def evaluate(model, data_loader, device, eval_file, max_len, use_squad_v2):
nll_meter = util.AverageMeter()
model.eval()
pred_dict = {}
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(data_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
cc_idxs = cc_idxs.to(device)
qc_idxs = qc_idxs.to(device)
batch_size = cw_idxs.size(0)
# Forward
log_p1, log_p2 = model(cc_idxs, qc_idxs, cw_idxs, qw_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, max_len, use_squad_v2)
# Log info
progress_bar.update(batch_size)
progress_bar.set_postfix(NLL=nll_meter.avg)
preds, _ = util.convert_tokens(gold_dict,
ids.tolist(),
starts.tolist(),
ends.tolist(),
use_squad_v2)
pred_dict.update(preds)
model.train()
results = util.eval_dicts(gold_dict, pred_dict, use_squad_v2)
results_list = [('NLL', nll_meter.avg),
('F1', results['F1']),
('EM', results['EM'])]
if use_squad_v2:
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
return results, pred_dict
def evaluate(model, data_loader, device, eval_file):
nll_meter = util.AverageMeter()
model.eval()
pred_dict = dict()
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(data_loader.dataset)) as progress_bar:
model.module.set_eval_data(gold_dict) # pass eval_data as model state
for cw_idxs, cc_idxs, \
qw_idxs, qc_idxs, \
start_idxs, end_idxs, \
counts, ids in data_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
cc_idxs = cc_idxs.to(device)
qw_idxs = qw_idxs.to(device)
qc_idxs = qc_idxs.to(device)
start_idxs = start_idxs.to(device)
end_idxs = end_idxs.to(device)
counts = counts.to(device)
ids = ids.to(device)
batch_size = cw_idxs.size(0)
# Forward
output_dict = model(cw_idxs, cc_idxs, qw_idxs, qc_idxs, ids,
start_idxs, end_idxs, counts)
loss = output_dict['loss']
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
# Log info
progress_bar.update(batch_size)
progress_bar.set_postfix(NLL=nll_meter.avg)
pred_dict.update(output_dict["predictions"]
) # Errato ogni volta penso sovrascriva
model.module.set_eval_data(None)
model.train()
eval_dict = eval_dicts(gold_dict, pred_dict)
results_list = [('Loss', nll_meter.avg), ('F1', eval_dict['F1']),
('EM', eval_dict['EM'])]
eval_dict = OrderedDict(results_list)
return eval_dict, pred_dict
def evaluate_bert(args):
eval_file_path: str = args.eval_file_path
with open(eval_file_path, 'r') as fh:
gold_dict = invert_golden(json_load(fh))
result_file_path: str = args.result_file_path
with open(result_file_path, 'r') as fh:
pred_dict = json_load(fh)
use_squad_v2: bool = True
results = util.eval_dicts(gold_dict, pred_dict, use_squad_v2)
if args.write_predictions_csv:
path, filename = os.path.split(eval_file_path)
filename = ".".join(filename.split(".")[:-1])
sub_path = os.path.join(path, filename) + ".csv"
print("Saving results to: %s" % sub_path)
with open(sub_path, 'w', newline='', encoding='utf-8') as csv_fh:
csv_writer = csv.writer(csv_fh, delimiter=',')
csv_writer.writerow(['Id', 'Predicted'])
for uuid in sorted(pred_dict):
csv_writer.writerow([uuid, pred_dict[uuid]])
return results
def test_model():
args = get_setup_args()
word_vectors = util.torch_from_json(args.word_emb_file)
with open(args.char2idx_file, "r") as f:
char2idx = json_load(f)
model = QANet(word_vectors, char2idx)
cw_idxs = torch.randint(2, 1000, (64, 374))
cc_idxs = torch.randint(2, 50, (64, 374, 200))
qw_idxs = torch.randint(2, 1000, (64, 70))
qc_idxs = torch.randint(2, 50, (64, 70, 200))
cw_idxs[:, 0] = 1
cw_idxs[3, -1] = 0
qw_idxs[:, 0] = 1
qw_idxs[3, -1] = 0
out = model(cw_idxs, cc_idxs, qw_idxs, qc_idxs)
print(out)
def inference(model):
# Get data loader
record_file = vars(args)[f'{args.split}_record_file']
dataset = SQuAD(record_file, args.use_squad_v2)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Evaluate
nll_meter = util.AverageMeter()
# pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = vars(args)[f'{args.split}_eval_file']
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
cc_idxs = cc_idxs.to(device)
qc_idxs = qc_idxs.to(device)
batch_size = cw_idxs.size(0)
# Forward
log_p1, log_p2 = model(cw_idxs, qw_idxs, cc_idxs, qc_idxs)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, args.max_ans_len, args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
idx2pred, uuid2pred = util.convert_tokens(gold_dict,
ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2)
# pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
return sub_dict
def test_input_embedding():
args = get_setup_args()
d_char = 200
word_dropout = 0.1
char_dropout = 0.05
with open(args.char2idx_file, "r") as f:
char2idx = json_load(f)
hidden_size = 500
highway_dropout = 0.1
word_vectors = util.torch_from_json(args.word_emb_file)
input_embedding = InputEmbedding(word_vectors, d_char, char2idx,
hidden_size, word_dropout, char_dropout,
highway_dropout)
word_inputs = torch.tensor([[1, 2, 0], [1, 2, 4]], dtype=torch.long)
char_inputs = torch.tensor([[[1, 2, 2, 0], [1, 3, 2, 3], [0, 0, 0, 0]],
[[1, 5, 2, 0], [1, 3, 6, 3], [3, 4, 2, 1]]],
dtype=torch.long)
emb = input_embedding(word_inputs, char_inputs)
pickle_in = open('input_emb.pickle', 'wb')
pickle.dump(emb, pickle_in)
assert emb.size() == (2, 3, 500)
return emb
def evaluate(model, data_loader, device, eval_file):
nll_meter = util.AverageMeter()
model.eval()
pred_dict = {}
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(data_loader.dataset)) as progress_bar:
for cw_idxs, cw_dep_paths, cc_idxs, qw_idxs, qw_dep_paths, qc_idxs, y1, y2, ids in data_loader:
# Setup for forward
batch_size = 1 # TODO:
# Forward
outputs = model(featues)
y = y.to(device)
loss = loss_fn(outputs, y) # TODO
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
# Log info
progress_bar.update(batch_size)
progress_bar.set_postfix(NLL=nll_meter.avg)
model.train()
results = util.eval_dicts(gold_dict, pred_dict, use_squad_v2)
results_list = [('NLL', nll_meter.avg), ('F1', results['F1']),
('EM', results['EM'])]
results = OrderedDict(results_list)
return results, pred_dict
from ujson import dump
from ujson import load as json_load
import numpy as np
with open('data/word2idx.json', 'r') as fh:
word2idx = json_load(fh)
with open('data/char2idx.json', 'r') as fh:
char2idx = json_load(fh)
with open('data/char_emb.json', 'r') as fh:
char_emb = json_load(fh)
char_emb_for_word = {}
for word in word2idx:
if word not in char_emb_for_word:
wid = word2idx[word]
emb = None
for c in word:
if emb is None:
emb = np.array(char_emb[char2idx[c]])
else:
emb += np.array(char_emb[char2idx[c]])
char_emb_for_word[int(wid)] = emb
with open("data/char_emb_for_word.json", "w") as ujson_file:
dump(char_emb_for_word, ujson_file)
def main(args):
# Set up logging
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=False)
log = util.get_logger(args.save_dir, args.name)
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
# Get model
log.info('Building model...')
if args.model == 'bidaf':
model = BiDAF(word_vectors=word_vectors, hidden_size=args.hidden_size)
elif args.model == 'bidafextra':
model = BiDAFExtra(word_vectors=word_vectors, args=args)
elif args.model == 'fusionnet':
model = FusionNet(word_vectors=word_vectors, args=args)
model = nn.DataParallel(model, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path}...')
model = util.load_model(model, args.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
# Get data loader
log.info('Building dataset...')
record_file = vars(args)[f'{args.split}_record_file']
dataset = SQuAD(record_file, args)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# print("*"*80)
# print(len(dataset.question_idxs))
# for question_idx in dataset.question_idxs:
# print(question_idx)
# print("*" * 80)
# print(self.question_idxs[question_idx])
# self.question_idxs[idx]
# print("data_loader: ",data_loader)
# Evaluate
log.info(f'Evaluating on {args.split} split...')
nll_meter = util.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = vars(args)[f'{args.split}_eval_file']
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
# create statistics
# print("*"*80)
# print(len(gold_dict))
# print(gold_dict['1']['question'])
count_questions_type = defaultdict(lambda: 0)
audit_trail_from_question_type = defaultdict(lambda: [])
list_of_interrogative_pronouns = [
"what", "whose", "why", "which", "where", "when", "how", "who", "whom"
]
for index in range(1, len(gold_dict)):
