def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
splits = ['train', 'valid'] #+ (['test'] if args.test else [])
datasets = OrderedDict()
for split in splits:
datasets[split] = PTB(
data_dir=args.data_dir,
split=split,
create_data=args.create_data,
max_sequence_length=args.max_sequence_length,
min_occ=args.min_occ
)
model = SentenceVAE(
vocab_size=datasets['train'].vocab_size,
sos_idx=datasets['train'].sos_idx,
eos_idx=datasets['train'].eos_idx,
pad_idx=datasets['train'].pad_idx,
max_sequence_length=args.max_sequence_length,
embedding_size=args.embedding_size,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
word_dropout=args.word_dropout,
latent_size=args.latent_size,
num_layers=args.num_layers,
bidirectional=args.bidirectional
)
if torch.cuda.is_available():
model = model.cuda()
if args.tensorboard_logging:
writer = SummaryWriter(os.path.join('./',args.logdir, expierment_name(args,ts)))
writer.add_text("model", str(model))
writer.add_text("args", str(args))
writer.add_text("ts", ts)
save_model_path = os.path.join('./',args.save_model_path,'VAE', ts)
os.makedirs(save_model_path)
def kl_anneal_function(anneal_function, step, k, x0):
if anneal_function == 'logistic':
return float(1/(1+np.exp(-k*(step-x0))))
elif anneal_function == 'linear':
return min(1, step/x0)
NLL = torch.nn.NLLLoss(size_average=False, ignore_index=datasets['train'].pad_idx)
def loss_fn(logp, target, length, mean, logv, anneal_function, step, k, x0):
# cut-off unnecessary padding from target, and flatten
target = target[:, :torch.max(length).data[0]].contiguous().view(-1)
logp = logp.view(-1, logp.size(2))
# Negative Log Likelihood
NLL_loss = NLL(logp, target)
NLL_w_avg = NLL_loss/torch.sum(length).float()
# KL Divergence
KL_loss = -0.5 * torch.sum(1 + logv - mean.pow(2) - logv.exp())
KL_weight = kl_anneal_function(anneal_function, step, k, x0)
return NLL_loss, KL_loss, KL_weight,NLL_w_avg
print(model)
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
tensor = torch.cuda.FloatTensor if torch.cuda.is_available() else torch.Tensor
step = 0
for epoch in range(args.epochs):
for split in splits:
data_loader = DataLoader(
dataset=datasets[split],
batch_size=args.batch_size,
shuffle=split=='train',
num_workers=cpu_count(),
pin_memory=torch.cuda.is_available()
)
tracker = defaultdict(tensor)
# Enable/Disable Dropout
if split == 'train':
model.train()
else:
model.eval()
for iteration, batch in enumerate(data_loader):
batch_size = batch['input'].size(0)
for k, v in batch.items():
if torch.is_tensor(v):
batch[k] = to_var(v)
# Forward pass
logp, mean, logv, z = model(batch['input'], batch['length'])
# loss calculation
NLL_loss, KL_loss, KL_weight,NLL_w_avg = loss_fn(logp, batch['target'],
batch['length'], mean, logv, args.anneal_function, step, args.k, args.x0)
loss = (NLL_loss + KL_weight * KL_loss)/batch_size
# backward + optimization
if split == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
# bookkeepeing
# Avoid the .cat error !!!
#print(loss.data)
#print(tracker['ELBO'])
loss_data = torch.tensor([loss.data.item()])
tracker['ELBO'] = torch.cat((tracker['ELBO'], loss_data)) #Orig: tracker['ELBO'] = torch.cat((tracker['ELBO'], loss.data),1)
if args.tensorboard_logging:
writer.add_scalar("%s/ELBO"%split.upper(), loss.data[0], epoch*len(data_loader) + iteration)
writer.add_scalar("%s/NLL Loss"%split.upper(), NLL_loss.data[0]/batch_size, epoch*len(data_loader) + iteration)
writer.add_scalar("%s/KL Loss"%split.upper(), KL_loss.data[0]/batch_size, epoch*len(data_loader) + iteration)
writer.add_scalar("%s/KL Weight"%split.upper(), KL_weight, epoch*len(data_loader) + iteration)
if iteration % args.print_every == 0 or iteration+1 == len(data_loader):
print("%s Batch %04d/%i, Loss %9.4f, NLL-Loss %9.4f, KL-Loss %9.4f, KL-Weight %6.3f, NLL-word-Loss %9.4f"
%(split.upper(), iteration, len(data_loader)-1, loss.data[0], NLL_loss.data[0]/batch_size, KL_loss.data[0]/batch_size, KL_weight,NLL_w_avg))
#split = 'invalid' #JUST TO DEBUG!!!
if split == 'valid':
if 'target_sents' not in tracker:
tracker['target_sents'] = list()
tracker['target_sents'] += idx2word(batch['target'].data, i2w=datasets['train'].get_i2w(), pad_idx=datasets['train'].pad_idx) #ERROR HERE!!!
tracker['z'] = torch.cat((tracker['z'], z.data), dim=0)
print("%s Epoch %02d/%i, Mean ELBO %9.4f"%(split.upper(), epoch, args.epochs, torch.mean(tracker['ELBO'])))
if args.tensorboard_logging:
writer.add_scalar("%s-Epoch/ELBO"%split.upper(), torch.mean(tracker['ELBO']), epoch)
# save a dump of all sentences and the encoded latent space
if split == 'valid':
dump = {'target_sents':tracker['target_sents'], 'z':tracker['z'].tolist()}
if not os.path.exists(os.path.join('./dumps', ts)):
os.makedirs('dumps/'+ts)
with open(os.path.join('./dumps/'+ts+'/valid_E%i.json'%epoch), 'w') as dump_file:
json.dump(dump,dump_file)
# save checkpoint
if split == 'train' and epoch %10 ==0 :
checkpoint_path = os.path.join(save_model_path, "E%i.pytorch"%(epoch))
torch.save(model.state_dict(), checkpoint_path)
print("Model saved at %s"%checkpoint_path)
def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
splits = ['train', 'valid'] + (['test'] if args.test else [])
datasets = OrderedDict()
for split in splits:
datasets[split] = PTB(data_dir=args.data_dir,
split=split,
create_data=args.create_data,
max_sequence_length=args.max_sequence_length,
min_occ=args.min_occ)
model = SentenceVAE(vocab_size=datasets['train'].vocab_size,
sos_idx=datasets['train'].sos_idx,
eos_idx=datasets['train'].eos_idx,
pad_idx=datasets['train'].pad_idx,
unk_idx=datasets['train'].unk_idx,
max_sequence_length=args.max_sequence_length,
embedding_size=args.embedding_size,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
word_dropout=args.word_dropout,
embedding_dropout=args.embedding_dropout,
latent_size=args.latent_size,
num_layers=args.num_layers,
bidirectional=args.bidirectional)
if torch.cuda.is_available():
model = model.cuda()
print(model)
if args.tensorboard_logging:
writer = SummaryWriter(
os.path.join(args.logdir, experiment_name(args, ts)))
writer.add_text("model", str(model))
writer.add_text("args", str(args))
writer.add_text("ts", ts)
save_model_path = os.path.join(args.save_model_path, ts)
os.makedirs(save_model_path)
def sigmoid(step):
x = step - 6569.5
if x < 0:
a = np.exp(x)
res = (a / (1 + a))
else:
res = (1 / (1 + np.exp(-x)))
return float(res)
def frange_cycle_linear(n_iter, start=0.0, stop=1.0, n_cycle=4, ratio=0.5):
L = np.ones(n_iter) * stop
period = n_iter / n_cycle
step = (stop - start) / (period * ratio) # linear schedule
for c in range(n_cycle):
v, i = start, 0
while v <= stop and (int(i + c * period) < n_iter):
L[int(i + c * period)] = v
v += step
i += 1
return L
n_iter = 0
for epoch in range(args.epochs):
split = 'train'
data_loader = DataLoader(dataset=datasets[split],
batch_size=args.batch_size,
shuffle=split == 'train',
num_workers=cpu_count(),
pin_memory=torch.cuda.is_available())
for iteration, batch in enumerate(data_loader):
n_iter += 1
print("Total no of iterations = " + str(n_iter))
L = frange_cycle_linear(n_iter)
def kl_anneal_function(anneal_function, step):
if anneal_function == 'identity':
return 1
if anneal_function == 'sigmoid':
return sigmoid(step)
if anneal_function == 'cyclic':
return float(L[step])
ReconLoss = torch.nn.NLLLoss(size_average=False,
ignore_index=datasets['train'].pad_idx)
def loss_fn(logp,
target,
length,
mean,
logv,
anneal_function,
step,
split='train'):
# cut-off unnecessary padding from target, and flatten
target = target[:, :torch.max(length).data[0]].contiguous().view(-1)
logp = logp.view(-1, logp.size(2))
# Negative Log Likelihood
recon_loss = ReconLoss(logp, target)
# KL Divergence
KL_loss = -0.5 * torch.sum(1 + logv - mean.pow(2) - logv.exp())
if split == 'train':
KL_weight = kl_anneal_function(anneal_function, step)
else:
KL_weight = 1
return recon_loss, KL_loss, KL_weight
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
tensor = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.Tensor
step = 0
for epoch in range(args.epochs):
for split in splits:
data_loader = DataLoader(dataset=datasets[split],
batch_size=args.batch_size,
shuffle=split == 'train',
num_workers=cpu_count(),
pin_memory=torch.cuda.is_available())
tracker = defaultdict(tensor)
# Enable/Disable Dropout
if split == 'train':
model.train()
else:
model.eval()
for iteration, batch in enumerate(data_loader):
batch_size = batch['input'].size(0)
for k, v in batch.items():
if torch.is_tensor(v):
batch[k] = to_var(v)
# Forward pass
logp, mean, logv, z = model(batch['input'], batch['length'])
# loss calculation
recon_loss, KL_loss, KL_weight = loss_fn(
logp, batch['target'], batch['length'], mean, logv,
args.anneal_function, step, split)
if split == 'train':
loss = (recon_loss + KL_weight * KL_loss) / batch_size
else:
# report complete elbo when validation
loss = (recon_loss + KL_loss) / batch_size
# backward + optimization
if split == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
# bookkeepeing
tracker['negELBO'] = torch.cat((tracker['negELBO'], loss.data))
if args.tensorboard_logging:
writer.add_scalar("%s/Negative_ELBO" % split.upper(),
loss.data[0],
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/Recon_Loss" % split.upper(),
recon_loss.data[0] / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL_Loss" % split.upper(),
KL_loss.data[0] / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL_Weight" % split.upper(),
KL_weight,
