def generate_probestatisticsretention_fixedfiller(
percentage_train_indistribution=100,
filler_distribution="add05",
normalize_fillerdistribution=True):
"""
NOTE: "outofdistribution", "indistribution", "middistribution" is according to special filler distribution.
"""
print("********************************************************")
print(
"generating probe statistics retention fixed filler with percentage_train_indistribution={percentage_train_indistribution}"
.format(
percentage_train_indistribution=percentage_train_indistribution))
NUM_PERSONS_PER_CATEGORY = 1000
NUM_DIMS = 50
NUM_TRAIN_EXAMPLES = 24000
NUM_TEST_EXAMPLES = 120
NUM_UNSEEN_TEST_EXAMPLES = 120
SAVE_PATH = os.path.join(
"/", "home", "cc27", "Thesis", "generalized_schema_learning", "data",
"probestatisticsretention_percentageindistribution{percentage_train_indistribution}_normalizefillerdistribution{normalize_fillerdistribution}"
.format(
percentage_train_indistribution=percentage_train_indistribution,
normalize_fillerdistribution=normalize_fillerdistribution))
print("Saving to {save_path}".format(save_path=SAVE_PATH))
STORY_FRAME = "begin subject sit subject friend announce emcee perform poet consume dessert drink goodbye".split(
" ")
QUESTIONS = ["QEmcee", "QFriend", "QPoet", "QSubject"]
ROLES = ["emcee", "friend", "poet", "subject"]
PADDING_WORD = "zzz"
num_fillers = len(ROLES)
# Get numerical representations of fillers.
num_fillers_per_category = NUM_PERSONS_PER_CATEGORY * num_fillers
fillers_indistribution = [i for i in range(num_fillers_per_category * 2)]
fillers_training = fillers_indistribution[:int(num_fillers_per_category)]
fillers_indistribution_unseen = fillers_indistribution[
int(num_fillers_per_category):]
fillers_outofdistribution_unseen = [
-1 * i for i in range(1, num_fillers_per_category + 1)
]
fillers_middistribution_unseen = [
i + np.max(fillers_indistribution)
for i in range(num_fillers_per_category)
]
fillers_indistribution = [int(i) for i in fillers_indistribution]
fillers_outofdistribution_unseen = [
int(i) for i in fillers_outofdistribution_unseen
]
fillers_middistribution_unseen = [
int(i) for i in fillers_middistribution_unseen
]
print("fillers_training", fillers_training)
print("fillers_indistribution_unseen", fillers_indistribution_unseen)
print("fillers_outofdistribution_unseen", fillers_outofdistribution_unseen)
print("fillers_middistribution_indices_unseen",
fillers_middistribution_unseen)
# Get wordslist.
wordslist = list(STORY_FRAME + QUESTIONS + fillers_indistribution +
fillers_outofdistribution_unseen +
fillers_middistribution_unseen)
wordslist.append(PADDING_WORD)
story_frame_matrix = np.expand_dims(np.expand_dims(np.array(
[wordslist.index(word) for word in STORY_FRAME]),
axis=1),
axis=0)
padding = np.reshape(np.array([wordslist.index(PADDING_WORD)]), (1, 1, 1))
# Get wordslist indices of fillers.
fillers_indices_training = [
wordslist.index(filler) for filler in fillers_training
]
fillers_indistribution_indices_unseen = [
wordslist.index(filler) for filler in fillers_indistribution_unseen
]
fillers_outofdistribution_indices_unseen = [
wordslist.index(filler) for filler in fillers_outofdistribution_unseen
]
fillers_middistribution_indices_unseen = [
wordslist.index(filler) for filler in fillers_middistribution_unseen
]
# Get fillers used in each role during training.
fillers_bytrainmissingrole = dict()
for i in range(num_fillers):
fillers_bytrainmissingrole[ROLES[i]] = fillers_indices_training[
i * NUM_PERSONS_PER_CATEGORY:(i + 1) * NUM_PERSONS_PER_CATEGORY]
fillers_bytrainrole = dict()
for i in range(num_fillers):
role = ROLES[i]
fillers_bytrainrole[role] = np.array(
list(
set(fillers_indices_training) -
set(fillers_bytrainmissingrole[role])))
# Get indices of certain words in wordslist and in story.
role_wordindices = dict()
for role in ROLES:
role_wordindices[role] = wordslist.index(role)
role_storyindices = dict()
for role in ROLES:
role_storyindices[role] = np.where(
np.squeeze(story_frame_matrix) == role_wordindices[role])[0]
question_wordindices = [
wordslist.index(question) for question in QUESTIONS
]
question_storyindices = {
question: STORY_FRAME.index(role)
for question, role in zip(question_wordindices, ROLES)
}
train_X = np.empty(
(0, story_frame_matrix.shape[1] + 2, story_frame_matrix.shape[2]))
train_y = np.empty((0, 1))
for i in range(NUM_TRAIN_EXAMPLES):
story = np.copy(story_frame_matrix)
for role in ROLES:
filler = np.random.choice(fillers_bytrainrole[role])
story[0, role_storyindices[role], 0] = filler
question = np.random.choice(question_wordindices)
answer = [story.squeeze()[question_storyindices[question]]]
story = np.concatenate(
(story, padding, np.reshape(question, (1, 1, 1))), axis=1)
train_X = np.concatenate((train_X, story), axis=0)
train_y = np.concatenate((train_y, np.reshape(np.array(answer),
(1, 1))),
axis=0)
if not os.path.exists(SAVE_PATH):
os.makedirs(SAVE_PATH)
with open(os.path.join(SAVE_PATH, "train.p"), "wb") as f:
pickle.dump([train_X, train_y], f)