# transform the question in lower case to simplify the analysis, thus losing the benefit of the capital letters
# possibly indicating the position of the interrogative pronoun in the sentence.
question_lower_case = gold_dict[str(index)]['question'].lower()
list_question_lower_case_with_punctuation = question_lower_case.translate(
{ord(i): " "
for i in "'"}).split()
#
question_lower_case = []
for item in list_question_lower_case_with_punctuation:
question_lower_case.append(
item.translate({ord(i): ""
for i in ",.<>!@£$%^&*()_-+=?"}))
# defining a variable for the first word
first_word_question_lower_case = question_lower_case[0]
# defining variable for the second word
second_word_question_lower_case = question_lower_case[1]
# defining variable for the first and second word
combined_first_and_second_words = first_word_question_lower_case + " " + second_word_question_lower_case
#printing on the screen test for debugging purpose
# Analyzing the sentence
if first_word_question_lower_case in list_of_interrogative_pronouns:
count_questions_type[first_word_question_lower_case] += 1
audit_trail_from_question_type[
first_word_question_lower_case].append(str(index))
# composed question starting by in
elif first_word_question_lower_case == "in":
if second_word_question_lower_case in list_of_interrogative_pronouns and second_word_question_lower_case != "whose":
count_questions_type[combined_first_and_second_words] += 1
audit_trail_from_question_type[
combined_first_and_second_words].append(str(index))
else:
pronoun = find_first_interrogative_pronoun(
list_of_interrogative_pronouns, question_lower_case)
count_questions_type[pronoun] += 1
audit_trail_from_question_type[pronoun].append(str(index))
# composed question starting by by
elif first_word_question_lower_case == "by":
if second_word_question_lower_case in list_of_interrogative_pronouns \
and second_word_question_lower_case !="whom"\
and second_word_question_lower_case !="which"\
and second_word_question_lower_case !="when"\
and second_word_question_lower_case !="how":
count_questions_type[combined_first_and_second_words] += 1
audit_trail_from_question_type[
combined_first_and_second_words].append(str(index))
else:
pronoun = find_first_interrogative_pronoun(
list_of_interrogative_pronouns, question_lower_case)
count_questions_type[pronoun] += 1
audit_trail_from_question_type[pronoun].append(str(index))
else:
pronoun = find_first_interrogative_pronoun(
list_of_interrogative_pronouns, question_lower_case)
#if pronoun =="":
# print(">>", question_lower_case)
# print("@@@", gold_dict[str(index)]['question'])
count_questions_type[pronoun] += 1
audit_trail_from_question_type[pronoun].append(str(index))
# if pronoun =="":
# print(">>", question_lower_case.split())
# print()
#if first_word_question_lower_case == "if":
# print(">>", question_lower_case.split())
# print(count_questions_type)
# if gold_dict[str(index)]['question'].lower().split()[0] == "in":
# print(gold_dict[str(index)]['question'])
reverse_dict_by_value = OrderedDict(
sorted(count_questions_type.items(), key=lambda x: x[1]))
# print(count_questions_type)
total_questions = sum(count_questions_type.values())
# print(reverse_dict)
#for k, v in reverse_dict_by_value.items():
# print( "%s: %s and in percentage: %s" % (k, v, 100*v/total_questions))
#print(audit_trail_from_question_type)
# exit()
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, cw_pos, cw_ner, cw_freq, cqw_extra, y1, y2, ids in data_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
# Forward
if args.model == 'bidaf':
log_p1, log_p2 = model(cw_idxs, qw_idxs)
else:
log_p1, log_p2 = model(cw_idxs, qw_idxs, cw_pos, cw_ner,
cw_freq, cqw_extra)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
if args.split != 'test':
# No labels for the test set, so NLL would be invalid
progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(gold_dict, ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
# Log results (except for test set, since it does not come with labels)
if args.split != 'test':
results = util.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
# Printing information for questions without interrogative pronouns
""""
print("len(gold_dict): ", len(gold_dict))
print("len(pred_dict): ", len(pred_dict))
print("Is gold_dict.keys() identical to pred_dict.keys(): ", gold_dict.keys()==pred_dict.keys())
if gold_dict.keys()!=pred_dict.keys():
for key in gold_dict.keys():
if key not in pred_dict.keys():
print("key ", key, " missing in pred_dict.keys(")
"""
results_list = [('NLL', nll_meter.avg), ('F1', results['F1']),
('EM', results['EM'])]
if args.use_squad_v2:
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
# Computing the F1 score for each type of question
#
# audit_trail_from_question_type[pronoun].append(str(index))
# create a list of the types of questions by extracting the keys from the dict audit_trail_from_question_type
types_of_questions = list(audit_trail_from_question_type.keys())
gold_dict_per_type_of_questions = defaultdict(lambda: [])
pred_dict_per_type_of_questions = {}
gold_dict_per_type_of_questions_start = {}
pred_dict_per_type_of_questions_start = {}
gold_dict_per_type_of_questions_middle = {}
pred_dict_per_type_of_questions_middle = {}
gold_dict_per_type_of_questions_end = {}
pred_dict_per_type_of_questions_end = {}
for type_of_questions in types_of_questions:
#gold_pred = {key: value for key, value in gold_dict.items() if key in audit_trail_from_question_type[type_of_questions]}
#lst_pred = {key: value for key, value in pred_dict.items() if key in audit_trail_from_question_type[type_of_questions]}
# Create two dictionnaries for each type of sentence for gold_dict_per_type_of_questions and pred_dict_per_type_of_questions
gold_dict_per_type_of_questions[type_of_questions] = {
key: value
for key, value in gold_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
pred_dict_per_type_of_questions[type_of_questions] = {
key: value
for key, value in pred_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
# print(type_of_questions," F1 score: ", util.eval_dicts(gold_dict_per_type_of_questions[type_of_questions], pred_dict_per_type_of_questions[type_of_questions], args.use_squad_v2)['F1'])
gold_dict_per_type_of_questions_start[type_of_questions] = {
key: value
for key, value in gold_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
pred_dict_per_type_of_questions_start[type_of_questions] = {
key: value
for key, value in pred_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
gold_dict_per_type_of_questions_middle[type_of_questions] = {
key: value
for key, value in gold_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
pred_dict_per_type_of_questions_middle[type_of_questions] = {
key: value
for key, value in pred_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
gold_dict_per_type_of_questions_end[type_of_questions] = {
key: value
for key, value in gold_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
pred_dict_per_type_of_questions_end[type_of_questions] = {
key: value
for key, value in pred_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
for key, value in gold_dict.items():
#if key in audit_trail_from_question_type[type_of_questions] and key in pred_dict.keys():
if key in audit_trail_from_question_type[
type_of_questions] and type_of_questions != "" and key in pred_dict_per_type_of_questions[
type_of_questions]:
"""
print("type_of_questions: ",type_of_questions)
print("key: ", key)
print("question: ", value["question"])
sub_index = value["question"].lower().find(type_of_questions)
print("sub_index: ",sub_index)
test_fc = value["question"].lower().find(type_of_questions)
print("present type of the var: ",type(test_fc))
#print("question: ", value["question"][str(key)])
print("length of the question: ", len(value["question"]))
print('Position of the interrogative pronoun in the question:', )
"""
# Create two dictionnaries for each type of sentence based at the start of the sentence
if value["question"].lower().find(
type_of_questions) == 1 or value["question"].lower(
).find(type_of_questions) == 0:
#print("BEGINNING")
if type_of_questions != "":
try:
del gold_dict_per_type_of_questions_middle[
type_of_questions][key]
except KeyError:
pass
try:
del pred_dict_per_type_of_questions_middle[
type_of_questions][key]
except KeyError:
pass
try:
del gold_dict_per_type_of_questions_end[
type_of_questions][key]
except KeyError:
pass
try:
del pred_dict_per_type_of_questions_end[
type_of_questions][key]
except KeyError:
pass
#pred_dict_per_type_of_questions_start[type_of_questions] = {key: pred_dict[key] for key in
# gold_dict_per_type_of_questions_start[
# type_of_questions].keys()}
elif value["question"].lower(
).find(type_of_questions) >= len(
value["question"]) - len(type_of_questions) - 5:
#print("END")
if type_of_questions != "":
try:
del gold_dict_per_type_of_questions_middle[
type_of_questions][key]
except KeyError:
pass
try:
del pred_dict_per_type_of_questions_middle[
type_of_questions][key]
except KeyError:
pass
try:
del gold_dict_per_type_of_questions_start[
type_of_questions][key]
except KeyError:
pass
try:
del pred_dict_per_type_of_questions_start[
type_of_questions][key]
except KeyError:
pass
#print("type_of_questions: ",type_of_questions)
#sub_index = value["question"].lower().find(type_of_questions)
#print("sub_index: ", sub_index)
#print("len(value['question']) - len(type_of_questions) - 2: ", len(value["question"])-len(type_of_questions)-2)
#start_string = len(value["question"])-len(type_of_questions)-6
#end_string = len(value["question"])-1
#print("extract at the end: ", value["question"][start_string:end_string])
else:
#print("MIDDLE")
if type_of_questions != "":
try:
del gold_dict_per_type_of_questions_start[
type_of_questions][key]
except KeyError:
pass
try:
del pred_dict_per_type_of_questions_start[
type_of_questions][key]
except KeyError:
pass
try:
del gold_dict_per_type_of_questions_end[
type_of_questions][key]
except KeyError:
pass
try:
del pred_dict_per_type_of_questions_end[
type_of_questions][key]
except KeyError:
pass
pass
"""
if type_of_questions != "":
gold_dict_per_type_of_questions_start[type_of_questions] = {key: value for key, value in gold_dict.items() if (key in audit_trail_from_question_type[type_of_questions] \
and (value["question"].lower().find(type_of_questions) <= 1) \
and key in pred_dict_per_type_of_questions[type_of_questions]) }
"""
"""
for key in gold_dict_per_type_of_questions_start[type_of_questions].keys():
print("key:: ", key )
print("type(key):: ", type(key) )
print("pred_dict[,key,] : ", pred_dict[key])
print("@@@@@@@@@@@@@@@@@@@@@@@@")
pred_dict_per_type_of_questions_start[type_of_questions] = {key: pred_dict[key] for key in gold_dict_per_type_of_questions_start[type_of_questions].keys()}
#pred_dict_per_type_of_questions_start[type_of_questions] = {key: value for key, value in pred_dict.items() if key in list(gold_dict_per_type_of_questions_start[type_of_questions].keys()) }
# Create two dictionnaries for each type of sentence based at the end of the sentence
gold_dict_per_type_of_questions_end[type_of_questions] = {key: value for key, value in gold_dict.items() if key in audit_trail_from_question_type[type_of_questions] \
and value["question"].lower().find(type_of_questions) >= len(value["question"])-len(type_of_questions)-2 \
and key in pred_dict_per_type_of_questions[type_of_questions]}
pred_dict_per_type_of_questions_end[type_of_questions] = {key: pred_dict[key] for key in list(gold_dict_per_type_of_questions_end[type_of_questions].keys())}
#print("*"*80)
# Create two dictionnaries for each type of sentence based at the middle of the sentencecount_questions_type
gold_dict_per_type_of_questions_middle[type_of_questions] = {key: value for key, value in gold_dict.items() if key not in list(gold_dict_per_type_of_questions_start[type_of_questions].keys()) \
and key not in list(gold_dict_per_type_of_questions_end[type_of_questions].keys())}
pred_dict_per_type_of_questions_middle[type_of_questions] = {key: pred_dict[key] for key in list(gold_dict_per_type_of_questions_end[type_of_questions].keys())}
else:
gold_dict_per_type_of_questions_start[""] = gold_dict_per_type_of_questions[""]
pred_dict_per_type_of_questions_start[""] = pred_dict_per_type_of_questions[""]
gold_dict_per_type_of_questions_end[""] = gold_dict_per_type_of_questions[""]
pred_dict_per_type_of_questions_end[""] = pred_dict_per_type_of_questions[""]
gold_dict_per_type_of_questions_middle[""] = gold_dict_per_type_of_questions[""]
pred_dict_per_type_of_questions_middle[""] = pred_dict_per_type_of_questions[""]
"""
positions_in_question = ["beginning", "middle", "end"]
# print(type_of_questions," F1 score: ", util.eval_dicts(gold_dict_per_type_of_questions[type_of_questions], pred_dict_per_type_of_questions[type_of_questions], args.use_squad_v2)['F1'])
list_beginning = [
util.eval_dicts(
gold_dict_per_type_of_questions_start[type_of_questions],
pred_dict_per_type_of_questions_start[type_of_questions],
args.use_squad_v2)['F1']
for type_of_questions in types_of_questions
]
list_middle = [
util.eval_dicts(
gold_dict_per_type_of_questions_middle[type_of_questions],
pred_dict_per_type_of_questions_middle[type_of_questions],
args.use_squad_v2)['F1']
for type_of_questions in types_of_questions
]
list_end = [
util.eval_dicts(
gold_dict_per_type_of_questions_end[type_of_questions],
pred_dict_per_type_of_questions_end[type_of_questions],
args.use_squad_v2)['F1']
for type_of_questions in types_of_questions
]
#for type_of_questions in types_of_questions:
# print("gold_dict_per_type_of_questions_start[type_of_questions]: ",gold_dict_per_type_of_questions_start[type_of_questions])
# print("pred_dict_per_type_of_questions[type_of_questions]: ",pred_dict_per_type_of_questions[type_of_questions])
F1 = np.array([list_beginning, list_middle, list_end])
m, n = F1.shape
value_to_ignore = []
for i in range(m):
for j in range(n):
if F1[i, j] == "NA" or F1[i, j] == 0:
value_to_ignore.append((i, j))
print("value to ignore: ", value_to_ignore)
#F1 = np.array([[0, 0, 0, 0, 0.1, 0, 0, 0, 0, 0, 0, 0, 0],
# [0, 0, 0, 0, 0.1, 0, 0, 0, 0, 0, 0, 0, 0],
# [0, 0, 0, 0, 0.1, 0, 0, 0, 0, 0, 0, 0, 0]])
data_label = copy.deepcopy(F1)
for row in data_label:
for column_idx in range(len(row)):
if row[column_idx] == "NA":
row[column_idx] = ""
# print question without interrogative pronoun required for the second part of the analysis:
for key, value in gold_dict.items():
if key in audit_trail_from_question_type[
""] and key in pred_dict.keys():
print("question: ", gold_dict_per_type_of_questions[''])
print("golden answers: ", )
print("prediction: ", pred_dict[key])
print()
fig, ax = plt.subplots()
types_of_questions[types_of_questions.index(
"")] = "Implicit question without interrogative pronoun"
im, cbar = heatmap(F1, positions_in_question, types_of_questions, ax=ax, \
cmap="YlGn", cbarlabel="F1 scores")
texts = annotate_heatmap(im,
data=data_label,
valfmt="{x:.1f}",
ignore=value_to_ignore)
fig.tight_layout()
plt.show()
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in results.items())
log.info(f'{args.split.title()} {results_str}')
# Log to TensorBoard
tbx = SummaryWriter(args.save_dir)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=0,
split=args.split,
num_visuals=args.num_visuals)
# Write submission file
sub_path = join(args.save_dir, args.split + '_' + args.sub_file)
log.info(f'Writing submission file to {sub_path}...')
