def train(model, args, epoch, dataset, logger, optimizer):
model.train()
total_loss = float(0)
with tqdm(desc='Training', total=len(dataset)) as pbar:
for i, (data, target, paths) in enumerate(dataset):
if True:
if i == args.stop_after:
break
pbar.update()
model.zero_grad()
output = model(data)
'''
target_var = Variable(maybe_cuda(torch.cat(target, 0), args.cuda), requires_grad=False)
loss = model.criterion(output, target_var)
'''
concate_target = torch.cat(target, 0)
new_target = torch.zeros(concate_target.size()[0], 2)
new_target[range(new_target.shape[0]), concate_target]=1
target_var = Variable(maybe_cuda(new_target, args.cuda), requires_grad=False)
target_var = target_var.type_as(output)
pos_weight = maybe_cuda(torch.FloatTensor([0.1, 1.0]))
loss = F.binary_cross_entropy_with_logits(output, target_var, reduction='sum', pos_weight=pos_weight)
loss.backward()
optimizer.step()
total_loss += loss.item()
pbar.set_description('Training, loss={:.4}'.format(loss.item()))
total_loss = total_loss / len(dataset)
logger.debug('Training Epoch: {}, Loss: {:.4}.'.format(epoch + 1, total_loss))
log_value('Training Loss', total_loss, epoch + 1)
def run_model_on_batch(encoder, decoder, config, use_teacher_forcing,
batch_data):
(x, x_len), (y, y_len) = batch_data
encoder_output, encoder_hidden = encoder(
x=maybe_cuda(Variable(x, requires_grad=False), cuda=config.cuda),
x_len=x_len,
hidden=None,
)
y_var = maybe_cuda(Variable(y, requires_grad=False), cuda=config.cuda)
target_length = y_var.size()[1]
decoder_input = maybe_cuda(Variable(torch.LongTensor([[0]] * x.size()[0])),
cuda=config.cuda)
decoder_hidden = encoder_hidden
decoder_output_ls = []
attn_ls = []
for di in range(target_length):
decoder_output, decoder_hidden, other_dict = decoder(
decoder_input, decoder_hidden, encoder_output)
decoder_output_ls.append(decoder_output)
if "attn" in other_dict:
attn_ls.append(other_dict["attn"])
if use_teacher_forcing:
decoder_input = y_var[:, di].unsqueeze(1)
else:
topv, topi = decoder_output.data.topk(1)
decoder_input = maybe_cuda(Variable(topi.squeeze(1)),
cuda=config.cuda)
full_decoder_output = torch.cat(decoder_output_ls, dim=1)
return full_decoder_output, attn_ls
def max_pooling_similarity_computing(self, doc_output, segment_idx):
similarities = Variable(maybe_cuda(torch.Tensor([])))
cos = nn.CosineSimilarity(dim=1, eps=1e-6)
#ln = nn.LayerNorm(512, elementwise_affine=False)
seg_outputs = []
index = 0
doc_output = F.softmax(doc_output)
for i, idx in enumerate(segment_idx):
if i == 0:
seg_output = doc_output[0:segment_idx[i] + 1, :]
elif i == len(segment_idx) - 1:
seg_output = doc_output[segment_idx[i - 1] + 1:, :]
else:
seg_output = doc_output[segment_idx[i - 1] + 1:segment_idx[i] +
1, :]
seg_outputs.append(seg_output)
maxes = Variable(
maybe_cuda(torch.zeros(len(seg_outputs), self.hidden * 2)))
for i in range(len(seg_outputs)):
maxes[i, :] = torch.max(seg_outputs[i], 0)[0]
#maxes[i, :] = torch.mean(seg_outputs[i], 0)
if len(seg_outputs) > 1:
tensor_1 = maxes[:-1, :]
tensor_2 = maxes[1:, :]
similarities = cos(tensor_1, tensor_2)
#similarity += (tensor_1 * tensor_2).sum()/(tensor_1.size()[0])
#similarities = torch.diag(torch.mm(tensor_1, tensor_2.permute(1,0)))
return similarities
def get_model(config):
encoder = maybe_cuda(Encoder(config), cuda=config.cuda)
if config.attn_method != "disabled":
decoder = maybe_cuda(AttnDecoder(config), cuda=config.cuda)
else:
decoder = maybe_cuda(Decoder(config), cuda=config.cuda)
return encoder, decoder
def similarity_computing_inner(self, doc_output, segment_idx):
similarities = []
cos = nn.CosineSimilarity(dim=1, eps=1e-6)
seg_outputs = []
index = 0
doc_output = F.softmax(doc_output)