# Generate test set with excluded role-filler pairs.
test_X = np.empty(
(0, story_frame_matrix.shape[1] + 2, story_frame_matrix.shape[2]))
test_y = np.empty((0, 1))
for i in range(NUM_TEST_EXAMPLES):
story = np.copy(story_frame_matrix)
for role in ROLES:
filler = np.random.choice(fillers_bytrainmissingrole[role])
story[0, role_storyindices[role], 0] = filler
question = np.random.choice(question_wordindices)
answer = [story.squeeze()[question_storyindices[question]]]
story = np.concatenate(
(story, padding, np.reshape(question, (1, 1, 1))), axis=1)
test_X = np.concatenate((test_X, story), axis=0)
test_y = np.concatenate((test_y, np.reshape(np.array(answer), (1, 1))),
axis=0)
with open(os.path.join(SAVE_PATH, "test.p"), "wb") as f:
pickle.dump([test_X, test_y], f)
# Generate split test set with excluded role-filler pairs.
for question in question_wordindices:
split_test_X = np.empty(
(0, story_frame_matrix.shape[1] + 2, story_frame_matrix.shape[2]))
split_test_y = np.empty((0, 1))
for i in range(NUM_TEST_EXAMPLES):
story = np.copy(story_frame_matrix)
for role in ROLES:
filler = np.random.choice(fillers_bytrainmissingrole[role])
story[0, role_storyindices[role], 0] = filler
answer = [story.squeeze()[question_storyindices[question]]]
story = np.concatenate(
(story, padding, np.reshape(question, (1, 1, 1))), axis=1)
split_test_X = np.concatenate((split_test_X, story), axis=0)
split_test_y = np.concatenate(
(split_test_y, np.reshape(np.array(answer), (1, 1))), axis=0)
print(wordslist[question], np.unique(split_test_y))
with open(os.path.join(SAVE_PATH, "test_%s.p" % wordslist[question]),
"wb") as f:
pickle.dump([split_test_X, split_test_y], f)
# Generate test set with unseen, in distribution fillers.
for question in question_wordindices:
split_testunseen_X = np.empty(
(0, story_frame_matrix.shape[1] + 2, story_frame_matrix.shape[2]))
split_testunseen_y = np.empty((0, 1))
for i in range(NUM_UNSEEN_TEST_EXAMPLES):
story = np.copy(story_frame_matrix)
for role in ROLES:
filler = np.random.choice(
fillers_indistribution_indices_unseen)
story[0, role_storyindices[role], 0] = filler
answer = [story.squeeze()[question_storyindices[question]]]
story = np.concatenate(
(story, padding, np.reshape(question, (1, 1, 1))), axis=1)
split_testunseen_X = np.concatenate((split_testunseen_X, story),
axis=0)
split_testunseen_y = np.concatenate(
(split_testunseen_y, np.reshape(np.array(answer), (1, 1))),
axis=0)
with open(
os.path.join(
SAVE_PATH,
"test_%s_unseen_indistribution.p" % wordslist[question]),
"wb") as f:
pickle.dump([split_testunseen_X, split_testunseen_y], f)
# Generate test set with unseen, out of distribution fillers.
for question in question_wordindices:
split_testunseen_X = np.empty(
(0, story_frame_matrix.shape[1] + 2, story_frame_matrix.shape[2]))
split_testunseen_y = np.empty((0, 1))
for i in range(NUM_UNSEEN_TEST_EXAMPLES):
story = np.copy(story_frame_matrix)
for role in ROLES:
filler = np.random.choice(
fillers_outofdistribution_indices_unseen)
story[0, role_storyindices[role], 0] = filler
answer = [story.squeeze()[question_storyindices[question]]]
story = np.concatenate(
(story, padding, np.reshape(question, (1, 1, 1))), axis=1)
split_testunseen_X = np.concatenate((split_testunseen_X, story),
axis=0)
split_testunseen_y = np.concatenate(
(split_testunseen_y, np.reshape(np.array(answer), (1, 1))),
axis=0)
with open(
os.path.join(
SAVE_PATH, "test_%s_unseen_outofdistribution.p" %
wordslist[question]), "wb") as f:
pickle.dump([split_testunseen_X, split_testunseen_y], f)
# Generate test set with unseen, mid distribution fillers.
for question in question_wordindices:
split_testunseen_X = np.empty(
(0, story_frame_matrix.shape[1] + 2, story_frame_matrix.shape[2]))
split_testunseen_y = np.empty((0, 1))
for i in range(NUM_UNSEEN_TEST_EXAMPLES):
story = np.copy(story_frame_matrix)
for role in ROLES:
filler = np.random.choice(
fillers_middistribution_indices_unseen)
story[0, role_storyindices[role], 0] = filler
answer = [story.squeeze()[question_storyindices[question]]]
story = np.concatenate(
(story, padding, np.reshape(question, (1, 1, 1))), axis=1)
split_testunseen_X = np.concatenate((split_testunseen_X, story),
axis=0)
split_testunseen_y = np.concatenate(
(split_testunseen_y, np.reshape(np.array(answer), (1, 1))),
axis=0)
with open(
os.path.join(
SAVE_PATH,
"test_%s_unseen_middistribution.p" % wordslist[question]),
"wb") as f:
pickle.dump([split_testunseen_X, split_testunseen_y], f)
