def ReadNNetResultsFile10k(rootpath):
file = rootpath + "/nnet.output.csv"
ga_noprogressfile = rootpath + "/output_test_lr_10k-iters_30-hiddenlayers_crowding.csv"
data = pd.read_csv(file)
ga_noprogress = pd.read_csv(ga_noprogressfile)
return data, ga_noprogress
ga_noprogress = u.FilterRows(data, lambda x: x['algo'] == 'ga')
data = u.FilterRows(data, lambda x: x['algo'] != 'ga')
ga_progress = pd.read_csv(
rootpath + "/output_test_lr_10k-iters_30-hiddenlayers_ga.csv")
data = pd.concat([data, ga_progress], ignore_index=True)
return data, ga_noprogress
def PlotAdaboostPerIterationCurves(file_template,
filter,
plot_output_file,
iters,
y_axis_name='F-Measure'):
ts = [20, 30, 40, 50, 60, 70, 80, 90, 100]
colors = u.GetColorCombinations()
y = []
for _ts in ts:
data = pd.read_csv(file_template.format(str(_ts)))
data = u.FilterRows(data, filter)
data = data.set_index('iter')
train_data = FilterRows(data, lambda x: x['istrain'] == 1)
test_data = FilterRows(data, lambda x: x['istrain'] == 0)
train_y = []
test_y = []
for iter in iters:
train_y.append(train_data.loc[iter]['m'])
test_y.append(test_data.loc[iter]['m'])
c = colors.pop()
y.append(
u.YSeries(train_y,
points_marker='o',
line_color=c['color'],
plot_legend_label=str(_ts) + "-train"))
y.append(
u.YSeries(test_y,
points_marker='x',
line_color=c['color'],
plot_legend_label=str(_ts) + "-validation"))
u.SaveDataPlotWithLegends(y,
iters,
plot_output_file,
x_axis_name="num of iters/weak learners",
y1_axis_name=y_axis_name)
def ComputeFinalResults(rootfolder):
clustering_stats = pd.read_csv(rootfolder+"/clustering.csv")
final_output_file = rootfolder+"/best_metrics.csv"
data = u.FilterRows(clustering_stats, lambda x : x['p'] == 2)
dim_red = ["ica","pca","rp","mi"]
clustering = ["kmeans","gmm"]
lines = []
lines.append("clustering,dim_red,k,p,ami_raw,ami_true,sc,bic")
raw_predictions = {}
for c in clustering:
d = data.loc[(data['dim_red_method'] == "raw") & (data['clustering'] == c),:]
d = d.loc[d['bic'] == np.min(d['bic']),:]
clusters_file = rootfolder + "/clustering_output/mthd={0}_k={1}_d=2_algo=raw.csv".format(c,d.iloc[0]['k'])
raw_predictions[c] = np.loadtxt(clusters_file,delimiter=',')
for dr in dim_red:
for c in clustering:
d = data.loc[(data['dim_red_method'] == dr) & (data['clustering'] == c),:]
d = d.loc[d['bic'] == np.min(d['bic']),:]
clusters_file = rootfolder + "/clustering_output/mthd={0}_k={1}_d=2_algo={2}.csv".format(c,d.iloc[0]['k'],dr)
predicted = np.loadtxt(clusters_file,delimiter=',')
ami = metrics.adjusted_mutual_info_score(raw_predictions[c],predicted)
lines.append(u.ConcatToStr(",",[c,dr,d.iloc[0]['k'],ami,d.iloc[0]['ami_true'],d.iloc[0]['sc'],d.iloc[0]['bic']]))
u.WriteTextArrayToFile(final_output_file,lines)
def PlotAvgRewardsPerEpisode(data,
totalpoints,
points_to_sample,
outputfile,
y_axis_name,
key_to_plot,
max_points=100000):
"""
cr : cum_rewards
ar : avg_rewards
len : episode_len
goal : reached_goal
"""
data_to_plot = u.FilterRows(
data, lambda x: (x['solver'] == 'q') & (x['gamma'] == 0.99) &
(x['alpha'] == 1) & (x['maxInnerVi'] == totalpoints))
x_to_take = np.arange(totalpoints)
x_to_take = x_to_take[x_to_take % points_to_sample == 0]
x_to_take = x_to_take[x_to_take < max_points]
ser = data_to_plot.apply(
lambda x: GetQRewardSeriesToPlot(x, x_to_take, key_to_plot), axis=1)
u.SaveDataPlotWithLegends(ser.values,
filename=outputfile,
x_axis_name="episodes",