for i, idx in enumerate(segment_idx):
if i == 0:
seg_output = doc_output[0:segment_idx[i] + 1, :]
elif i == len(segment_idx) - 1:
seg_output = doc_output[segment_idx[i - 1] + 1:, :]
else:
seg_output = doc_output[segment_idx[i - 1] + 1:segment_idx[i] +
1, :]
seg_outputs.append(seg_output)
for i in range(len(seg_outputs)):
sent_idx = maybe_cuda(
torch.LongTensor([k for k in range(seg_outputs[i].size()[0])]))
if seg_outputs[i].size()[0] > 1:
pairs = torch.combinations(sent_idx)
pair_sims = []
for p in pairs:
pair_sims.append(
cos(seg_outputs[i][p[0], :].unsqueeze(0),
seg_outputs[i][p[1], :].unsqueeze(0)))
similarities.append(sum(pair_sims) / len(pair_sims))
else:
continue
return Variable(maybe_cuda(torch.Tensor(similarities)))
def train(model, args, epoch, dataset, logger, optimizer):
model.train()
total_loss = float(0)
with tqdm(desc='Training', total=len(dataset)) as pbar:
for i, (data, target, paths, sent_bert_vec, target_idx) in enumerate(dataset):
if True:
if i == args.stop_after:
break
pbar.update()
model.zero_grad()
output, sims = model(data, sent_bert_vec, target_idx)
target_var = Variable(maybe_cuda(torch.cat(target, 0), args.cuda), requires_grad=False)
# generate gold label for coherence scores...
target_list = target_var.data.cpu().numpy()
target_coh_list = []
for t in target_list:
if t == 0:
target_coh_list.append(torch.LongTensor([1]))
else:
target_coh_list.append(torch.LongTensor([0]))
target_coh_var = Variable(maybe_cuda(torch.cat(target_coh_list, 0), args.cuda), requires_grad=False)
loss = supervised_cross_entropy(output, sims, target_var, target_coh_var)
#loss = model.criterion(output, target_var)
#sim = sim_.data.cpu()
#total_sim += sim
#concate_target = torch.cat(target, 0)
#total_1.append(sum(sim.squeeze(1)[concate_target == 1])/len(sim.squeeze(1)[concate_target == 1]))
#total_0.append(sum(sim.squeeze(1)[concate_target == 0])/len(sim.squeeze(1)[concate_target == 0]))
'''
new_target = torch.zeros(concate_target.size()[0], 2)
new_target[range(new_target.shape[0]), concate_target]=1
target_var = Variable(maybe_cuda(new_target, args.cuda), requires_grad=False)
target_var = target_var.type_as(output)
pos_weight = maybe_cuda(torch.FloatTensor([0.1, 1.0]))
loss = F.binary_cross_entropy_with_logits(output, target_var, reduction='sum', pos_weight=pos_weight) + 10*sim_
'''
loss.backward()
optimizer.step()
total_loss += loss.item()
# logger.debug('Batch %s - Train error %7.4f', i, loss.data[0])
pbar.set_description('Training, loss={:.4}'.format(loss.item()))
# except Exception as e:
# logger.info('Exception "%s" in batch %s', e, i)
# logger.debug('Exception while handling batch with file paths: %s', paths, exc_info=True)
# pass
#print('The similarity between the segs: ', total_sim/len(dataset))
total_loss = total_loss / len(dataset)
logger.debug('Training Epoch: {}, Loss: {:.4}.'.format(epoch + 1, total_loss))
log_value('Training Loss', total_loss, epoch + 1)
def validate(model, args, epoch, dataset, logger):
model.eval()
with tqdm(desc='Validatinging', total=len(dataset)) as pbar:
acc = Accuracies()
for i, (data, target, paths) in enumerate(dataset):
if True:
if i == args.stop_after:
break
pbar.update()
output = model(data)
output_softmax = F.softmax(output, 1)
targets_var = Variable(maybe_cuda(torch.cat(target, 0),
args.cuda),
requires_grad=False)
output_seg = output.data.cpu().numpy().argmax(axis=1)
target_seg = targets_var.data.cpu().numpy()
preds_stats.add(output_seg, target_seg)
acc.update(output_softmax.data.cpu().numpy(), target)
# except Exception as e:
# # logger.info('Exception "%s" in batch %s', e, i)
# logger.debug('Exception while handling batch with file paths: %s', paths, exc_info=True)
# pass
epoch_pk, epoch_windiff, threshold = acc.calc_accuracy()
