allInOneTrainingPlots.py

Go to the documentation of this file.

import os, sys
import re
import pandas as pd
import argparse
import matplotlib.pyplot as plt
import numpy as np
from ROOT import *

#gROOT.Macro( os.path.expanduser( '$ROOTSYS/macros/rootlogon.C') )

parser = argparse.ArgumentParser()
parser.add_argument('--input', type=str, required=True)
parser.add_argument('--output', type=str, required=True)
args = parser.parse_args()
log1=args.input

def extractTrainLoss(logPath):
    lossPattern = r"Iteration (?P<iter_num>\d+) .*, loss = (?P<loss_val>[+-]?(\d+(\.\d*)?|\.\d+)([eE][+-]?\d+)?)"

    iterations = []
    losses = []

    for pair in re.findall(lossPattern, open(logPath, 'r').read()):
        iterations.append(int(pair[0]))
        losses.append(float(pair[1]))

    return pd.DataFrame({"iteration":iterations, "loss":losses})

df = extractTrainLoss(log1)

print df.head()

df.plot("iteration", "loss")

def extractTestLoss(logPath):
    iterationPattern = r"Iteration (?P<iter_num>\d+), Testing net"
    iterations = [int(r) for r
                  in re.findall(iterationPattern,
                                open(logPath, 'r').read())]

    testPattern =  r"Test net output #0: loss3/loss3 = (?P<loss_val>[+-]?(\d+(\.\d*)?|\.\d+)([eE][+-]?\d+)?)"
    tests = [float(r[0]) for r
             in re.findall(testPattern, open(logPath, 'r').read())]

    accPattern =  r" Test net output #1: loss3/top-1 = (?P<loss_val>[+-]?(\d+(\.\d*)?|\.\d+)([eE][+-]?\d+)?)"
    accs = [float(r[0]) for r
            in re.findall(accPattern, open(logPath, 'r').read())]
    print len(iterations), len(tests), len(accs)

    return pd.DataFrame({"iteration":iterations, "loss":tests, "accuracy":accs})

TGaxis.SetMaxDigits(8)

tdf = extractTestLoss(log1)

print tdf.head()

ax = df.plot("iteration", "loss", label="Train Loss")
tdf.plot("iteration", "loss", ax=ax, label="Test Loss" )

df["rollavg"] = pd.rolling_window(df.loss, window=200, win_type="boxcar")

ax = df.plot("iteration", "rollavg", label="Train Loss")
tdf.plot("iteration", "loss", ax=ax, label="Test Loss" )

tdf.plot("iteration", "accuracy", label="Test Accuracy" )

tdf["rollacc"] = pd.rolling_window(tdf.accuracy, window=10, win_type="boxcar")
ax = tdf.plot("iteration", "rollacc", label="Test Accuracy" )

trainLoss = TGraph()
trainLoss.SetName("trainLoss")

testLoss = TGraph()
testLoss.SetName("testLoss")

testAccuracy = TGraph()
testAccuracy.SetName("testAccuracy")

trainLoss.SetLineColor(kGreen+2);
testLoss.SetLineColor(kBlue);
testAccuracy.SetLineColor(kRed);

trainLoss.SetLineWidth(2);
testLoss.SetLineWidth(2);
testAccuracy.SetLineWidth(2);
           
iPoint = 0
for num, row in tdf.iterrows():
    testLoss.SetPoint(iPoint, row.iteration, row.loss)
    testAccuracy.SetPoint(iPoint, row.iteration, row.accuracy)
    iPoint += 1
    
iPoint = 0
for num, row in df.iterrows():
    if not np.isnan(row.rollavg):
        trainLoss.SetPoint(iPoint, row.iteration, row.rollavg)
        iPoint += 1

c1 = TCanvas("c1","gerrors2")
pad = TPad("pad","",0,0,1,1)
pad.SetFillColor(kWhite)
pad.SetTickx(False)
pad.SetTicky(False)
pad.Draw()
pad.cd()

hr = c1.DrawFrame(0,2.0,2000000,4.0)
hr.SetYTitle("Loss")
hr.SetXTitle("Number of Training Iterations")
hr.GetXaxis().CenterTitle()
hr.GetYaxis().SetTitleSize(.04)
hr.GetYaxis().CenterTitle()
hr.GetXaxis().SetTitleOffset(1.3)
hr.GetYaxis().SetTitleOffset(1.2)
hr.GetXaxis().SetNoExponent(kFALSE)
hr.GetXaxis().SetNdivisions(5)
hr.GetXaxis().SetLabelOffset(.02)
hr.SetFillColor(kWhite)

pad.GetFrame().SetFillColor(kWhite)
pad.GetFrame().SetBorderSize(12)

testLoss.Draw("l")
trainLoss.Draw("l same")
c1.cd()

overlay = TPad("overlay","",0,0,1,1)
overlay.SetFillStyle(4000)
overlay.SetFillColor(kWhite)
overlay.SetFrameFillStyle(4000)
overlay.SetTickx(False)
overlay.SetTicky(False)
overlay.Draw()
overlay.cd()
                                                      
xmin = pad.GetUxmin()
ymin = 0.1
xmax = pad.GetUxmax()
ymax = 0.4
hframe = overlay.DrawFrame(xmin,ymin,xmax,ymax)
hframe.SetFillColor(kWhite)
hframe.GetXaxis().SetLabelOffset(99)
hframe.GetYaxis().SetLabelOffset(99)
hframe.GetXaxis().SetTickLength(0)
hframe.GetYaxis().SetTickLength(0)

testAccuracy.Draw("l")

mylegend = TLegend(0.63,0.67,0.86,0.89)
mylegend.AddEntry(trainLoss,"Training Loss","L")
mylegend.AddEntry(testLoss,"Test Loss","L")
mylegend.AddEntry(testAccuracy,"Test Accuracy","L")
mylegend.SetLineWidth(0)
mylegend.SetBorderSize(0)
mylegend.SetFillStyle(0)
mylegend.Draw()

axis = TGaxis(xmax,ymin,xmax, ymax,ymin,ymax,510,"+L")
axis.SetLineColor(kBlack)
axis.SetTitle("Accuracy")
axis.SetTitleFont(42)
axis.SetTitleOffset(1.1)
axis.SetTitleColor(kBlack)
axis.SetLabelFont(42)
axis.CenterTitle()
axis.SetLabelColor(kBlack)
axis.SetLabelSize(.035)
axis.Draw()

c1.cd()
c1.Modified()
c1.Update()

outName = "learning_" + str(args.output) + ".png"

#c1.cd()
c1.Print(outName)

#raw_input("Press Enter to Exit")