# Generate embedding.
embedding = []
fillers_training_indistribution = []
fillers_training_outofdistribution = []
for i in range(num_fillers):
fillers_training_subset = fillers_training[i *
NUM_PERSONS_PER_CATEGORY:
(i + 1) *
NUM_PERSONS_PER_CATEGORY]
fillers_training_indistribution += fillers_training_subset[:int(
NUM_PERSONS_PER_CATEGORY * percentage_train_indistribution /
100.0)]
fillers_training_outofdistribution += fillers_training_subset[
int(NUM_PERSONS_PER_CATEGORY * percentage_train_indistribution /
100.0):]
print("fillers training indistribution", fillers_training_indistribution)
print("fillers_training_outofdistribution",
fillers_training_outofdistribution)
for i in range(len(wordslist)):
word = wordslist[i]
word_embedding = {}
word_embedding['index'] = i
word_embedding['word'] = word
if word in fillers_training_indistribution:
print(word, "train filler, in distribution")
word_embedding['vector'] = create_word_vector(
"add05",
normalize_fillerdistribution=normalize_fillerdistribution)
elif word in fillers_training_outofdistribution:
print(word, "train filler, out of distribution")
word_embedding['vector'] = create_word_vector()
elif word in fillers_indistribution_unseen:
print(word, "in distribution")
word_embedding['vector'] = create_word_vector(
"add05",
normalize_fillerdistribution=normalize_fillerdistribution)
elif word in fillers_middistribution_unseen:
print(word, "mid distribution")
word_embedding['vector'] = create_word_vector(
"add025",
normalize_fillerdistribution=normalize_fillerdistribution)
else:
word_embedding['vector'] = create_word_vector()
embedding.append(word_embedding)
with open(os.path.join(SAVE_PATH, "embedding.p"), "wb") as f:
pickle.dump(embedding, f)
with open(os.path.join(SAVE_PATH, "wordslist.p"), "wb") as f:
pickle.dump(wordslist, f)
def generate_experiments(num_dims=50,
num_train_examples=12000,
num_test_examples=120,
num_train_fillers_per_category=10000,
num_test_fillers_per_category=1000,
normalize_filler_distribution=True,
dims=50,
percentage=80):
save_dir = os.path.join(
"/", "home", "cc27", "Thesis", "generalized_schema_learning", "data",
"probe_role_statistic_recall_normalize_%d" % percentage)
print("Saving to {save_dir}".format(save_dir=save_dir))
if not os.path.exists(save_dir):
os.mkdir(save_dir)
STORY_FRAME = "begin subject sit subject friend announce emcee perform poet consume dessert drink goodbye".split(
" ")
ROLES = ["emcee", "friend", "poet", "subject"]
NOISE_WORD = "zzz"
questions = [get_question_for_role(role) for role in ROLES]
# Get wordlist.
add_05_fillers_train = [
'add_05_train' + str(filler_index)
for filler_index in range(num_train_fillers_per_category)
]
add_05_fillers_test = [
'add_05_test' + str(filler_index)
for filler_index in range(num_test_fillers_per_category)
]
add_05_fillers = add_05_fillers_train + add_05_fillers_test
subtract_05_fillers_train = [
'subtract_05_train' + str(filler_index)
for filler_index in range(num_train_fillers_per_category)
]
subtract_05_fillers_test = [
'subtract_05_test' + str(filler_index)
for filler_index in range(num_test_fillers_per_category)
]
subtract_05_fillers = subtract_05_fillers_train + subtract_05_fillers_test
no_addition_fillers = [
str(filler_index)
for filler_index in range(num_test_fillers_per_category)
]
fillers = add_05_fillers + subtract_05_fillers + no_addition_fillers
wordlist = list(STORY_FRAME + questions + fillers)
wordlist.append(NOISE_WORD)
story_frame_matrix = np.expand_dims(np.expand_dims(np.array(
[wordlist.index(word) for word in STORY_FRAME]),
axis=1),
axis=0)
distribution_A_fillers_train = list(
np.random.choice(
add_05_fillers_train,
size=num_train_fillers_per_category * percentage // 100,
replace=False)) + list(
np.random.choice(subtract_05_fillers_train,
size=num_train_fillers_per_category *
(100 - percentage) // 100,
replace=False))
distribution_A_fillers_test = list(
np.random.choice(
add_05_fillers_test,
size=num_test_fillers_per_category * percentage // 100,
replace=False)) + list(
np.random.choice(subtract_05_fillers_test,
size=num_test_fillers_per_category *
(100 - percentage) // 100,
replace=False))
distribution_B_fillers_train = list(
np.random.choice(
subtract_05_fillers_train,
size=num_train_fillers_per_category * percentage // 100,
replace=False)) + list(
np.random.choice(add_05_fillers_train,
size=num_train_fillers_per_category *
(100 - percentage) // 100,
replace=False))
distribution_B_fillers_test = list(
np.random.choice(
subtract_05_fillers_test,
size=num_test_fillers_per_category * percentage // 100,
replace=False)) + list(
np.random.choice(add_05_fillers_test,
size=num_test_fillers_per_category *
(100 - percentage) // 100,
replace=False))
# Get indices of words.
role_wordlist_indices = dict()
for role in ROLES:
role_wordlist_indices[role] = wordlist.index(role)
role_story_indices = dict()
for role in ROLES:
role_story_indices[role] = np.where(np.array(STORY_FRAME) == role)[0]
question_wordlist_indices = {
role: wordlist.index(get_question_for_role(role))
for role in ROLES
}
noise_wordlist_index = wordlist.index(NOISE_WORD)
distribution_A_fillers_indices_train = [
wordlist.index(filler) for filler in distribution_A_fillers_train
]
distribution_B_fillers_indices_train = [
wordlist.index(filler) for filler in distribution_B_fillers_train
]
distribution_A_fillers_indices_test = [
wordlist.index(filler) for filler in distribution_A_fillers_test
]
distribution_B_fillers_indices_test = [
wordlist.index(filler) for filler in distribution_B_fillers_test
]
no_addition_fillers_indices = [
wordlist.index(filler) for filler in no_addition_fillers
]
# Generate training set.
train_fillers_by_role_dict = {
'emcee': distribution_B_fillers_indices_train,
'friend': distribution_B_fillers_indices_train,
'poet': distribution_A_fillers_indices_train,
'subject': distribution_A_fillers_indices_train,
}
write_examples(fillers_by_role_dict=train_fillers_by_role_dict,
story_frame_matrix=story_frame_matrix,
num_examples=num_train_examples,
roles=ROLES,
role_story_indices=role_story_indices,
question_wordlist_indices=question_wordlist_indices,
noise_wordlist_index=noise_wordlist_index,
save_path=os.path.join(save_dir, "train.p"))