logger.info(
'Validating Epoch: {}, accuracy: {:.4}, Pk: {:.4}, Windiff: {:.4}, F1: {:.4} . '
.format(epoch + 1, preds_stats.get_accuracy(), epoch_pk,
epoch_windiff, preds_stats.get_f1()))
preds_stats.reset()
return epoch_pk, threshold
def forward(self, batch, return_attns=False):
batch_size = len(batch)
sentences_per_doc = []
all_batch_sentences = []
for document in batch:
all_batch_sentences.extend(document)
sentences_per_doc.append(len(document))
lengths = [s.size()[0] for s in all_batch_sentences]
sort_order = np.argsort(lengths)[::-1]
sorted_sentences = [all_batch_sentences[i] for i in sort_order]
sorted_lengths = [s.size()[0] for s in sorted_sentences]
max_length = max(lengths)
logger.debug('Num sentences: %s, max sentence length: %s',
sum(sentences_per_doc), max_length)
padded_sentences = [self.pad(s, max_length) for s in sorted_sentences]
big_tensor = torch.cat(padded_sentences, 1) # (max_length, batch size, 300)
processed_tensor = pack_padded_sequence(big_tensor, sorted_lengths)
encoded_sentences = self.sentence_encoder(processed_tensor)
unsort_order = Variable(maybe_cuda(torch.LongTensor(unsort(sort_order))))
unsorted_encodings = encoded_sentences.index_select(0, unsort_order)
index = 0
encoded_documents = []
for sentences_count in sentences_per_doc:
end_index = index + sentences_count
encoded_documents.append(unsorted_encodings[index : end_index, :])
index = end_index
doc_sizes = [doc.size()[0] for doc in encoded_documents]
max_doc_size = np.max(doc_sizes)
ordered_document_idx = np.argsort(doc_sizes)[::-1]
ordered_doc_sizes = sorted(doc_sizes)[::-1]
ordered_documents = [encoded_documents[idx] for idx in ordered_document_idx]
padded_docs = [self.pad_document(d, max_doc_size) for d in ordered_documents]
padded_docs = torch.stack(padded_docs).squeeze(2) # turn the tensor list into tensor
##############################################################
pos_emb = self.PositionEncoder.pe[:, :padded_docs.size()[1]].expand(padded_docs.size())
inputs = padded_docs+pos_emb
sent_mask = generate_mask(ordered_doc_sizes, max_doc_size)
non_pad_mask = sent_mask.unsqueeze(-1)
slf_attn_mask = (1-sent_mask).unsqueeze(1).expand(-1,sent_mask.size()[1],-1).type(torch.bool)
outputs = self.Transformer(inputs,non_pad_mask, slf_attn_mask)
outputs = self.Dropoutlayer(outputs)
outputs = self.Decoderlayer(outputs) # batch * length * 1
doc_outputs = []
for i, doc_len in enumerate(ordered_doc_sizes):
doc_outputs.append(outputs[i, 0:doc_len - 1, :]) # -1 to remove last prediction
unsorted_doc_outputs = [doc_outputs[i] for i in unsort(ordered_document_idx)]
x = torch.cat(unsorted_doc_outputs, 0)
return x
def train(model, args, epoch, dataset, logger, optimizer):
model.train()
total_loss = float(0)
with tqdm(desc='Training', total=len(dataset)) as pbar:
for i, (data, target, paths) in enumerate(dataset):
if True:
if i == args.stop_after:
break
pbar.update()
model.zero_grad()
output = model(data)
target_var = Variable(maybe_cuda(torch.cat(target, 0),
args.cuda),
requires_grad=False)
loss = model.criterion(output, target_var)
loss.backward()
optimizer.step()
# total_loss += loss.data[0]
total_loss += loss.item()
# logger.debug('Batch %s - Train error %7.4f', i, loss.data[0])
pbar.set_description('Training, loss={:.4}'.format(
loss.item()))
# except Exception as e:
# logger.info('Exception "%s" in batch %s', e, i)
# logger.debug('Exception while handling batch with file paths: %s', paths, exc_info=True)
# pass
total_loss = total_loss / len(dataset)
logger.debug('Training Epoch: {}, Loss: {:.4}.'.format(