with open(sub_path, 'w', newline='', encoding='utf-8') as csv_fh:
csv_writer = csv.writer(csv_fh, delimiter=',')
csv_writer.writerow(['Id', 'Predicted'])
for uuid in sorted(sub_dict):
csv_writer.writerow([uuid, sub_dict[uuid]])
def evaluate(model, data_loader, device, eval_file, max_len, use_squad_v2, model_name, gpu_ids):
nll_meter = util.AverageMeter()
model.eval()
pred_dict = {}
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(data_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, cpos_idxs, cner_idxs, cw_ems, cw_tfs, qw_idxs, qc_idxs, qpos_idxs, qner_idxs, qw_ems, qw_tfs, y1, y2, ids in data_loader: # NEW
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
# Forward
if 'baseline' in model_name:
log_p1, log_p2 = model(cw_idxs, qw_idxs)
elif model_name == 'BiDAF_char':
# Additional setup for forward
cc_idxs = cc_idxs.to(device)
qc_idxs = qc_idxs.to(device)
log_p1, log_p2 = model(cw_idxs, qw_idxs, cc_idxs, qc_idxs)
elif (model_name == 'BiDAF_tag') or (model_name == 'BiDAF_tag_unfrozen') or (model_name == 'BiDAF_tag_loss') or (model_name == 'BiDAF_tag_unfrozen_loss'):
# Additional setup for forward
cc_idxs = cc_idxs.to(device)
cpos_idxs = cpos_idxs.to(device)
cner_idxs = cner_idxs.to(device)
qc_idxs = qc_idxs.to(device)
qpos_idxs = qpos_idxs.to(device)
qner_idxs = qner_idxs.to(device)
log_p1, log_p2 = model(cw_idxs, qw_idxs, cc_idxs, qc_idxs, cpos_idxs, qpos_idxs, cner_idxs, qner_idxs)
elif (model_name == 'BiDAF_tag_ext') or (model_name == 'BiDAF_tag_ext_unfrozen'):
# Additional setup for forward
cc_idxs = cc_idxs.to(device)
cpos_idxs = cpos_idxs.to(device)
cner_idxs = cner_idxs.to(device)
cw_ems = cw_ems.to(device)
cw_tfs = cw_tfs.to(device)
qc_idxs = qc_idxs.to(device)
qpos_idxs = qpos_idxs.to(device)
qner_idxs = qner_idxs.to(device)
qw_ems = qw_ems.to(device)
qw_tfs = qw_tfs.to(device)
log_p1, log_p2 = model(cw_idxs, qw_idxs, cc_idxs, qc_idxs, cpos_idxs, qpos_idxs, cner_idxs, qner_idxs, cw_ems, qw_ems, cw_tfs, qw_tfs)
elif args.name == 'coattn':
max_c_len = cw_idxs.size(1)
max_q_len = qw_idxs.size(1)
c_len = []
q_len = []
for i in range(cw_idxs.size(0)):
if len((cw_idxs[i] == 0).nonzero()) != 0:
c_len_i = (cw_idxs[i] == 0).nonzero()[0].item()
else:
c_len_i = cw_idxs.size(1)
if len((qw_idxs[i] == 0).nonzero()) != 0:
q_len_i = (qw_idxs[i] == 0).nonzero()[0].item()
else:
q_len_i = qw_idxs.size(1)
c_len.append(int(c_len_i))
q_len.append(int(q_len_i))
c_len = torch.Tensor(c_len).int()
q_len = torch.Tensor(q_len).int()
num_examples = int(cw_idxs.size(0) / len(gpu_ids))
log_p1, log_p2 = model(max_c_len, max_q_len, cw_idxs, qw_idxs, c_len, q_len, num_examples, True, False)
else: # default: run baseline
log_p1, log_p2 = model(cw_idxs, qw_idxs)
y1, y2 = y1.to(device), y2.to(device)
#if model_name == 'coattn':
# loss = nn.CrossEntropyLoss()(log_p1, y1) + nn.CrossEntropyLoss()(log_p2, y2)
#else:
# loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
#if model_name != 'coattn':
# p1, p2 = log_p1.exp(), log_p2.exp()
#else:
# p1, p2 = log_p1, log_p2
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, max_len, use_squad_v2)
# Log info
progress_bar.update(batch_size)
progress_bar.set_postfix(NLL=nll_meter.avg)
preds, _ = util.convert_tokens(gold_dict,
ids.tolist(),
starts.tolist(),
ends.tolist(),
use_squad_v2)
pred_dict.update(preds)
model.train()
results = util.eval_dicts(gold_dict, pred_dict, use_squad_v2)
results_list = [('NLL', nll_meter.avg),
('F1', results['F1']),
('EM', results['EM'])]
if use_squad_v2:
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
return results, pred_dict
def load(path):
from ujson import load as json_load
with open(path, 'r') as pf:
return GridSearch(json_load(pf))
def train_QaNet(args):
device, args.gpu_ids = util.get_available_devices()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
word_mat = util.torch_from_json(args.word_emb_file)
char_mat = util.torch_from_json(args.char_emb_file)
with open(args.dev_eval_file, 'r') as fh:
dev_eval_file = json_load(fh)
print("Building model...")
train_dataset = SQuAD(args.train_record_file, args.use_squad_v2)
train_dataset = data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
dev_dataset = SQuAD(args.dev_record_file, args.use_squad_v2)
dev_dataset = data.DataLoader(dev_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
lr = args.lr
base_lr = 1
lr_warm_up_num = args.lr_warm_up_num
model = QaNet(word_mat,
char_mat,
args.connector_dim,
args.glove_dim,
args.char_dim,
args.drop_prob,
args.dropout_char,
args.num_heads,
args.c_len,
args.q_len).to(device)
ema = util.EMA(model, args.ema_decay)
parameters = filter(lambda param: param.requires_grad, model.parameters())
optimizer = optim.Adam(lr=base_lr, betas=(0.9, 0.999), eps=1e-7, weight_decay=5e-8, params=parameters)
cr = lr / math.log2(lr_warm_up_num)
scheduler = optim.lr_scheduler.LambdaLR(
optimizer,
lr_lambda=lambda ee: cr * math.log2(ee + 1) if ee < lr_warm_up_num else lr)
best_f1 = 0
best_em = 0
patience = 0
unused = False
for iter in range(args.num_epochs):
train(model, optimizer, scheduler, train_dataset, dev_dataset, dev_eval_file, iter, ema, device)
ema.assign(model)
metrics = test(model, dev_dataset, dev_eval_file, (iter + 1) * len(train_dataset))
dev_f1 = metrics["f1"]
dev_em = metrics["exact_match"]
if dev_f1 < best_f1 and dev_em < best_em:
patience += 1
if patience > args.early_stop:
break
else:
patience = 0
best_f1 = max(best_f1, dev_f1)
best_em = max(best_em, dev_em)
fn = os.path.join(args.save_dir, "model.pt")
torch.save(model, fn)
ema.resume(model)
def main(args):
args.save_dir = util.get_save_dir(args.save_dir,
"exp1_training",
training=False)
log = get_logger(args.logging_dir, "exp1_training")
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
# Get model
log.info('Building model...')
model = BiDAF(word_vectors=word_vectors, hidden_size=args.hidden_size)
model = nn.DataParallel(model, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path}...')
model = util.load_model(model, c.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
# Get data loader
log.info('Building dataset...')
dataset = SQuAD(args.test_record_file, True)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Evaluate
log.info(f'Evaluating on {args.datasplit} split...')
nll_meter = util.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
with open(args.test_eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
# Forward
log_p1, log_p2 = model(cw_idxs, qw_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, c.max_ans_len, True)
# Log info
progress_bar.update(batch_size)
# Not using the unlabeled test set
# if args.split != 'test':
# # No labels for the test set, so NLL would be invalid
# progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(gold_dict, ids.tolist(),
starts.tolist(),
ends.tolist(), True)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
# Log results (except for test set, since it does not come with labels)
results = util.eval_dicts(gold_dict, pred_dict, True)
results_list = [('NLL', nll_meter.avg), ('F1', results['F1']),
('EM', results['EM'])]
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in results.items())
log.info(f'{args.datasplit} {results_str}')
# Log to TensorBoard
tbx = SummaryWriter(c.save_dir)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.test_eval_file,
step=0,
split=args.datasplit,
num_visuals=args.num_visuals)
def main(args):
# Set up logging
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=False)
log = util.get_logger(args.save_dir, args.name)
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
ch_vectors = util.torch_from_json(args.char_emb_file)
# Get model
log.info('Building model...')