# Generate in distribution test set.
test_fillers_by_role_dict = {
'emcee': distribution_B_fillers_indices_test,
'friend': distribution_B_fillers_indices_test,
'poet': distribution_A_fillers_indices_test,
'subject': distribution_A_fillers_indices_test,
}
write_examples(fillers_by_role_dict=test_fillers_by_role_dict,
story_frame_matrix=story_frame_matrix,
num_examples=num_test_examples,
roles=ROLES,
role_story_indices=role_story_indices,
question_wordlist_indices=question_wordlist_indices,
noise_wordlist_index=noise_wordlist_index,
save_path=os.path.join(save_dir, "test.p"))
# Generate out of distribution test set. (draw each filler from the opposite test pool)
test_flipped_distribution_fillers_by_role_dict = {
'emcee': distribution_A_fillers_indices_train,
'friend': distribution_A_fillers_indices_train,
'poet': distribution_B_fillers_indices_train,
'subject': distribution_B_fillers_indices_train,
}
write_examples(
fillers_by_role_dict=test_flipped_distribution_fillers_by_role_dict,
story_frame_matrix=story_frame_matrix,
num_examples=num_test_examples,
roles=ROLES,
role_story_indices=role_story_indices,
question_wordlist_indices=question_wordlist_indices,
noise_wordlist_index=noise_wordlist_index,
save_path=os.path.join(save_dir, "test_flipped_distribution.p"))
# Generate unseen flipped_distribution test set.
unseen_flipped_distribution_fillers_by_role_dict = {
'emcee': distribution_A_fillers_indices_test,
'friend': distribution_A_fillers_indices_test,
'poet': distribution_B_fillers_indices_test,
'subject': distribution_B_fillers_indices_test,
}
write_examples(
fillers_by_role_dict=unseen_flipped_distribution_fillers_by_role_dict,
story_frame_matrix=story_frame_matrix,
num_examples=num_test_examples,
roles=ROLES,
role_story_indices=role_story_indices,
question_wordlist_indices=question_wordlist_indices,
noise_wordlist_index=noise_wordlist_index,
save_path=os.path.join(save_dir, "test_unseen_flipped_distribution.p"))
# Generate unseen ood test set.
unseen_no_addition_fillers_by_role_dict = {
'emcee': no_addition_fillers_indices,
'friend': no_addition_fillers_indices,
'poet': no_addition_fillers_indices,
'subject': no_addition_fillers_indices,
}
write_examples(
fillers_by_role_dict=unseen_no_addition_fillers_by_role_dict,
story_frame_matrix=story_frame_matrix,
num_examples=num_test_examples,
roles=ROLES,
role_story_indices=role_story_indices,
question_wordlist_indices=question_wordlist_indices,
noise_wordlist_index=noise_wordlist_index,
save_path=os.path.join(save_dir,
"test_unseen_no_addition_distribution.p"))
# Generate ambiguous test set. (replace the queried filler with padding)
write_examples(fillers_by_role_dict=train_fillers_by_role_dict,
story_frame_matrix=story_frame_matrix,
num_examples=num_test_examples,
roles=ROLES,
role_story_indices=role_story_indices,
question_wordlist_indices=question_wordlist_indices,
noise_wordlist_index=noise_wordlist_index,
save_path=os.path.join(save_dir,
"test_ambiguous_queried_role.p"),
ambiguous='queried_role')
# Generate fully ambiguous test set. (replace the entire story frame with padding)
write_examples(fillers_by_role_dict=train_fillers_by_role_dict,
story_frame_matrix=story_frame_matrix,
num_examples=num_test_examples,
roles=ROLES,
role_story_indices=role_story_indices,
question_wordlist_indices=question_wordlist_indices,
noise_wordlist_index=noise_wordlist_index,
save_path=os.path.join(save_dir, "test_ambiguous_all.p"),
ambiguous='all')
# Generate embedding.
embedding = []
for i in range(len(wordlist)):
word = wordlist[i]
word_embedding = {}
word_embedding['index'] = i
word_embedding['word'] = word
if "add_05" in word:
word_embedding['vector'] = create_word_vector(
"add05",
normalize_filler_distribution=normalize_filler_distribution)
elif "subtract_05" in word:
word_embedding['vector'] = create_word_vector(
"subtract05",
normalize_filler_distribution=normalize_filler_distribution)
elif word == 'zzz':
print('generating zzz vector')
word_embedding['vector'] = np.zeros(dims)
else:
word_embedding['vector'] = create_word_vector()
embedding.append(word_embedding)
with open(os.path.join(save_dir, "embedding.p"), "wb") as f:
pickle.dump(embedding, f)
with open(os.path.join(save_dir, "wordlist.p"), "wb") as f:
pickle.dump(wordlist, f)
def generate_train3roles_testnewrole(num_persons_per_category,
filler_distribution=None):
NUM_TRAIN_EXAMPLES = 24000
NUM_TEST_EXAMPLES = 120
NUM_UNSEEN_TEST_EXAMPLES = 120
NUM_UNSEEN_FILLERS = 100
SAVE_PATH = os.path.join(
"/", "home", "cc27", "Thesis", "generalized_schema_learning", "data",
"generate_train3roles_testnewrole_withunseentestfillers_%dpersonspercategory_%dtrain_%dtest"
% (num_persons_per_category, NUM_TRAIN_EXAMPLES, NUM_TEST_EXAMPLES))
if not os.path.exists(SAVE_PATH):
os.mkdir(SAVE_PATH)
STORY_FRAME = "begin subject sit subject friend announce emcee perform poet consume dessert drink goodbye".split(
" ")
QUESTIONS = ["QEmcee", "QFriend", "QPoet", "QSubject"]
ROLES = ["emcee", "friend", "poet", "subject", "dessert", "drink"]
PADDING_WORD = "zzz"
NUM_PERSON_FILLERS = 4
filler_indices = []
for role in ROLES:
filler_indices += list(np.where(np.array(STORY_FRAME) == role)[0])
person_fillers = [
str(i) for i in range(num_persons_per_category * NUM_PERSON_FILLERS)
]
person_fillers_unseenintraining = [
str(-1 * i) for i in range(1, NUM_UNSEEN_FILLERS + 1)
]
wordslist = list(
set(STORY_FRAME + QUESTIONS + person_fillers +
person_fillers_unseenintraining))
wordslist.append(PADDING_WORD)
story_frame_matrix = np.expand_dims(np.expand_dims(np.array(
[wordslist.index(word) for word in STORY_FRAME]),
axis=1),
axis=0)
person_fillers_indices = [
wordslist.index(filler) for filler in person_fillers
]
person_fillers_unseenintraining_indices = [
wordslist.index(filler) for filler in person_fillers_unseenintraining
]
person_fillers_bytrainmissingrole = {}
for i in range(NUM_PERSON_FILLERS):
person_fillers_bytrainmissingrole[ROLES[i]] = person_fillers_indices[
i * num_persons_per_category:(i + 1) * num_persons_per_category]
person_fillers_bytrainrole = {}
for i in range(NUM_PERSON_FILLERS):
role = ROLES[i]
person_fillers_bytrainrole[role] = np.array(
list(
set(person_fillers_indices) -
set(person_fillers_bytrainmissingrole[role])))