epoch + 1, total_loss))
log_value('Training Loss', total_loss, epoch + 1)
def validate(model, args, epoch, dataset, logger):
model.eval()
with tqdm(desc='Validatinging', total=len(dataset)) as pbar:
acc = Accuracies()
for i, (data, target, paths) in enumerate(dataset):
if True:
if i == args.stop_after:
break
pbar.update()
output = model(data)
output_softmax = F.softmax(output, 1)
targets_var = Variable(maybe_cuda(torch.cat(target, 0), args.cuda), requires_grad=False)
output_seg = output.data.cpu().numpy().argmax(axis=1)
target_seg = targets_var.data.cpu().numpy()
preds_stats.add(output_seg, target_seg)
acc.update(output_softmax.data.cpu().numpy(), target)
epoch_pk, epoch_windiff, threshold = acc.calc_accuracy()
logger.info('Validating Epoch: {}, Pk: {:.4} . '.format(epoch + 1, epoch_pk))
preds_stats.reset()
return epoch_pk, threshold
def prepare_tensor(sentences):
tensored_data = []
for sentence in sentences:
if len(sentence) > 0:
tensored_data.append(utils.maybe_cuda(torch.FloatTensor(np.concatenate(sentence))))
return tensored_data
def pad(self, s, max_length):
s_length = s.size()[0]
v = Variable(maybe_cuda(s.unsqueeze(0).unsqueeze(0)))
padded = F.pad(
v, (0, 0, 0, max_length - s_length)) # (1, 1, max_length, 300)
shape = padded.size()
return padded.view(shape[2], 1, shape[3]) # (max_length, 1, 300)
def main(args):
sys.path.append(str(Path(__file__).parent))
checkpoint_path = Path(args.checkpoint_dir)
checkpoint_path.mkdir(exist_ok=True)
logger = utils.setup_logger(__name__, os.path.join(args.checkpoint_dir, 'train.log'))
utils.read_config_file(args.config)
utils.config.update(args.__dict__)
# logger.debug('Running with config %s', utils.config)
# configure(os.path.join('runs', args.expname))
word2vec = None
if not args.infer:
dataset_path = utils.config['choidataset']
train_dataset = ChoiDataset(dataset_path, word2vec)
dev_dataset = ChoiDataset(dataset_path, word2vec)
test_dataset = ChoiDataset(dataset_path, word2vec)
train_dl = DataLoader(train_dataset, batch_size=args.bs, collate_fn=collate_fn, shuffle=True,
num_workers=args.num_workers)
dev_dl = DataLoader(dev_dataset, batch_size=args.test_bs, collate_fn=collate_fn, shuffle=False,
num_workers=args.num_workers)
test_dl = DataLoader(test_dataset, batch_size=args.test_bs, collate_fn=collate_fn, shuffle=False,
num_workers=args.num_workers)
assert bool(args.model) ^ bool(args.load_from) # exactly one of them must be set
if args.model:
model = import_model(args.model)
elif args.load_from:
with open(args.load_from, 'rb') as f:
model = torch.load(f)
model.train()
model = maybe_cuda(model)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
if not args.infer:
best_val_pk = 1.0
for j in range(args.epochs):
train(model, args, j, train_dl, logger, optimizer)
with (checkpoint_path / 'model{:03d}.t7'.format(j)).open('wb') as f:
torch.save(model, f)
val_pk, threshold = validate(model, args, j, dev_dl, logger)
if val_pk < best_val_pk:
test_pk = test(model, args, j, test_dl, logger, threshold)
logger.debug(
colored(
'Current best model from epoch {} with p_k {} and threshold {}'.format(j, test_pk, threshold),
'green'))
best_val_pk = val_pk
with (checkpoint_path / 'Meilleur_model.t7'.format(j)).open('wb') as f:
torch.save(model, f)
def resnet_init(self):
end_time = time.time()
from utils import maybe_cuda
from glico_model.img_to_vec import Img2Vec
img2vec = Img2Vec()
# Matrix to hold the image vectors
for i_loader, dloader in enumerate(self.data_loader):
for i, batch in enumerate(dloader): # batch_size=1, no shuffle
idx = maybe_cuda(batch[0])
imgs = maybe_cuda(batch[1])
labels = batch[2]
if i_loader < 1:
vec = img2vec.get_vec(imgs.squeeze().detach().cpu().numpy())
