def AverageProbs(args, p_start, p_end, id_quest, gold_dict):
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
nbr_model = len(p_start)
nbr_batch = len(p_start[0])
for j in range(nbr_batch):
p1 = np.zeros(p_start[0][j].shape)
p2 = np.zeros(p_start[0][j].shape)
for model in range(nbr_model):
p1 += p_start[model][j]
p2 += p_end[model][j]
p1 /= nbr_model
p2 /= nbr_model
ids = id_quest[j]
p1 = torch.from_numpy(p1).float().cuda()
p2 = torch.from_numpy(p2).float().cuda()
ids = ids.cuda()
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
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)
results = util.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
results_list = [('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())
print(f'{args.split.title()} {results_str}')
# Write submission file
sub_path = join("../save/test/" + split + "/submission.csv")
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):
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)
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, 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 MaxProbs(args, p_start, p_end, id_quest, gold_dict):
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
nbr_model = len(p_start)
nbr_batch = len(p_start[0])
for j in range(nbr_batch):
prob = torch.zeros(nbr_model, p_start[0][j].shape[0])
start = -1 * torch.ones(
(nbr_model, p_start[0][j].shape[0]), dtype=torch.long)
end = -1 * torch.ones(
(nbr_model, p_start[0][j].shape[0]), dtype=torch.long)
ids = id_quest[j]
for model in range(nbr_model):
cur_p1 = torch.from_numpy(p_start[model][j]).float().cuda()
cur_p2 = torch.from_numpy(p_end[model][j]).float().cuda()
ids = ids.cuda()
s_model, e_model, probs = discretizeAndMax(cur_p1, cur_p2,
args.max_ans_len,
args.use_squad_v2)
start[model, :] = s_model
end[model, :] = e_model
prob[model, :] = probs
index_s_e = torch.argmax(prob, dim=0)
starts = torch.gather(start, 0, index_s_e.unsqueeze(0)).squeeze()
ends = torch.gather(end, 0, index_s_e.unsqueeze(0)).squeeze()
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)
results = util.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
results_list = [('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())
print(f'{args.split.title()} {results_str}')
def evaluate(self,
model,
data_loader,
data_dict,
return_preds=False,
split='validation'):
device = self.device
model.eval()
pred_dict = {}
all_start_logits = []
all_end_logits = []
with torch.no_grad(), \
tqdm(total=len(data_loader.dataset)) as progress_bar:
for batch in data_loader:
# Setup for forward
input_ids = batch['input_ids'].to(device)
attention_mask = batch['attention_mask'].to(device)
batch_size = len(input_ids)
outputs = model(input_ids, attention_mask=attention_mask)
# Forward
start_logits, end_logits = outputs.start_logits, outputs.end_logits
# TODO: compute loss
all_start_logits.append(start_logits)
all_end_logits.append(end_logits)
progress_bar.update(batch_size)
# Get F1 and EM scores
start_logits = torch.cat(all_start_logits).cpu().numpy()
end_logits = torch.cat(all_end_logits).cpu().numpy()
preds = util.postprocess_qa_predictions(data_dict,
data_loader.dataset.encodings,
(start_logits, end_logits))
if split == 'validation':
results = util.eval_dicts(data_dict, preds)
results_list = [('F1', results['F1']), ('EM', results['EM'])]
wandb.log({"F1": results["F1"]})
wandb.log({"EM": results["EM"]})
else:
results_list = [('F1', -1.0), ('EM', -1.0)]
results = OrderedDict(results_list)
if return_preds:
return preds, results
return results
def evaluate(model, eval_loader, eval_size, gold_dict):
pred_dict = {}
with torch.no_grad(), tqdm(total=eval_size) as progress_bar:
nll_meter = util.AverageMeter()
for sample in eval_loader:
loss, batch_size, preds = process_sample(
sample, model, gold_dict)
nll_meter.update(loss.item(), batch_size)
pred_dict.update(preds)
progress_bar.update(batch_size)
progress_bar.set_postfix(NLL=nll_meter.avg)
results = {
**util.eval_dicts(gold_dict, pred_dict, args.use_squad_v2),
**{
'NLL': nll_meter.avg
}
}
return results, pred_dict
def evaluate(self, Experts, gate, data_loader, data_dict, return_preds=False, split='validation'):
