source: trunk/bin/train @ 10

#!/usr/bin/env python3
#-*- mode: Python;-*-

import sys
import os
import time
import random
import statistics
import functools
import argparse
import pprint
import json


VERSION = "{DEVELOPMENT}"
if VERSION == "{DEVELOPMENT}":
    script_dir = '.'
    try:
        script_dir = os.path.dirname(os.path.realpath(__file__))
    except:
        try:
            script_dir = os.path.dirname(os.path.abspath(sys.argv[0]))
        except:
            pass
    sys.path.append("%s/../lib" % script_dir)

from nanownlib import *
from nanownlib.stats import *
from nanownlib.parallel import WorkerThreads
import nanownlib.storage



parser = argparse.ArgumentParser(
    description="")
#parser.add_argument('-c', dest='cases', type=str, default='{"short":10000,"long":1010000}',
#                    help='JSON representation of echo timing cases. Default: {"short":10000,"long":1010000}')
parser.add_argument('--retrain', action='append', default=[], help='Force a classifier to be retrained.  May be specified multiple times.')
parser.add_argument('--retest', action='append', default=[], help='Force a classifier to be retested.  May be specified multiple times.')
parser.add_argument('session_data', default=None,
                    help='Database file storing session information')
options = parser.parse_args()


def trainBoxTest(db, unusual_case, greater, num_observations):
    db.resetOffsets()
    
    def trainAux(low,high,num_trials):
        estimator = functools.partial(multiBoxTest, {'low':low, 'high':high}, greater)
        estimates = bootstrap3(estimator, db, 'train', unusual_case, num_observations, num_trials)
        null_estimates = bootstrap3(estimator, db, 'train_null', unusual_case, num_observations, num_trials)

        bad_estimates = len([e for e in estimates if e != 1])
        bad_null_estimates = len([e for e in null_estimates if e != 0])
        
        false_negatives = 100.0*bad_estimates/num_trials
        false_positives = 100.0*bad_null_estimates/num_trials
        return false_positives,false_negatives

    #start = time.time()
    wt = WorkerThreads(2, trainAux)
    
    num_trials = 200
    width = 1.0
    performance = []
    for low in range(0,50):
        wt.addJob(low, (low,low+width,num_trials))
    wt.wait()
    while not wt.resultq.empty():
        job_id,errors = wt.resultq.get()
        fp,fn = errors
        performance.append(((fp+fn)/2.0, job_id, fn, fp))
    performance.sort()
    #pprint.pprint(performance)
    #print(time.time()-start)
    
    num_trials = 200
    lows = [p[1] for p in performance[0:5]]
    widths = [w/10.0 for w in range(5,65,5)]
    performance = []
    for width in widths:
        false_positives = []
        false_negatives = []
        for low in lows:
            wt.addJob(low,(low,low+width,num_trials))
        wt.wait()
        while not wt.resultq.empty():
            job_id,errors = wt.resultq.get()
            fp,fn = errors
            false_negatives.append(fn)
            false_positives.append(fp)

        #print(width, false_negatives)
        #print(width, false_positives)
        #performance.append(((statistics.mean(false_positives)+statistics.mean(false_negatives))/2.0,
        #                    width, statistics.mean(false_negatives), statistics.mean(false_positives)))
        performance.append((abs(statistics.mean(false_positives)-statistics.mean(false_negatives)),
                            width, statistics.mean(false_negatives), statistics.mean(false_positives)))
    performance.sort()
    #pprint.pprint(performance)
    good_width = performance[0][1]
    #print("good_width:",good_width)


    num_trials = 500
    performance = []
    for low in lows:
        wt.addJob(low, (low,low+good_width,num_trials))
    wt.wait()
    while not wt.resultq.empty():
        job_id,errors = wt.resultq.get()
        fp,fn = errors
        performance.append(((fp+fn)/2.0, job_id, fn, fp))
    performance.sort()
    #pprint.pprint(performance)
    best_low = performance[0][1]
    #print("best_low:", best_low)