y1_axis_name=y_axis_name)
def PlotClusteringMetrics(rootfolder, data, k,dim='raw',p=2):
filter = lambda x : x['dim_red_method'] == dim and x['p'] == p
filtered_data = u.FilterRows(data,filter)
metrics = ["ami_raw","ami_true","sc","bic"]
gmm_data = filtered_data.loc[filtered_data['clustering'] == "gmm",:]
kmeans_data = filtered_data.loc[filtered_data['clustering'] == "kmeans",:]
d = {"kmeans":('o','b','kmeans'),"gmm":('x','r','gmm')}
for metric in metrics:
outputfile = u.PreparePath(rootfolder + "/plots/metrics/{0}_{1}_p={2}.png".format(metric,dim,str(p)))
kmeans_ser = u.YSeries(kmeans_data[metric],xvalues=kmeans_data["k"],points_marker = d["kmeans"][0],line_color=d["kmeans"][1],plot_legend_label=d["kmeans"][2])
gmm_ser = u.YSeries(gmm_data[metric],xvalues=gmm_data["k"],points_marker = d["gmm"][0],line_color=d["gmm"][1],plot_legend_label=d["gmm"][2])
u.SaveDataPlotWithLegends([kmeans_ser,gmm_ser],x_axis_name="number of clusters",y1_axis_name=metric,filename=outputfile)
def GetBestRawClustering(rootfolder):
clustering_stats = pd.read_csv(rootfolder+"/clustering.csv")
final_output_file = rootfolder+"/best_metrics.csv"
data = u.FilterRows(clustering_stats, lambda x : x['p'] == 2)
dim_red = ["ica","pca","rp","mi"]
clustering = ["kmeans","gmm"]
lines = []
lines.append("clustering,dim_red,k,p,ami_raw,ami_true,sc,bic")
raw_predictions = {}
for c in clustering:
d = data.loc[(data['dim_red_method'] == "raw") & (data['clustering'] == c),:]
d = d.loc[d['bic'] == np.min(d['bic']),:]
clusters_file = rootfolder + "/clustering_output/mthd={0}_k={1}_d=2_algo=raw.csv".format(c,d.iloc[0]['k'])
raw_predictions[c] = (np.loadtxt(clusters_file,delimiter=','),d.copy())
return raw_predictions
def PlotClusteringMetricsForDimsAndDimRed(rootfolder, data, dims, dim_reds,k):
colors = {"ica":'r','pca':'b','rp':'g','mi':'k','raw':'orange'}
markers = {"kmeans":'o',"gmm":'x'}
metrics = ["ami_raw","ami_true","sc","bic"]
for _k in k:
filter = lambda x : x['k'] == _k
filtered_data = u.FilterRows(data,filter)
for metric in metrics:
ser = []
outputfile = u.PreparePath(rootfolder + "/plots/metrics/dr_{0}_k={1}.png".format(metric,str(_k)))
for dim_red in dim_reds:
d = data.loc[(data['dim_red_method'] == dim_red) & (data['k'] == _k) & (data['clustering'] == 'kmeans') ,:]
ser.append(u.YSeries(d[metric],xvalues=d['p'],line_color=colors[dim_red],points_marker=markers['kmeans'],plot_legend_label="{0}-{1}".format(dim_red,'kmeans')))
d = data.loc[(data['dim_red_method'] == dim_red) & (data['k'] == _k) & (data['clustering'] == 'gmm') ,:]
ser.append(u.YSeries(d[metric],xvalues=d['p'],line_color=colors[dim_red],points_marker=markers['gmm'],plot_legend_label="{0}-{1}".format(dim_red,'gmm')))
u.SaveDataPlotWithLegends(ser,x_axis_name="dimensions",y1_axis_name=metric,filename = outputfile)
def ComputeFinalResults1(rootfolder,clusters,dims):
raw_results = GetBestRawClustering(rootfolder)
clustering_results = pd.read_csv(rootfolder+"/clustering.csv")
final_output_file = rootfolder+"/best_metrics.csv"
best_raw_clustering_output = rootfolder+"/best_raw_clustering.csv"
o = pd.concat([raw_results["kmeans"][1], raw_results["gmm"][1]])
o.to_csv(best_raw_clustering_output)
dim_red = ["mi","pca","ica","rp"]
lines = []
lines.append("clustering,dim_red,k,p,ami_raw,ami_true,sc,bic")
raw_predictions = {}
output = None
for dr in dim_red:
data = u.FilterRows(clustering_results,lambda x : x["dim_red_method"] == dr)
