pr(ax, 0, precisions, recalls, average_precision)
            mean_average_precision.append(average_precision)
        else:
            fig, axes = plt.subplots(nrows=grid, ncols=grid)
            mean_average_precision = []
            # TODO: data structure not optimal for this operation...
            for i, ax in enumerate(axes.flat):
                if i > self.n_class - 1:
                    break
                precisions, recalls = self.compute_precision_recall_(i, interpolated)
                average_precision = self.compute_ap(precisions, recalls)
                self.plot_pr(ax, i, precisions, recalls, average_precision)
                mean_average_precision.append(average_precision)

to get it work, but I think it can be more precise .

[MathGaron]

Hi, thanks for the bug report! I will fix this as soon as possible!
I am quite busy right now but it should be pushed somewhere around next week.

[TrungHieu-Le]

Hi YuCosine,
The source code that you fixed it run, but the result still wrong when we only predict one class.
For example:

pred_bb1 = np.array([])
pred_cls1 = np.array([],dtype=int)
pred_conf1 = np.array([])
gt_bb1 = np.array([[0.86132812, 0.48242188, 0.97460938, 0.6171875],
[0.18554688, 0.234375, 0.36132812, 0.41601562],
[0., 0.47265625, 0.0703125, 0.62109375],
[0.47070312, 0.3125, 0.77929688, 0.78125],
[0.8, 0.1, 0.9, 0.2]])
gt_cls1 = np.array([1, 1, 1, 1, 1,1 ])

Because the model didn't predict object so Mean Average Precision = 0, but we get Mean Average Precision = 1

[YuCosine]

@TrungHieu-Le
With only one class, the gt of class label should be 0.
I modify the last line to gt_cls1 = np.array([0, 0, 0, 0, 0,0 ]) and get MAP=0

[TrungHieu-Le]

@MathGaron
Thank for your help.
Have a nice weekend!

[MathGaron]

Actually @YuCosine did all the work for now! Thanks again!