epoch * len(data_loader) + iteration)
if iteration % args.print_every == 0 or iteration + 1 == len(
data_loader):
logger.info(
"%s Batch %04d/%i, Loss %9.4f, Recon-Loss %9.4f, KL-Loss %9.4f, KL-Weight %6.3f"
% (split.upper(), iteration, len(data_loader) - 1,
loss.data[0], recon_loss.data[0] / batch_size,
KL_loss.data[0] / batch_size, KL_weight))
if split == 'valid':
if 'target_sents' not in tracker:
tracker['target_sents'] = list()
tracker['target_sents'] += idx2word(
batch['target'].data,
i2w=datasets['train'].get_i2w(),
pad_idx=datasets['train'].pad_idx)
tracker['z'] = torch.cat((tracker['z'], z.data), dim=0)
logger.info("%s Epoch %02d/%i, Mean Negative ELBO %9.4f" %
(split.upper(), epoch, args.epochs,
torch.mean(tracker['negELBO'])))
if args.tensorboard_logging:
writer.add_scalar("%s-Epoch/NegELBO" % split.upper(),
torch.mean(tracker['negELBO']), epoch)
# save a dump of all sentences and the encoded latent space
if split == 'valid':
dump = {
'target_sents': tracker['target_sents'],
'z': tracker['z'].tolist()
}
if not os.path.exists(os.path.join('dumps', ts)):
os.makedirs('dumps/' + ts)
with open(
os.path.join('dumps/' + ts +
'/valid_E%i.json' % epoch),
'w') as dump_file:
json.dump(dump, dump_file)
# save checkpoint
if split == 'train':
checkpoint_path = os.path.join(save_model_path,
"E%i.pytorch" % (epoch))
torch.save(model.state_dict(), checkpoint_path)
logger.info("Model saved at %s" % checkpoint_path)
def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
splits = ['train', 'valid'] + (['test'] if args.test else [])
datasets = OrderedDict()
for split in splits:
datasets[split] = PTB(data_dir=args.data_dir,
split=split,
create_data=args.create_data,
max_sequence_length=args.max_sequence_length,
min_occ=args.min_occ)
log_file = open("res.txt", "a")
log_file.write(expierment_name(args, ts))
log_file.write("\n")
graph_file = open("elbo-graph.txt", "a")
graph_file.write(expierment_name(args, ts))
graph_file.write("\n")
model = SentenceVAE(vocab_size=datasets['train'].vocab_size,
sos_idx=datasets['train'].sos_idx,
eos_idx=datasets['train'].eos_idx,
pad_idx=datasets['train'].pad_idx,
unk_idx=datasets['train'].unk_idx,
max_sequence_length=args.max_sequence_length,
embedding_size=args.embedding_size,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
word_dropout=args.word_dropout,
embedding_dropout=args.embedding_dropout,
latent_size=args.latent_size,
num_layers=args.num_layers,
bidirectional=args.bidirectional)
if torch.cuda.is_available():
model = model.cuda()
print(model)
if args.tensorboard_logging:
writer = SummaryWriter(
os.path.join(args.logdir, expierment_name(args, ts)))
writer.add_text("model", str(model))
writer.add_text("args", str(args))
writer.add_text("ts", ts)
save_model_path = os.path.join(args.save_model_path, ts)
os.makedirs(save_model_path)
def kl_anneal_function(anneal_function, step, k, x0):
if anneal_function == 'logistic':
return float(1 / (1 + np.exp(-k * (step - x0))))
elif anneal_function == 'linear':
return min(1, step / x0)
elif anneal_function == "softplus":
return min(1, np.log(1 + np.exp(k * step)))
elif anneal_function == "no":
return 1
NLL = torch.nn.NLLLoss(size_average=False,
ignore_index=datasets['train'].pad_idx)
def loss_fn(logp, target, length, mean, logv, anneal_function, step, k,
x0):
# cut-off unnecessary padding from target, and flatten
target = target[:, :torch.max(length).data[0]].contiguous().view(-1)
logp = logp.view(-1, logp.size(2))
# Negative Log Likelihood
NLL_loss = NLL(logp, target)
# KL Divergence
KL_loss = -0.5 * torch.sum(1 + logv - mean.pow(2) - logv.exp())
KL_weight = kl_anneal_function(anneal_function, step, k, x0)
return NLL_loss, KL_loss, KL_weight
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
tensor = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.Tensor
step = 0
val_lowest_elbo = 5000
val_accu_epoch = 0
val_min_epoch = 0
split_elbo = {"train": [], "valid": []}
if args.test:
split_elbo["test"] = []
split_loss = {"train": [], "valid": []}
if args.test:
split_loss["test"] = []
for epoch in range(args.epochs):
for split in splits:
data_loader = DataLoader(dataset=datasets[split],
batch_size=args.batch_size,
shuffle=split == 'train',
num_workers=cpu_count(),
pin_memory=torch.cuda.is_available())
tracker = defaultdict(tensor)
# Enable/Disable Dropout
if split == 'train':
model.train()
else:
model.eval()
for iteration, batch in enumerate(data_loader):
batch_size = batch['input'].size(0)
for k, v in batch.items():
if torch.is_tensor(v):
batch[k] = to_var(v)
# Forward pass
logp, mean, logv, z = model(batch['input'], batch['length'])
# loss calculation
NLL_loss, KL_loss, KL_weight = loss_fn(logp, batch['target'],
batch['length'], mean,
logv,
args.anneal_function,
step, args.k, args.x0)
if split != 'train':
KL_weight = 1.0
loss = (NLL_loss + KL_weight * KL_loss) / batch_size
# backward + optimization
if split == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
# bookkeepeing
tracker['ELBO'] = torch.cat((tracker['ELBO'], loss.data))
if args.tensorboard_logging:
writer.add_scalar("%s/ELBO" % split.upper(), loss.data[0],
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/NLL Loss" % split.upper(),
NLL_loss.data[0] / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL Loss" % split.upper(),
KL_loss.data[0] / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL Weight" % split.upper(),
KL_weight,
epoch * len(data_loader) + iteration)
if iteration % args.print_every == 0 or iteration + 1 == len(
data_loader):
print(
"%s Batch %04d/%i, Loss %9.4f, NLL-Loss %9.4f, KL-Loss %9.4f, KL-Weight %6.3f"
% (split.upper(), iteration, len(data_loader) - 1,
loss.data[0], NLL_loss.data[0] / batch_size,
KL_loss.data[0] / batch_size, KL_weight))
split_loss[split].append([
loss.data[0], NLL_loss.data[0] / batch_size,
KL_loss.data[0] / batch_size
])
if split == 'valid':
if 'target_sents' not in tracker:
tracker['target_sents'] = list()
tracker['target_sents'] += idx2word(
batch['target'].data,
i2w=datasets['train'].get_i2w(),
pad_idx=datasets['train'].pad_idx)
tracker['z'] = torch.cat((tracker['z'], z.data), dim=0)
print("%s Epoch %02d/%i, Mean ELBO %9.4f" %
(split.upper(), epoch, args.epochs,
torch.mean(tracker['ELBO'])))
split_elbo[split].append([torch.mean(tracker["ELBO"])])
if args.tensorboard_logging:
writer.add_scalar("%s-Epoch/ELBO" % split.upper(),
torch.mean(tracker['ELBO']), epoch)
# save a dump of all sentences and the encoded latent space
if split == 'valid':
dump = {
'target_sents': tracker['target_sents'],
'z': tracker['z'].tolist()
}
if not os.path.exists(os.path.join('dumps', ts)):
os.makedirs('dumps/' + ts)
with open(
os.path.join('dumps/' + ts +
'/valid_E%i.json' % epoch),
'w') as dump_file:
json.dump(dump, dump_file)
# save checkpoint
if split == 'train':
checkpoint_path = os.path.join(save_model_path,
"E%i.pytorch" % (epoch))
torch.save(model.state_dict(), checkpoint_path)
print("Model saved at %s" % checkpoint_path)
if split == 'valid':
if torch.mean(tracker['ELBO']) < val_lowest_elbo:
val_lowest_elbo = torch.mean(tracker['ELBO'])
val_accu_epoch = 0
val_min_epoch = epoch
else:
val_accu_epoch += 1
if val_accu_epoch >= 3:
if not args.test:
exp_str = ""
exp_str += "train_ELBO={}\n".format(
split_elbo["train"][val_min_epoch])
exp_str += "valid_ELBO={}\n".format(
split_elbo["valid"][val_min_epoch])
exp_str += "==========\n"
log_file.write(exp_str)
log_file.close()
print(exp_str)
graph_file.write("ELBO\n")
line = ""
for s in splits:
for i in split_loss[s]:
line += "{},".format(i[0])
line += "\n"
graph_file.write(line)
graph_file.write("NLL\n")
line = ""
for s in splits:
for i in split_loss[s]:
line += "{},".format(i[1])
line += "\n"
graph_file.write(line)
graph_file.write("KL\n")
line = ""
for s in splits:
for i in split_loss[s]:
line += "{},".format(i[2])
line += "\n"
graph_file.write(line)
graph_file.close()
exit()
elif split == 'test' and val_accu_epoch >= 3:
exp_str = ""
exp_str += "train_ELBO={}\n".format(
split_elbo["train"][val_min_epoch])
exp_str += "valid_ELBO={}\n".format(
split_elbo["valid"][val_min_epoch])
exp_str += "test_ELBO={}\n".format(
split_elbo["test"][val_min_epoch])
exp_str += "==========\n"
log_file.write(exp_str)
log_file.close()
print(exp_str)
graph_file.write("ELBO\n")
line = ""
for s in splits:
for i in split_loss[s]:
line += "{},".format(i[0])
line += "\n"
for s in splits:
for i in split_elbo[s]:
line += "{},".format(i[0])
line += "\n"
graph_file.write(line)
graph_file.write("NLL\n")
line = ""
for s in splits:
for i in split_loss[s]:
line += "{},".format(i[1])
line += "\n"
graph_file.write(line)
graph_file.write("KL\n")
line = ""
for s in splits:
for i in split_loss[s]:
line += "{},".format(i[2])
line += "\n"
graph_file.write(line)
graph_file.close()
exit()
if epoch == args.epochs - 1:
exp_str = ""
exp_str += "train_ELBO={}\n".format(
split_elbo["train"][val_min_epoch])
exp_str += "valid_ELBO={}\n".format(
split_elbo["valid"][val_min_epoch])
if args.test:
exp_str += "test_ELBO={}\n".format(
split_elbo["test"][val_min_epoch])
exp_str += "==========\n"
log_file.write(exp_str)
log_file.close()
print(exp_str)
graph_file.write("ELBO\n")
line = ""
for s in splits:
for i in split_loss[s]:
line += "{},".format(i[0])
line += "\n"
graph_file.write(line)
graph_file.write("NLL\n")
line = ""
for s in splits:
for i in split_loss[s]:
line += "{},".format(i[1])
line += "\n"
graph_file.write(line)
graph_file.write("KL\n")
line = ""
for s in splits:
for i in split_loss[s]:
line += "{},".format(i[2])
line += "\n"
graph_file.write(line)
graph_file.close()
exit()
def main(args):
# Load the vocab
with open(args.data_dir+'/ptb.vocab.json', 'r') as file:
vocab = json.load(file)
w2i, i2w = vocab['w2i'], vocab['i2w']
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
splits = ['train', 'valid'] + (['test'] if args.test else [])
# Initialize semantic loss
sl = Semantic_Loss()
datasets = OrderedDict()
for split in splits:
datasets[split] = PTB(
data_dir=args.data_dir,
split=split,
create_data=args.create_data,
max_sequence_length=args.max_sequence_length,
min_occ=args.min_occ
)
params = dict(
vocab_size=datasets['train'].vocab_size,
sos_idx=datasets['train'].sos_idx,
eos_idx=datasets['train'].eos_idx,
pad_idx=datasets['train'].pad_idx,
unk_idx=datasets['train'].unk_idx,
max_sequence_length=args.max_sequence_length,
embedding_size=args.embedding_size,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
word_dropout=args.word_dropout,
embedding_dropout=args.embedding_dropout,
latent_size=args.latent_size,
num_layers=args.num_layers,
bidirectional=args.bidirectional
)
model = SentenceVAE(**params)
if torch.cuda.is_available():
model = model.cuda()
print(model)
if args.tensorboard_logging:
writer = SummaryWriter(os.path.join(args.logdir, expierment_name(args, ts)))
writer.add_text("model", str(model))
writer.add_text("args", str(args))
writer.add_text("ts", ts)
save_model_path = os.path.join(args.save_model_path, ts)
os.makedirs(save_model_path)
with open(os.path.join(save_model_path, 'model_params.json'), 'w') as f:
json.dump(params, f, indent=4)
def kl_anneal_function(anneal_function, step, k, x0):
if anneal_function == 'logistic':
return float(1/(1+np.exp(-k*(step-x0))))
elif anneal_function == 'linear':
return min(1, step/x0)
def perplexity_anneal_function(anneal_function, step, k, x0):
if anneal_function == 'logistic':
return float(1/ 1+np.exp(-k*(step-x0)))
elif anneal_function == 'linear':
return min(1, (step/x0))
NLL = torch.nn.NLLLoss(ignore_index=datasets['train'].pad_idx, reduction='sum')
def loss_fn(logp, target, length, mean, logv, anneal_function, step, k, x0, \
batch_perplexity, perplexity_anneal_function):
# cut-off unnecessary padding from target, and flatten
target = target[:, :torch.max(length).item()].contiguous().view(-1)
logp = logp.view(-1, logp.size(2))
# Negative Log Likelihood
NLL_loss = NLL(logp, target)
# KL Divergence
KL_loss = -0.5 * torch.sum(1 + logv - mean.pow(2) - logv.exp())
KL_weight = kl_anneal_function(anneal_function, step, k, x0)
# Perplexity
perp_loss = batch_perplexity
perp_weight = perplexity_anneal_function(anneal_function, step, k, x0)
return NLL_loss, KL_loss, KL_weight, perp_loss, perp_weight
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
tensor = torch.cuda.FloatTensor if torch.cuda.is_available() else torch.Tensor
step = 0
for epoch in range(args.epochs):
# Keep track of epoch loss
epoch_loss = []
for split in splits:
data_loader = DataLoader(
dataset=datasets[split],
batch_size=args.batch_size,
shuffle=split=='train',
num_workers=cpu_count(),
pin_memory=torch.cuda.is_available()
)
tracker = defaultdict(tensor)
# Enable/Disable Dropout
if split == 'train':
model.train()
else:
model.eval()
batch_t_start = None
for iteration, batch in enumerate(data_loader):
if batch_t_start:
batch_run_time = time.time() - batch_t_start
# print("Batch run time: " + str(batch_run_time))
batch_t_start = time.time()
batch_size = batch['input_sequence'].size(0)
for k, v in batch.items():
if torch.is_tensor(v):
batch[k] = to_var(v)
# Get the original sentences in this batch
batch_sentences = idx2word(batch['input_sequence'], i2w=i2w, pad_idx=w2i['<pad>'])
# Remove the first tag
batch_sentences = [x.replace("<sos>", "") for x in batch_sentences]
# Forward pass
(logp, mean, logv, z), states = model(**batch)
# Choose some random pairs of samples within the batch
# to get latent representations for
batch_index_pairs = list(itertools.combinations(np.arange(batch_size), 2))
random.shuffle(batch_index_pairs)
batch_index_pairs = batch_index_pairs[:args.perplexity_samples_per_batch]
batch_perplexity = []
# If we start the perplexity
start_perplexity = epoch > 10
# If we should have perplexity loss
if start_perplexity and args.perplexity_loss:
# For each pair, get the intermediate representations in the latent space
for index_pair in batch_index_pairs:
with torch.no_grad():
z1_hidden = states['z'][index_pair[0]].cpu()
z2_hidden = states['z'][index_pair[1]].cpu()
z_hidden = to_var(torch.from_numpy(interpolate(start=z1_hidden, end=z2_hidden, steps=1)).float())
if args.rnn_type == "lstm":
with torch.no_grad():
z1_cell_state = states['z_cell_state'].cpu().squeeze()[index_pair[0]]
z2_cell_state = states['z_cell_state'].cpu().squeeze()[index_pair[1]]
z_cell_states = \
to_var(torch.from_numpy(interpolate(start=z1_cell_state, end=z2_cell_state, steps=1)).float())
samples, _ = model.inference(z=z_hidden, z_cell_state=z_cell_states)
else:
samples, _ = model.inference(z=z_hidden, z_cell_state=None)
# Check interpolated sentences
interpolated_sentences = idx2word(samples, i2w=i2w, pad_idx=w2i['<pad>'])
# For each sentence, get the perplexity and show it
perplexities = []
for sentence in interpolated_sentences:
perplexities.append(sl.get_perplexity(sentence))
avg_sample_perplexity = sum(perplexities) / len(perplexities)
batch_perplexity.append(avg_sample_perplexity)
# Calculate batch perplexity
avg_batch_perplexity = sum(batch_perplexity) / len(batch_perplexity)
# loss calculation
NLL_loss, KL_loss, KL_weight, perp_loss, perp_weight = loss_fn(logp, batch['target'],
batch['length'], mean, logv, args.anneal_function, step, \
args.k, args.x0, avg_batch_perplexity, perplexity_anneal_function)
loss = ((NLL_loss + KL_weight * KL_loss) / batch_size) + (perp_loss * perp_weight)
else: # Epochs < X, so train without perplexity
# loss calculation
NLL_loss, KL_loss, KL_weight, perp_loss, perp_weight = loss_fn(logp, batch['target'],
batch['length'], mean, logv, args.anneal_function, step, \
args.k, args.x0, 0, perplexity_anneal_function)
loss = (NLL_loss + KL_weight * KL_loss) / batch_size
# Turn model back into train, since inference changed to eval
if split == 'train':
model.train()
else:
model.eval()
# backward + optimization
if split == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
# Add loss
epoch_loss.append(loss.item())
# bookkeepeing
tracker['ELBO'] = torch.cat((tracker['ELBO'], loss.data.view(1, -1)), dim=0)
if args.tensorboard_logging:
writer.add_scalar("%s/ELBO" % split.upper(), loss.item(), epoch*len(data_loader) + iteration)
writer.add_scalar("%s/NLL Loss" % split.upper(), NLL_loss.item() / batch_size,
epoch*len(data_loader) + iteration)
writer.add_scalar("%s/KL Loss" % split.upper(), KL_loss.item() / batch_size,
epoch*len(data_loader) + iteration)
writer.add_scalar("%s/KL Weight" % split.upper(), KL_weight,
epoch*len(data_loader) + iteration)
if iteration % args.print_every == 0 or iteration+1 == len(data_loader):
print("%s Batch %04d/%i, Loss %9.4f, NLL-Loss %9.4f, KL-Loss %9.4f, KL-Weight %6.3f, Perp-loss %9.4f, Perp-weight %6.3f"
% (split.upper(), iteration, len(data_loader)-1, loss.item(), NLL_loss.item()/batch_size,
KL_loss.item()/batch_size, KL_weight, perp_loss, perp_weight))
if split == 'valid':
if 'target_sents' not in tracker:
tracker['target_sents'] = list()
tracker['target_sents'] += idx2word(batch['target'].data, i2w=datasets['train'].get_i2w(),
pad_idx=datasets['train'].pad_idx)
tracker['z'] = torch.cat((tracker['z'], z.data), dim=0)
print("%s Epoch %02d/%i, Mean ELBO %9.4f" % (split.upper(), epoch, args.epochs, tracker['ELBO'].mean()))
if args.tensorboard_logging:
writer.add_scalar("%s-Epoch/ELBO" % split.upper(), torch.mean(tracker['ELBO']), epoch)
# save a dump of all sentences and the encoded latent space
if split == 'valid':
dump = {'target_sents': tracker['target_sents'], 'z': tracker['z'].tolist()}
if not os.path.exists(os.path.join('dumps', ts)):
os.makedirs('dumps/'+ts)
with open(os.path.join('dumps/'+ts+'/valid_E%i.json' % epoch), 'w') as dump_file:
json.dump(dump,dump_file)
# save checkpoint
if split == 'train':
checkpoint_path = os.path.join(save_model_path, "E%i.pytorch" % epoch)
torch.save(model.state_dict(), checkpoint_path)
print("Model saved at %s" % checkpoint_path)
def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
splits = ['train', 'valid'] + (['test'] if args.test else [])
RANDOM_SEED = 42
dataset = load_dataset("yelp_polarity", split="train")
TRAIN_SIZE = len(dataset) - 2_000
VALID_SIZE = 1_000
TEST_SIZE = 1_000
train_test_split = dataset.train_test_split(train_size=TRAIN_SIZE,
seed=RANDOM_SEED)
train_dataset = train_test_split["train"]
test_val_dataset = train_test_split["test"].train_test_split(
train_size=VALID_SIZE, test_size=TEST_SIZE, seed=RANDOM_SEED)
val_dataset, test_dataset = test_val_dataset["train"], test_val_dataset[
"test"]
tokenizer = AutoTokenizer.from_pretrained(args.model_name, use_fast=True)
datasets = OrderedDict()
datasets['train'] = TextDataset(train_dataset, tokenizer,
args.max_sequence_length,
not args.disable_sent_tokenize)
datasets['valid'] = TextDataset(val_dataset, tokenizer,
args.max_sequence_length,
not args.disable_sent_tokenize)
if args.test:
datasets['text'] = TextDataset(test_dataset, tokenizer,
args.max_sequence_length,
not args.disable_sent_tokenize)
print(
f"Loading {args.model_name} model. Setting {args.trainable_layers} trainable layers."