self.emb.weight.data[idx] = torch.tensor(vec)
self.label2idx[labels.item()].add(idx.item())
self.idx2label[idx.item()] = labels.item()
self.get_norm()
print(f"init time: {(time.time() - end_time) / 60:8.2f}m")
def forward(self, batch):
batch_size = len(batch)
sentences_per_doc = []
all_batch_sentences = []
for document in batch:
all_batch_sentences.extend(document)
sentences_per_doc.append(len(document))
lengths = [s.size()[0] for s in all_batch_sentences]
sort_order = np.argsort(lengths)[::-1]
sorted_sentences = [all_batch_sentences[i] for i in sort_order]
sorted_lengths = [s.size()[0] for s in sorted_sentences]
max_length = max(lengths)
logger.debug('Num sentences: %s, max sentence length: %s',
sum(sentences_per_doc), max_length)
padded_sentences = [self.pad(s, max_length) for s in sorted_sentences]
big_tensor = torch.cat(padded_sentences,
1) # (max_length, batch size, 300)
packed_tensor = pack_padded_sequence(big_tensor, sorted_lengths)
encoded_sentences = self.sentence_encoder(packed_tensor)
unsort_order = torch.tensor(
maybe_cuda(torch.LongTensor(unsort(sort_order))))
unsorted_encodings = encoded_sentences.index_select(0, unsort_order)
index = 0
encoded_documents = []
for sentences_count in sentences_per_doc:
end_index = index + sentences_count
encoded_documents.append(unsorted_encodings[index:end_index, :])
index = end_index
doc_sizes = [doc.size()[0] for doc in encoded_documents]
max_doc_size = np.max(doc_sizes)
ordered_document_idx = np.argsort(doc_sizes)[::-1]
ordered_doc_sizes = sorted(doc_sizes)[::-1]
ordered_documents = [
encoded_documents[idx] for idx in ordered_document_idx
]
padded_docs = [
self.pad_document(d, max_doc_size) for d in ordered_documents
]
docs_tensor = torch.cat(padded_docs, 1)
packed_docs = pack_padded_sequence(docs_tensor, ordered_doc_sizes)
sentence_lstm_output, _ = self.sentence_lstm(
packed_docs, zero_state(self, batch_size=batch_size))
padded_x, _ = pad_packed_sequence(
sentence_lstm_output) # (max sentence len, batch, 256)
doc_outputs = []
for i, doc_len in enumerate(ordered_doc_sizes):
doc_outputs.append(padded_x[0:doc_len - 1,
i, :]) # -1 to remove last prediction
unsorted_doc_outputs = [
doc_outputs[i] for i in unsort(ordered_document_idx)
]
sentence_outputs = torch.cat(unsorted_doc_outputs, 0)
x = self.h2s(sentence_outputs)
return x
def forward(self, x):
batch_size = x.batch_sizes[0]
s = zero_state(self, batch_size)
packed_output, _ = self.lstm(x, s)
padded_output, lengths = pad_packed_sequence(
packed_output) # (max sentence len, batch, 256)
maxes = torch.tensor(
maybe_cuda(torch.zeros(batch_size, padded_output.size(2))))
for i in range(batch_size):
maxes[i, :] = torch.max(padded_output[:lengths[i], i, :], 0)[0]
return maxes
def cube_init(self):
import itertools
from utils import maybe_cuda
end_time = time.time()
vertices = itertools.product('01', repeat=self.nz) # cartesian product
# self.vertices = [tuple(int(s) for s in v) for v in vertices]
self.vertices = sample_from_iter(vertices, self.num_classes)
print(f"Init Cube: num classes= {self.num_classes}")
for i_loader, dloader in enumerate(self.data_loader):
for i, batch in enumerate(dloader): # batch_size=1, no shuffle
idx = maybe_cuda(batch[0])
_ = maybe_cuda(batch[1])
labels = batch[2]
if i_loader < 1:
vertex = self.vertices[labels]
self.emb.weight.data[idx] = self.emb.weight.data[idx] + torch.tensor(vertex).float()
self.label2idx[labels.item()].add(idx.item())
self.idx2label[idx.item()] = labels.item()
self.get_norm()
print(f"init time: {(time.time() - end_time) / 60:8.2f}m")
def get_attn_key_pad_mask(seq_k, seq_q, lengths):
''' For masking out the padding part of key sequence. '''