    
    num_trials = 500
    widths = [good_width+(x/100.0) for x in range(-70,75,5) if good_width+(x/100.0) > 0.0]
    performance = []
    for width in widths:
        wt.addJob(width, (best_low,best_low+width,num_trials))
    wt.wait()
    while not wt.resultq.empty():
        job_id,errors = wt.resultq.get()
        fp,fn = errors
        #performance.append(((fp+fn)/2.0, job_id, fn, fp))
        performance.append((abs(fp-fn), job_id, fn, fp))
    performance.sort()
    #pprint.pprint(performance)
    best_width=performance[0][1]
    #print("best_width:",best_width)
    #print("final_performance:", performance[0][0])

    wt.stop()
    params = json.dumps({"low":best_low,"high":best_low+best_width})
    return {'trial_type':"train",
            'num_observations':num_observations,
            'num_trials':num_trials,
            'params':params,
            'false_positives':performance[0][3],
            'false_negatives':performance[0][2]}


def trainSummary(summaryFunc, db, unusual_case, greater, num_observations):
    db.resetOffsets()
    stest = functools.partial(summaryTest, summaryFunc)
    
    def trainAux(distance, threshold, num_trials):
        estimator = functools.partial(stest, {'distance':distance,'threshold':threshold}, greater)
        estimates = bootstrap3(estimator, db, 'train', unusual_case, num_observations, num_trials)
        null_estimates = bootstrap3(estimator, db, 'train_null', unusual_case, num_observations, num_trials)

        bad_estimates = len([e for e in estimates if e != 1])
        bad_null_estimates = len([e for e in null_estimates if e != 0])
        
        false_negatives = 100.0*bad_estimates/num_trials
        false_positives = 100.0*bad_null_estimates/num_trials
        return false_positives,false_negatives

    #determine expected delta based on differences
    mean_diffs = [s['unusual_case']-s['other_cases'] for s in db.subseries('train', unusual_case)]
    threshold = summaryFunc(mean_diffs)/2.0
    #print("init_threshold:", threshold)
    
    wt = WorkerThreads(2, trainAux)
    
    num_trials = 500
    performance = []
    for distance in range(1,50):
        wt.addJob(distance, (distance,threshold,num_trials))
    wt.wait()
    while not wt.resultq.empty():
        job_id,errors = wt.resultq.get()
        fp,fn = errors
        performance.append(((fp+fn)/2.0, job_id, fn, fp))
    
    performance.sort()
    #pprint.pprint(performance)
    good_distance = performance[0][1]
    #print("good_distance:",good_distance)

    
    num_trials = 500
    performance = []
    for t in range(80,122,2):
        wt.addJob(threshold*(t/100.0), (good_distance,threshold*(t/100.0),num_trials))
    wt.wait()
    while not wt.resultq.empty():
        job_id,errors = wt.resultq.get()
        fp,fn = errors
        #performance.append(((fp+fn)/2.0, job_id, fn, fp))
        performance.append((abs(fp-fn), job_id, fn, fp))
    performance.sort()
    #pprint.pprint(performance)
    good_threshold = performance[0][1]
    #print("good_threshold:", good_threshold)

    
    num_trials = 500
    performance = []
    for d in [good_distance+s for s in range(-4,5) if good_distance+s > -1]:
        wt.addJob(d, (d,good_threshold,num_trials))
    wt.wait()
    while not wt.resultq.empty():
        job_id,errors = wt.resultq.get()
        fp,fn = errors
        performance.append(((fp+fn)/2.0, job_id, fn, fp))
    performance.sort()
    #pprint.pprint(performance)
    best_distance = performance[0][1]
    #print("best_distance:",best_distance)