)
encoder = AutoModel.from_pretrained(args.model_name, return_dict=True)
if not args.train_embeddings:
for p in encoder.embeddings.parameters():
p.requires_grad = False
encoder_layers = encoder.encoder.layer
if args.trainable_layers > len(encoder_layers):
warnings.warn(
f"You are asking to train {args.trainable_layers} layers, but this model has only {len(encoder_layers)}"
)
for layer in range(len(encoder_layers) - args.trainable_layers):
for p in encoder_layers[layer].parameters():
p.requires_grad = False
params = dict(vocab_size=datasets['train'].vocab_size,
embedding_size=args.embedding_size,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
word_dropout=args.word_dropout,
embedding_dropout=args.embedding_dropout,
latent_size=args.latent_size,
num_layers=args.num_layers,
bidirectional=args.bidirectional,
max_sequence_length=args.max_sequence_length)
model = SentenceVAE(encoder=encoder, tokenizer=tokenizer, **params)
if torch.cuda.is_available():
model = model.cuda()
print(model)
if args.tensorboard_logging:
writer = SummaryWriter(
os.path.join(args.logdir, expierment_name(args, ts)))
writer.add_text("model", str(model))
writer.add_text("args", str(args))
writer.add_text("ts", ts)
save_model_path = os.path.join(args.save_model_path, ts)
os.makedirs(save_model_path)
with open(os.path.join(save_model_path, 'model_params.json'), 'w') as f:
json.dump(params, f, indent=4)
with open(os.path.join(save_model_path, 'train_args.json'), 'w') as f:
json.dump(vars(args), f, indent=4)
def kl_anneal_function(anneal_function, step, k, x0):
if step <= x0:
return args.initial_kl_weight
if anneal_function == 'logistic':
return float(1 / (1 + np.exp(-k * (step - x0 - 2500))))
elif anneal_function == 'linear':
return min(1, step / x0)
NLL = torch.nn.NLLLoss(ignore_index=datasets['train'].pad_idx,
reduction='sum')
def loss_fn(logp, target, length, mean, logv, anneal_function, step, k,
x0):
# cut-off unnecessary padding from target, and flatten
target = target[:, :torch.max(length).item()].contiguous().view(-1)
logp = logp.view(-1, logp.size(2))
# Negative Log Likelihood
NLL_loss = NLL(logp, target)
# KL Divergence
KL_loss = -0.5 * torch.sum(1 + logv - mean.pow(2) - logv.exp())
KL_weight = kl_anneal_function(anneal_function, step, k, x0)
return NLL_loss, KL_loss, KL_weight
params = [{
'params': model.encoder.parameters(),
'lr': args.encoder_learning_rate
}, {
'params': [
*model.decoder_rnn.parameters(), *model.hidden2mean.parameters(),
*model.hidden2logv.parameters(), *model.latent2hidden.parameters(),
*model.outputs2vocab.parameters()
]
}]
optimizer = torch.optim.Adam(params,
lr=args.learning_rate,
weight_decay=args.weight_decay)
tensor = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.Tensor
step = 0
for epoch in range(args.epochs):
for split in splits:
data_loader = DataLoader(dataset=datasets[split],
batch_size=args.batch_size,
shuffle=(split == 'train'),
num_workers=cpu_count(),
pin_memory=torch.cuda.is_available(),
collate_fn=DataCollator(tokenizer))
tracker = defaultdict(tensor)
# Enable/Disable Dropout
if split == 'train':
model.train()
else:
model.eval()
for iteration, batch in enumerate(data_loader):
batch_size = batch['input'].size(0)
for k, v in batch.items():
if torch.is_tensor(v):
batch[k] = to_var(v)
# Forward pass
logp, mean, logv, z = model(batch['input'],
batch['attention_mask'],
batch['length'])
# loss calculation
NLL_loss, KL_loss, KL_weight = loss_fn(logp, batch['target'],
batch['length'], mean,
logv,
args.anneal_function,
step, args.k, args.x0)
loss = (NLL_loss + KL_weight * KL_loss) / batch_size
# backward + optimization
if split == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
# bookkeepeing
tracker['ELBO'] = torch.cat(
(tracker['ELBO'], loss.data.view(1, -1)), dim=0)
if args.tensorboard_logging:
writer.add_scalar("%s/ELBO" % split.upper(), loss.item(),
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/NLL Loss" % split.upper(),
NLL_loss.item() / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL Loss" % split.upper(),
KL_loss.item() / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL Weight" % split.upper(),
KL_weight,
epoch * len(data_loader) + iteration)
if iteration % args.print_every == 0 or iteration + 1 == len(
data_loader):
print(
"%s Batch %04d/%i, Loss %9.4f, NLL-Loss %9.4f, KL-Loss %9.4f, KL-Weight %6.3f"
% (split.upper(), iteration, len(data_loader) - 1,
loss.item(), NLL_loss.item() / batch_size,
KL_loss.item() / batch_size, KL_weight))
if split == 'valid':
if 'target_sents' not in tracker:
tracker['target_sents'] = list()
tracker['target_sents'] += idx2word(
batch['target'].tolist(), tokenizer=tokenizer)
tracker['z'] = torch.cat((tracker['z'], z.data), dim=0)
print("%s Epoch %02d/%i, Mean ELBO %9.4f" %
(split.upper(), epoch, args.epochs, tracker['ELBO'].mean()))
if args.tensorboard_logging:
writer.add_scalar("%s-Epoch/ELBO" % split.upper(),
torch.mean(tracker['ELBO']), epoch)
# save a dump of all sentences, the encoded latent space and generated sequences
if split == 'valid':
samples, _ = model.inference(z=tracker['z'])
generated_sents = idx2word(samples.tolist(), tokenizer)
sents = [{
'original': target,
'generated': generated
} for target, generated in zip(tracker['target_sents'],
generated_sents)]
dump = {'sentences': sents, 'z': tracker['z'].tolist()}
if not os.path.exists(os.path.join('dumps', ts)):
os.makedirs('dumps/' + ts)
with open(
os.path.join('dumps/' + ts +
'/valid_E%i.json' % epoch),
'w') as dump_file:
json.dump(dump, dump_file, indent=3)
# save checkpoint
if split == 'train':
checkpoint_path = os.path.join(save_model_path,
"E%i.pytorch" % epoch)
torch.save(model.state_dict(), checkpoint_path)
print("Model saved at %s" % checkpoint_path)
def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
splits = ['train', 'valid']
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
filename=os.path.join(args.logdir,
experiment_name(args, ts) + ".log"))
logger = logging.getLogger(__name__)
datasets = OrderedDict()
for split in splits:
datasets[split] = PTB(data_dir=args.data_dir,
split=split,
create_data=args.create_data,
max_sequence_length=args.max_sequence_length,
min_occ=args.min_occ)
model = SentenceVAE(vocab_size=datasets['train'].vocab_size,
sos_idx=datasets['train'].sos_idx,
eos_idx=datasets['train'].eos_idx,
pad_idx=datasets['train'].pad_idx,
unk_idx=datasets['train'].unk_idx,
max_sequence_length=args.max_sequence_length,
embedding_size=args.embedding_size,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
word_dropout=args.word_dropout,
embedding_dropout=args.embedding_dropout,
latent_size=args.latent_size,
num_layers=args.num_layers,
bidirectional=args.bidirectional)
if torch.cuda.is_available():
model = model.cuda()
print(model)
if args.tensorboard_logging:
writer = SummaryWriter(
os.path.join(args.logdir, experiment_name(args, ts)))
writer.add_text("model", str(model))
writer.add_text("args", str(args))
writer.add_text("ts", ts)
save_model_path = os.path.join(args.save_model_path, ts)
os.makedirs(save_model_path)
total_step = int(args.epochs * 42000.0 / args.batch_size)
def kl_anneal_function(anneal_function, step):
if anneal_function == 'half':
return 0.5
if anneal_function == 'identity':
return 1
if anneal_function == 'double':
return 2
if anneal_function == 'quadra':
return 4
if anneal_function == 'sigmoid':
return 1 / (1 + np.exp((0.5 * total_step - step) / 200))
if anneal_function == 'monotonic':
beta = step * 4 / total_step
if beta > 1:
beta = 1.0
return beta
if anneal_function == 'cyclical':
t = total_step / 4
beta = 4 * (step % t) / t
if beta > 1:
beta = 1.0
return beta
ReconLoss = torch.nn.NLLLoss(reduction='sum',
ignore_index=datasets['train'].pad_idx)
def loss_fn(logp, target, length, mean, logv, anneal_function, step):
# cut-off unnecessary padding from target, and flatten
target = target[:, :torch.max(length).item()].contiguous().view(-1)
logp = logp.view(-1, logp.size(2))
# Negative Log Likelihood
recon_loss = ReconLoss(logp, target)
# KL Divergence
KL_loss = -0.5 * torch.sum(1 + logv - mean.pow(2) - logv.exp())
KL_weight = kl_anneal_function(anneal_function, step)
return recon_loss, KL_loss, KL_weight
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
tensor = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.Tensor
step = 0
train_loss = []
test_loss = []
for epoch in range(args.epochs):
for split in splits:
data_loader = DataLoader(dataset=datasets[split],
batch_size=args.batch_size,
shuffle=split == 'train',
num_workers=cpu_count(),
pin_memory=torch.cuda.is_available())
tracker = defaultdict(list)
# Enable/Disable Dropout
if split == 'train':
model.train()
else:
model.eval()
for iteration, batch in enumerate(data_loader):
batch_size = batch['input'].size(0)
for k, v in batch.items():
if torch.is_tensor(v):
batch[k] = to_var(v)
# Forward pass
logp, mean, logv, z = model(batch['input'], batch['length'])
# loss calculation
recon_loss, KL_loss, KL_weight = loss_fn(
logp, batch['target'], batch['length'], mean, logv,
args.anneal_function, step)
if split == 'train':
loss = (recon_loss + KL_weight * KL_loss) / batch_size
else:
# report complete elbo when validation
loss = (recon_loss + KL_loss) / batch_size
# backward + optimization
if split == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
# bookkeepeing
# tracker['negELBO'] = torch.cat((tracker['negELBO'], loss.data))
tracker["negELBO"].append(loss.item())
tracker["recon_loss"].append(recon_loss.item() / batch_size)
tracker["KL_Loss"].append(KL_loss.item() / batch_size)
tracker["KL_Weight"].append(KL_weight)
if args.tensorboard_logging:
writer.add_scalar("%s/Negative_ELBO" % split.upper(),
loss.item(),
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/Recon_Loss" % split.upper(),
recon_loss.item() / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL_Loss" % split.upper(),
KL_loss.item() / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL_Weight" % split.upper(),
KL_weight,
epoch * len(data_loader) + iteration)
if iteration % args.print_every == 0 or iteration + 1 == len(
