analysisPlotFunctions.py

import numpy as np
import matplotlib.pyplot as plt
from matplotlib.pyplot import figure
import seaborn as sns
from collections import defaultdict

def plotClassAccuracy(df1, df2, label):
    '''Plot Accuracy per class for given test-set (e.g. Pred_Blurred_15_Label), for both German (df1) and Eur (df2) datasets,
    as well as the original label and original test-set predictions without augmentation for comparison
    '''
    numClasses = 43
    priorClass1 = [len(df1.loc[df1['Orig_Label'] == i]) for i in range(43)]
    priorClass2 = [len(df1.loc[df2['Orig_Label'] == i]) for i in range(43)]
    origAccGer = np.zeros(numClasses)
    origAccEur = np.zeros(numClasses)
    accuracies1 = np.zeros(numClasses)
    accuracies2 = np.zeros(numClasses)
    for idx,row in df1.iterrows():
        key = row['Orig_Label']
        if row[label] == key:
            accuracies1[key] += 1/priorClass1[key]
        if row['Pred_Label'] == key:
            origAccGer[key] += 1/priorClass1[key]
    for idx,row in df2.iterrows():
        key = row['Orig_Label']
        if row[label] == key:
            accuracies2[key] += 1/priorClass2[key]
        if row['Pred_Label'] == key:
            origAccEur[key] += 1/priorClass2[key]
    figure(num=None, figsize=(20, 10), dpi=80, facecolor='w', edgecolor='k')
    sns.set()
    w=0.2
    plt.bar(np.arange(0,43,1)-1.5*w, origAccGer, width=w, label = "Orig Test Set added Eur", color = 'blue')
    plt.bar(np.arange(0,43,1)-0.5*w, origAccEur, width=w, label = "Orig Test Set added Augmented Eur", color = 'red')
    plt.bar(np.arange(0,43,1)+0.5*w, accuracies1, width=w, label = "Augmented Test Set added Eur", color = 'orange')
    plt.bar(np.arange(0,43,1)+1.5*w, accuracies2, width=w, label = "Augmented Test Set augmented Eur", color = 'g')
    plt.xticks(np.arange(0,43,1))
    plt.xlabel("Class", size =14)
    plt.ylabel("Accuracy", size= 14)
    plt.legend(prop={'size': 14})
    plt.title("Class ~ Accuracy {}".format(label), size= 14)
    fig = plt.gcf()
    fig.savefig("C:/Users/alexm/Desktop/AnalysisFigsAugment/classErr_{}.png".format(label), quality = 95, bbox_inches = 'tight')
    plt.show()
    # print(accuracies[0])

def plotBothFeat(df1, df2, feat):
    '''Plots count of prediction per class, for a given feature (e.g. Pred_Blurred_15_Label), as well as the
    True Label, Original prediction for both German (df1) and European (df2)
    '''
    numClasses = 43
    orig = np.zeros(numClasses)
    predOrigGer = np.zeros(numClasses)
    predOrigEur = np.zeros(numClasses)
    predGer = np.zeros(numClasses)
    predEur = np.zeros(numClasses)
    for i,row in df1.iterrows():
        orig[row['Orig_Label']] += 1
        predGer[row[feat]] += 1
        predOrigGer[row['Pred_Label']] += 1
    for i,row in df2.iterrows():
        predEur[row[feat]] += 1
        predOrigEur[row['Pred_Label']] += 1
    x = np.arange(0,43,1)
    w = 0.15
    figure(num=None, figsize=(20, 10), dpi=80, facecolor='w', edgecolor='k')
    sns.set()
    plt.bar(x-2*w, orig, width =w, label = "Original")
    plt.bar(x-1*w, predOrigGer, width =w, label = "Orig pred Eur")
    plt.bar(x, predOrigEur, width =w, label = "Orig pred Aug")
    plt.bar(x+w, predGer, width=w, label = "{} pred Eur".format(feat))
    plt.bar(x+2*w, predEur, width=w, label = "{} pred Aug".format(feat))
    plt.xticks(np.arange(0,43,1))
    plt.xlabel("Class", size =14)
    plt.ylabel("Prediction count", size= 14)
    plt.legend(prop={'size': 14})
    plt.title("Class ~ {} Predictions".format(feat), size= 14)
    fig = plt.gcf()
    # fig.savefig("C:/Users/alexm/Desktop/AnalysisFigsAugment/{}_err.png".format(feat), quality = 95, bbox_inches = 'tight')
    plt.show()

