from plasma.primitives.hyperparameters import (

# CategoricalHyperparam, ContinuousHyperparam,

# LogContinuousHyperparam,

HyperparamExperiment

)

import matplotlib.pylab as plt

# from pprint import pprint

# import yaml, datetime, uuid

import sys

import os

import getpass

# import shutil, pandas

# import subprocess as sp

import numpy as np

import plasma.conf

dir_path = "/{}/{}/hyperparams/".format(

plasma.conf.conf['fs_path'], getpass.getuser())

if len(sys.argv) <= 1:

dir_path = dir_path + os.listdir(dir_path)[0] + '/'

print("using default dir {}".format(dir_path))

else:

dir_path = sys.argv[1]

def get_experiments(path, verbose=0):

experiments = []

num_tot = 0

num_finished = 0

num_success = 0

for name in sorted(os.listdir(path)):

if os.path.isdir(os.path.join(path, name)):

print(os.path.join(path, name))

exp = HyperparamExperiment(os.path.join(path, name))

num_finished += 1 if exp.finished else 0

num_success += 1 if exp.success else 0

num_tot += 1

experiments.append(exp)

if verbose:

print("Read {} experiments, {} finished ({} success)".format(

num_tot, num_finished, num_success))

return experiments

experiments = sorted(get_experiments(dir_path))

best_experiments = np.argsort(

np.array([e.get_maximum(False)[0] for e in experiments]))

best = []

for e in np.array(experiments)[best_experiments][-5:]:

best.append(e.get_number())

bigdict = {}

for base in best:

f = "/{}/{}/changed_params.out".format(dir_path, base)

data = open(f).readlines()

for line in data:

tuples = line.split(":")

# if len(tuples) == 2:

key, values = tuples[-2:]

key = key.strip()

try:

bigdict[key] += [values]

except KeyError:

bigdict[key] = [values]

def make_comparison_plot(key, tunable, trial):

values, edges = tunable

trial = list(map(lambda x: eval(x), trial))

trial_values, _ = np.histogram(trial, bins=edges)

total = trial_values.sum()

values_percentages = list(map(lambda x: x*100.0/total, trial_values))

plt.bar(edges[:-1], values_percentages,

width=np.diff(edges), ec="k", align="edge")

plt.xlabel(key, fontsize=20)

# plt.yscale('log')

plt.ylabel('Fraction of trials [%]/bin', fontsize=20)

plt.bar(edges[:-1], values, width=np.diff(edges), ec="k", align="edge")

plt.savefig(key+".png")

# plt.show()

plt.clf()

# default tunables:

defaults = {}

defaults['lr'] = np.histogram([1e-7, 1e-4])

defaults['lr_decay'] = np.histogram([0.97, 0.985, 1.0])

defaults['positive_example_penalty'] = np.histogram([1.0, 4.0, 16.0])

# defaults['target'] = np.histogram([50,50,50],bins=['hinge','ttdinv','ttd'])

defaults['batch_size'] = np.histogram([64, 256, 1024])

defaults['dropout_prob'] = np.histogram([0.1, 0.3, 0.5])

defaults['rnn_layers'] = np.histogram([1, 4])

defaults['rnn_size'] = np.histogram([100, 200, 300])

defaults['num_conv_filters'] = np.histogram([5, 10])

defaults['num_conv_layers'] = np.histogram([2, 4])

defaults['T_warning'] = np.histogram([0.256, 1.024, 10.024])

defaults['cut_shot_ends'] = np.histogram([False, True])

# Histogram it

for key, trial in bigdict.items():

if key == 'target':

continue

make_comparison_plot(key, defaults[key], trial)