show_errors.py

from tidecv import TIDE, datasets
import argparse
import seaborn as sns
import matplotlib.pyplot as plt
import json
import pandas as pd
import numpy as np
import math
from matplotlib.ticker import MultipleLocator

def arg():
    parser = argparse.ArgumentParser()
    parser.add_argument('--annotation', '-a', type=str, default='examples/instances_val2017.json')
    parser.add_argument('--result', '-r', default='examples/coco_instances_results.json', type=str)
    parser.add_argument('--name', '-n', default='', type=str)
    parser.add_argument('--show', type=bool, default=True)
    parser.add_argument('--normalize', type=bool, default=False)
    args = parser.parse_args()
    return args


def main():
    args = arg()

    tide = TIDE()
    gt = datasets.COCO(args.annotation)
    bbox_results = datasets.COCOResult(args.result)
    #tide.evaluate(datasets.COCO(), datasets.COCOResult('path/to/your/results/file'), mode=TIDE.BOX) # Use TIDE.MASK for masks
    run = tide.evaluate(gt, bbox_results, mode=TIDE.BOX, name=args.name) # Use TIDE.MASK for masks
    tide.summarize()
    tide.plot('./result') 

    errors = tide.get_main_per_class_errors()
    error_names = ['Cls', 'Loc', 'Both', 'Dupe', 'Bkg', 'Miss']
    total_table = []
    for idx in sorted(errors[''][error_names[0]].keys()):
        rows_items = [gt.classes[idx], ]
        rows_items += [run.ap_data.objs[idx].get_ap()]
        rows_items += [ errors[''][name][idx] for name in error_names]
        total_table.append(rows_items)
        
    df = pd.DataFrame(total_table, columns=['Name', 'AP']+error_names)
    maximum_dap = math.ceil(df[error_names].max().max())

    sns.set(font_scale=0.4)
    g = sns.PairGrid(df.sort_values("AP", ascending=False),
                 x_vars=['AP']+error_names,
                 y_vars=["Name"],
                 height=8, aspect=.25)
    g.map(sns.stripplot, size=8, orient="h", jitter=False,
          palette="flare_r", linewidth=1, edgecolor="w")
    g.axes[0,0].set_xlim(0,100)
    for idx in range(1, g.axes.shape[1]):
        g.axes[0,idx].set_xlim(0,maximum_dap)
        g.axes[0,idx].xaxis.set_major_locator(MultipleLocator(5))

    titles = ['AP'] + error_names
    for ax, title in zip(g.axes.flat, titles):
        ax.set(title=title)
        ax.xaxis.grid(False)
        ax.yaxis.grid(True)

    sns.despine(left=True, bottom=True)
    plt.subplots_adjust(left=0.05, top=0.98)
    plt.savefig('per_accuracy_and_errors.png')

    ret = tide.get_confusion_matrix()
    dat = ret[''].T

    if args.normalize:
        dat = dat / dat.astype(np.float).sum(axis=0)
    cm = pd.DataFrame(data=dat,
                      index=gt.classes.values(),
                      columns=gt.classes.values())
    sns.set(font_scale=0.4)
    fig, axes = plt.subplots(figsize=(10,8))
    sns.heatmap(cm, square=True, cbar=True, annot=False, cmap='Blues',
            xticklabels=True, yticklabels=True,
            linewidths=.5
            )
    plt.xlabel("Predict", fontsize=13)
    plt.ylabel("GT", fontsize=13)
    fig.subplots_adjust(bottom=0.15)
    plt.savefig('class_error_confusion_matrix.png')


if __name__ == '__main__':
    main()