def plotAccDiff(files, names):
    "Plots the accuracy difference per class from base to all other networks as described in names"
    # Orig, Blur, Bright, Dark, Occ
    # files = ["output/VGG16ADDEUR/2019-05-22_12_23_40_log0.html", "output/VGG16ADDBLUR/2019-05-27_00_20_22_log0.html", "output/VGG16ADDBRIGHT/2019-05-27_01_06_00_log0.html",
    # "output/VGG16ADDDARK/2019-05-27_00_37_32_log0.html", "output/VGG16ADDOCC/2019-05-27_01_25_52_log0.html"]
    # files = ["output/DENSENETADDEUR/2019-06-01_14_30_28_log0.html", "output/DENSENETADDBLUR/2019-06-02_13_45_25_log0.html", "output/DENSENETADDLIGHT/2019-06-02_22_15_08_log0.html",
    # "output/DENSENETADDDARK/2019-06-02_14_11_11_log0.html", "output/DENSENETADDOCCL/2019-05-27_01_25_52_log0.html"]
    # names = ["DENSE_EUR", "DENSE_BLUR", "DENSE_BRIGHT", "DENSE_DARK", "DENSE_OCCL"]
    accuracies = defaultdict(list)
    diffAcc = defaultdict(list)
    for name in files:
        key = name
        f = open(name)
        for x in f:
            if "Number" in x.split() and "wrong" in x.split():
                accuracies[key].append(int(x.split()[-1]))
        diffAcc[key] = [(1 - (float(accuracies[key][i] / 12631))) - (1 - (float(accuracies[files[0]][i] / 12631))) for i in range(len(accuracies[key]))]
    x = np.arange(0,20,1)
    w = 1/len(files)
    figure(num=None, figsize=(20, 10), dpi=80, facecolor='w', edgecolor='k')
    sns.set()
    i = 0
    print(accuracies[files[0]])
    for idx,name in enumerate(files[1:]):
        plt.bar(x+(idx-2)*w, diffAcc[name], width=w, align='center', label = names[idx])
    xlabs = ["Orig", "Blur5", "Blur10", "Blur15", "Blur20", "Dark1", "Dark2", "Dark3", "Bright1", "Bright2", "Bright3", "Bright4","Bright5","Bright6", "Bright7", "Occl5", "Occl10", "Occl15", "Occl20", "Ocll25"]
    plt.xticks(ticks= x, labels = xlabs, size = 12)
    plt.xlabel("Test Set", size = 20)
    plt.ylabel("Accuracy difference", size = 20)
    plt.axhline(y = 0.0018, color = 'b', label = 'Significance threshold')
    plt.axhline(y = -0.0018, color = 'b')
    plt.legend(prop={'size': 14})
    fig = plt.gcf()
    fig.savefig("C:/Users/alexm/Desktop/AnalysisFigsAugment/AccuracyDifference.png", quality = 95, bbox_inches = 'tight')
    plt.show()

def plotPredFeat(df, feat, lab1, lab2, lab3, levels):
    '''Not currently called from plots.py. I used this initially  for one dataframe.
    You can use this to plot 3 features (e.g. Pred_Blurred_15_Label, Pred_Blurred_10_Label and Pred_Blurred_5_Label)
    for a single dataframe (e.g. German). Specify label (Blurred here), the 3 features (as above), as well as their levels [5,10,15] in this case
    '''
    numClasses = 43
    orig = np.zeros(numClasses)
    pred = np.zeros(numClasses)
    predb5 = np.zeros(numClasses)
    predb10 = np.zeros(numClasses)
    predb15 = np.zeros(numClasses)

    for i,row in df.iterrows():
        orig[row['Orig_Label']] += 1
        pred[row['Pred_Label']] += 1
        predb5[row[lab1]] += 1
        predb10[row[lab2]] += 1
        predb15[row[lab3]] += 1

    x = np.arange(0,43,1)
    w = 0.15
    sns.set()
    figure(num=None, figsize=(20, 10), dpi=80, facecolor='w', edgecolor='k')
    sns.set()
    ax = plt.subplot(111)
    plt.bar(x-2*w, orig, width=w, color='b', align='center', label = "True Label")
    plt.bar(x-w, pred, width=w, color='g', align='center', label = "Original Prediction")
    plt.bar(x, predb5, width=w, color='r', align='center', label = "{} {} Prediction".format(feat, levels[0]))
    plt.bar(x+w, predb10, width=w, color='y', align='center', label = "{} {} Prediction".format(feat, levels[1]))
    plt.bar(x+2*w, predb15, width=w, color='m', align='center', label = "{} {} Prediction".format(feat, levels[2]))
    plt.xticks(x)
    plt.xlabel("Class", size = 14)
    plt.ylabel("Prediction Count", size = 14)
    plt.title("Class predictions for various {} levels.".format(feat), size = 14)
    box = ax.get_position()
    ax.set_position([box.x0, box.y0, box.width * 0.8, box.height])
    plt.legend(prop={'size': 14}, loc = 'center left', bbox_to_anchor=(1, 0.5),
        fancybox=True, shadow=True)
    fig = plt.gcf()
    # fig.savefig("C:/Users/alexm/Desktop/AnalysisFigsAugment/{}_predictions.png".format(feat), quality = 95, bbox_inches = 'tight')
    plt.show()