data_loader):
logger.info(
"\tStep\t%s\t%04d\t%i\t%9.4f\t%9.4f\t%9.4f\t%6.3f" %
(split.upper(), iteration, len(data_loader) - 1,
loss.item(), recon_loss.item() / batch_size,
KL_loss.item() / batch_size, KL_weight))
print(
"%s Batch %04d/%i, Loss %9.4f, Recon-Loss %9.4f, KL-Loss %9.4f, KL-Weight %6.3f"
% (split.upper(), iteration, len(data_loader) - 1,
loss.item(), recon_loss.item() / batch_size,
KL_loss.item() / batch_size, KL_weight))
if split == 'valid':
if 'target_sents' not in tracker:
tracker['target_sents'] = list()
tracker['target_sents'] += idx2word(
batch['target'].data,
i2w=datasets['train'].get_i2w(),
pad_idx=datasets['train'].pad_idx)
tracker['z'].append(z.data.tolist())
logger.info(
"\tEpoch\t%s\t%02d\t%i\t%9.4f\t%9.4f\t%9.4f\t%6.3f" %
(split.upper(), epoch, args.epochs,
sum(tracker['negELBO']) / len(tracker['negELBO']),
1.0 * sum(tracker['recon_loss']) / len(tracker['recon_loss']),
1.0 * sum(tracker['KL_Loss']) / len(tracker['KL_Loss']),
1.0 * sum(tracker['KL_Weight']) / len(tracker['KL_Weight'])))
print("%s Epoch %02d/%i, Mean Negative ELBO %9.4f" %
(split.upper(), epoch, args.epochs,
sum(tracker['negELBO']) / len(tracker['negELBO'])))
if args.tensorboard_logging:
writer.add_scalar(
"%s-Epoch/NegELBO" % split.upper(),
1.0 * sum(tracker['negELBO']) / len(tracker['negELBO']),
epoch)
writer.add_scalar(
"%s-Epoch/recon_loss" % split.upper(), 1.0 *
sum(tracker['recon_loss']) / len(tracker['recon_loss']),
epoch)
writer.add_scalar(
"%s-Epoch/KL_Loss" % split.upper(),
1.0 * sum(tracker['KL_Loss']) / len(tracker['KL_Loss']),
epoch)
writer.add_scalar(
"%s-Epoch/KL_Weight" % split.upper(), 1.0 *
sum(tracker['KL_Weight']) / len(tracker['KL_Weight']),
epoch)
if split == 'train':
train_loss.append(1.0 * sum(tracker['negELBO']) /
len(tracker['negELBO']))
else:
test_loss.append(1.0 * sum(tracker['negELBO']) /
len(tracker['negELBO']))
# save a dump of all sentences and the encoded latent space
if split == 'valid':
dump = {
'target_sents': tracker['target_sents'],
'z': tracker['z']
}
if not os.path.exists(os.path.join('dumps', ts)):
os.makedirs('dumps/' + ts)
with open(
os.path.join('dumps/' + ts +
'/valid_E%i.json' % epoch),
'w') as dump_file:
json.dump(dump, dump_file)
# save checkpoint
if split == 'train':
checkpoint_path = os.path.join(save_model_path,
"E%i.pytorch" % (epoch))
torch.save(model.state_dict(), checkpoint_path)
print("Model saved at %s" % checkpoint_path)
sns.set(style="whitegrid")
df = pd.DataFrame()
df["train"] = train_loss
df["test"] = test_loss
ax = sns.lineplot(data=df, legend=False)
ax.set(xlabel='Epoch', ylabel='Loss')
plt.legend(title='Split', loc='upper right', labels=['Train', 'Test'])
plt.savefig(os.path.join(args.logdir,
experiment_name(args, ts) + ".png"),
transparent=True,
dpi=300)
def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
splits = ['train', 'valid'] + (['test'] if args.test else [])
datasets = OrderedDict()
for split in splits:
if args.dataset == 'ptb':
Dataset = PTB
elif args.dataset == 'twitter':
Dataset = PoliticianTweets
else:
print("Invalid dataset. Exiting")
exit()
datasets[split] = Dataset(
data_dir=args.data_dir,
split=split,
create_data=args.create_data,
max_sequence_length=args.max_sequence_length,
min_occ=args.min_occ
)
model = SentenceVAE(
vocab_size=datasets['train'].vocab_size,
sos_idx=datasets['train'].sos_idx,
eos_idx=datasets['train'].eos_idx,
pad_idx=datasets['train'].pad_idx,
unk_idx=datasets['train'].unk_idx,
max_sequence_length=args.max_sequence_length,
embedding_size=args.embedding_size,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
word_dropout=args.word_dropout,
embedding_dropout=args.embedding_dropout,
latent_size=args.latent_size,
num_layers=args.num_layers,
bidirectional=args.bidirectional
)
# if args.from_file != "":
# model = torch.load(args.from_file)
#
if torch.cuda.is_available():
model = model.cuda()
print(model)
if args.tensorboard_logging:
writer = SummaryWriter(os.path.join(args.logdir, experiment_name(args,ts)))
writer.add_text("model", str(model))
writer.add_text("args", str(args))
writer.add_text("ts", ts)
save_model_path = os.path.join(args.save_model_path, ts)
os.makedirs(save_model_path)
if 'sigmoid' in args.anneal_function and args.dataset=='ptb':
linspace = np.linspace(-5,5,13160) # 13160 = number of training examples in ptb
elif 'sigmoid' in args.anneal_function and args.dataset=='twitter':
linspace = np.linspace(-5, 5, 25190) #6411/25190? = number of training examples in short version of twitter
def kl_anneal_function(anneal_function, step, param_dict=None):
if anneal_function == 'identity':
return 1
elif anneal_function == 'sigmoid' or anneal_function=='sigmoid_klt':
s = 1/(len(linspace))
return(float((1)/(1+np.exp(-param_dict['ag']*(linspace[step])))))
NLL = torch.nn.NLLLoss(size_average=False, ignore_index=datasets['train'].pad_idx)
def loss_fn(logp, target, length, mean, logv, anneal_function, step, param_dict=None):
# cut-off unnecessary padding from target, and flatten
target = target[:, :torch.max(length).data[0]].contiguous().view(-1)
logp = logp.view(-1, logp.size(2))
# Negative Log Likelihood
NLL_loss = NLL(logp, target)
# KL Divergence
KL_loss = -0.5 * torch.sum(1 + logv - mean.pow(2) - logv.exp())
if args.anneal_function == 'sigmoid_klt':
if float(KL_loss)/args.batch_size < param_dict['kl_threshold']:
# print("KL_loss of %s is below threshold %s. Returning this threshold instead"%(float(KL_loss)/args.batch_size,param_dict['kl_threshold']))
KL_loss = to_var(torch.Tensor([param_dict['kl_threshold']*args.batch_size]))
KL_weight = kl_anneal_function(anneal_function, step, {'ag': args.anneal_aggression})
return NLL_loss, KL_loss, KL_weight
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
tensor = torch.cuda.FloatTensor if torch.cuda.is_available() else torch.Tensor
step = 0
for epoch in range(args.epochs):
for split in splits:
data_loader = DataLoader(
dataset=datasets[split],
batch_size=args.batch_size,
shuffle=split=='train',
num_workers=0,
pin_memory=torch.cuda.is_available()
)
tracker = defaultdict(tensor)
# Enable/Disable Dropout
if split == 'train':
model.train()
else:
model.eval()
for iteration, batch in enumerate(data_loader):
batch_size = batch['input'].size(0)
if split == 'train' and batch_size != args.batch_size:
print("WARNING: Found different batch size\nargs.batch_size= %s, input_size=%s"%(args.batch_size, batch_size))
for k, v in batch.items():
if torch.is_tensor(v):
batch[k] = to_var(v)
# Forward pass
logp, mean, logv, z = model(batch['input'], batch['length'])
# loss calculation
NLL_loss, KL_loss, KL_weight = loss_fn(logp, batch['target'],
batch['length'], mean, logv, args.anneal_function, step, {'kl_threshold': args.kl_threshold})
loss = (NLL_loss + KL_weight * KL_loss)/batch_size
# backward + optimization
if split == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
# print(step)
# bookkeepeing
tracker['ELBO'] = torch.cat((tracker['ELBO'], loss.data))
if args.tensorboard_logging:
writer.add_scalar("%s/ELBO"%split.upper(), loss.data[0], epoch*len(data_loader) + iteration)
writer.add_scalar("%s/NLL_Loss"%split.upper(), NLL_loss.data[0]/batch_size, epoch*len(data_loader) + iteration)
writer.add_scalar("%s/KL_Loss"%split.upper(), KL_loss.data[0]/batch_size, epoch*len(data_loader) + iteration)
# print("Step %s: %s"%(epoch*len(data_loader) + iteration, KL_weight))
writer.add_scalar("%s/KL_Weight"%split.upper(), KL_weight, epoch*len(data_loader) + iteration)
if iteration % args.print_every == 0 or iteration+1 == len(data_loader):
logger.info("%s Batch %04d/%i, Loss %9.4f, NLL-Loss %9.4f, KL-Loss %9.4f, KL-Weight %6.3f"
%(split.upper(), iteration, len(data_loader)-1, loss.data[0], NLL_loss.data[0]/batch_size, KL_loss.data[0]/batch_size, KL_weight))
if split == 'valid':
if 'target_sents' not in tracker:
tracker['target_sents'] = list()
tracker['target_sents'] += idx2word(batch['target'].data, i2w=datasets['train'].get_i2w(), pad_idx=datasets['train'].pad_idx)
tracker['z'] = torch.cat((tracker['z'], z.data), dim=0)
logger.info("%s Epoch %02d/%i, Mean ELBO %9.4f"%(split.upper(), epoch, args.epochs, torch.mean(tracker['ELBO'])))
if args.tensorboard_logging:
writer.add_scalar("%s-Epoch/ELBO"%split.upper(), torch.mean(tracker['ELBO']), epoch)
# save a dump of all sentences and the encoded latent space
if split == 'valid':
dump = {'target_sents':tracker['target_sents'], 'z':tracker['z'].tolist()}
if not os.path.exists(os.path.join('dumps', ts)):
os.makedirs('dumps/'+ts)
with open(os.path.join('dumps/'+ts+'/valid_E%i.json'%epoch), 'w') as dump_file:
json.dump(dump,dump_file)
# save checkpoint
if split == 'train':
checkpoint_path = os.path.join(save_model_path, "E%i.pytorch"%(epoch))
torch.save(model.state_dict(), checkpoint_path)
logger.info("Model saved at %s"%checkpoint_path)
torch.save(model, f"model-{args.dataset}-{ts}.pickle")
def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
splits = ['train', 'valid'] + (['test'] if args.test else [])
datasets = OrderedDict()
for split in splits:
datasets[split] = PTB(data_dir=args.data_dir,
split=split,
create_data=args.create_data,
max_sequence_length=args.max_sequence_length,
min_occ=args.min_occ)
model = SentenceVAE(vocab_size=datasets['train'].vocab_size,
sos_idx=datasets['train'].sos_idx,
eos_idx=datasets['train'].eos_idx,
pad_idx=datasets['train'].pad_idx,
unk_idx=datasets['train'].unk_idx,
max_sequence_length=args.max_sequence_length,
embedding_size=args.embedding_size,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
word_dropout=args.word_dropout,
embedding_dropout=args.embedding_dropout,
latent_size=args.latent_size,
num_layers=args.num_layers,
bidirectional=args.bidirectional)
if torch.cuda.is_available():
model = model.cuda()
print(model)
if args.tensorboard_logging:
writer = SummaryWriter(
os.path.join(args.logdir, experiment_name(args, ts)))
writer.add_text("model", str(model))
writer.add_text("args", str(args))
writer.add_text("ts", ts)
save_model_path = os.path.join(args.save_model_path, ts)
os.makedirs(save_model_path)