dim_best_val = None
dim_result = None
for dim in dims:
best = {} # {p,k,ami}
for cluster_mthd in ["kmeans","gmm"]:
for cluster in clusters:
print("{0},{1},{2},{3}".format(dr,str(dim),cluster_mthd,str(cluster)))
d = data.loc[(data['clustering'] == cluster_mthd)&(data['k'] == cluster) & (data['p'] == dim)]
row = d.head(1).copy()
if(cluster_mthd not in best or best[cluster_mthd]['bic'].iloc[0] > row['bic'].iloc[0]):
best[cluster_mthd] = row
curr_val = (best["kmeans"]['ami_true'].iloc[0] + best["gmm"]['ami_true'].iloc[0]) / 2
#curr_val = (best["kmeans"]['ami_true'].iloc[0] + best["gmm"]['ami_true'].iloc[0]) / 2
#curr_val = np.minimum(best["kmeans"]['ami_true'].iloc[0], best["gmm"]['ami_true'].iloc[0])
if(dim_best_val is None or dim_best_val < curr_val):
dim_best_val = curr_val
dim_result = best.copy()
for c in ["kmeans","gmm"]:
ami_raw = GetAmiWithRawPredictions(rootfolder,raw_results,dr,dim_result[c].iloc[0]["p"],dim_result[c].iloc[0]["k"],c)
lines.append("{0},{1},{2},{3},{4},{5},{6},{7}".format(c,str(dim_result[c].iloc[0]["dim_red_method"]),str(dim_result[c].iloc[0]["k"]),str(dim_result[c].iloc[0]["p"]),str(ami_raw[c]),str(dim_result[c].iloc[0]["ami_true"]),str(dim_result[c].iloc[0]["sc"]),str(dim_result[c].iloc[0]["bic"])))
#if(output is None):
# output = pd.concat([dim_result["kmeans"],dim_result["gmm"]])
#else:
# output = pd.concat([output,dim_result["kmeans"],dim_result["gmm"]])
u.WriteTextArrayToFile(final_output_file,lines)
def PlotPerIterationCurves(rootFolder, outputfolder):
mimic = pd.read_csv(rootFolder + "/mimic.csv")
sa = pd.read_csv(rootFolder + "/sa.csv")
rhc = pd.read_csv(rootFolder + "/rhc.csv")
ga = pd.read_csv(rootFolder + "/ga.csv")
sizes = np.array(mimic['size'].unique())
algo_decoration = {
'mimic': ('r', 'o', 'mimic', mimic),
'ga': ('g', 's', 'genetic algo', ga),
'sa': ('b', '+', 'sim annealing', sa),
'rhc': ('k', '*', 'rhc', rhc)
}
def f(data, name):
x = data['iters']
y = data['fn_value']
deco = algo_decoration[name]
return u.YSeries(y,
xvalues=x,
points_marker='.',
plot_legend_label=deco[2],
legend_marker='o',
line_color=deco[0])
for size in sizes:
size_root = u.PreparePath(outputfolder + "/itercurves_" + str(size) +
".png")
y_ser = []
for key in algo_decoration.keys():
d = u.FilterRows(algo_decoration[key][3],
lambda x: x['size'] == size).head(10000)
y_ser.append(f(d, key))
u.SaveDataPlotWithLegends(y_ser,
x_axis_name="iters",
x=None,
y1_axis_name="fn value",
filename=size_root)
def PlotTempVariationCurvesForSa(rootfolder, algoname, temperatures):
"""
"""
rhcdata = u.FilterRows(
pd.read_csv(rootfolder + "/" + algoname + "/stats_agg.csv"),
lambda x: x['algo'] == 'rhc')
y_Ser = []
y_Ser.append(
u.YSeries(rhcdata['converged_iters'],
xvalues=rhcdata['size'],
points_marker="*",
line_color="k",
plot_legend_label="rhc"))
data_dict = {}
deco = {
'0': ("r", "x"),
'90': ("b", "o"),
'95': ("g", "+"),
'99': ("orange", ">")
}
for t in temperatures:
path = rootfolder + "/" + algoname + "_" + t
CompteStats(path, ["sa.csv"])
data_dict[t] = pd.read_csv(path + "/stats_agg.csv")
y_Ser.append(
u.YSeries(data_dict[t]['converged_iters'],
xvalues=data_dict[t]['size'],
points_marker=deco[t][1],
line_color=deco[t][0],
plot_legend_label="sa_" + t))
outputfile = rootfolder + "/" + algoname + "/plots/sa_temperatures.png"
u.SaveDataPlotWithLegends(y_Ser,
y1_axis_name="iterations to converge",
x_axis_name="size",