device = self.device
for expert in Experts:
expert.eval()
pred_dict = {}
all_start_logits = []
all_end_logits = []
with torch.no_grad(), tqdm(total=len(data_loader.dataset)) as progress_bar:
for batch in data_loader:
# Setup for forward
input_ids = batch['input_ids'].to(device)
attention_mask = batch['attention_mask'].to(device)
# Forward
outputs = gate(input_ids)
selected_expert = torch.argmax(outputs)
batch_size = len(input_ids)
# total_loss, start_logits, end_logits, distilbert_output.hidden_states
_, start_logits, end_logits,_ = Experts[selected_expert.item()](input_ids, attention_mask=attention_mask)
# TODO: compute loss
all_start_logits.append(start_logits)
all_end_logits.append(end_logits)
progress_bar.update(batch_size)
# Get F1 and EM scores
start_logits = torch.cat(all_start_logits).cpu().numpy()
end_logits = torch.cat(all_end_logits).cpu().numpy()
preds = util.postprocess_qa_predictions(data_dict,
data_loader.dataset.encodings,
(start_logits, end_logits))
if split == 'validation':
results = util.eval_dicts(data_dict, preds)
results_list = [('F1', results['F1']),
('EM', results['EM'])]
else:
results_list = [('F1', -1.0),
('EM', -1.0)]
results = OrderedDict(results_list)
if return_preds:
return preds, results
return results
def WeightedAverage(args, p_start, p_end, weigths, id_quest, gold_dict):
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
nbr_model = len(p_start)
nbr_batch = len(p_start[0])
for j in range(nbr_batch):
p1 = np.zeros(p_start[0][j].shape)
p2 = np.zeros(p_start[0][j].shape)
for model in range(nbr_model):
p1 += weigths[model] * p_start[model][j]
p2 += weigths[model] * p_end[model][j]
ids = id_quest[j]
p1 = torch.from_numpy(p1).float().cuda()
p2 = torch.from_numpy(p2).float().cuda()
ids = ids.cuda()
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
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)
results = util.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
results_list = [('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())
print(f'{args.split.title()} {results_str}')
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 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
def evaluate(model, data_loader, device, eval_file, max_len, use_squad_v2,
enable_EM, enable_posner):
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, cwf, lemma_indicators, c_posner, q_posner in data_loader:
# Setup for forward
cc_idxs = cc_idxs.to(device)
qc_idxs = qc_idxs.to(device)
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
if enable_EM:
cwf = cwf.to(device)
#qwf = qwf.to(device)
lemma_indicators = lemma_indicators.to(device)
else:
cwf = None
lemma_indicators = None
if enable_posner:
c_posner = c_posner.to(device)
q_posner = q_posner.to(device)
else:
q_posner = None
c_posner = None
batch_size = cw_idxs.size(0)
# Forward
log_p1, log_p2 = model(cc_idxs, qc_idxs, cw_idxs, qw_idxs, cwf,
lemma_indicators, c_posner, q_posner)
y1, y2 = y1.to(device), y2.to(device)
weight = torch.ones(log_p1.size(1), device=device)
weight[0] = 5
loss = F.nll_loss(log_p1, y1, weight=weight) + F.nll_loss(
log_p2, y2, weight=weight)
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
all_results3 = pickle.load(f)
with open('save/saves/xxlarge_123_{}.pickle'.format(d), 'rb') as f:
all_results4 = pickle.load(f)
all_results = [all_results1, all_results2, all_results3, all_results4]
pred_dict = util.compute_predictions_logits_ensemble(
examples,
features,
all_results,
5,
15,
True,
'save/temp/predict_temp.json',
'save/temp/nbest_temp.json',
'save/temp/nlog_odd.log',
False,
True,
0,
tokenizer,
)
with open('save/saves/{}_submission.csv'.format(d),
'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]])
if d == 'dev':
results = util.eval_dicts(gold_dict, pred_dict, True)
print(results)
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]])
examples,
features,
result,
5,
15,
True,
'save/temp/predict_temp.json',
'save/temp/nbest_temp.json',
'save/temp/nlog_odd.log',
False,
True,
0,
tokenizer,
)
if d == 'dev':
acc = util.eval_dicts(gold_dict, pred_dict, True)
print(acc)
all_preds.append(pred_dict)
ensembled_pred = util.compute_predictions_logits_ensemble(
examples,
features,