    
    num_trials = 500
    performance = []
    for t in range(90,111):
        wt.addJob(good_threshold*(t/100.0), (best_distance,good_threshold*(t/100.0),num_trials))
    wt.wait()
    while not wt.resultq.empty():
        job_id,errors = wt.resultq.get()
        fp,fn = errors
        #performance.append(((fp+fn)/2.0, job_id, fn, fp))
        performance.append((abs(fp-fn), job_id, fn, fp))
    performance.sort()
    #pprint.pprint(performance)
    best_threshold = performance[0][1]
    #print("best_threshold:", best_threshold)

    wt.stop()
    params = json.dumps({'distance':best_distance,'threshold':best_threshold})
    return {'trial_type':"train",
            'num_observations':num_observations,
            'num_trials':num_trials,
            'params':params,
            'false_positives':performance[0][3],
            'false_negatives':performance[0][2]}


def trainKalman(db, unusual_case, greater, num_observations):
    db.resetOffsets()

    def trainAux(params, num_trials):
        estimator = functools.partial(kalmanTest, params, greater)
        estimates = bootstrap3(estimator, db, 'train', unusual_case, num_observations, num_trials)
        null_estimates = bootstrap3(estimator, db, 'train_null', unusual_case, num_observations, num_trials)
        
        bad_estimates = len([e for e in estimates if e != 1])
        bad_null_estimates = len([e for e in null_estimates if e != 0])
        
        false_negatives = 100.0*bad_estimates/num_trials
        false_positives = 100.0*bad_null_estimates/num_trials
        return false_positives,false_negatives
    
    mean_diffs = [s['unusual_case']-s['other_cases'] for s in db.subseries('train', unusual_case)]
    good_threshold = kfilter({},mean_diffs)['est'][-1]/2.0

    wt = WorkerThreads(2, trainAux)
    num_trials = 200
    performance = []
    for t in range(90,111):
        params = {'threshold':good_threshold*(t/100.0)}
        wt.addJob(good_threshold*(t/100.0), (params,num_trials))
    wt.wait()
    while not wt.resultq.empty():
        job_id,errors = wt.resultq.get()
        fp,fn = errors
        #performance.append(((fp+fn)/2.0, job_id, fn, fp))
        performance.append((abs(fp-fn), job_id, fn, fp))
    performance.sort()
    #pprint.pprint(performance)
    best_threshold = performance[0][1]
    #print("best_threshold:", best_threshold)
    params = {'threshold':best_threshold}

    wt.stop()
    
    return {'trial_type':"train",
            'num_observations':num_observations,
            'num_trials':num_trials,
            'params':json.dumps(params),
            'false_positives':performance[0][3],
            'false_negatives':performance[0][2]}

    
    #determine expected delta based on differences
classifiers = {'boxtest':{'train':trainBoxTest, 'test':multiBoxTest, 'train_results':[]},
               'midsummary':{'train':functools.partial(trainSummary, midsummary), 'test':midsummaryTest, 'train_results':[]},
               #'ubersummary':{'train':functools.partial(trainSummary, ubersummary), 'test':ubersummaryTest, 'train_results':[]},
               'quadsummary':{'train':functools.partial(trainSummary, quadsummary), 'test':quadsummaryTest, 'train_results':[]},
               'kalman':{'train':trainKalman, 'test':kalmanTest, 'train_results':[]},
               #'_trimean':{'train':None, 'test':trimeanTest, 'train_results':[]},
              }


db = nanownlib.storage.db(options.session_data)

import cProfile

def trainClassifier(db, unusual_case, greater, classifier, retrain=False):
    if retrain:
        print("Dropping stored training results...")
        db.deleteClassifierResults(classifier, 'train')
    
    trainer = classifiers[classifier]['train']
    threshold = 5.0 # in percent
    num_obs = 1000
    max_obs = int(db.populationSize('train')/5)
    result = None
    while num_obs < max_obs:
        num_obs = min(int(num_obs*1.5), max_obs)
        result = db.fetchClassifierResult(classifier, 'train', num_obs)
        if result != None:
            train_time = "(stored)"
        else:
            start = time.time()
            result = trainer(db,unusual_case,greater,num_obs)
            result['classifier'] = classifier
            train_time = "%f" % (time.time()-start)
            