model = BiDAF(word_vectors=word_vectors,
ch_vectors=ch_vectors,
hidden_size=args.hidden_size)
model = nn.DataParallel(model, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path}...')
model = util.load_model(model, args.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
# Get data loader
log.info('Building dataset...')
record_file = vars(args)[f'{args.split}_record_file']
dataset = SQuAD(record_file, args.use_squad_v2)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Evaluate
log.info(f'Evaluating on {args.split} split...')
nll_meter = util.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = vars(args)[f'{args.split}_eval_file']
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
cc_idxs = cc_idxs.to(device)
qc_idxs = qc_idxs.to(device)
batch_size = cw_idxs.size(0)
# Forward
log_p1, log_p2 = model(cw_idxs, qw_idxs, cc_idxs, qc_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
if args.split != 'test':
# No labels for the test set, so NLL would be invalid
progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(gold_dict, ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
# Log results (except for test set, since it does not come with labels)
if args.split != 'test':
results = util.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
results_list = [('NLL', nll_meter.avg), ('F1', results['F1']),
('EM', results['EM'])]
if args.use_squad_v2:
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in results.items())
log.info(f'{args.split.title()} {results_str}')
# Log to TensorBoard
tbx = SummaryWriter(args.save_dir)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=0,
split=args.split,
num_visuals=args.num_visuals)
# Write submission file
sub_path = join(args.save_dir, args.split + '_' + args.sub_file)
log.info(f'Writing submission file to {sub_path}...')
with open(sub_path, 'w', newline='', encoding='utf-8') as csv_fh:
csv_writer = csv.writer(csv_fh, delimiter=',')
csv_writer.writerow(['Id', 'Predicted'])
for uuid in sorted(sub_dict):
csv_writer.writerow([uuid, sub_dict[uuid]])
def main(args):
# Set up logging
args.save_dir = utilz.get_save_dir(args.save_dir,
args.name,
training=False)
log = utilz.get_logger(args.save_dir, args.name)
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
device, gpu_ids = utilz.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# Get embeddings
log.info('Loading embeddings...')
word_vectors = utilz.torch_from_json(args.word_emb_file)
# Get model
log.info('Building model...')
#model = BiDAF(word_vectors=word_vectors,
# hidden_size=args.hidden_size)
## QANet
# load word vectors
wv_tensor = torch.FloatTensor(
np.array(pickle_load_large_file('./data/processed/SQuAD/word_emb.pkl'),
dtype=np.float32))
cv_tensor = torch.FloatTensor(
np.array(pickle_load_large_file('./data/processed/SQuAD/char_emb.pkl'),
dtype=np.float32))
wv_word2ix = pickle_load_large_file('./data/processed/SQuAD/word_dict.pkl')
# construct model
model = QANet(wv_tensor,
cv_tensor,
400,
50,
128,
num_head=8,
train_cemb=False,
pad=wv_word2ix["<PAD>"])
## QANet End
# model = nn.DataParallel(model, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path}...')
model = utilz.load_model(model, args.load_path, gpu_ids, return_step=False)
#model = model.to(device)
#ema = EMA(0.9999)
#ema.assign(model)
model = model.to(device)
model.eval()
# Get data loader
log.info('Building dataset...')
record_file = vars(args)[f'{args.split}_record_file']
dataset = SQuAD(record_file, args.use_squad_v2)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Evaluate
log.info(f'Evaluating on {args.split} split...')
nll_meter = utilz.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = vars(args)[f'{args.split}_eval_file']
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
cc_idxs = cc_idxs.to(device)
qc_idxs = qc_idxs.to(device)
batch_size = cw_idxs.size(0)
# Forward
p1, p2 = model(cw_idxs, cc_idxs, qw_idxs, qc_idxs)
p1 = F.softmax(p1, dim=1)
p2 = F.softmax(p2, dim=1)
y1, y2 = y1.to(device), y2.to(device)
loss1 = torch.nn.CrossEntropyLoss()(p1, y1)
loss2 = torch.nn.CrossEntropyLoss()(p2, y2)
loss = torch.mean(loss1 + loss2)
nll_meter.update(loss.item(), batch_size)
#starts, ends = utilz.discretize(p1, p2, args.max_ans_len, args.use_squad_v2)
#outer = torch.matmul(p1.unsqueeze(2), p2.unsqueeze(1))
#for j in range(outer.size()[0]):
# outer[j] = torch.triu(outer[j])
# outer[j] = torch.tril(outer[j], self.args.ans_limit)
#a1, _ = torch.max(outer, dim=2)
#a2, _ = torch.max(outer, dim=1)
#ymin = torch.argmax(a1, dim=1)
#iymax = torch.argmax(a2, dim=1)
#idx2pred, uuid2pred = utilz.convert_tokens(gold_dict, ids.tolist(), ymin.tolist(), ymax.tolist(),args.use_squad_v2)
#idx2pred = {}
#uuid2pred = {}
#for qid, p1, p2 in zip(ids.tolist(), starts.tolist(), ends.tolist()):
# context = gold_dict[str(qid)]["context"]
# spans = gold_dict[str(qid)]["spans"]
# uuid = gold_dict[str(qid)]["uuid"]
# if args.use_squad_v2 and (p1 == 0 or p2 == 0):
# idx2pred[str(qid)] = ''
# uuid2pred[uuid] = ''
# else:
# p1, p2 = p1-1, p2-1
# start_idx = spans[p1][0]
# end_idx = spans[p2][1]
# idx2pred[str(qid)] = context[start_idx: end_idx]
# uuid2pred[uuid] = context[start_idx: end_idx]
# Get F1 and EM scores
starts, ends = utilz.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
if args.split != 'test':
# No labels for the test set, so NLL would be invalid
progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = utilz.convert_tokens(gold_dict, ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
# Log results (except for test set, since it does not come with labels)
if args.split != 'test':
results = utilz.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
results_list = [('NLL', nll_meter.avg), ('F1', results['F1']),
('EM', results['EM'])]
if args.use_squad_v2:
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in results.items())
log.info(f'{args.split.title()} {results_str}')
# Log to TensorBoard
tbx = SummaryWriter(args.save_dir)
utilz.visualize(tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=0,
split=args.split,
num_visuals=args.num_visuals)
# Write submission file
sub_path = join(args.save_dir, args.split + '_' + args.sub_file)
log.info(f'Writing submission file to {sub_path}...')
with open(sub_path, 'w', newline='', encoding='utf-8') as csv_fh:
csv_writer = csv.writer(csv_fh, delimiter=',')
csv_writer.writerow(['Id', 'Predicted'])
for uuid in sorted(sub_dict):
csv_writer.writerow([uuid, sub_dict[uuid]])
def evaluate(model, data_loader, device, eval_file, max_len, use_squad_v2, model_name="", a1=0.5, a2=0.5):
meter = util.AverageMeter()
# setup losses
bceLoss = nn.BCELoss()
ceLoss = nn.CrossEntropyLoss()
model.eval()
pred_dict = {}
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(data_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_loader:
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
cc_idxs = cc_idxs.to(device)
qc_idxs = qc_idxs.to(device)
batch_size = cw_idxs.size(0)
# Forward
y1, y2 = y1.to(device), y2.to(device)
if model_name == 'sketchy':
yi = model(cw_idxs, qw_idxs, cc_idxs, qc_idxs)
loss = bceLoss(yi, torch.where(
y1 == 0, 0, 1).type(torch.FloatTensor))
meter.update(loss.item(), batch_size)
starts, ends = [[0 if yi[i] == 0 else 1 for i, y in enumerate(
y1)], [0 if yi[i] == 0 else 2 for i, y in enumerate(y2)]]
elif model_name == 'intensive':
yi, log_p1, log_p2 = model(
cw_idxs, qw_idxs, cc_idxs, qc_idxs)
loss = a1 * bceLoss(yi, torch.where(y1 == 0, 0, 1).type(
torch.FloatTensor)) + a2 * (F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2))
#loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1 = log_p1.exp()
p2 = log_p2.exp()
# print(p1[0,:])
# print(p1)
# print(p2[0,:])
# print(p2)
starts, ends = util.discretize(p1, p2, max_len, use_squad_v2)
starts, ends = starts.tolist(), ends.tolist()
elif model_name == 'retro':
log_p1, log_p2 = model(cw_idxs, qw_idxs, cc_idxs, qc_idxs)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
meter.update(loss.item(), batch_size)
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, max_len, use_squad_v2)
starts, ends = starts.tolist(), ends.tolist()
else:
raise ValueError(
'invalid --model_name, sketchy or intensive required')
print("starts: ", starts, "Truth", y1)
print("ends: ", ends, "Truth: ", y2)
# Log info
progress_bar.update(batch_size)
progress_bar.set_postfix(loss_calc=meter.avg)
preds, _ = util.convert_tokens(gold_dict,
ids.tolist(),
starts,
ends,
use_squad_v2)
pred_dict.update(preds)
model.train()
results = util.eval_dicts(gold_dict, pred_dict, use_squad_v2)
results_list = [('Loss', meter.avg),
('F1', results['F1']),
('EM', results['EM'])]
if use_squad_v2:
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
return results, pred_dict
def main(args):
# Set up logging
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=False)
#log = util.get_logger(args.save_dir, args.name)
#log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# Get embeddings
print('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
# Get model
print('Building model...')
model = BiDAF(word_vectors=word_vectors, hidden_size=args.hidden_size)
model = nn.DataParallel(model, gpu_ids)
print(f'Loading checkpoint from {args.load_path}...')
model = util.load_model(model, args.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
# Get data loader
print('Building dataset...')
