def test_drift_from_original(path):
"""Tests to ensure the behavior of NDChild has not changed.
The directory tests/run_data contains a set of logs generated by NDChild
runs on random languages, using randomly generated learning rates. Each
file represents snapshots from a single child on a single language. The
language, and learning rates are contained in the filename.
Each of these files is a gzipped jsonl file. Each line in the file is a
snapshot of the learner's grammar after consuming a sentence.
"""
lang, rate, cons = re.search(r'lang(\d+):rate(\d+.\d+):cons(\d+.\d+)',
path).groups()
rate, cons = float(rate), float(cons)
child = NDChild(rate, cons, lang)
cached_child = CachedChild(rate, cons, lang)
with gzip.open(path) as fh:
for line in fh:
data = json.loads(line)
s = DOMAIN.sentences[data['sentence_id']]
expected = data['grammar']
child.consumeSentence(s)
cached_child.consumeSentence(s)
assert child.grammar == expected
assert child.grammar == cached_child.grammar
def run_trial(params: TrialParameters):
""" Runs a single echild simulation and reports the results """
logging.debug('running echild with %s', params)
language = params.language
noise_level = params.noise
child = NDChild(params.rate, params.conservativerate, language)
then = datetime.now()
for i in range(params.numberofsentences):
if random() < noise_level:
s = DOMAIN.get_sentence_not_in_language(grammar_id=language)
else:
s = DOMAIN.get_sentence_in_language(grammar_id=language)
child.consumeSentence(s)
now = datetime.now()
result = NDResult(trial_params=params,
timestamp=now,
duration=now - then,
language=child.target_language,
grammar=child.grammar)
logging.debug('experiment result: %s', result)
return result
def run_traced_trial(params: TrialParameters, output_directory):
""" Runs a single echild simulation and reports the results """
logging.debug('running echild with %s', params)
language = params.language
noise_level = params.noise
child = NDChild(params.rate, params.conservativerate, language)
history = []
then = datetime.now()
sample_period = params.numberofsentences / 1000
for i in range(params.numberofsentences):
if random() < noise_level:
s = DOMAIN.get_sentence_not_in_language(grammar_id=language)
else:
s = DOMAIN.get_sentence_in_language(grammar_id=language)
child.consumeSentence(s)
if i % sample_period == 0:
history.append(dict(child.grammar))
now = datetime.now()
result = NDResult(trial_params=params,
timestamp=now,
duration=now - then,
language=child.target_language,
grammar=child.grammar)
logging.debug('experiment result: %s', result)
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
plt.rcParams['figure.figsize'] = 15, 10
df = pd.DataFrame(history)
df = df + (np.arange(len(df.columns)) / 500)
df.index = df.index * 1000
ax = df.plot(title='lang={}, noise={}'.format(params.language,
params.noise),
sharey=True,
subplots=True)
# ax.set_xscale('log')
plt.savefig(
os.path.join(output_directory,
'{}-{}.png'.format(language, noise_level)))
return result
def childLearnsLanguage(ndr, languageDomain, language):
ndr.resetThresholdDict()
aChild = NDChild(rate, conservativerate, language)
for j in xrange(numberofsentences):
s = pickASentence(languageDomain)
aChild.consumeSentence(s)
# If a parameter value <= to the threshold for the first time,
# this is recorded in ndr for writing output
ndr.checkIfParametersMeetThreshold(threshold, aChild.grammar, j)
return [aChild.grammar, ndr.thresholdDict]
def test_ndchild_parity(language):
"""Compares the behavior of NDChild and CachedChild to ensure identical
behavior
"""
child = NDChild(0.9, 0.005, 2253)
cached_child = CachedChild(0.9, 0.005, 2253)
num_sents = int(50000)
for i in range(num_sents):
s = DOMAIN.get_sentence_in_language(grammar_id=language)
child.pre = dict(child.grammar)
cached_child.pre = dict(child.grammar)
child.consumeSentence(s)
cached_child.consumeSentence(s)
assert child.grammar == cached_child.grammar
def run_simulations(params: ExperimentParameters):
"""Runs echild simulations according to `params`.
Returns a generator that yields one result dictionary (as returned by
run_trial) for each simulation run.
"""
trials = (TrialParameters(
language=lang,
noise=noise,
rate=params.learningrate,
numberofsentences=params.num_sentences,
conservativerate=params.conservative_learningrate)
for lang in params.languages for noise in params.noise_levels
for _ in range(params.num_echildren))
# for reporting progress during the run
num_trials = (params.num_echildren * len(params.languages) *
len(params.noise_levels))
DOMAIN.init_from_flatfile()
# compute all "static" triggers once for each sentence in the domain and
# store the cached value.
NDChild.precompute_domain(DOMAIN)
with multiprocessing.Pool(params.num_procs) as p:
# run trials across processors (this doesn't actually start them
# running- `results` is a generator. the actual computation is deferred
# until somebody iterates through the generator object.
if params.trace:
results = p.imap_unordered(run_traced_trial, trials)
else:
results = p.imap_unordered(run_trial, trials)
# report output
results = progress_bar(results,
total=num_trials,
desc="running simulations")
yield from results
def childLearnsLanguage(ndr, languageDomain, language, numberofsentences):
ndr.resetThresholdDict()
global distribution
sets = [[[], []], [[], []], [[], []], [[], []], [[], []], [[], []], [[],
[]],
[[], []], [[], []], [[], []], [[], []], [[], []], [[], []]]
#print(sets[1][0])
aChild = NDChild(rate, conservativerate, language, sets)
#print numberofsentences
#probs = []
#for l in languageDomain:
#probs.append(1 / len(languageDomain))
for j in xrange(numberofsentences):
s = pickASentence(languageDomain) #,probs)
distribution.append(s.sentenceList)
#print(distribution)
aChild.consumeSentence(s)
# If a parameter value <= to the threshold for the first time,
# this is recorded in ndr for writing output
ndr.checkIfParametersMeetThreshold(threshold, aChild.grammar, j)
#print(sets)
#return [aChild.grammar, ndr.thresholdDict]
return [[aChild.grammar, ndr.thresholdDict], distribution, sets]