total_steps = (len(datasets["train"]) // args.batch_size) * args.epochs
print("Train dataset size", total_steps)
def kl_anneal_function(anneal_function, step):
if anneal_function == 'identity':
return 1
if anneal_function == 'linear':
if args.warmup is None:
return 1 - (total_steps - step) / total_steps
else:
warmup_steps = (total_steps / args.epochs) * args.warmup
return 1 - (warmup_steps - step
) / warmup_steps if step < warmup_steps else 1.0
ReconLoss = torch.nn.NLLLoss(size_average=False,
ignore_index=datasets['train'].pad_idx)
def loss_fn(logp, target, length, mean, logv, anneal_function, step):
# cut-off unnecessary padding from target, and flatten
target = target[:, :torch.max(length).data[0]].contiguous().view(-1)
logp = logp.view(-1, logp.size(2))
# Negative Log Likelihood
recon_loss = ReconLoss(logp, target)
# KL Divergence
KL_loss = -0.5 * torch.sum(1 + logv - mean.pow(2) - logv.exp())
KL_weight = kl_anneal_function(anneal_function, step)
return recon_loss, KL_loss, KL_weight
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
tensor = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.Tensor
step = 0
for epoch in range(args.epochs):
for split in splits:
data_loader = DataLoader(dataset=datasets[split],
batch_size=args.batch_size,
shuffle=split == 'train',
num_workers=cpu_count(),
pin_memory=torch.cuda.is_available())
tracker = defaultdict(tensor)
# Enable/Disable Dropout
if split == 'train':
model.train()
else:
model.eval()
for iteration, batch in enumerate(data_loader):
batch_size = batch['input'].size(0)
for k, v in batch.items():
if torch.is_tensor(v):
batch[k] = to_var(v)
# Forward pass
logp, mean, logv, z = model(batch['input'], batch['length'])
# loss calculation
recon_loss, KL_loss, KL_weight = loss_fn(
logp, batch['target'], batch['length'], mean, logv,
args.anneal_function, step)
if split == 'train':
loss = (recon_loss + KL_weight * KL_loss) / batch_size
else:
# report complete elbo when validation
loss = (recon_loss + KL_loss) / batch_size
# backward + optimization
if split == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
# bookkeepeing
tracker['negELBO'] = torch.cat(
(tracker['negELBO'], loss.data.unsqueeze(0)))
if args.tensorboard_logging:
neg_elbo = (recon_loss + KL_loss) / batch_size
writer.add_scalar("%s/Negative_ELBO" % split.upper(),
neg_elbo.data[0],
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/Recon_Loss" % split.upper(),
recon_loss.data[0] / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL_Loss" % split.upper(),
KL_loss.data[0] / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL_Weight" % split.upper(),
KL_weight,
epoch * len(data_loader) + iteration)
if iteration % args.print_every == 0 or iteration + 1 == len(
data_loader):
logger.info(
"%s Batch %04d/%i, Loss %9.4f, Recon-Loss %9.4f, KL-Loss %9.4f, KL-Weight %6.3f"
% (split.upper(), iteration, len(data_loader) - 1,
loss.data[0], recon_loss.data[0] / batch_size,
KL_loss.data[0] / batch_size, KL_weight))
if split == 'valid':
if 'target_sents' not in tracker:
tracker['target_sents'] = list()
tracker['target_sents'] += idx2word(
batch['target'].data,
i2w=datasets['train'].get_i2w(),
pad_idx=datasets['train'].pad_idx)
tracker['z'] = torch.cat((tracker['z'], z.data), dim=0)
logger.info("%s Epoch %02d/%i, Mean Negative ELBO %9.4f" %
(split.upper(), epoch, args.epochs,
torch.mean(tracker['negELBO'])))
if args.tensorboard_logging:
writer.add_scalar("%s-Epoch/NegELBO" % split.upper(),
torch.mean(tracker['negELBO']), epoch)
# save a dump of all sentences and the encoded latent space
if split == 'valid':
dump = {
'target_sents': tracker['target_sents'],
'z': tracker['z'].tolist()
}
if not os.path.exists(os.path.join('dumps', ts)):
os.makedirs('dumps/' + ts)
with open(
os.path.join('dumps/' + ts +
'/valid_E%i.json' % epoch),
'w') as dump_file:
json.dump(dump, dump_file)
# save checkpoint
if split == 'train':
checkpoint_path = os.path.join(save_model_path,
"E%i.pytorch" % (epoch))
torch.save(model.state_dict(), checkpoint_path)
logger.info("Model saved at %s" % checkpoint_path)
if args.num_samples:
torch.cuda.empty_cache()
model.eval()
with torch.no_grad():
print(f"Generating {args.num_samples} samples")
generations, _ = model.inference(n=args.num_samples)
vocab = datasets["train"].i2w
print(
"Sampled latent codes from z ~ N(0, I), generated sentences:")
for i, generation in enumerate(generations, start=1):
sentence = [vocab[str(word.item())] for word in generation]
print(f"{i}:", " ".join(sentence))
def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
splits = ['train', 'valid'] + (['test'] if args.test else [])
datasets = OrderedDict()
for split in splits:
datasets[split] = PTB(data_dir=args.data_dir,
split=split,
create_data=args.create_data,
max_sequence_length=args.max_sequence_length,
min_occ=args.min_occ)
model = SentenceVAE(vocab_size=datasets['train'].vocab_size,
sos_idx=datasets['train'].sos_idx,
eos_idx=datasets['train'].eos_idx,
pad_idx=datasets['train'].pad_idx,
unk_idx=datasets['train'].unk_idx,
max_sequence_length=args.max_sequence_length,
embedding_size=args.embedding_size,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
word_dropout=args.word_dropout,
embedding_dropout=args.embedding_dropout,
latent_size=args.latent_size,
num_layers=args.num_layers,
bidirectional=args.bidirectional)
if torch.cuda.is_available():
model = model.cuda()
print(model)
if args.tensorboard_logging:
writer = SummaryWriter(
os.path.join(args.logdir, experiment_name(args, ts)))
writer.add_text("model", str(model))
writer.add_text("args", str(args))
writer.add_text("ts", ts)
save_model_path = os.path.join(args.save_model_path, ts)
os.makedirs(save_model_path)
def kl_anneal_function(anneal_function, step, x1, x2):
if anneal_function == 'identity':
return 1
elif anneal_function == 'linear':
return min(1, step / x1)
elif anneal_function == 'logistic':
return float(1 / (1 + np.exp(-x2 * (step - x1))))
elif anneal_function == 'cyclic_log':
return float(1 / (1 + np.exp(-x2 * ((step % (3 * x1)) - x1))))
elif anneal_function == 'cyclic_lin':
return min(1, (step % (3 * x1)) / x1)
ReconLoss = torch.nn.NLLLoss(size_average=False,
ignore_index=datasets['train'].pad_idx)
def loss_fn(logp, target, length, mean, logv, anneal_function, step, x1,
x2):
# cut-off unnecessary padding from target, and flatten
target = target[:, :torch.max(length).item()].contiguous().view(-1)
logp = logp.view(-1, logp.size(2))
# Negative Log Likelihood
recon_loss = ReconLoss(logp, target)
# KL Divergence
KL_loss = -0.5 * torch.sum(1 + logv - mean.pow(2) - logv.exp())
KL_weight = kl_anneal_function(anneal_function, step, x1, x2)
return recon_loss, KL_loss, KL_weight
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
tensor = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.Tensor
step = 0
early_stopping = EarlyStopping(history=10)
for epoch in range(args.epochs):
early_stopping_flag = False
for split in splits:
data_loader = DataLoader(dataset=datasets[split],
batch_size=args.batch_size,
shuffle=split == 'train',
num_workers=cpu_count(),
pin_memory=torch.cuda.is_available())
# tracker = defaultdict(tensor)
tracker = defaultdict(list)
# Enable/Disable Dropout
if split == 'train':
model.train()
else:
model.eval()
for iteration, batch in enumerate(data_loader):
batch_size = batch['input'].size(0)
for k, v in batch.items():
if torch.is_tensor(v):
batch[k] = to_var(v)
# Forward pass
logp, mean, logv, z = model(batch['input'], batch['length'])
# loss calculation
recon_loss, KL_loss, KL_weight = loss_fn(
logp, batch['target'], batch['length'], mean, logv,
args.anneal_function, step, args.x1, args.x2)
if split == 'train':
loss = (recon_loss + KL_weight * KL_loss) / batch_size
else:
# report complete elbo when validation
loss = (recon_loss + KL_loss) / batch_size
# backward + optimization
if split == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
# bookkeepeing
tracker['negELBO'].append(loss.item())
if args.tensorboard_logging:
writer.add_scalar("%s/Negative_ELBO" % split.upper(),
loss.item(),
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/Recon_Loss" % split.upper(),
recon_loss.item() / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL_Loss" % split.upper(),
KL_loss.item() / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL_Weight" % split.upper(),
KL_weight,
epoch * len(data_loader) + iteration)
if iteration % args.print_every == 0 or iteration + 1 == len(
data_loader):
# print(step)
# logger.info("Step = %d"%step)
logger.info(
"%s Batch %04d/%i, Loss %9.4f, Recon-Loss %9.4f, KL-Loss %9.4f, KL-Weight %6.3f"
% (split.upper(), iteration, len(data_loader) - 1,
loss.item(), recon_loss.item() / batch_size,
KL_loss.item() / batch_size, KL_weight))
if split == 'valid':
if 'target_sents' not in tracker:
tracker['target_sents'] = list()
tracker['target_sents'] += idx2word(
batch['target'].data,
i2w=datasets['train'].get_i2w(),
pad_idx=datasets['train'].pad_idx)
# tracker['z'] = torch.cat((tracker['z'], z.data), dim=0)
# print(z.data.shape)
tracker['z'].append(z.data.tolist())
mean_loss = sum(tracker['negELBO']) / len(tracker['negELBO'])
logger.info("%s Epoch %02d/%i, Mean Negative ELBO %9.4f" %
(split.upper(), epoch, args.epochs, mean_loss))
# print(mean_loss)
if args.tensorboard_logging:
writer.add_scalar("%s-Epoch/NegELBO" % split.upper(),
mean_loss, epoch)
# save a dump of all sentences and the encoded latent space
if split == 'valid':
dump = {
'target_sents': tracker['target_sents'],
'z': tracker['z']
}
if not os.path.exists(os.path.join('dumps', ts)):
os.makedirs('dumps/' + ts)
with open(
os.path.join('dumps/' + ts +
'/valid_E%i.json' % epoch),
'w') as dump_file:
json.dump(dump, dump_file)
if (args.early_stopping):
if (early_stopping.check(mean_loss)):
early_stopping_flag = True
# save checkpoint
if split == 'train':
checkpoint_path = os.path.join(save_model_path,
"E%i.pytorch" % (epoch))
torch.save(model.state_dict(), checkpoint_path)
logger.info("Model saved at %s" % checkpoint_path)
if (early_stopping_flag):
print("Early stopping trigerred. Training stopped...")