filename=outputfile)
def RunDecisionTreesWithOptimalInst(datasets_root_folder,
weka_jar_path,
cv_results_file,
use_arff_files=True):
file_extn = "arff" if use_arff_files else ".csv"
testfiles = glob.glob("{0}/*.test.{1}".format(datasets_root_folder,
file_extn))
cv_results = pd.read_csv(datasets_root_folder + "/" + cv_results_file)
for dataset_dir in u.Get_Subdirectories(datasets_root_folder):
trainfile = glob.glob("{0}/*.train.{1}".format(dataset_dir,
file_extn))[0]
paramfile = glob.glob("{0}/*.params.txt".format(dataset_dir))[0]
dt_root = u.PreparePath(dataset_dir + "/dt", is_file=False)
filter_name, filter_val = GetFilterOptions(dataset_dir)
config_gen = ParameterGrid({'prune': [True, False]})
for config in config_gen:
filter = lambda x: (x['prune'] == False) & (x[
filter_name] == filter_val) & (x[
'istrain'] == 1) # this will output on the held out set
filtered_rows = u.FilterRows(cv_results, filter)
a = filtered_rows['m']
if (len(a) == 0):
print("ignoring : {0}".format(dataset_dir))
continue
b = np.max(filtered_rows['m'])
indxs = np.isclose(a, b)
best_insts = filtered_rows[indxs]
best_insts = best_insts.iloc[0]['inst']
config['inst'] = best_insts
id = GetIdForOptConfig(config)
params_info = u.ReadLinesFromFile(paramfile)
params_info_dict = sl.GetDictionary(params_info)
run_output_dir = u.PreparePath("{0}/{1}".format(dt_root, id),
is_file=False)
params_output_file = u.PreparePath("{0}/{1}.params.txt".format(
run_output_dir, id))
model_output_file = u.PreparePath("{0}/{1}.model".format(
run_output_dir, id))
train_output_file = u.PreparePath(
"{0}/{1}.train.predictions.csv".format(run_output_dir, id))
test_output_file = u.PreparePath(
"{0}/{1}.test.predictions.csv".format(run_output_dir, id))
# if(os.path.isfile(train_output_file)):
# continue
config['random_state'] = params_info_dict['random_state']
config["wekajar"] = weka_jar_path
config["trainset"] = trainfile
config["class"] = "last"
config["trainpredictionoutputfile"] = train_output_file
config["predictionoutputfile"] = config[
"trainpredictionoutputfile"]
config["modeloutputfile"] = model_output_file
config["testpredictionoutputfile"] = test_output_file
# for every config there has to be a train prediction and test prediction
cmd = GetWekaCommandLineForConfig(config, False, False)
config["modelbuildtimesecs"] = timeit.timeit(
lambda: sl.RunCmdWithoutConsoleWindow(cmd), number=1)
# now for test set
config["predictionoutputfile"] = test_output_file
config["testset"] = testfiles[0]
cmd = GetWekaCommandLineForConfig(config, True, False)
config["modelevaltimesecs"] = timeit.timeit(
lambda: sl.RunCmdWithoutConsoleWindow(cmd), number=1)
config.pop('random_state',
None) # since we already have that in params_info
for k in config:
params_info.append("{0}={1}".format(k, config[k]))
u.WriteTextArrayToFile(params_output_file, params_info)
print("done dataset : " + dataset_dir)
def PlotPiViConvergenceForSmallAndLargeMdp(outputfolder, datafile, gamma):
data = pd.read_csv(datafile)
decorations = {1: 'g', 10: 'k', 10000: 'r'}
pi_sweeps = [1, 10, 10000]
ser = []
ser1 = []
vi_added = False
for sweep in pi_sweeps:
data_vi = u.FilterRows(
data, lambda x: (x['mdp'] == 'LargeMdpRwTraps50') &
(x['solver'] == 'vi') & (x['gamma'] == gamma))
data_pi = u.FilterRows(
data, lambda x:
(x['mdp'] == 'LargeMdpRwTraps50') & (x['solver'] == 'pi') &
(x['gamma'] == gamma) & (x['maxSweepsPerIteration'] == sweep))
