def learn_unigram(data, verbose=True):
"""Learns a unigram model from data.train.
It also evaluates the model on data.dev and data.test, along with generating
some sample sentences from the model.
"""
from lm import Unigram
unigram = Unigram()
unigram.fit_corpus(data.train)
if verbose:
print("vocab:", len(unigram.vocab()))
# evaluate on train, test, and dev
print("train:", unigram.perplexity(data.train))
print("dev :", unigram.perplexity(data.dev))
print("test :", unigram.perplexity(data.test))
from generator import Sampler
sampler = Sampler(unigram)
print("sample 1: ", " ".join(str(x) for x in sampler.sample_sentence(['The', 'president'])))
print("sample 2: ", " ".join(str(x) for x in sampler.sample_sentence(['This', 'university'])))
print("sample 3: ", " ".join(str(x) for x in sampler.sample_sentence(['The', 'United', 'States'])))
print("sample 4: ", " ".join(str(x) for x in sampler.sample_sentence(['An', 'explosion'])))
print("sample 5: ", " ".join(str(x) for x in sampler.sample_sentence(['To', 'be', 'or', 'to'])))
print("sample 6: ", " ".join(str(x) for x in sampler.sample_sentence(['This', 'is', 'awesome'])))
print("sample 7: ", " ".join(str(x) for x in sampler.sample_sentence(['I', 'am', 'sorry'])))
print("sample 8: ", " ".join(str(x) for x in sampler.sample_sentence(['Today', 'the', 'chair', 'of'])))
print("sample 9: ", " ".join(str(x) for x in sampler.sample_sentence(['Hello', 'I', 'came', 'from'])))
print("sample 10: ", " ".join(str(x) for x in sampler.sample_sentence(['I', 'major', 'in', 'Computer', 'Science'])))
return unigram
def learn_unigram(data):
"""Learns a unigram model from data.train.
It also evaluates the model on data.dev and data.test, along with generating
some sample sentences from the model.
"""
from lm import Unigram
unigram = Unigram()
unigram.fit_corpus(data.train)
print("vocab:", len(unigram.vocab()))
# evaluate on train, test, and dev
# print("train:", unigram.perplexity(data.train))
# print("dev :", unigram.perplexity(data.dev))
# print("test :", unigram.perplexity(data.test))
# from generator import Sampler
# sampler = Sampler(unigram)
# print("sample: ", " ".join(str(x) for x in sampler.sample_sentence([])))
# print("sample: ", " ".join(str(x) for x in sampler.sample_sentence([])))
# print("sample: ", " ".join(str(x) for x in sampler.sample_sentence([])))
return unigram
def learn_unigram(data, verbose=True):
"""Learns a unigram model from data.train.
It also evaluates the model on data.dev and data.test, along with generating
some sample sentences from the model.
"""
from lm import Unigram
unigram = Unigram()
unigram.fit_corpus(data.train)
if verbose:
print("vocab:", len(unigram.vocab()))
# evaluate on train, test, and dev
print("train:", unigram.perplexity(data.train))
print("dev :", unigram.perplexity(data.dev))
print("test :", unigram.perplexity(data.test))
return unigram
incl_eos determines whether the space of words should include EOS or not.
"""
wps = []
tot = -np.inf # this is the log (total mass)
for w in self.lm.vocab():
if not incl_eos and w == "END_OF_SENTENCE":
continue
lp = self.lm.cond_logprob(w, prev, 0)
wps.append([w, lp / self.temp])
tot = np.logaddexp2(lp / self.temp, tot)
p = random.random()
word = random.choice(wps)[0]
s = -np.inf # running mass
for w, lp in wps:
s = np.logaddexp2(s, lp)
if p < pow(2, s - tot):
word = w
break
return word
if __name__ == "__main__":
from lm import Unigram
unigram = Unigram()
corpus = [["sam", "i", "am"]]
unigram.fit_corpus(corpus)
print(unigram.model)
sampler = Sampler(unigram)
for i in range(10):
print(i, ":", " ".join(str(x) for x in sampler.sample_sentence([])))
from data import run_model
from lm import Unigram, NgramNoUnk, NgramUnk
import sys
def check_params(params):
if (
(len(params) <= 1)
or (params[1] not in ('unigram', 'ngram'))
or (params[1] == 'ngram' and len(params) < 4)
):
print(f'Usage: {params[0]} <unigram|ngram> [n λ [voc_ratio]]')
sys.exit(-1)
if __name__ == '__main__':
params = sys.argv
check_params(params)
model = params[1]
if model == 'unigram':
run_model(lambda: Unigram(), 'results/unigram')
else:
n = int(params[2])
λ = float(params[3])
if len(params) == 5:
voc_ratio = float(params[4])
run_model(lambda: NgramUnk(n, λ, voc_ratio), f'results/ngram_n={n}_l={λ}_voc={voc_ratio}')
else:
run_model(lambda: NgramNoUnk(n, λ), f'results/ngram_n={n}_l={λ}')