# Expand to fit the shape of key query attention matrix.
batch, len_q, len_k = seq_k.size(0), seq_q.size(1), seq_k.size(1)
padding_mask = maybe_cuda(torch.ByteTensor([[0 if i<clen else 1 for i in range(len_k) ] for clen in lengths]) )
padding_mask = padding_mask.unsqueeze(1).expand(-1, len_q, -1) # b x lq x lk
# len_q = seq_q.size(1)
# padding_mask = seq_k.eq(Constants.PAD)
# padding_mask = padding_mask.unsqueeze(1).expand(-1, len_q, -1) # b x lq x lk
return padding_mask
def load_model(model_path=None, is_cuda=None):
if model_path is None:
model_path = utils.config['model']
with open(model_path, 'r') as f:
model = torch.load(f)
model.eval()
if is_cuda is None:
is_cuda = utils.config['cuda']
return utils.maybe_cuda(model, is_cuda)
def test(model, args, epoch, dataset, logger, threshold):
model.eval()
with tqdm(desc='Testing', total=len(dataset)) as pbar:
acc = accuracy.Accuracy()
for i, (data, target, paths) in enumerate(dataset):
if True:
if i == args.stop_after:
break
pbar.update()
output = model(data)
output_softmax = F.softmax(output, 1)
targets_var = Variable(maybe_cuda(torch.cat(target, 0),
args.cuda),
requires_grad=False)
output_seg = output.data.cpu().numpy().argmax(axis=1)
target_seg = targets_var.data.cpu().numpy()
preds_stats.add(output_seg, target_seg)
current_idx = 0
for k, t in enumerate(target):
document_sentence_count = len(t)
to_idx = int(current_idx + document_sentence_count)
output = ((output_softmax.data.cpu().numpy()[
current_idx:to_idx, :])[:, 1] > threshold)
h = np.append(output, [1])
tt = np.append(t, [1])
acc.update(h, tt)
current_idx = to_idx
# acc.update(output_softmax.data.cpu().numpy(), target)
#
# except Exception as e:
# # logger.info('Exception "%s" in batch %s', e, i)
# logger.debug('Exception while handling batch with file paths: %s', paths, exc_info=True)
epoch_pk, epoch_windiff = acc.calc_accuracy()
logger.debug(
'Testing Epoch: {}, accuracy: {:.4}, Pk: {:.4}, Windiff: {:.4}, F1: {:.4} . '
.format(epoch + 1, preds_stats.get_accuracy(), epoch_pk,
epoch_windiff, preds_stats.get_f1()))
preds_stats.reset()
return epoch_pk
def accuracy(model, test_data, batch_size, topk=(1, 5), aug_param=None):
"""Computes the accuracy over the k top predictions for the specified values of k"""
model.eval()
if aug_param is None:
aug_param = {
'std': None,
'mean': None,
'rand_crop': 32,
'image_size': 32
}
top1 = AverageMeter('Acc@1', ':6.2f')
top5 = AverageMeter('Acc@5', ':6.2f')
with torch.no_grad():
test_loader = get_loader_with_idx(test_data,
batch_size=batch_size,
**aug_param,
num_workers=8,
shuffle=False,
eval=True)
for i, batch in enumerate(test_loader):
imgs = maybe_cuda(batch[1])
targets = maybe_cuda(batch[2])
output = maybe_cuda(model(imgs))
maxk = max(topk)
batch_size = targets.size(0)
# target = validate_loader_consistency(batch_size, idx, target, test_data)
_, pred = output.topk(maxk, 1, True, True)
pred = pred.t()
correct = pred.eq(targets.view(1, -1).expand_as(pred))
res = []
for k in topk:
correct_k = correct[:k].view(-1).float().sum(0, keepdim=True)
res.append(correct_k.mul_(100.0 / batch_size))
top1.update(res[0].item())
top5.update(res[1].item())
return top1.avg, top5.avg
def main(args):
sys.path.append(str(Path(__file__).parent))
logger = utils.setup_logger(__name__, 'cross_validate_choi.log')
utils.read_config_file(args.config)
utils.config.update(args.__dict__)
logger.debug('Running with config %s', utils.config)
configure(os.path.join('runs', args.expname))
if not args.test:
word2vec = gensim.models.KeyedVectors.load_word2vec_format(utils.config['word2vecfile'], binary=True)
else:
word2vec = None
dataset_path = Path(args.flat_choi)
with open(args.load_from, 'rb') as f:
model = torch.load(f)
model.eval()
model = maybe_cuda(model)
test_accuracy = accuracy.Accuracy()
for j in range(5):
validate_folder_numbers = range(5)
validate_folder_numbers.remove(j)
validate_folder_names = [dataset_path.joinpath(str(num)) for num in validate_folder_numbers]
dev_dataset = ChoiDataset(dataset_path , word2vec, folder=True, folders_paths=validate_folder_names)
test_dataset = ChoiDataset(dataset_path, word2vec, folder=True, folders_paths=[dataset_path.joinpath(str(j))])
dev_dl = DataLoader(dev_dataset, batch_size=args.test_bs, collate_fn=collate_fn, shuffle=False,
num_workers=args.num_workers)
test_dl = DataLoader(test_dataset, batch_size=args.test_bs, collate_fn=collate_fn, shuffle=False,
num_workers=args.num_workers)
_, threshold = validate(model, args, j, dev_dl, logger)
test_pk = test(model, args, j, test_dl, logger, threshold, test_accuracy)
logger.debug(colored('Cross validation section {} with p_k {} and threshold {}'.format(j, test_pk, threshold),'green'))
cross_validation_pk, _ = test_accuracy.calc_accuracy()
print ('Final cross validaiton Pk is: ' + str(cross_validation_pk))
logger.debug(
colored('Final cross validaiton Pk is: {}'.format(cross_validation_pk), 'green'))
def validate_step(engine, batch):
encoder.eval()
decoder.eval()
(x, x_len), (y, y_len) = batch
full_decoder_output, attn_ls = run_model_on_batch(
encoder=encoder,
decoder=decoder,
config=config,
use_teacher_forcing=False,
batch_data=batch,
)
y_var = maybe_cuda(Variable(y, requires_grad=False), cuda=config.cuda)
batch_loss = criterion(
full_decoder_output.view(-1,
full_decoder_output.size()[2]),
y_var.view(-1))
return batch_loss
def __init__(self,
len_max_seq,
d_word_vec,
n_layers,
n_head,
d_k,
d_v,
d_model,
d_inner,
dropout=0.1):
super(SentenceEncoding, self).__init__()
self.CLS = maybe_cuda(torch.randn(1, 300))
n_position = len_max_seq + 1
self.d_model = d_model
self.src_emb = nn.Linear(300, d_model)
self.position_enc = nn.Embedding.from_pretrained( \
get_sinusoid_encoding_table(n_position, d_word_vec), \
freeze=True)
self.layers = nn.ModuleList([ \
EncoderLayer(d_model, d_inner, n_head, d_k, d_v, dropout=dropout) for _ in range(n_layers)])
self.dropout_func = nn.Dropout(dropout)
def test(model, args, epoch, dataset, logger, test_threshold, test_acc):
model.eval()
with tqdm(desc='Testing', total=len(dataset)) as pbar:
for i, (data, target, paths) in enumerate(dataset):
if True:
if i == args.stop_after:
break
pbar.update()
output = model(data)
output_softmax = F.softmax(output, 1)
targets_var = Variable(maybe_cuda(torch.cat(target, 0), args.cuda), requires_grad=False)
output_seg = output.data.cpu().numpy().argmax(axis=1)
target_seg = targets_var.data.cpu().numpy()
preds_stats.add(output_seg, target_seg)
current_idx = 0
for k, t in enumerate(target):
document_sentence_count = len(t)
to_idx = int(current_idx + document_sentence_count)
output = ((output_softmax.data.cpu().numpy()[current_idx: to_idx, :])[:, 1] > test_threshold)
h = np.append(output, [1])
tt = np.append(t, [1])
test_acc.update(h, tt)
current_idx = to_idx
test_pk, epoch_windiff = test_acc.calc_accuracy()
logger.debug('Testing validation section: {}, accuracy: {:.4}, Pk: {:.4}, Windiff: {:.4}, F1: {:.4} . '.format(epoch + 1,
preds_stats.get_accuracy(),
test_pk,
epoch_windiff,
preds_stats.get_f1()))
preds_stats.reset()
return test_pk
def forward(self, src_seq, max_length, lengths, return_attns=False):
enc_slf_attn_list = []
# -- Prepare masks
slf_attn_mask = get_attn_key_pad_mask(seq_k=src_seq, seq_q=src_seq, lengths=lengths)
non_pad_mask = get_non_pad_mask(src_seq, lengths)