# Generate train data.
PERSON_ROLES = ["emcee", "friend", "poet", "subject"]
person_wordindices = {}
for role in PERSON_ROLES:
person_wordindices[role] = wordslist.index(role)
person_storyindices = {}
for role in PERSON_ROLES:
person_storyindices[role] = np.where(
np.squeeze(story_frame_matrix) == person_wordindices[role])[0]
role_wordindices = {}
for role in ROLES:
role_wordindices[role] = wordslist.index(role)
role_storyindices = {}
for role in ROLES:
role_storyindices[role] = np.where(
np.squeeze(story_frame_matrix) == role_wordindices[role])[0]
question_wordindices = [
wordslist.index(question) for question in QUESTIONS
]
question_storyindices = {
question: STORY_FRAME.index(role)
for question, role in zip(question_wordindices, ROLES)
}
padding = np.reshape(np.array([wordslist.index(PADDING_WORD)]), (1, 1, 1))
train_X = np.empty(
(0, story_frame_matrix.shape[1] + 2, story_frame_matrix.shape[2]))
train_y = np.empty((0, 1))
for i in range(NUM_TRAIN_EXAMPLES):
story = np.copy(story_frame_matrix)
for role in PERSON_ROLES:
filler = np.random.choice(person_fillers_bytrainrole[role])
story[0, person_storyindices[role], 0] = filler
question = np.random.choice(question_wordindices)
answer = [story.squeeze()[question_storyindices[question]]]
story = np.concatenate(
(story, padding, np.reshape(question, (1, 1, 1))), axis=1)
train_X = np.concatenate((train_X, story), axis=0)
train_y = np.concatenate((train_y, np.reshape(np.array(answer),
(1, 1))),
axis=0)
with open(os.path.join(SAVE_PATH, "train.p"), "wb") as f:
pickle.dump([train_X, train_y], f)
test_X = np.empty(
(0, story_frame_matrix.shape[1] + 2, story_frame_matrix.shape[2]))
test_y = np.empty((0, 1))
for i in range(NUM_TEST_EXAMPLES):
story = np.copy(story_frame_matrix)
for role in PERSON_ROLES:
filler = np.random.choice(person_fillers_bytrainmissingrole[role])
story[0, person_storyindices[role], 0] = filler
question = np.random.choice(question_wordindices)
answer = [story.squeeze()[question_storyindices[question]]]
story = np.concatenate(
(story, padding, np.reshape(question, (1, 1, 1))), axis=1)
test_X = np.concatenate((test_X, story), axis=0)
test_y = np.concatenate((test_y, np.reshape(np.array(answer), (1, 1))),
axis=0)
with open(os.path.join(SAVE_PATH, "test.p"), "wb") as f:
pickle.dump([test_X, test_y], f)
for question in question_wordindices:
split_testunseen_X = np.empty(
(0, story_frame_matrix.shape[1] + 2, story_frame_matrix.shape[2]))
split_testunseen_y = np.empty((0, 1))
for i in range(NUM_UNSEEN_TEST_EXAMPLES):
story = np.copy(story_frame_matrix)
for role in ROLES:
filler = np.random.choice(
person_fillers_unseenintraining_indices)
story[0, role_storyindices[role], 0] = filler
answer = [story.squeeze()[question_storyindices[question]]]
story = np.concatenate(
(story, padding, np.reshape(question, (1, 1, 1))), axis=1)
split_testunseen_X = np.concatenate((split_testunseen_X, story),
axis=0)
split_testunseen_y = np.concatenate(
(split_testunseen_y, np.reshape(np.array(answer), (1, 1))),
axis=0)
with open(
os.path.join(SAVE_PATH,
"test_%s_unseen.p" % wordslist[question]),
"wb") as f:
pickle.dump([split_testunseen_X, split_testunseen_y], f)
embedding = []
for i in range(len(wordslist)):
word = wordslist[i]
word_embedding = {}
word_embedding['index'] = i
word_embedding['word'] = word
if word in person_fillers or word in person_fillers_unseenintraining:
print(word, filler_distribution)
word_embedding['vector'] = create_word_vector(filler_distribution)
else:
word_embedding['vector'] = create_word_vector()
embedding.append(word_embedding)
with open(os.path.join(SAVE_PATH, "embedding.p"), "wb") as f:
pickle.dump(embedding, f)
with open(os.path.join(SAVE_PATH, "wordslist.p"), "wb") as f:
pickle.dump(wordslist, f)
def generate_onefillerperrole():
NUM_TRAIN_EXAMPLES = 24000
NUM_TEST_EXAMPLES = 120
NUM_DIMS = 50
SAVE_PATH = os.path.join("/", "home", "cc27", "Thesis", "generalized_schema_learning", "data", "storyv2_train20000_AllQs")
STORY_FRAME = "begin subject sit subject friend announce emcee perform poet consume dessert drink goodbye".split(" ")
QUESTIONS = ["QEmcee", "QFriend", "QPoet", "QSubject", "QDessert", "QDrink"]
ROLES = ["emcee", "friend", "poet", "subject", "dessert", "drink"]
PADDING_WORD = "zzz"
filler_indices = []
for role in ROLES:
filler_indices += list(np.where(np.array(STORY_FRAME) == role)[0])
num_questions = len(QUESTIONS)
wordslist = list(set(STORY_FRAME + QUESTIONS))
wordslist.append(PADDING_WORD)
story_frame_matrix = np.expand_dims(np.expand_dims(np.array([wordslist.index(word) for word in STORY_FRAME]), axis=1), axis=0)