all_results,
5,
15,
True,
'save/temp/predict_temp.json',
'save/temp/nbest_temp.json',
'save/temp/nlog_odd.log',
False,
True,
0,
def evaluate(self,
model,
discriminator,
data_loader,
data_dict,
return_preds=False,
split='validation'):
device = self.device
model.eval()
pred_dict = {}
all_start_logits = []
all_end_logits = []
all_dis_logits = []
all_ground_truth_data_set_ids = []
with torch.no_grad(), \
tqdm(total=len(data_loader.dataset)) as progress_bar:
for batch in data_loader:
# Setup for forward
input_ids = batch['input_ids'].to(device)
attention_mask = batch['attention_mask'].to(device)
data_set_ids = batch['data_set_id'].to(device)
batch_size = len(input_ids)
outputs = model(input_ids,
attention_mask=attention_mask,
output_hidden_states=True)
# Forward
start_logits, end_logits = outputs.start_logits, outputs.end_logits
hidden_states = outputs.hidden_states[-1]
_, dis_logits = self.forward_discriminator(
discriminator,
hidden_states,
data_set_ids,
full_adv=self.full_adv)
# TODO: compute loss
all_start_logits.append(start_logits)
all_end_logits.append(end_logits)
all_dis_logits.append(dis_logits)
all_ground_truth_data_set_ids.append(data_set_ids)
progress_bar.update(batch_size)
# Get F1 and EM scores
start_logits = torch.cat(all_start_logits).cpu().numpy()
end_logits = torch.cat(all_end_logits).cpu().numpy()
dis_logits = torch.cat(all_dis_logits).cpu().numpy()
ground_truth_data_set_ids = torch.cat(
all_ground_truth_data_set_ids).cpu().numpy()
preds = util.postprocess_qa_predictions(data_dict,
data_loader.dataset.encodings,
(start_logits, end_logits))
if split == 'validation':
discriminator_eval_results = util.eval_discriminator(
data_dict, ground_truth_data_set_ids, dis_logits)
results = util.eval_dicts(data_dict, preds)
results_list = [('F1', results['F1']), ('EM', results['EM']),
('discriminator_precision',
discriminator_eval_results['precision'])]
else:
results_list = [('F1', -1.0), ('EM', -1.0),
('discriminator_precision', -1.0)]
results = OrderedDict(results_list)
if return_preds:
return preds, results
return results
def evaluate(args, model, dev_dataset, data_loader, gold_dict, tokenizer,
device, max_len, use_squad_v2):
dataset, examples, features = dev_dataset.dataset, dev_dataset.examples, dev_dataset.features
model.eval()
all_results = []
with torch.no_grad(), \
tqdm(total=len(data_loader.dataset)) as progress_bar:
for batch in data_loader:
batch = tuple(t.to(device) for t in batch)
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"token_type_ids": batch[2],
}
if args.bidaf:
inputs['char_ids'] = batch[6]
example_indices = batch[3]
batch_size = inputs["input_ids"].size(0)
# Forward
outputs = model(**inputs)
for i, example_index in enumerate(example_indices):
eval_feature = features[example_index.item()]
unique_id = int(eval_feature.unique_id)
output = [to_list(output[i]) for output in outputs]
start_logits, end_logits = output
result = SquadResult(unique_id, start_logits, end_logits)
all_results.append(result)
# Log info
progress_bar.update(batch_size)
if args.dev_logits_save_file is not None:
with open(args.dev_logits_save_file, 'wb') as f:
pickle.dump(all_results, f)
pred_dict = compute_predictions_logits(
examples,
features,
all_results,
1,
args.max_ans_len,
True,
'save/temp/predict_temp.json',
'save/temp/nbest_temp.json',
'save/temp/nlog_odd.log',
False,
args.use_squad_v2,
args.null_score_diff_threshold,
tokenizer,
)
model.train()
results = util.eval_dicts(gold_dict, pred_dict, use_squad_v2)
results_list = [('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, 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
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 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)
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
# Filter out bert_pred_file
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
sub_dict[uuid] = value
results = util.eval_dicts(gold_dict, pred_dict, no_answer=True)
results = util.eval_dicts(gold_dict, pred_dict, no_answer=True)
print(results)
def main(args, actions=None):
""""
actions is a tuple (action, number of actions to be taken)
action can be either: "substitute", "delete" or "add".