position = maybe_cuda(torch.tensor([[i if i<clen else 0 for i in range(max_length) ] for clen in lengths]) )
# -- Forward
enc_output = src_seq + self.position_enc(position)
for enc_layer in self.layer_stack:
enc_output, enc_slf_attn = enc_layer(
enc_output,
non_pad_mask=non_pad_mask,
slf_attn_mask=slf_attn_mask)
if return_attns:
enc_slf_attn_list += [enc_slf_attn]
if return_attns:
return enc_output, enc_slf_attn_list
return enc_output
def train(model, args, epoch, dataset, logger, optimizer):
model.train()
total_loss = float(0)
with tqdm(desc='Training', total=len(dataset)) as pbar:
for i, (data, target, paths) in enumerate(dataset):
if True :
if i == args.stop_after:
break
pbar.update()
model.zero_grad()
output = model(data)
target_var = Variable(maybe_cuda(torch.cat(target, 0), args.cuda), requires_grad=False)
loss = model.criterion(output, target_var)
loss.backward()
optimizer.step()
# total_loss += loss.data
total_loss += float(loss.item())
# pbar.set_description('Training : LOSS = {:.4}'.format(float(loss.item())))
torch.cuda.empty_cache()
del data, target, output, target_var, loss
gc.collect()
def test(model, args, epoch, dataset, logger, threshold):
model.eval()
with tqdm(desc='Testing', total=len(dataset)) as pbar:
acc = Accuracy() # accuracies class is not needed
for i, (data, target, paths) in enumerate(dataset):
if True:
if i == args.stop_after:
break
pbar.update()
output = model(data)
output_softmax = F.softmax(output, 1)
targets_var = torch.tensor(maybe_cuda(torch.cat(target, 0), args.cuda), requires_grad=False)
output_seg = output.data.cpu().numpy().argmax(axis=1)
target_seg = targets_var.data.cpu().numpy()
preds_stats.add(output_seg, target_seg)
current_idx = 0
for k, t in enumerate(target):
path = paths[k]
document_sentence_count = len(t)
to_idx = int(current_idx + document_sentence_count)
output = ((output_softmax.data.cpu().numpy()[current_idx: to_idx, :])[:, 1] > threshold)
h = np.append(output, [1])
tt = np.append(t, [1])
acc.update(h, tt, path)
current_idx = to_idx
epoch_pk, epoch_windiff, epoch_b, epoch_s = acc.calc_accuracy()
logger.info('Testing Epoch: {}, accuracy: {:.4}, Pk: {:.4}, Windiff: {:.4}, B: {:.4}, S: {:.4} ,F1: {:.4} . '.format(
epoch + 1,
preds_stats.get_accuracy(),
epoch_pk,
epoch_windiff,
epoch_b,
epoch_s,
preds_stats.get_f1()))
preds_stats.reset()
epoch_result = acc.all_test_result
return epoch_pk, epoch_result
def train_step(engine, batch):
encoder.train()
decoder.train()
encoder_optimizer.zero_grad()
decoder_optimizer.zero_grad()
(x, x_len), (y, y_len) = batch
full_decoder_output, attn_ls = run_model_on_batch(
encoder=encoder,
decoder=decoder,
config=config,
use_teacher_forcing=use_teacher_forcing,
batch_data=batch,
)
y_var = maybe_cuda(Variable(y, requires_grad=False), cuda=config.cuda)
batch_loss = criterion(
full_decoder_output.view(-1,
full_decoder_output.size()[2]),
y_var.view(-1))
batch_loss.backward()
encoder_optimizer.step()
decoder_optimizer.step()
return batch_loss
def train(model, args, epoch, dataset, logger, optimizer):
model.train()
total_loss = float(0)
with tqdm(desc='Training', total=len(dataset)) as pbar:
for i, (data, target, paths) in enumerate(dataset):
if True:
if i == args.stop_after:
break
pbar.update()
model.zero_grad()
output = model(data)
target_var = torch.tensor(maybe_cuda(torch.cat(target, 0),
args.cuda),
requires_grad=False)
loss = model.criterion(output, target_var)
loss.backward()
optimizer.step()
total_loss += loss.item()
pbar.set_description('Training, loss={:.4}'.format(
loss.item()))
total_loss = total_loss / len(dataset)
logger.debug('Training Epoch: {}, Loss: {:.4}.'.format(
epoch + 1, total_loss))
log_value('Training Loss', total_loss, epoch + 1)