def generate_data(num_examples, questions, roles):
num_questions = len(questions)
stories = np.repeat(story_frame_matrix, num_questions, axis=0)
padding = np.reshape(np.repeat([wordslist.index(PADDING_WORD)], num_questions), (num_questions, 1, 1))
queries = np.reshape(np.array([wordslist.index(question) for question in questions]), (num_questions, 1, 1))
stories = np.concatenate((stories, padding, queries), axis=1)
answers = np.reshape(np.array([wordslist.index(role) for role in roles]), (num_questions, 1))
num_repeats = num_examples // num_questions
stories = np.repeat(stories, num_repeats, axis=0)
answers = np.repeat(answers, num_repeats, axis=0)
return stories, answers
train_X, train_y = generate_data(NUM_TRAIN_EXAMPLES, QUESTIONS, ROLES)
test_X, test_y = generate_data(NUM_TEST_EXAMPLES, QUESTIONS, ROLES)
with open(os.path.join(SAVE_PATH, "train.p"), "wb") as f:
pickle.dump([train_X, train_y], f)
with open(os.path.join(SAVE_PATH, "test.p"), "wb") as f:
pickle.dump([test_X, test_y], f)
for question, role in zip(QUESTIONS, ROLES):
split_test_X, split_test_y = generate_data(NUM_TEST_EXAMPLES, [question], [role])
with open(os.path.join(SAVE_PATH, "test_%s.p" % question), "wb") as f:
pickle.dump([split_test_X, split_test_y], f)
import sys
sys.path.append("../")
from directories import base_dir
from embedding_util import create_word_vector
embedding = []
for i in range(len(wordslist)):
word = wordslist[i]
word_embedding = {}
word_embedding['index'] = i
word_embedding['word'] = word
word_embedding['vector'] = create_word_vector()
embedding.append(word_embedding)
with open(os.path.join(SAVE_PATH, "embedding.p"), "wb") as f:
pickle.dump(embedding, f)
with open(os.path.join(SAVE_PATH, "wordslist.p"), "wb") as f:
pickle.dump(wordslist, f)
def write_csw_experiment(experiment_name, num_examples_per_frame, num_unseen_examples_per_frame):
print(experiment_name)
"""Create train and test sets for a role-filler binding experiment.
Assumes story files have been written by Coffee Shop world.
Args:
experiment_name: Name of folder in which stories are stored.
Assumes stories are stored in the directory
home_dir + "narrative/story/", where home_dir is
defined in directories.py.
Saves (in the directory base_dir + "data/experiment_name/", where
base_dir is defined in directories.py):
train.p: A pickle file containing:
X: [num_train_examples x num_words_per_story x 1] matrix of train inputs.
y: [num_train_examples x 1] matrix of correct train outputs.
test.p: A pickle file containing:
X: [num_test_examples x num_words_per_story x 1] matrix of test inputs.
y: [num_test_examples x 1] matrix of correct test outputs.
Xy_english.txt: A file containing human-readable versions of the inputs,
correct outputs, and the word list used in the experiment.
(Each X and y matrix represents words by their index in
the word list.)
"""
experiment_name += "_AllQs"
experiment_data_path = os.path.join(base_dir, "data", experiment_name)
query_delimiter = "?"
query_starter = "Q"
padding_word = "zzz"
distributions_dict = {"DESSERT": "A",
"DRINK": "B",
"EMCEE": "A",
"FRIEND": "B",
"POET": "A",
"SUBJECT": "B"}
if not os.path.exists(experiment_data_path):
os.makedirs(experiment_data_path)
# Create frames.
with open('story_frame.json', 'r') as f:
story_frame_info = json.load(f)
transitions = story_frame_info['transitions']
state_contents = story_frame_info['state_contents']
role_types = story_frame_info['role_types']
state_sequences = construct_all_state_sequences(transitions)
assert(len(state_sequences) == 24)
frames = [flatten_arrays([state_contents[state] for state in state_sequence]) for state_sequence in state_sequences]
num_examples = len(frames) * num_examples_per_frame
num_unseen_examples = len(frames) * num_unseen_examples_per_frame
if 'variablefiller' in experiment_name:
dummy_instances = {role: ['%sFILLER' % role] for role in role_types.keys()}
train_instances, test_instances = dummy_instances, dummy_instances
elif 'fixedfiller' in experiment_name:
train_instances, test_instances = hard_coded_things.fixed_train_instances, hard_coded_things.fixed_test_instances
query_choices = role_types.keys()
wordslist = flatten_arrays(state_contents.values()) + flatten_arrays(train_instances.values()) + flatten_arrays(test_instances.values()) + [padding_word, query_delimiter]
for query_choice in query_choices:
wordslist.append(query_starter + query_choice)
wordslist = list(set(wordslist))
# Determine experiment information.
max_story_length = max([len(frame) for frame in frames])
input_dims = max_story_length + 3 # +2 for query delimiter and the actual query. +1 for padding at end.