#record_file = vars(args)[f'{args.split}_record_file']
dataset = SQuAD("./data/my_test.npz", args.use_squad_v2)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Evaluate
print(f'Evaluating on {args.split} split...')
nll_meter = util.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
#eval_file = vars(args)[f'{args.split}_eval_file']
with open("./data/my_test_eval.json", 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_loader:
print("viewing the dataset")
print(cw_idxs, cc_idxs, qw_idxs, qc_idxs)
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
# Forward
log_p1, log_p2 = model(cw_idxs, qw_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
#if args.split != 'test':
# No labels for the test set, so NLL would be invalid
#progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(gold_dict, ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
print("my evaluation ....")
for el in pred_dict:
print(el, pred_dict[el])
for el in sub_dict:
print(el, sub_dict[el])
def get_bpe(args):
bpe = BPE()
with open(args.bpe_file, "r") as file:
bpe.load_state_dict(json_load(file))
add_special_tokens(args)
return bpe
from args import get_train_args
from collections import OrderedDict
from json import dumps
from models import BiDAF
from tensorboardX import SummaryWriter
from tqdm import tqdm
from ujson import load as json_load
from util import collate_fn, SQuAD
starter_gold_file = "/Users/zhangyue/Desktop/starter_gold.json"
starter_pred_file = "/Users/zhangyue/Desktop/starter_pred.json"
bert_pred_file = "/Users/zhangyue/Desktop/All/18-19_Winter/CS224N/Project/squad_cs224n/bert_train/bert_eval/predictions.json"
with open(starter_gold_file, 'r') as fh:
starter_gold_dict = json_load(fh)
with open(starter_pred_file, 'r') as fh:
starter_pred_dict = json_load(fh)
with open(bert_pred_file, 'r') as fh:
bert_pred_file = json_load(fh)
bert_gold_dict = {}
gold_dict = {}
pred_dict = {}
sub_dict = {}
# Use uuid as dictionary key to construct the gold_dict for bert
for key, value in starter_gold_dict.items():
bert_gold_dict[value['uuid']] = value
def main(args):
# Load TF-IDF from pickle
scorer = TFIDF([])
scorer.get_from_pickle()
# Set up logging
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=False)
log = util.get_logger(args.save_dir, args.name)
log.info('Args: {}'.format(dumps(vars(args), indent=4, sort_keys=True)))
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
# Get data loader
log.info('Building dataset...')
record_file = vars(args)['{}_record_file'.format(args.split)]
dataset = SQuAD(record_file, args.use_squad_v2)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Get model
log.info('Building model...')
model = BiDAF(word_vectors=word_vectors,
char_vocab_size= 1376,
hidden_size=args.hidden_size)
model = nn.DataParallel(model, gpu_ids)
log.info('Loading checkpoint from {}...'.format(args.load_path))
model = util.load_model(model, args.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
# Evaluate
log.info('Evaluating on {} split...'.format(args.split))
nll_meter = util.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = vars(args)['{}_eval_file'.format(args.split)]
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
# Forward
log_p1, log_p2 = model(cw_idxs, qw_idxs, cc_idxs,qc_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, args.max_ans_len, args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
if args.split != 'test':
# No labels for the test set, so NLL would be invalid
progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(gold_dict,
ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
if (args.use_tfidf):
# Apply TF-IDF filtering to pred_dict
tf_idf_threshold = 2
tf_idf_common_threshold = 1
for key, value in pred_dict.items():
if value != "":
tf_idf_score = scorer.normalized_additive_idf_ignore_common_words(
value, threshold_frequency=tf_idf_common_threshold)
if tf_idf_score < tf_idf_threshold:
pred_dict[key] = ''
pass
# print ("pred_dict: {}, pruned".format(tf_idf_score))
else:
pass
# print ("pred_dict: {}, kept".format(tf_idf_score))
# Log results (except for test set, since it does not come with labels)
if args.split != 'test':
results = util.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
results_list = [('NLL', nll_meter.avg),
('F1', results['F1']),
('EM', results['EM'])]
if args.use_squad_v2:
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
# Log to console
results_str = ', '.join('{}: {:05.2f}'.format(k, v)
for k, v in results.items())
log.info('{} {}'.format(args.split.title(), results_str))
# Log to TensorBoard
tbx = SummaryWriter(args.save_dir)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=0,
split=args.split,
num_visuals=args.num_visuals)
# Write submission file
sub_path = join(args.save_dir, args.split + '_' + args.sub_file)
log.info('Writing submission file to {}...'.format(sub_path))
with open(sub_path, 'w') as csv_fh:
csv_writer = csv.writer(csv_fh, delimiter=',')
csv_writer.writerow(['Id', 'Predicted'])
for uuid in sorted(sub_dict):
csv_writer.writerow([uuid, sub_dict[uuid]])
def load_eval_file(args, eval_file):
with open(preprocessed_path(eval_file, args.data_dir, args.dataset),
'r') as fh:
from ujson import load as json_load
return json_load(fh)
def eval_results(all_examples, all_features, all_results, eval_gold_file,
n_best_size, max_answer_length, do_lower_case,
verbose_logging, version_2_with_negative,
null_score_diff_threshold):
"""Write final predictions to the json file and log-odds of null if needed."""
example_index_to_features = collections.defaultdict(list)
for feature in all_features:
example_index_to_features[feature.example_index].append(feature)
unique_id_to_result = {}
for result in all_results:
unique_id_to_result[result.unique_id] = result
_PrelimPrediction = collections.namedtuple( # pylint: disable=invalid-name
"PrelimPrediction", [
"feature_index", "start_index", "end_index", "start_logit",
"end_logit"
])
all_predictions = collections.OrderedDict()
all_nbest_json = collections.OrderedDict()
scores_diff_json = collections.OrderedDict()
for (example_index, example) in enumerate(all_examples):
features = example_index_to_features[example_index]
prelim_predictions = []
# keep track of the minimum score of null start+end of position 0
score_null = 1000000 # large and positive
min_null_feature_index = 0 # the paragraph slice with min mull score
null_start_logit = 0 # the start logit at the slice with min null score
null_end_logit = 0 # the end logit at the slice with min null score
for (feature_index, feature) in enumerate(features):
result = unique_id_to_result[feature.unique_id]
start_indexes = _get_best_indexes(result.start_logits, n_best_size)
end_indexes = _get_best_indexes(result.end_logits, n_best_size)
# if we could have irrelevant answers, get the min score of irrelevant
if version_2_with_negative:
feature_null_score = result.start_logits[
0] + result.end_logits[0]
if feature_null_score < score_null:
score_null = feature_null_score
min_null_feature_index = feature_index
null_start_logit = result.start_logits[0]
null_end_logit = result.end_logits[0]
for start_index in start_indexes:
for end_index in end_indexes:
# We could hypothetically create invalid predictions, e.g., predict
# that the start of the span is in the question. We throw out all
# invalid predictions.
if start_index >= len(feature.tokens):
continue
if end_index >= len(feature.tokens):
continue
if start_index not in feature.token_to_orig_map:
continue
if end_index not in feature.token_to_orig_map:
continue
if not feature.token_is_max_context.get(
start_index, False):
continue
if end_index < start_index:
continue
length = end_index - start_index + 1
if length > max_answer_length:
continue
prelim_predictions.append(
_PrelimPrediction(
feature_index=feature_index,
start_index=start_index,
end_index=end_index,
start_logit=result.start_logits[start_index],
end_logit=result.end_logits[end_index]))
if version_2_with_negative:
prelim_predictions.append(
_PrelimPrediction(feature_index=min_null_feature_index,
start_index=0,
end_index=0,
start_logit=null_start_logit,
end_logit=null_end_logit))
prelim_predictions = sorted(prelim_predictions,
key=lambda x:
(x.start_logit + x.end_logit),
reverse=True)
_NbestPrediction = collections.namedtuple( # pylint: disable=invalid-name
"NbestPrediction", ["text", "start_logit", "end_logit"])
seen_predictions = {}
nbest = []
for pred in prelim_predictions:
if len(nbest) >= n_best_size:
break
feature = features[pred.feature_index]
if pred.start_index > 0: # this is a non-null prediction
tok_tokens = feature.tokens[pred.start_index:(pred.end_index +
1)]
orig_doc_start = feature.token_to_orig_map[pred.start_index]
orig_doc_end = feature.token_to_orig_map[pred.end_index]
orig_tokens = example.doc_tokens[orig_doc_start:(orig_doc_end +
1)]
tok_text = " ".join(tok_tokens)
# De-tokenize WordPieces that have been split off.
tok_text = tok_text.replace(" ##", "")
tok_text = tok_text.replace("##", "")
# Clean whitespace
tok_text = tok_text.strip()
tok_text = " ".join(tok_text.split())
orig_text = " ".join(orig_tokens)
final_text = get_final_text(tok_text, orig_text, do_lower_case,
verbose_logging)
if final_text in seen_predictions:
continue
seen_predictions[final_text] = True
else:
final_text = ""
seen_predictions[final_text] = True
nbest.append(
_NbestPrediction(text=final_text,
start_logit=pred.start_logit,
end_logit=pred.end_logit))
# if we didn't include the empty option in the n-best, include it
if version_2_with_negative:
if "" not in seen_predictions:
nbest.append(
_NbestPrediction(text="",
start_logit=null_start_logit,
end_logit=null_end_logit))
# In very rare edge cases we could have no valid predictions. So we
# just create a nonce prediction in this case to avoid failure.
if not nbest:
nbest.append(
_NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))
assert len(nbest) >= 1
total_scores = []
best_non_null_entry = None
for entry in nbest:
total_scores.append(entry.start_logit + entry.end_logit)
if not best_non_null_entry:
if entry.text:
best_non_null_entry = entry
probs = _compute_softmax(total_scores)
nbest_json = []
for (i, entry) in enumerate(nbest):
output = collections.OrderedDict()
output["text"] = entry.text
output["probability"] = probs[i]
output["start_logit"] = entry.start_logit
output["end_logit"] = entry.end_logit
nbest_json.append(output)
assert len(nbest_json) >= 1
if not version_2_with_negative:
all_predictions[example.qas_id] = nbest_json[0]["text"]
else:
# predict "" iff the null score - the score of best non-null > threshold
score_diff = score_null - best_non_null_entry.start_logit - (
best_non_null_entry.end_logit)
scores_diff_json[example.qas_id] = score_diff
if score_diff > null_score_diff_threshold:
all_predictions[example.qas_id] = ""
else:
all_predictions[example.qas_id] = best_non_null_entry.text
all_nbest_json[example.qas_id] = nbest_json
with open(eval_gold_file, "r") as fh:
starter_gold_dict = json_load(fh)
bert_gold_dict = {}
gold_dict = {}
pred_dict = {}
# Use uuid as dictionary key to construct the gold_dict for bert
for key, value in starter_gold_dict.items():
bert_gold_dict[value['uuid']] = value
# Filter out bert_pred_file
bert_pred_file = all_predictions
counter = 1
for key, value in bert_pred_file.items():
if key in bert_gold_dict.keys():
pred_dict[str(counter)] = value
gold_dict[str(counter)] = bert_gold_dict[key]
counter += 1
results = eval_dicts(gold_dict, pred_dict, no_answer=True)
return results
log.info('Args: {}'.format(dumps(vars(args), indent=4, sort_keys=True)))
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
# Get model
log.info('Building model...')