number of actions to be taken: the number of words to apply the "substitute", "delete" or "add" action.
"""
# check that actions parameters received
#print("actions: ",actions)
# 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)
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
# 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)
count_questions_type[pronoun] += 1
audit_trail_from_question_type[pronoun].append(str(index))
reverse_dict_by_value = OrderedDict(
sorted(count_questions_type.items(), key=lambda x: x[1]))
total_questions = sum(count_questions_type.values())
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
# **********************************************************
#
# >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
#
# Where we make the modif:
if actions[0] == "substitute":
# substitute to random token in each question of the batch (substitution is made within the same sentence:
batch_size = cw_idxs.size()[0]
number_of_actions = actions[1]
for _ in range(number_of_actions):
length_index_batch = cw_idxs.size()[1]
for i in range(batch_size):
tensor_with_zero_value = ((
cw_idxs[i] == 0).nonzero()).squeeze()
try:
first_zero_value = torch.min(
tensor_with_zero_value)
except:
first_zero_value = length_index_batch
if first_zero_value > 2:
select_item_idx_1 = random.randint(
0, first_zero_value - 1)
select_item_idx_2 = random.randint(
0, first_zero_value - 1)
save_value_1 = copy.deepcopy(
cw_idxs[i, select_item_idx_1])
cw_idxs[i, select_item_idx_1] = cw_idxs[
i, select_item_idx_2]
cw_idxs[i, select_item_idx_2] = save_value_1
elif actions[0] == "delete":
# substitute to random token in each question of the batch (substitution is made within the same sentence:
batch_size = cw_idxs.size()[0]
number_of_actions = actions[1]
for _ in range(number_of_actions):
length_index_batch = cw_idxs.size()[1]
for i in range(batch_size):
tensor_with_zero_value = ((
cw_idxs[i] == 0).nonzero()).squeeze()
try:
first_zero_value = torch.min(
tensor_with_zero_value)
except:
first_zero_value = length_index_batch
if first_zero_value > 2:
print("debug:", i)
print("> size before amendment of cw_idxs",
cw_idxs.size())
print(cw_idxs[i])
select_item_idx_1 = random.randint(
0, first_zero_value - 1)
#print("section 1 :", cw_idxs[i, 0:select_item_idx_1)
#print("remove:", cw_idxs[i,0:select_item_idx_1)
#print("section 2 :", cw_idxs[i, 0:select_item_idx_1)
cw_idxs[i, :] = torch.cat(
(cw_idxs[i, 0:select_item_idx_1],
cw_idxs[i, select_item_idx_1 + 1:],
torch.tensor([0])), -1)
print("> size before amendment of cw_idxs",
cw_idxs.size())
print(cw_idxs[i])
elif actions[0] == "add":
batch_size = cw_idxs.size()[0]
number_of_actions = actions[1]
for _ in range(number_of_actions):
length_index_batch = cw_idxs.size()[1]
for i in range(batch_size):
tensor_with_zero_value = ((
cw_idxs[i] == 0).nonzero()).squeeze()
try:
first_zero_value = torch.min(
tensor_with_zero_value)
except:
first_zero_value = length_index_batch
if first_zero_value > 2:
select_item_idx_1 = random.randint(
0, first_zero_value - 1)
cw_idxs[i, select_item_idx_1] = random.randint(
1, 50000)
elif actions[0] == "add2":
# substitute to random token in each question of the batch (substitution is made within the same sentence:
batch_size = cw_idxs.size()[0]
number_of_actions = actions[1]
for _ in range(number_of_actions):
length_index_batch = cw_idxs.size()[1]
for i in range(batch_size):
tensor_with_zero_value = ((
cw_idxs[i] == 0).nonzero()).squeeze()
try:
first_zero_value = torch.min(