X = np.zeros([num_examples, input_dims, 1], dtype=np.int32)
y = np.zeros([num_examples, 1], dtype=np.int32)
test_unseen_X = np.zeros([num_unseen_examples, input_dims, 1], dtype=np.int32)
test_unseen_y = np.zeros([num_unseen_examples, 1], dtype=np.int32)
# Generate inputs and correct outputs from stories.
for frame_index, frame in enumerate(frames):
print('Generating for frame ', frame)
padding_size = max_story_length - len(frame)
frame_roles = [role for role in role_types.keys() if role in frame]
for example_index in range(num_examples_per_frame):
if example_index % 1000 == 0:
print(example_index)
story = copy.deepcopy(frame)
role_assignments = {}
for role in frame_roles:
if 'fixedfiller' in experiment_name:
role_assignment = np.random.choice(train_instances[role_types[role]])
while role_assignment in role_assignments.values():
role_assignment = np.random.choice(train_instances[role_types[role]])
elif 'variablefiller' in experiment_name:
role_assignment = '%sFILLER' % role
role_assignments[role] = role_assignment
story = [role_assignments[word] if word in role_assignments else word for word in story]
queried_role = np.random.choice(list(role_assignments.keys()))
query = query_starter + queried_role
response = role_assignments[queried_role]
# If necessary, add padding to end of story (ensures that inputs are all the same length).
story += [padding_word] * (padding_size + 1) # so we can shift all stories later.
story += [query_delimiter, query]
outputs = [response]
# Convert to numerical representation and add to X and y.
data_index = (num_examples_per_frame * frame_index) + example_index
X[data_index, :, :] = np.expand_dims([wordslist.index(storyword) for storyword in story], axis=1)
y[data_index, :] = [wordslist.index(output_word) for output_word in outputs]
if 'fixedfiller' in experiment_name:
for example_index in range(num_unseen_examples_per_frame):
story = copy.deepcopy(frame)
role_assignments = {}
for role in frame_roles:
role_assignment = np.random.choice(test_instances[role_types[role]])
while role_assignment in role_assignments.values():
role_assignment = np.random.choice(test_instances[role_types[role]])
role_assignments[role] = role_assignment
story = [role_assignments[word] if word in role_assignments else word for word in story]
queried_role = np.random.choice(list(role_assignments.keys()))
query = query_starter + queried_role
response = role_assignments[queried_role]
# If necessary, add padding to end of story (ensures that inputs are all the same length).
story += [padding_word] * (padding_size + 1) # so we can shift all stories later.
story += [query_delimiter, query]
outputs = [response]
# Convert to numerical representation and add to X and y.
data_index = (num_unseen_examples_per_frame * frame_index) + example_index
test_unseen_X[data_index, :, :] = np.expand_dims([wordslist.index(storyword) for storyword in story], axis=1)
test_unseen_y[data_index, :] = [wordslist.index(output_word) for output_word in outputs]
# Assert no repeated stories.
X, unique_seen_indices = np.unique(X, axis=0, return_index=True)
y = y[unique_seen_indices]
if 'fixedfiller' in experiment_name:
num_train = int(4 * len(X) / 5)
train_indices = np.random.choice(len(X), num_train, replace=False)
test_indices = np.array([idx for idx in range(len(X)) if idx not in train_indices])
train_X = X[train_indices, :, :]
train_y = y[train_indices, :]
test_X = X[test_indices, :, :]
test_y = y[test_indices, :]
test_unseen_X, unique_unseen_indices = np.unique(test_unseen_X, axis=0, return_index=True)
test_unseen_y = test_unseen_y[unique_unseen_indices]
elif 'variablefiller' in experiment_name:
train_X, train_y = X, y
test_X, test_y = X, y
test_unseen_X, test_unseen_y = X, y
# Save data into pickle files.
if not os.path.exists(experiment_data_path):
os.makedirs(experiment_data_path)
print(experiment_data_path)
with open(os.path.join(experiment_data_path, 'train.p'), 'wb') as f:
pickle.dump([train_X, train_y], f)
with open(os.path.join(experiment_data_path, 'test.p'), 'wb') as f:
pickle.dump([test_X, test_y], f)
with open(os.path.join(experiment_data_path, 'test_unseen.p'), 'wb') as f:
pickle.dump([test_unseen_X, test_unseen_y], f)
with open(os.path.join(experiment_data_path, 'wordslist.p'), 'wb') as f:
pickle.dump(wordslist, f)
with open('../experiment_parameters.json', 'r') as f:
experiment_parameters = json.load(f)
experiment_parameters['input_dims'][experiment_name] = input_dims
fillers = list(set(flatten_arrays(train_instances.values()) + flatten_arrays(test_instances.values())))
experiment_parameters['filler_indices'][experiment_name] = [wordslist.index(filler) for filler in fillers]
experiment_parameters['padding_indices'][experiment_name] = wordslist.index(padding_word)
if 'variablefiller' in experiment_name:
experiment_parameters['query_to_filler_index'][experiment_name] = {wordslist.index(query_starter + role): [wordslist.index(filler) for filler in dummy_instances[role]] for role in role_types.keys()}
filler_distributions_dict = {wordslist.index(dummy_instances[role][0]): distributions_dict[role] for role in role_types.keys()}
experiment_parameters['filler_distributions'][experiment_name] = filler_distributions_dict
with open('../experiment_parameters.json', 'w') as f:
json.dump(experiment_parameters, f)
embedding = []
for i in range(len(wordslist)):
word = wordslist[i]
word_embedding = {}
word_embedding['index'] = i
word_embedding['word'] = word
word_embedding['vector'] = create_word_vector()
embedding.append(word_embedding)
with open(os.path.join(experiment_data_path, "embedding.p"), 'wb') as f:
pickle.dump(embedding, f)
def generate_batch(X, y, FLAGS, embedding, do_shift_inputs=True, noise_proportion=0.1, zero_vector_noise=False):
"""Generate a train batch.