break
def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
splits = ['train', 'valid'] + (['test'] if args.test else [])
datasets = OrderedDict()
for split in splits:
# datasets[split] = BGoogle(
# data_dir=args.data_dir,
# split=split,
# create_data=args.create_data,
# batch_size=args.batch_size ,
# max_sequence_length=args.max_sequence_length,
# min_occ=args.min_occ
# )
datasets[split] = Amazon(data_dir=args.data_dir,
split=split,
create_data=args.create_data,
batch_size=args.batch_size,
max_sequence_length=args.max_sequence_length,
min_occ=args.min_occ)
model = SentenceVAE(vocab_size=datasets['train'].vocab_size,
sos_idx=datasets['train'].sos_idx,
eos_idx=datasets['train'].eos_idx,
pad_idx=datasets['train'].pad_idx,
unk_idx=datasets['train'].unk_idx,
max_sequence_length=args.max_sequence_length,
embedding_size=args.embedding_size,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
word_dropout=args.word_dropout,
embedding_dropout=args.embedding_dropout,
latent_size=args.latent_size,
num_layers=args.num_layers,
bidirectional=args.bidirectional)
if torch.cuda.is_available():
model = model.cuda()
print(model)
tokenizer = TweetTokenizer(preserve_case=False)
vocab_file = "amazon.vocab.json"
with open(os.path.join(args.data_dir, vocab_file), 'r') as file:
vocab = json.load(file)
w2i, i2w = vocab['w2i'], vocab['i2w']
if args.tensorboard_logging:
writer = SummaryWriter(
os.path.join(args.logdir, expierment_name(args, ts)))
writer.add_text("model", str(model))
writer.add_text("args", str(args))
writer.add_text("ts", ts)
# save_model_path = os.path.join(args.save_model_path, ts)
save_model_path = args.save_model_path
if not os.path.exists(save_model_path):
os.makedirs(save_model_path)
def kl_anneal_function(anneal_function, step, k, x0):
if anneal_function == 'logistic':
return float(1 / (1 + np.exp(-k * (step - x0))))
elif anneal_function == 'linear':
return min(1, step / x0)
NLL = torch.nn.NLLLoss(size_average=False,
ignore_index=datasets['train'].pad_idx)
def loss_fn(logp, target, length, mean, logv, anneal_function, step, k,
x0):
# cut-off unnecessary padding from target, and flatten
target = target[:, :torch.max(length).data].contiguous().view(-1)
logp = logp.view(-1, logp.size(2))
# Negative Log Likelihood
NLL_loss = NLL(logp, target)
# KL Divergence
KL_loss = -0.5 * torch.sum(1 + logv - mean.pow(2) - logv.exp())
KL_weight = kl_anneal_function(anneal_function, step, k, x0)
return NLL_loss, KL_loss, KL_weight
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
tensor = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.Tensor
step = 0
save_mode = True
last_ELBO = 1e32
for epoch in range(args.epochs):
print("+" * 20)
# f_test_example(model, tokenizer, w2i, i2w)
for split in splits:
# data_loader = DataLoader(
# dataset=datasets[split],
# batch_size=args.batch_size,
# shuffle=split=='train',
# num_workers=cpu_count(),
# pin_memory=torch.cuda.is_available()
# )
batch_size = args.batch_size
tracker = defaultdict(tensor)
# Enable/Disable Dropout
if split == 'train':
model.train()
else:
model.eval()
# for iteration, batch in enumerate(data_loader):
iteration = 0
iteration_total = datasets[split].batch_num
print("batch_num", iteration_total)
for input_batch_tensor, target_batch_tensor, length_batch_tensor in datasets[
split]:
if torch.is_tensor(input_batch_tensor):
input_batch_tensor = to_var(input_batch_tensor)
if torch.is_tensor(target_batch_tensor):
target_batch_tensor = to_var(target_batch_tensor)
if torch.is_tensor(length_batch_tensor):
length_batch_tensor = to_var(length_batch_tensor)
# batch_size = batch['input'].size(0)
# for k, v in batch.items():
# if torch.is_tensor(v):
# batch[k] = to_var(v)
# Forward pass
# logp, mean, logv, z = model(batch['input'], batch['length'])
logp, mean, logv, z = model(input_batch_tensor,
length_batch_tensor)
# loss calculation
NLL_loss, KL_loss, KL_weight = loss_fn(
logp, target_batch_tensor, length_batch_tensor, mean, logv,
args.anneal_function, step, args.k, args.x0)
loss = (NLL_loss + KL_weight * KL_loss) / batch_size
# backward + optimization
if split == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
iteration += 1
# bookkeepeing
# print("elbo", tracker['ELBO'])
# print("loss", loss)
if iteration == 0:
tracker['ELBO'] = loss.data
tracker['ELBO'] = tracker['ELBO'].view(1)
else:
tracker['ELBO'] = torch.cat(
(tracker['ELBO'], loss.view(1)))
if args.tensorboard_logging:
# print(loss.data)
writer.add_scalar("%s/ELBO" % split.upper(),
loss.data.item(),
epoch * iteration_total + iteration)
writer.add_scalar("%s/NLL Loss" % split.upper(),
NLL_loss.data.item() / batch_size,
epoch * iteration_total + iteration)
writer.add_scalar("%s/KL Loss" % split.upper(),
KL_loss.data.item() / batch_size,
epoch * iteration_total + iteration)
writer.add_scalar("%s/KL Weight" % split.upper(),
KL_weight,
epoch * iteration_total + iteration)
if iteration % args.print_every == 0 or iteration + 1 == iteration_total:
print(
"%s Batch %04d/%i, Loss %9.4f, NLL-Loss %9.4f, KL-Loss %9.4f, KL-Weight %6.3f"
% (split.upper(), iteration, iteration_total - 1,
loss.data.item(), NLL_loss.data.item() / batch_size,
KL_loss.data.item() / batch_size, KL_weight))
# if split == 'valid':
# if 'target_sents' not in tracker:
# tracker['target_sents'] = list()
# tracker['target_sents'] += idx2word(batch['target'].data, i2w=datasets['train'].get_i2w(), pad_idx=datasets['train'].pad_idx)
# # print("z", tracker['z'], z)
# tracker['z'] = torch.cat((tracker['z'], z.data), dim=0)
# break
print("%s Epoch %02d/%i, Mean ELBO %9.4f" %
(split.upper(), epoch, args.epochs,
torch.mean(tracker['ELBO'])))
cur_ELBO = torch.mean(tracker['ELBO'])
if args.tensorboard_logging:
writer.add_scalar("%s-Epoch/ELBO" % split.upper(), cur_ELBO,
epoch)
if split == "valid":
if cur_ELBO < last_ELBO:
save_mode = True
else:
save_mode = False
last_ELBO = cur_ELBO
# save a dump of all sentences and the encoded latent space
# if split == 'valid':
# dump = {'target_sents':tracker['target_sents'], 'z':tracker['z'].tolist()}
# if not os.path.exists(os.path.join('dumps', ts)):
# os.makedirs('dumps/'+ts)
# with open(os.path.join('dumps/'+ts+'/valid_E%i.json'%epoch), 'w') as dump_file:
# json.dump(dump,dump_file)
# save checkpoint
if split == 'train':
# checkpoint_path = os.path.join(save_model_path, "E%i.pytorch"%(epoch))
checkpoint_path = os.path.join(save_model_path, "best.pytorch")
if save_mode == True:
torch.save(model.state_dict(), checkpoint_path)
print("Model saved at %s" % checkpoint_path)
def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.localtime())
splits = ['train', 'valid'] + (['test'] if args.test else [])
datasets = OrderedDict()
for split in splits:
datasets[split] = PTB(data_dir=args.data_dir,
split=split,
create_data=args.create_data,
max_sequence_length=args.max_sequence_length,
min_occ=args.min_occ,
use_bert=args. False)
model = SentenceVAE(alphabet_size=datasets['train'].alphabet_size,
vocab_size=datasets['train'].vocab_size,
sos_idx=datasets['train'].sos_idx,
eos_idx=datasets['train'].eos_idx,
pad_idx=datasets['train'].pad_idx,
unk_idx=datasets['train'].unk_idx,
max_sequence_length=args.max_sequence_length,
embedding_size=args.embedding_size,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
word_dropout=args.word_dropout,
embedding_dropout=args.embedding_dropout,
latent_size=args.latent_size,
num_layers=args.num_layers,
bidirectional=args.bidirectional)
if torch.cuda.is_available():
model = model.cuda()
print(model)
if args.tensorboard_logging:
writer = SummaryWriter(
os.path.join(args.logdir, expierment_name(args, ts)))
writer.add_text("model", str(model))
writer.add_text("args", str(args))
writer.add_text("ts", ts)
save_model_path = os.path.join(args.save_model_path, ts)
os.makedirs(save_model_path)
print("Saving model to directory: " + save_model_path)
def kl_anneal_function(anneal_function, step, k, x0):
if anneal_function == 'logistic':
return float(1 / (1 + np.exp(-k * (step - x0))))
elif anneal_function == 'linear':
return min(1, step / x0)
def word_weight_function(step, k, x0):
return float(1 / (1 + np.exp(-k * (step - x0))))
NLL = torch.nn.NLLLoss(reduction='sum',
ignore_index=datasets['train'].pad_idx)
def loss_fn(def_logp, word_logp, def_target, def_length, word_target,
word_length, mean, logv):
# cut-off unnecessary padding from target definition, and flatten
def_target = def_target[:, :torch.max(def_length).item()].contiguous(
).view(-1)
def_logp = def_logp.view(-1, def_logp.size(2))
# Negative Log Likelihood
def_NLL_loss = NLL(def_logp, def_target)
# cut off padding for words
word_target = word_target[:, :torch.max(word_length).item(
)].contiguous().view(-1)
word_logp = word_logp.view(-1, word_logp.size(2))
# Word NLL
word_NLL_loss = NLL(word_logp, word_target)
# KL Divergence
KL_loss = -0.5 * torch.sum(1 + logv - mean.pow(2) - logv.exp())
return def_NLL_loss, word_NLL_loss, KL_loss
def get_weights(anneal_function, step, k, x0):
# for logistic function, k = growth rate
KL_weight = kl_anneal_function(anneal_function, step, k, x0)
word_weight = word_weight_function(step, k, x0)
return {'def': 1, 'word': word_weight, 'kl': KL_weight}
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
tensor = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.Tensor
step = 0
for epoch in range(args.epochs):
for split in splits:
data_loader = DataLoader(dataset=datasets[split],
batch_size=args.batch_size,
shuffle=split == 'train',
num_workers=cpu_count(),
pin_memory=torch.cuda.is_available())
tracker = defaultdict(tensor)
# Enable/Disable Dropout
if split == 'train':
model = model.train()
else:
model = model.eval()
for iteration, batch in enumerate(data_loader):
batch_size = batch['input'].size(0)
for k, v in batch.items():
if torch.is_tensor(v):
batch[k] = to_var(v)
# Forward pass
[def_logp,
word_logp], mean, logv, z = model(batch['input'],
batch['length'],
batch['word_length'])
# loss calculation
def_NLL_loss, word_NLL_loss, KL_loss = loss_fn(
def_logp, word_logp, batch['target'], batch['length'],
batch['word'], batch['word_length'], mean, logv)
weights = get_weights(args.anneal_function, step, args.k,
args.x0)
loss = (weights['def'] * def_NLL_loss + weights['word'] *
word_NLL_loss + weights['kl'] * KL_loss) / batch_size
mean_logv = torch.mean(logv)
# backward + optimization
if split == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
# bookkeepeing
tracker['ELBO'] = torch.cat(
(tracker['ELBO'], loss.detach().unsqueeze(0)))
if args.tensorboard_logging:
writer.add_scalar("%s/ELBO" % split.upper(), loss.item(),
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/Def NLL Loss" % split.upper(),
def_NLL_loss.item() / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/Word NLL Loss" % split.upper(),
word_NLL_loss.item() / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL Loss" % split.upper(),
KL_loss.item() / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL Weight" % split.upper(),
weights['kl'],
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/Word Weight" % split.upper(),