model = BiDAF(word_vectors=word_vectors, hidden_size=args.hidden_size)
model = nn.DataParallel(model, gpu_ids)
log.info('Loading checkpoint from {}...'.format(args.load_path))
model = util.load_model(model, args.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
y_pred = model(cw_idxs, qn_idxs)
return y_pred
if __name__ == '__main__':
context = input("Enter the context: ")
question = input("Enter a question: ")
data = [context, question]
word2idx_dict = json_load(".data/word2idx.json")
char2idx_dict = json_load(".data/char2idx.json")
is_test = True
context_idxs, context_char_idxs, ques_idxs, ques_char_idxs = convert_to_features(
get_test_args(), data, word2idx_dict, char2idx_dict, is_test)
predict(get_test_args(), context_idxs, ques_idxs)
def main(args):
# Set up
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=False)
log = util.get_logger(args.save_dir, args.name)
device, gpu_ids = util.get_available_devices()
# Get embeddings
log.info('Loading embeddings...')
with open('data/meta.msgpack', 'rb') as f:
meta = msgpack.load(f, encoding='utf8')
embedding = torch.Tensor(meta['embedding'])
opt = vars(args)
opt['pretrained_words'] = True
opt['vocab_size'] = embedding.size(0)
opt['embedding_dim'] = embedding.size(1)
opt['pos_size'] = len(meta['vocab_tag'])
opt['ner_size'] = len(meta['vocab_ent'])
# Get model
log.info('Building model...')
model = DRQA(opt, embedding=embedding)
model = nn.DataParallel(model, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path}...')
model = util.load_model(model, args.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
BatchGen.pos_size = opt['pos_size']
BatchGen.ner_size = opt['ner_size']
# Get data loader
log.info('Building dataset...')
with open(opt['data_file'], 'rb') as f:
data = msgpack.load(f, encoding='utf8')
test = data['test']
# Evaluate
log.info(f'Evaluating on test split...')
batches = BatchGen(test,
batch_size=args.batch_size,
evaluation=True,
is_test=True,
gpu=args.cuda)
nll_meter = util.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = './data/test_eval.json'
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(test)) as progress_bar:
model.eval()
for i, batch in enumerate(batches):
# Setup for forward
inputs = [e.to(device) for e in batch[:7]]
ids = batch[-1]
# Forward
with torch.no_grad():
score_s, score_e = model(*inputs)
p1, p2 = score_s, score_e
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
# Log info
progress_bar.update(args.batch_size)
idx2pred, uuid2pred = util.convert_tokens(gold_dict, ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
# Log results (except for test set, since it does not come with labels)
if args.split != 'test':
results = util.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
results_list = [('NLL', nll_meter.avg), ('F1', results['F1']),
('EM', results['EM'])]
if args.use_squad_v2:
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in results.items())
log.info(f'{args.split.title()} {results_str}')
# Log to TensorBoard
tbx = SummaryWriter(args.save_dir)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=0,
split=args.split,
num_visuals=args.num_visuals)
# Write submission file
sub_path = join(args.save_dir, args.split + '_' + args.sub_file)
log.info(f'Writing submission file to {sub_path}...')
with open(sub_path, 'w', newline='', encoding='utf-8') as csv_fh:
csv_writer = csv.writer(csv_fh, delimiter=',')
csv_writer.writerow(['Id', 'Predicted'])
for uuid in sorted(sub_dict):
csv_writer.writerow([uuid, sub_dict[uuid]])
def preprocess_eval(eval_file):
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
return gold_dict
def main(args):
# Set up logging
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=False)
log = util.get_logger(args.save_dir, args.name)
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
char_vectors = util.torch_from_json(args.char_emb_file)
# Get model
log.info('Building model...')
nbr_model = 0
if (args.load_path_baseline):
model_baseline = Baseline(word_vectors=word_vectors, hidden_size=100)
model_baseline = nn.DataParallel(model_baseline, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path_baseline}...')
model_baseline = util.load_model(model_baseline,
args.load_path_baseline,
gpu_ids,
return_step=False)
model_baseline = model_baseline.to(device)
model_baseline.eval()
nll_meter_baseline = util.AverageMeter()
nbr_model += 1
save_prob_baseline_start = []
save_prob_baseline_end = []
if (args.load_path_bidaf):
model_bidaf = BiDAF(word_vectors=word_vectors,
char_vectors=char_vectors,
char_emb_dim=args.char_emb_dim,
hidden_size=args.hidden_size)
model_bidaf = nn.DataParallel(model_bidaf, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path_bidaf}...')
model_bidaf = util.load_model(model_bidaf,
args.load_path_bidaf,
gpu_ids,
return_step=False)
model_bidaf = model_bidaf.to(device)
model_bidaf.eval()
nll_meter_bidaf = util.AverageMeter()
nbr_model += 1
save_prob_bidaf_start = []
save_prob_bidaf_end = []
if (args.load_path_bidaf_fusion):
model_bidaf_fu = BiDAF_fus(word_vectors=word_vectors,
char_vectors=char_vectors,
char_emb_dim=args.char_emb_dim,
hidden_size=args.hidden_size)
model_bidaf_fu = nn.DataParallel(model_bidaf_fu, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path_bidaf_fusion}...')
model_bidaf_fu = util.load_model(model_bidaf_fu,
args.load_path_bidaf_fusion,
gpu_ids,
return_step=False)
model_bidaf_fu = model_bidaf_fu.to(device)
model_bidaf_fu.eval()
nll_meter_bidaf_fu = util.AverageMeter()
nbr_model += 1
save_prob_bidaf_fu_start = []
save_prob_bidaf_fu_end = []
if (args.load_path_qanet):
model_qanet = QANet(word_vectors=word_vectors,
char_vectors=char_vectors,
char_emb_dim=args.char_emb_dim,
hidden_size=args.hidden_size,
n_heads=args.n_heads,
n_conv_emb_enc=args.n_conv_emb,
n_conv_mod_enc=args.n_conv_mod,
n_emb_enc_blocks=args.n_emb_blocks,
n_mod_enc_blocks=args.n_mod_blocks,
divisor_dim_kqv=args.divisor_dim_kqv)
model_qanet = nn.DataParallel(model_qanet, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path_qanet}...')
model_qanet = util.load_model(model_qanet,
args.load_path_qanet,
gpu_ids,
return_step=False)
model_qanet = model_qanet.to(device)
model_qanet.eval()
nll_meter_qanet = util.AverageMeter()
nbr_model += 1
save_prob_qanet_start = []
save_prob_qanet_end = []
if (args.load_path_qanet_old):
model_qanet_old = QANet_old(word_vectors=word_vectors,
char_vectors=char_vectors,
device=device,
char_emb_dim=args.char_emb_dim,
hidden_size=args.hidden_size,
n_heads=args.n_heads,
n_conv_emb_enc=args.n_conv_emb,
n_conv_mod_enc=args.n_conv_mod,
n_emb_enc_blocks=args.n_emb_blocks,
n_mod_enc_blocks=args.n_mod_blocks)
model_qanet_old = nn.DataParallel(model_qanet_old, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path_qanet_old}...')
model_qanet_old = util.load_model(model_qanet_old,
args.load_path_qanet_old,
gpu_ids,
return_step=False)
model_qanet_old = model_qanet_old.to(device)
model_qanet_old.eval()
nll_meter_qanet_old = util.AverageMeter()
nbr_model += 1
save_prob_qanet_old_start = []
save_prob_qanet_old_end = []
if (args.load_path_qanet_inde):
model_qanet_inde = QANet_independant_encoder(
word_vectors=word_vectors,
char_vectors=char_vectors,
char_emb_dim=args.char_emb_dim,
hidden_size=args.hidden_size,
n_heads=args.n_heads,
n_conv_emb_enc=args.n_conv_emb,
n_conv_mod_enc=args.n_conv_mod,
n_emb_enc_blocks=args.n_emb_blocks,
n_mod_enc_blocks=args.n_mod_blocks,
divisor_dim_kqv=args.divisor_dim_kqv)
model_qanet_inde = nn.DataParallel(model_qanet_inde, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path_qanet_inde}...')
model_qanet_inde = util.load_model(model_qanet_inde,
args.load_path_qanet_inde,
gpu_ids,
return_step=False)
model_qanet_inde = model_qanet_inde.to(device)
model_qanet_inde.eval()
nll_meter_qanet_inde = util.AverageMeter()
nbr_model += 1
save_prob_qanet_inde_start = []
save_prob_qanet_inde_end = []
if (args.load_path_qanet_s_e):
model_qanet_s_e = QANet_S_E(word_vectors=word_vectors,
char_vectors=char_vectors,
char_emb_dim=args.char_emb_dim,
hidden_size=args.hidden_size,
n_heads=args.n_heads,
n_conv_emb_enc=args.n_conv_emb,
n_conv_mod_enc=args.n_conv_mod,
n_emb_enc_blocks=args.n_emb_blocks,
n_mod_enc_blocks=args.n_mod_blocks,
divisor_dim_kqv=args.divisor_dim_kqv)
model_qanet_s_e = nn.DataParallel(model_qanet_s_e, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path_qanet_s_e}...')
model_qanet_s_e = util.load_model(model_qanet_s_e,
args.load_path_qanet_s_e,
gpu_ids,
return_step=False)
model_qanet_s_e = model_qanet_s_e.to(device)
model_qanet_s_e.eval()
nll_meter_qanet_s_e = util.AverageMeter()
nbr_model += 1
save_prob_qanet_s_e_start = []
save_prob_qanet_s_e_end = []
# Get data loader
log.info('Building dataset...')
record_file = vars(args)[f'{args.split}_record_file']
dataset = SQuAD(record_file, args.use_squad_v2)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Evaluate
log.info(f'Evaluating on {args.split} split...')