        error = statistics.mean([result['false_positives'],result['false_negatives']])
        print("number of observations: %d | error: %f | false_positives: %f | false_negatives: %f | train time: %s | params: %s"
              % (num_obs, error, result['false_positives'],result['false_negatives'], train_time, result['params']))
        db.addClassifierResults(result)
        classifiers[classifier]['train_results'].append(result)

        if error < threshold:
            break

    return result



def testClassifier(db, unusual_case, greater, classifier, retest=False):
    target_error = 5.0 # in percent
    num_trials = 1000
    max_obs = int(db.populationSize('test')/5)

    tester = classifiers[classifier]['test']
    
    def testAux(params, num_trials, num_observations):
        estimator = functools.partial(tester, params, greater)
        estimates = bootstrap3(estimator, db, 'test', unusual_case, num_observations, num_trials)
        null_estimates = bootstrap3(estimator, db, 'train_null', unusual_case, num_observations, num_trials)

        bad_estimates = len([e for e in estimates if e != 1])
        bad_null_estimates = len([e for e in null_estimates if e != 0])
        
        false_negatives = 100.0*bad_estimates/num_trials
        false_positives = 100.0*bad_null_estimates/num_trials
        print("testAux:", num_observations, false_positives, false_negatives, params)
        return false_positives,false_negatives


    if retest:
        print("Dropping stored test results...")
        db.deleteClassifierResults(classifier, 'test')


    test_results = []
    lte = math.log(target_error/100.0)
    for tr in classifiers[classifier]['train_results']:
        db.resetOffsets()
        params = json.loads(tr['params'])
        num_obs = tr['num_observations']
    
        print("initial test")
        fp,fn = testAux(params, num_trials, num_obs)
        error = (fp+fn)/2.0
        print("walking up")
        while (error > target_error) and (num_obs < max_obs):
            increase_factor = 1.5 * lte/math.log(error/100.0) # don't ask how I came up with this
            #print("increase_factor:", increase_factor)
            num_obs = min(int(increase_factor*num_obs), max_obs)
            fp,fn = testAux(params, num_trials, num_obs)
            error = (fp+fn)/2.0

        print("walking down")
        while (num_obs > 0):
            current_best = (num_obs,error,params,fp,fn)
            num_obs = int(0.95*num_obs)
            fp,fn = testAux(params, num_trials, num_obs)
            error = (fp+fn)/2.0
            if error > target_error:
                break
        
        test_results.append(current_best)

    test_results.sort()
    best_obs,error,best_params,fp,fn = test_results[0]
    
    return {'classifier':classifier,
            'trial_type':"test",
            'num_observations':best_obs,
            'num_trials':num_trials,
            'params':best_params,
            'false_positives':fp,
            'false_negatives':fn}


start = time.time()
unusual_case,unusual_diff = findUnusualTestCase(db)
greater = (unusual_diff > 0)
print("unusual_case:", unusual_case)
print("unusual_diff:", unusual_diff)
end = time.time()
print(":", end-start)

for c in sorted(classifiers.keys()):
    if classifiers[c]['train'] == None:
        continue
    start = time.time()
    print("Training %s..." % c)
    result = trainClassifier(db, unusual_case, greater, c, c in options.retrain)
    print("%s result:" % c)
    pprint.pprint(result)
    print("completed in:", time.time()-start)

db.clearCache()

for c in sorted(classifiers.keys()):
    start = time.time()
    print("Testing %s..." % c)
    result = testClassifier(db, unusual_case, greater, c, c in options.retest)
    print("%s result:" % c)
    pprint.pprint(result)
    classifiers[c]['test_error'] = (result['false_positives']+result['false_negatives'])/2.0
    print("completed in:", time.time()-start)