tensor_with_zero_value)
except:
first_zero_value = length_index_batch
if first_zero_value > 2:
select_item_idx_1 = random.randint(
0, first_zero_value - 1)
select_item_idx_2 = random.randint(
0, first_zero_value - 1)
cw_idxs[i, select_item_idx_1] = random.randint(
1, 50000)
cw_idxs[i, select_item_idx_2] = random.randint(
1, 50000)
else:
print("Incorrect command: exiting")
exit()
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
print("for ", actions, ": ", results['F1'])
# create a list of the types of questions by extracting the keys from the dict audit_trail_from_question_type
# 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]])
return results['F1']
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):
# 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 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 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 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 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('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)
char_vectors = util.torch_from_json(args.char_emb_file)
# NEW : load the tag embeddings
pos_vectors = util.torch_from_json(args.pos_emb_file)
ner_vectors = util.torch_from_json(args.ner_emb_file)
# Choose model
log.info('Building model {}...'.format(args.name))
if 'baseline' in args.name:
model = BiDAF(word_vectors=word_vectors, hidden_size=args.hidden_size)
elif args.name == 'BiDAF_char':
model = BiDAF_char(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=args.hidden_size)
# NEW
elif (args.name == 'BiDAF_tag') or (args.name == 'BiDAF_tag_unfrozen') or (
args.name == 'BiDAF_tag_loss') or (args.name
== 'BiDAF_tag_unfrozen_loss'):
model = BiDAF_tag(word_vectors=word_vectors,
char_vectors=char_vectors,
pos_vectors=pos_vectors,
ner_vectors=ner_vectors,
hidden_size=args.hidden_size)
elif (args.name == 'BiDAF_tag_ext') or (args.name
== 'BiDAF_tag_ext_unfrozen'):
model = BiDAF_tag_ext(word_vectors=word_vectors,
char_vectors=char_vectors,
pos_vectors=pos_vectors,
ner_vectors=ner_vectors,
hidden_size=args.hidden_size)
elif args.name == 'coattn':
model = CoattentionModel(hidden_dim=args.hidden_size,
embedding_matrix=word_vectors,
train_word_embeddings=False,
dropout=0.35,
pooling_size=16,
number_of_iters=4,
number_of_layers=2,
device=device)
else:
raise NameError('No model named ' + args.name)
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()
# 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)
# 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, 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 args.name:
log_p1, log_p2 = model(cw_idxs, qw_idxs)
elif args.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 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)
# NEW
elif (args.name
== 'BiDAF_tag') or (args.name == 'BiDAF_tag_unfrozen') or (
args.name
== 'BiDAF_tag_loss') or (args.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 (args.name
== 'BiDAF_tag_ext') or (args.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)
else:
raise NameError('No model named ' + args.name)
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('{}: {:05.2f}'.format(k, v)
for k, v in results.items())
log.info('{} {}'.format(args.split.title(), results_str))
# More detailed error analysis
results = util.analysis_dicts(gold_dict, pred_dict, args.use_squad_v2)
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', 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]])