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = vars(args)[f'{args.split}_eval_file']
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
cc_idxs = cc_idxs.to(device)
qc_idxs = qc_idxs.to(device)
batch_size = cw_idxs.size(0)
y1, y2 = y1.to(device), y2.to(device)
l_p1, l_p2 = [], []
# Forward
if (args.load_path_baseline):
log_p1_baseline, log_p2_baseline = model_baseline(
cw_idxs, cc_idxs)
loss_baseline = F.nll_loss(log_p1_baseline, y1) + F.nll_loss(
log_p2_baseline, y2)
nll_meter_baseline.update(loss_baseline.item(), batch_size)
l_p1 += [log_p1_baseline.exp()]
l_p2 += [log_p2_baseline.exp()]
if (args.save_probabilities):
save_prob_baseline_start += [
log_p1_baseline.exp().detach().cpu().numpy()
]
save_prob_baseline_end += [
log_p2_baseline.exp().detach().cpu().numpy()
]
if (args.load_path_qanet):
log_p1_qanet, log_p2_qanet = model_qanet(
cw_idxs, cc_idxs, qw_idxs, qc_idxs)
loss_qanet = F.nll_loss(log_p1_qanet, y1) + F.nll_loss(
log_p2_qanet, y2)
nll_meter_qanet.update(loss_qanet.item(), batch_size)
# Get F1 and EM scores
l_p1 += [log_p1_qanet.exp()]
l_p2 += [log_p2_qanet.exp()]
if (args.save_probabilities):
save_prob_qanet_start += [
log_p1_qanet.exp().detach().cpu().numpy()
]
save_prob_qanet_end += [
log_p2_qanet.exp().detach().cpu().numpy()
]
if (args.load_path_qanet_old):
log_p1_qanet_old, log_p2_qanet_old = model_qanet_old(
cw_idxs, cc_idxs, qw_idxs, qc_idxs)
loss_qanet_old = F.nll_loss(log_p1_qanet_old, y1) + F.nll_loss(
log_p2_qanet_old, y2)
nll_meter_qanet_old.update(loss_qanet_old.item(), batch_size)
# Get F1 and EM scores
l_p1 += [log_p1_qanet_old.exp()]
l_p2 += [log_p2_qanet_old.exp()]
if (args.save_probabilities):
save_prob_qanet_old_start += [
log_p1_qanet_old.exp().detach().cpu().numpy()
]
save_prob_qanet_old_end += [
log_p2_qanet_old.exp().detach().cpu().numpy()
]
if (args.load_path_qanet_inde):
log_p1_qanet_inde, log_p2_qanet_inde = model_qanet_inde(
cw_idxs, cc_idxs, qw_idxs, qc_idxs)
loss_qanet_inde = F.nll_loss(
log_p1_qanet_inde, y1) + F.nll_loss(log_p2_qanet_inde, y2)
nll_meter_qanet_inde.update(loss_qanet_inde.item(), batch_size)
# Get F1 and EM scores
l_p1 += [log_p1_qanet_inde.exp()]
l_p2 += [log_p2_qanet_inde.exp()]
if (args.save_probabilities):
save_prob_qanet_inde_start += [
log_p1_qanet_inde.exp().detach().cpu().numpy()
]
save_prob_qanet_inde_end += [
log_p2_qanet_inde.exp().detach().cpu().numpy()
]
if (args.load_path_qanet_s_e):
log_p1_qanet_s_e, log_p2_qanet_s_e = model_qanet_s_e(
cw_idxs, cc_idxs, qw_idxs, qc_idxs)
loss_qanet_s_e = F.nll_loss(log_p1_qanet_s_e, y1) + F.nll_loss(
log_p2_qanet_s_e, y2)
nll_meter_qanet_s_e.update(loss_qanet_s_e.item(), batch_size)
# Get F1 and EM scores
l_p1 += [log_p1_qanet_s_e.exp()]
l_p2 += [log_p2_qanet_s_e.exp()]
if (args.save_probabilities):
save_prob_qanet_s_e_start += [
log_p1_qanet_s_e.exp().detach().cpu().numpy()
]
save_prob_qanet_s_e_end += [
log_p2_qanet_s_e.exp().detach().cpu().numpy()
]
if (args.load_path_bidaf):
log_p1_bidaf, log_p2_bidaf = model_bidaf(
cw_idxs, cc_idxs, qw_idxs, qc_idxs)
loss_bidaf = F.nll_loss(log_p1_bidaf, y1) + F.nll_loss(
log_p2_bidaf, y2)
nll_meter_bidaf.update(loss_bidaf.item(), batch_size)
l_p1 += [log_p1_bidaf.exp()]
l_p2 += [log_p2_bidaf.exp()]
if (args.save_probabilities):
save_prob_bidaf_start += [
log_p1_bidaf.exp().detach().cpu().numpy()
]
save_prob_bidaf_end += [
log_p2_bidaf.exp().detach().cpu().numpy()
]
if (args.load_path_bidaf_fusion):
log_p1_bidaf_fu, log_p2_bidaf_fu = model_bidaf_fu(
cw_idxs, cc_idxs, qw_idxs, qc_idxs)
loss_bidaf_fu = F.nll_loss(log_p1_bidaf_fu, y1) + F.nll_loss(
log_p2_bidaf_fu, y2)
nll_meter_bidaf_fu.update(loss_bidaf_fu.item(), batch_size)
l_p1 += [log_p1_bidaf_fu.exp()]
l_p2 += [log_p2_bidaf_fu.exp()]
if (args.save_probabilities):
save_prob_bidaf_fu_start += [
log_p1_bidaf_fu.exp().detach().cpu().numpy()
]
save_prob_bidaf_fu_end += [
log_p2_bidaf_fu.exp().detach().cpu().numpy()
]
p1, p2 = l_p1[0], l_p2[0]
for i in range(1, nbr_model):
p1 += l_p1[i]
p2 += l_p2[i]
p1 /= nbr_model
p2 /= nbr_model
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
if args.split != 'test':
# No labels for the test set, so NLL would be invalid
if (args.load_path_qanet):
progress_bar.set_postfix(NLL=nll_meter_qanet.avg)
elif (args.load_path_bidaf):
progress_bar.set_postfix(NLL=nll_meter_bidaf.avg)
elif (args.load_path_bidaf_fusion):
progress_bar.set_postfix(NLL=nll_meter_bidaf_fu.avg)
elif (args.load_path_qanet_old):
progress_bar.set_postfix(NLL=nll_meter_qanet_old.avg)
elif (args.load_path_qanet_inde):
progress_bar.set_postfix(NLL=nll_meter_qanet_inde.avg)
elif (args.load_path_qanet_s_e):
progress_bar.set_postfix(NLL=nll_meter_qanet_s_e.avg)
else:
progress_bar.set_postfix(NLL=nll_meter_baseline.avg)
idx2pred, uuid2pred = util.convert_tokens(gold_dict, ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
if (args.save_probabilities):
if (args.load_path_baseline):
with open(args.save_dir + "/probs_start", "wb") as fp: #Pickling
pickle.dump(save_prob_baseline_start, fp)
with open(args.save_dir + "/probs_end", "wb") as fp: #Pickling
pickle.dump(save_prob_baseline_end, fp)
if (args.load_path_bidaf):
with open(args.save_dir + "/probs_start", "wb") as fp: #Pickling
pickle.dump(save_prob_bidaf_start, fp)
with open(args.save_dir + "/probs_end", "wb") as fp: #Pickling
pickle.dump(save_prob_bidaf_end, fp)
if (args.load_path_bidaf_fusion):
with open(args.save_dir + "/probs_start", "wb") as fp: #Pickling
pickle.dump(save_prob_bidaf_fu_start, fp)
with open(args.save_dir + "/probs_end", "wb") as fp: #Pickling
pickle.dump(save_prob_bidaf_fu_end, fp)
if (args.load_path_qanet):
with open(args.save_dir + "/probs_start", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_start, fp)
with open(args.save_dir + "/probs_end", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_end, fp)
if (args.load_path_qanet_old):
with open(args.save_dir + "/probs_start", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_old_start, fp)
with open(args.save_dir + "/probs_end", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_old_end, fp)
if (args.load_path_qanet_inde):
with open(args.save_dir + "/probs_start", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_inde_start, fp)
with open(args.save_dir + "/probs_end", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_inde_end, fp)
if (args.load_path_qanet_s_e):
with open(args.save_dir + "/probs_start", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_s_e_start, fp)
with open(args.save_dir + "/probs_end", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_s_e_end, fp)
# Log results (except for test set, since it does not come with labels)
if args.split != 'test':
results = util.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
if (args.load_path_qanet):
meter_avg = nll_meter_qanet.avg
elif (args.load_path_bidaf):
meter_avg = nll_meter_bidaf.avg
elif (args.load_path_bidaf_fusion):
meter_avg = nll_meter_bidaf_fu.avg
elif (args.load_path_qanet_inde):
meter_avg = nll_meter_qanet_inde.avg
elif (args.load_path_qanet_s_e):
meter_avg = nll_meter_qanet_s_e.avg
elif (args.load_path_qanet_old):
meter_avg = nll_meter_qanet_old.avg
else:
meter_avg = nll_meter_baseline.avg
results_list = [('NLL', meter_avg), ('F1', results['F1']),
('EM', results['EM'])]
if args.use_squad_v2:
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in results.items())
log.info(f'{args.split.title()} {results_str}')
# Log to TensorBoard
tbx = SummaryWriter(args.save_dir)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=0,
split=args.split,
num_visuals=args.num_visuals)
# Write submission file
sub_path = join(args.save_dir, args.split + '_' + args.sub_file)
log.info(f'Writing submission file to {sub_path}...')
with open(sub_path, 'w', newline='', encoding='utf-8') as csv_fh:
csv_writer = csv.writer(csv_fh, delimiter=',')
csv_writer.writerow(['Id', 'Predicted'])
for uuid in sorted(sub_dict):
csv_writer.writerow([uuid, sub_dict[uuid]])
def test(args):
# Set up logging
log = util.get_logger(args.save_dir, args.name)
log.info(f"Args: {dumps(vars(args), indent=4, sort_keys=True)}")
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# Get embeddings
log.info("Loading embeddings...")
word_vectors = util.torch_from_json(args.word_emb_file)
# TODO: Hardcode padding_idx
padding_idx = 0
# Get model
log.info("Building model...")
model = WordTransformerQA(
dim=args.dim,
n_heads=args.n_heads,
ff_dim=args.ff_dim,
activation=args.activation,
dropout=args.dropout,
attn_dropout=args.attn_dropout,
act_dropout=args.act_dropout,
n_layers=args.n_layers,
max_positions=args.max_positions,
word_vectors=word_vectors,
)
model = nn.DataParallel(model, gpu_ids)
log.info(f"Loading checkpoint from {args.load_path}...")
model = util.load_model(model, args.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
# Get data loader
log.info("Building dataset...")
record_file = vars(args)[f"{args.split}_record_file"]
dataset = SQuAD(record_file, args.use_squad_v2)
data_loader = data.DataLoader(
dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn,
)
# Evaluate
log.info(f"Evaluating on {args.split} split...")
nll_meter = stats.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = vars(args)[f"{args.split}_eval_file"]
with open(eval_file, "r") as fh:
gold_dict = json_load(fh)
with torch.no_grad(), tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_loader:
batch_size = cw_idxs.size(0)
_, loss_val, scores = forward(cw_idxs, qw_idxs, y1, y2,
padding_idx, args, device, model)
nll_meter.update(loss_val, batch_size)
# Get F1 and EM scores
p1, p2 = model.module.get_prob(scores).split(1, dim=-1)
p1, p2 = p1.squeeze(-1), p2.squeeze(-1)
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
if args.split != "test":
# No labels for the test set, so NLL would be invalid
progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(
gold_dict,
ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2,
)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
# Log results (except for test set, since it does not come with labels)
if args.split != "test":
results = eval.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
results_list = [
("NLL", nll_meter.avg),
("F1", results["F1"]),
("EM", results["EM"]),
]
if args.use_squad_v2:
results_list.append(("AvNA", results["AvNA"]))
results = OrderedDict(results_list)
# Log to console
results_str = ", ".join(f"{k}: {v:05.2f}" for k, v in results.items())
log.info(f"{args.split.title()} {results_str}")
# Log to TensorBoard
tbx = SummaryWriter(args.save_dir)
util.visualize(
tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=0,
split=args.split,
num_visuals=args.num_visuals,
)
# Write submission file
if args.split == "dev":
sub_path = join(args.save_dir, "val" + "_" + args.sub_file)
else:
sub_path = join(args.save_dir, args.split + "_" + args.sub_file)
log.info(f"Writing submission file to {sub_path}...")
eval.write_submission(sub_path, sub_dict)
def main(args):
# Set up logging and devices
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=True)
log = util.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir)
device, args.gpu_ids = util.get_available_devices()
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
args.batch_size *= max(1, len(args.gpu_ids))
# Set random seed
log.info(f'Using random seed {args.seed}...')