Constructs batches using one of three possible representations (specified by
FLAGS.filler_type):
fixed_filler: Each word vector is specified by the embedding argument.
variable_filler: Each non-filler word vector is
specified by the embedding argument.
Each filler word (manually specified
for each experiment) represented by a
new randomly generated vector in each
story.
Args:
X: [num_examples x num_words_per_input x 1] matrix of inputs.
y: [num_examples x 1] matrix of correct outputs.
FLAGS: Parameters object of experiment information.
embedding: [num_words x embedding_dims] matrix of word embeddings.
NOTE: irrelevant if using one-hot embedding.
Returns:
A generator containing batch_size examples, each of which contains:
X: [batch_size x num_words_per_input x num_dimensions_per_word] matrix
of inputs.
y: [batch_size x num_dimensions_per_word] matrix of correct outputs.
embedding: [num_words_in_corpus x num_dimensions_per_word] matrix
of vectors representing words in the batch.
"""
batch_size, filler_type = FLAGS.batch_size, FLAGS.filler_type
data_size = len(X)
num_batches = (data_size // batch_size)
for batch_num in range(num_batches):
start_index = batch_num * batch_size
end_index = min((batch_num + 1) * batch_size, data_size)
if filler_type == "fixed_filler":
if do_shift_inputs:
yield embedding[shift_inputs(X[start_index:end_index].squeeze(), FLAGS.experiment_name)], embedding[y[start_index:end_index].squeeze()], embedding
else:
yield embedding[X[start_index:end_index]].squeeze(), embedding[y[start_index:end_index].squeeze()], embedding
elif "variable_filler" in filler_type:
# NOTE: Filler indices manually determined using word list saved by experiment creators.
if "distributions" in filler_type:
filler_indices_and_distributions = experiment_parameters["filler_distributions"][FLAGS.experiment_name]
query_to_filler_indices = experiment_parameters["query_to_filler_index"][FLAGS.experiment_name]
filler_indices = list(filler_indices_and_distributions.keys())
filler_distributions = [filler_indices_and_distributions[filler_index] for filler_index in filler_indices]
filler_indices = [int(index) for index in filler_indices]
else:
filler_indices = experiment_parameters["filler_indices"][FLAGS.experiment_name]
batchX, batchy = X[start_index:end_index].squeeze(), y[start_index:end_index].squeeze()
if FLAGS.function != "analyze" and do_shift_inputs: # Don't randomly shift inputs for decoding analysis.
batchX = shift_inputs(batchX, FLAGS.experiment_name)
embeddingX, embeddingy = embedding[batchX], embedding[batchy]
padding_index = experiment_parameters['padding_indices'][FLAGS.experiment_name]
padding_vector = embedding[padding_index]
epoch_embedding = embedding
for examplenum in range(batch_size):
# Create new random embedding for each filler.
num_fillers = len(filler_indices)
new_filler_embedding = np.empty((num_fillers, embedding_size))
if "distributions" in filler_type:
for j, filler_distribution in enumerate(filler_distributions):
if "variable_filler_distributions_no_subtract" in filler_type:
new_filler_embedding[j, :] = embedding_util.create_word_vector(filler_distribution="C")
elif "variable_filler_distributions_one_distribution" in filler_type:
new_filler_embedding[j, :] = embedding_util.create_word_vector(filler_distribution=filler_distribution, dominant_distribution_proportion=1)
elif "variable_filler_distributions_all_randn_distribution" in filler_type:
new_filler_embedding[j, :] = embedding_util.create_word_vector(filler_distribution="randn")
elif "variable_filler_distributions_A" in filler_type:
new_filler_embedding[j, :] = embedding_util.create_word_vector(filler_distribution="A", dominant_distribution_proportion=1)
elif "variable_filler_distributions_B" in filler_type:
new_filler_embedding[j, :] = embedding_util.create_word_vector(filler_distribution="B", dominant_distribution_proportion=1)
elif "variable_filler_distributions_5050_AB" in filler_type:
new_filler_embedding[j, :] = embedding_util.create_word_vector(filler_distribution="B", dominant_distribution_proportion=0.5)
else:
new_filler_embedding[j, :] = embedding_util.create_word_vector(filler_distribution=filler_distribution)
else:
for j in range(num_fillers):
new_filler_embedding[j, :] = embedding_util.create_word_vector()
# Replace filler embedding with new random embedding.
filler_ix_X = np.where(np.isin(batchX[examplenum], filler_indices))
new_embedding_ix_X = [filler_indices.index(i) for i in batchX[examplenum, filler_ix_X][0]]
embeddingX[examplenum, filler_ix_X] = new_filler_embedding[new_embedding_ix_X]
if "noise" in filler_type:
if np.random.rand() < noise_proportion:
print('noise trial')
queried_filler_index = query_to_filler_indices[str(batchX[examplenum, -1])]
queried_filler_indices = np.where(batchX[examplenum] == queried_filler_index)
if zero_vector_noise:
print('zero vector noise')
embeddingX[examplenum, queried_filler_indices] = np.zeros(padding_vector.shape)
else:
embeddingX[examplenum, queried_filler_indices] = padding_vector
new_embedding_ix_y = [filler_indices.index(batchy[examplenum])]
embeddingy[examplenum] = new_filler_embedding[new_embedding_ix_y]
# Append embedding to original embedding identifying response.
epoch_embedding = np.concatenate((epoch_embedding, new_filler_embedding), axis=0)
yield embeddingX, embeddingy, epoch_embedding