assert (len(data_vi) == 1)
assert (len(data_pi) == 1)
data_vi_qchange = np.array(
[float(s) for s in data_vi.iloc[0]['cum_rewards'].split(';')])
data_vi_value = np.array([
float(s) for s in data_vi.iloc[0]['ran_to_completion'].split(';')
])
data_pi_qchange = np.array(
[float(s) for s in data_pi.iloc[0]['cum_rewards'].split(';')])
data_pi_value = np.array([
float(s) for s in data_pi.iloc[0]['ran_to_completion'].split(';')
])
if (vi_added == False):
s_vi = u.YSeries(data_vi_qchange,
xvalues=np.arange(len(data_vi_qchange)) + 1,
line_color='b',
plot_legend_label='VI')
ser.append(s_vi)
s_pi = u.YSeries(data_pi_qchange,
xvalues=np.arange(len(data_pi_qchange)) + 1,
line_color=decorations[sweep],
plot_legend_label='PI_' + str(sweep))
ser.append(s_pi)
if (vi_added == False):
s_vi = u.YSeries(data_vi_value,
xvalues=np.arange(len(data_vi_value)) + 1,
line_color='b',
plot_legend_label='VI')
ser1.append(s_vi)
s_pi = u.YSeries(data_pi_value,
xvalues=np.arange(len(data_pi_value)) + 1,
line_color=decorations[sweep],
plot_legend_label='PI_' + str(sweep))
ser1.append(s_pi)
vi_added = True
outputfile = u.PreparePath(outputfolder + "/plots/large_qchange_gamma=" +
str(gamma) + ".png")
u.SaveDataPlotWithLegends(ser,
filename=outputfile,
x_axis_name="iterations",
y1_axis_name="Max change in state value")
outputfile = u.PreparePath(outputfolder + "/plots/large_value_gamma=" +
str(gamma) + ".png")
u.SaveDataPlotWithLegends(ser1,
filename=outputfile,
x_axis_name="iterations",
y1_axis_name="Total value accross states")
ser = []
ser1 = []
vi_added = False
for sweep in pi_sweeps:
data_vi = u.FilterRows(
data, lambda x: (x['mdp'] == 'SmallMdpRwTraps') &
(x['solver'] == 'vi') & (x['gamma'] == gamma))
data_pi = u.FilterRows(
data, lambda x:
(x['mdp'] == 'SmallMdpRwTraps') & (x['solver'] == 'pi') &
(x['gamma'] == gamma) & (x['maxSweepsPerIteration'] == sweep))
assert (len(data_vi) == 1)
assert (len(data_pi) == 1)
data_vi_qchange = np.array(
[float(s) for s in data_vi.iloc[0]['cum_rewards'].split(';')])
data_vi_value = np.array([
float(s) for s in data_vi.iloc[0]['ran_to_completion'].split(';')
])
data_pi_qchange = np.array(
[float(s) for s in data_pi.iloc[0]['cum_rewards'].split(';')])
data_pi_value = np.array([
float(s) for s in data_pi.iloc[0]['ran_to_completion'].split(';')
])
if (vi_added == False):
s_vi = u.YSeries(data_vi_qchange,
xvalues=np.arange(len(data_vi_qchange)) + 1,
line_color='b',
plot_legend_label='VI')
ser.append(s_vi)
s_pi = u.YSeries(data_pi_qchange,
xvalues=np.arange(len(data_pi_qchange)) + 1,
line_color=decorations[sweep],
plot_legend_label='PI_' + str(sweep))
ser.append(s_pi)
if (vi_added == False):
s_vi = u.YSeries(data_vi_value,
xvalues=np.arange(len(data_vi_value)) + 1,
line_color='b',
plot_legend_label='VI')
ser1.append(s_vi)
s_pi = u.YSeries(data_pi_value,
xvalues=np.arange(len(data_pi_value)) + 1,
line_color=decorations[sweep],
plot_legend_label='PI_' + str(sweep))
ser1.append(s_pi)
vi_added = True
outputfile = u.PreparePath(outputfolder + "/plots/small_qchange_gamma=" +
str(gamma) + ".png")
u.SaveDataPlotWithLegends(ser,
filename=outputfile,
x_axis_name="iterations",
y1_axis_name="Max change in state value")
outputfile = u.PreparePath(outputfolder + "/plots/small_value_gamma=" +
str(gamma) + ".png")
u.SaveDataPlotWithLegends(ser1,
filename=outputfile,
x_axis_name="iterations",
y1_axis_name="Total value accross states")