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
# Get size of char vocab
with open(args.char2idx_file, 'r') as fh:
char_vocab_size = len(json_load(fh))
# Get model
log.info('Building model...')
model = QANet(word_vectors=word_vectors,
hidden_size=args.hidden_size,
char_vocab_size = char_vocab_size,
char_emb_size = args.char_emb_size,
word_char_emb_size = args.word_char_emb_size,
drop_prob=args.drop_prob,
num_blocks_embd = args.num_blocks_embd,
num_conv_embd = args.num_conv_embd,
kernel_size = args.kernel_size,
num_heads = args.num_heads,
num_blocks_model = args.num_blocks_model,
num_conv_model = args.num_conv_model,
dropout_char = args.dropout_char,
dropout_word = args.dropout_word,
survival_prob = args.survival_prob)
model = nn.DataParallel(model, args.gpu_ids)
if args.load_path:
log.info(f'Loading checkpoint from {args.load_path}...')
model, step = util.load_model(model, args.load_path, args.gpu_ids)
else:
step = 0
model = model.to(device)
model.train()
ema = util.EMA(model, args.ema_decay)
# Get saver
saver = util.CheckpointSaver(args.save_dir,
max_checkpoints=args.max_checkpoints,
metric_name=args.metric_name,
maximize_metric=args.maximize_metric,
log=log)
# Get optimizer and scheduler
#params = filter(lambda param: param.requires_grad, model.parameters())
optimizer = optim.Adam(lr=1, betas=(args.beta1, args.beta2), eps=args.adam_eps, weight_decay=args.l2_wd, params=model.parameters())
cr = args.lr / math.log2(args.warm_up)
scheduler = optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda ee: cr * math.log2(ee + 1) if ee < args.warm_up else args.lr)
# Get data loader
log.info('Building dataset...')
train_dataset = SQuAD(args.train_record_file, args.use_squad_v2)
train_loader = data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
dev_dataset = SQuAD(args.dev_record_file, args.use_squad_v2)
dev_loader = data.DataLoader(dev_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Train
log.info('Training...')
steps_till_eval = args.eval_steps
epoch = step // len(train_dataset)
torch.autograd.set_detect_anomaly(True)
while epoch != args.num_epochs:
epoch += 1
log.info(f'Starting epoch {epoch}...')
with torch.enable_grad(), \
tqdm(total=len(train_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in train_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
cc_idxs=cc_idxs.to(device)
qc_idxs=qc_idxs.to(device)
batch_size = cw_idxs.size(0)
optimizer.zero_grad()
# Forward
log_p1, log_p2 = model(cw_idxs, qw_idxs, cc_idxs, qc_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
loss_val = loss.item()
# Backward
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step(step // batch_size)
ema(model, step // batch_size)
# Log info
step += batch_size
progress_bar.update(batch_size)
progress_bar.set_postfix(epoch=epoch,
NLL=loss_val)
tbx.add_scalar('train/NLL', loss_val, step)
tbx.add_scalar('train/LR',
optimizer.param_groups[0]['lr'],
step)
steps_till_eval -= batch_size
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
log.info(f'Evaluating at step {step}...')
ema.assign(model)
results, pred_dict = evaluate(model, dev_loader, device,
args.dev_eval_file,
args.max_ans_len,
args.use_squad_v2)
saver.save(step, model, results[args.metric_name], device)
ema.resume(model)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in results.items())
log.info(f'Dev {results_str}')
# Log to TensorBoard
log.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar(f'dev/{k}', v, step)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.dev_eval_file,
step=step,
split='dev',
num_visuals=args.num_visuals)
def main(args):
# Set up logging
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=False)
log = util.get_logger(args.save_dir, args.name)
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
models = {}
if args.use_ensemble:
total_models = 0
for model_name in ['bidaf', 'bidafextra', 'fusionnet']:
models_list = []
for model_file in glob.glob(
f'{args.load_path}/{model_name}-*/{args.ensemble_models}'):
# Get model
log.info('Building model...')
if model_name == 'bidaf':
model = BiDAF(word_vectors=word_vectors,
hidden_size=args.hidden_size)
elif model_name == 'bidafextra':
model = BiDAFExtra(word_vectors=word_vectors, args=args)
elif model_name == 'fusionnet':
model = FusionNet(word_vectors=word_vectors, args=args)
model = nn.DataParallel(model, gpu_ids)
log.info(f'Loading checkpoint from {model_file}...')
model = util.load_model(model,
model_file,
gpu_ids,
return_step=False)
# Load each model on CPU (have plenty of RAM ...)
model = model.cpu()
model.eval()
models_list.append(model)
models[model_name] = models_list
total_models += len(models_list)
log.info(f'Using an ensemble of {total_models} models')
else:
device, gpu_ids = util.get_available_devices()
# Get model
log.info('Building model...')
if args.model == 'bidaf':
model = BiDAF(word_vectors=word_vectors,
hidden_size=args.hidden_size)
elif args.model == 'bidafextra':
model = BiDAFExtra(word_vectors=word_vectors, args=args)
elif args.model == 'fusionnet':
model = FusionNet(word_vectors=word_vectors, args=args)
model = nn.DataParallel(model, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path}...')
model = util.load_model(model,
args.load_path,
gpu_ids,
return_step=False)
model = model.to(device)
model.eval()
models[args.model] = [model]
# Get data loader
log.info('Building dataset...')
record_file = vars(args)[f'{args.split}_record_file']
dataset = SQuAD(record_file, args)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Evaluate
log.info(f'Evaluating on {args.split} split...')
nll_meter = util.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = vars(args)[f'{args.split}_eval_file']
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, cw_pos, cw_ner, cw_freq, cqw_extra, y1, y2, ids in data_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
p1s = []
p2s = []
for model_name in models:
for model in models[model_name]:
# Move model to GPU to evaluate
model = model.to(device)
# Forward
if model_name == 'bidaf':
log_p1, log_p2 = model.to(device)(cw_idxs, qw_idxs)
else:
log_p1, log_p2 = model.to(device)(cw_idxs, qw_idxs,
cw_pos, cw_ner,
cw_freq, cqw_extra)
log_p1, log_p2 = log_p1.cpu(), log_p2.cpu()
if not args.use_ensemble:
y1, y2 = y1.to(device), y2.to(device)
log_p1, log_p2 = log_p1.to(device), log_p2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
nll_meter.update(loss.item(), batch_size)
# Move model back to CPU to release GPU memory
model = model.cpu()
# Get F1 and EM scores
p1, p2 = log_p1.exp().unsqueeze(
-1).cpu(), log_p2.exp().unsqueeze(-1).cpu()
p1s.append(p1), p2s.append(p2)
best_ps = torch.max(
torch.cat([
torch.cat(p1s, -1).unsqueeze(-1),
torch.cat(p2s, -1).unsqueeze(-1)
], -1), -2)[0]
p1, p2 = best_ps[:, :, 0], best_ps[:, :, 1]
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
if args.split != 'test':
# No labels for the test set, so NLL would be invalid
progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(gold_dict, ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
# Log results (except for test set, since it does not come with labels)
if args.split != 'test':
results = util.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
results_list = [('NLL', nll_meter.avg), ('F1', results['F1']),
('EM', results['EM'])]
if args.use_squad_v2:
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in results.items())
log.info(f'{args.split.title()} {results_str}')
# Log to TensorBoard
tbx = SummaryWriter(args.save_dir)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=0,
split=args.split,
num_visuals=args.num_visuals)
# Write submission file
sub_path = join(args.save_dir, args.split + '_' + args.sub_file)
log.info(f'Writing submission file to {sub_path}...')
with open(sub_path, 'w', newline='', encoding='utf-8') as csv_fh:
csv_writer = csv.writer(csv_fh, delimiter=',')
csv_writer.writerow(['Id', 'Predicted'])
for uuid in sorted(sub_dict):
csv_writer.writerow([uuid, sub_dict[uuid]])
def test(args):
# Set up logging
log = util.get_logger(args.save_dir, args.name)
log.info(f"Args: {dumps(vars(args), indent=4, sort_keys=True)}")
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# Get embeddings
log.info("Loading embeddings...")
word_vectors = util.torch_from_json(args.word_emb_file)
# Get model
log.info("Building model...")
model = BiDAF(
word_vectors=word_vectors,
hidden_size=args.hidden_size,
use_glove=args.use_glove,
)
model = nn.DataParallel(model, gpu_ids)
log.info(f"Loading checkpoint from {args.load_path}...")
model = util.load_model(model, args.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
# Get data loader
log.info("Building dataset...")
record_file = vars(args)[f"{args.split}_record_file"]
dataset = SQuAD(record_file, args.use_squad_v2)
data_loader = data.DataLoader(
dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn,
)
# Evaluate
log.info(f"Evaluating on {args.split} split...")
nll_meter = stats.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = vars(args)[f"{args.split}_eval_file"]
with open(eval_file, "r") as fh:
gold_dict = json_load(fh)
with torch.no_grad(), tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
# Forward
log_p1, log_p2 = model(cw_idxs, qw_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
if args.split != "test":
# No labels for the test set, so NLL would be invalid
progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(
gold_dict,
ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2,
)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
# Log results (except for test set, since it does not come with labels)
if args.split != "test":
results = {"NLL": nll_meter.avg}
results.update(eval.eval_dicts(gold_dict, pred_dict,
args.use_squad_v2))
# Log to console
results_str = ", ".join(f"{k}: {v:05.2f}" for k, v in results.items())
log.info(f"{args.split.title()} {results_str}")
# Log to TensorBoard
tbx = SummaryWriter(args.save_dir)
util.visualize(
tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=0,
split=args.split,
num_visuals=args.num_visuals,
)
# Write submission file
if args.split == "dev":
sub_path = join(args.save_dir, "val" + "_" + args.sub_file)
else:
sub_path = join(args.save_dir, args.split + "_" + args.sub_file)
log.info(f"Writing submission file to {sub_path}...")
eval.write_submission(sub_path, sub_dict)