weights['word'],
epoch * len(data_loader) + iteration)
if iteration % args.print_every == 0 or iteration + 1 == len(
data_loader):
print(
"%s Batch %04d/%i, Loss %9.4f, Def NLL-Loss %9.4f, Word NLL-Loss %9.4f Word-Weight %6.3f, KL-Loss %9.4f, KL-Weight %6.3f KL-VAL %9.4f"
% (split.upper(), iteration, len(data_loader) - 1,
loss.item(), def_NLL_loss.item() / batch_size,
word_NLL_loss.item() / batch_size, weights['word'],
KL_loss.item() / batch_size, weights['kl'],
mean_logv))
if split == 'valid':
if 'target_sents' not in tracker:
tracker['target_sents'] = list()
tracker['target_sents'] += idx2word(
batch['target'],
i2w=datasets['train'].get_i2w(),
pad_idx=datasets['train'].pad_idx)
tracker['z'] = torch.cat((tracker['z'], z.data), dim=0)
print("%s Epoch %02d/%i, Mean ELBO %9.4f" %
(split.upper(), epoch, args.epochs,
torch.mean(tracker['ELBO'])))
if args.tensorboard_logging:
writer.add_scalar("%s-Epoch/ELBO" % split.upper(),
torch.mean(tracker['ELBO']), epoch)
# save a dump of all sentences and the encoded latent space
if split == 'valid':
dump = {
'target_sents': tracker['target_sents'],
'z': tracker['z'].tolist()
}
if not os.path.exists(os.path.join('dumps', ts)):
os.makedirs('dumps/' + ts)
with open(
os.path.join('dumps/' + ts +
'/valid_E%i.json' % epoch),
'w') as dump_file:
json.dump(dump, dump_file)
# save checkpoint
if split == 'train':
checkpoint_path = os.path.join(save_model_path,
"E%i.pytorch" % (epoch))
torch.save(model.state_dict(), checkpoint_path)
print("Model saved at %s" % checkpoint_path)
def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
ptb = PTB(vocab_file=args.vocab_file,
train_file=args.train_file,
train_with_vocab=False,
create_data=args.create_data,
max_sequence_length=args.max_sequence_length,
min_occ=args.min_occ)
datasets = PTBDataset(ptb)
print('done preprocessing data')
model = SentenceVAE(vocab_size=datasets.vocab_size,
sos_idx=datasets.sos_idx,
eos_idx=datasets.eos_idx,
pad_idx=datasets.pad_idx,
unk_idx=datasets.unk_idx,
max_sequence_length=args.max_sequence_length,
embedding_size=args.embedding_size,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
word_dropout=args.word_dropout,
embedding_dropout=args.embedding_dropout,
latent_size=args.latent_size,
num_layers=args.num_layers,
bidirectional=args.bidirectional)
model.ptb = ptb
if torch.cuda.is_available():
model = model.cuda()
print(model)
if args.tensorboard_logging:
writer = SummaryWriter(
os.path.join(args.logdir, expierment_name(args, ts)))
writer.add_text("model", str(model))
writer.add_text("args", str(args))
writer.add_text("ts", ts)
save_model_path = os.path.join(args.save_model_path, ts)
os.makedirs(save_model_path)
def kl_anneal_function(anneal_function, step, k, x0):
if anneal_function == 'logistic':
return float(1 / (1 + np.exp(-k * (step - x0))))
elif anneal_function == 'linear':
return min(1, step / x0)
NLL = torch.nn.NLLLoss(size_average=False, ignore_index=datasets.pad_idx)
def loss_fn(logp, target, length, mean, logv, anneal_function, step, k,
x0):
# cut-off unnecessary padding from target, and flatten
#target = target[:, :torch.max(length).data[0]].contiguous().view(-1)
target = target[:, :torch.max(length).data].contiguous().view(-1)
logp = logp.view(-1, logp.size(2))
# Negative Log Likelihood
NLL_loss = NLL(logp, target)
# KL Divergence
KL_loss = -0.5 * torch.sum(1 + logv - mean.pow(2) - logv.exp())
KL_weight = kl_anneal_function(anneal_function, step, k, x0)
return NLL_loss, KL_loss, KL_weight
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
tensor = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.Tensor
step = 0
model.train()
split = 'train'
for epoch in range(args.epochs):
data_loader = DataLoader(dataset=datasets,
batch_size=args.batch_size,
shuffle=True,
num_workers=cpu_count(),
pin_memory=torch.cuda.is_available())
tracker = defaultdict(list)
for iteration, batch in enumerate(data_loader):
batch_size = batch['input'].size(0)
for k, v in batch.items():
if torch.is_tensor(v):
batch[k] = to_var(v)
# Forward pass
#logp, mean, logv, z = model(batch['input'], batch['length'])
logp, mean, logv, z, encoder_last = model(batch['input'],
batch['length'])
# loss calculation
NLL_loss, KL_loss, KL_weight = loss_fn(logp, batch['target'],
batch['length'], mean, logv,
args.anneal_function, step,
args.k, args.x0)
loss = (NLL_loss + KL_weight * KL_loss) / batch_size
# backward + optimization
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
# bookkeepeing
tracker['ELBO'].append(loss.data.cpu().numpy().tolist())
if args.tensorboard_logging:
writer.add_scalar("%s/ELBO" % split.upper(), loss.data,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/NLL Loss" % split.upper(),
NLL_loss.data / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL Loss" % split.upper(),
KL_loss.data / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL Weight" % split.upper(), KL_weight,
epoch * len(data_loader) + iteration)
if iteration % args.print_every == 0 or iteration + 1 == len(
data_loader):
print(
"%s Batch %04d/%i, Loss %9.4f, NLL-Loss %9.4f, KL-Loss %9.4f, KL-Weight %6.3f"
% (split.upper(), iteration, len(data_loader) - 1,
loss.data, NLL_loss.data / batch_size,
KL_loss.data / batch_size, KL_weight))
if split == 'valid':
if 'target_sents' not in tracker:
tracker['target_sents'] = list()
tracker['target_sents'] += idx2word(batch['target'].data,
i2w=datasets.get_i2w(),
pad_idx=datasets.pad_idx)
tracker['z'].append(z.data)
print("%s Epoch %02d/%i, Mean ELBO %9.4f" %
(split.upper(), epoch, args.epochs, np.mean(tracker['ELBO'])))
if args.tensorboard_logging:
writer.add_scalar("%s-Epoch/ELBO" % split.upper(),
np.mean(tracker['ELBO']), epoch)
'''
# save a dump of all sentences and the encoded latent space
if split == 'valid':
dump = {'target_sents':tracker['target_sents'], 'z':tracker['z']}
if not os.path.exists(os.path.join('dumps', ts)):
os.makedirs('dumps/'+ts)
with open(os.path.join('dumps/'+ts+'/valid_E%i.json'%epoch), 'w') as dump_file:
json.dump(dump,dump_file)
'''
# save checkpoint
if split == 'train':
checkpoint_path = os.path.join(save_model_path,
"E%i.pytorch" % (epoch))
torch.save(model.state_dict(), checkpoint_path)
joblib.dump(model.cpu(), checkpoint_path)
print("Model saved at %s" % checkpoint_path)
if torch.cuda.is_available():
model.cuda()
def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
print("Loading Vocab", args.vocab_path)
vocab = WordVocab.load_vocab(args.vocab_path)
print("Vocab Size: ", len(vocab))
print("Loading Train Dataset", args.train_dataset)
train_dataset = BERTDataset(args.train_dataset, vocab, seq_len=args.max_sequence_length,
corpus_lines=args.corpus_lines, on_memory=args.on_memory)
print("Loading Test Dataset", args.test_dataset)
test_dataset = BERTDataset(args.test_dataset, vocab, seq_len=args.max_sequence_length, on_memory=args.on_memory) \
if args.test_dataset is not None else None
print("Creating Dataloader")
train_data_loader = DataLoader(train_dataset, batch_size=args.batch_size, num_workers=args.num_workers)
test_data_loader = DataLoader(test_dataset, batch_size=args.batch_size, num_workers=args.num_workers) \
if test_dataset is not None else None
splits = ['train', 'test']
data_loaders = {
'train': train_data_loader,
'test': test_data_loader
}
model = SentenceVAE(
vocab_size=len(vocab),
sos_idx=vocab.sos_index,
eos_idx=vocab.eos_index,
pad_idx=vocab.pad_index,
unk_idx=vocab.unk_index,
max_sequence_length=args.max_sequence_length,
embedding_size=args.embedding_size,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
word_dropout=args.word_dropout,
embedding_dropout=args.embedding_dropout,
latent_size=args.latent_size,
num_layers=args.num_layers,
bidirectional=args.bidirectional
)
if torch.cuda.is_available():
model = model.cuda()
print(model)
if args.tensorboard_logging:
writer = SummaryWriter(os.path.join(args.logdir, expierment_name(args,ts)))
writer.add_text("model", str(model))
writer.add_text("args", str(args))
writer.add_text("ts", ts)
save_model_path = os.path.join(args.save_model_path)
if not os.path.exists(save_model_path):
os.makedirs(save_model_path)
def kl_anneal_function(anneal_function, step, k, x0):
if anneal_function == 'logistic':
return float(1/(1+np.exp(-k*(step-x0))))
elif anneal_function == 'linear':
return min(1, step/x0)
NLL = torch.nn.NLLLoss(size_average=False, ignore_index=vocab.pad_index)
def loss_fn(logp, target, length, mean, logv, anneal_function, step, k, x0):
# cut-off unnecessary padding from target, and flatten
# Negative Log Likelihood
NLL_loss = NLL(logp, target)
# KL Divergence
KL_loss = -0.5 * torch.sum(1 + logv - mean.pow(2) - logv.exp())
KL_weight = kl_anneal_function(anneal_function, step, k, x0)
return NLL_loss, KL_loss, KL_weight
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
tensor = torch.cuda.FloatTensor if torch.cuda.is_available() else torch.Tensor
step = 0
for epoch in range(args.epochs):
for split in splits:
data_loader = data_loaders[split]
tracker = defaultdict(tensor)
# Enable/Disable Dropout
if split == 'train':
model.train()
else:
model.eval()
correct = 0
close = 0
total = 0
for iteration, batch in enumerate(data_loader):
batch_size = batch['input'].size(0)
for k, v in batch.items():
if torch.is_tensor(v):
batch[k] = to_var(v)
# Forward pass
logp, mean, logv, z = model(batch['input'], batch['raw_length'])
# loss calculation
NLL_loss, KL_loss, KL_weight = loss_fn(logp, batch['target'],
batch['raw_length'], mean, logv, args.anneal_function, step, args.k, args.x0)
loss = (NLL_loss + KL_weight * KL_loss)/batch_size
# backward + optimization
if split == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
correct += logp.argmax(dim=1).eq(batch['target']).sum().item()
close += torch.mul(logp.argmax(dim=1).ge(batch["target"]-10), logp.argmax(dim=1).le(batch["target"]+10)).sum().item()
total += batch['target'].nelement()
# bookkeepeing
tracker['ELBO'] = torch.cat((tracker['ELBO'], loss.view(1,)))
if args.tensorboard_logging:
writer.add_scalar("%s/ELBO"%split.upper(), loss.data[0], epoch*len(data_loader) + iteration)
writer.add_scalar("%s/NLL Loss"%split.upper(), NLL_loss.data[0]/batch_size, epoch*len(data_loader) + iteration)
writer.add_scalar("%s/KL Loss"%split.upper(), KL_loss.data[0]/batch_size, epoch*len(data_loader) + iteration)
writer.add_scalar("%s/KL Weight"%split.upper(), KL_weight, epoch*len(data_loader) + iteration)
if iteration % args.print_every == 0 or iteration+1 == len(data_loader):
print("%s Batch %04d/%i, Loss %9.4f, NLL-Loss %9.4f, KL-Loss %9.4f, KL-Weight %6.3f"
%(split.upper(), iteration, len(data_loader)-1, loss.item(), NLL_loss.item()/batch_size, KL_loss.item()/batch_size, KL_weight))
if split == 'valid':
if 'target_sents' not in tracker:
tracker['target_sents'] = list()
tracker['target_sents'] += idx2word(batch['raw'].data, i2w=datasets['train'].get_i2w(), pad_idx=datasets['train'].pad_idx)
tracker['z'] = torch.cat((tracker['z'], z.data), dim=0)
print("%s Epoch %02d/%i, Mean ELBO %9.4f, acc %f, clo %f"%(split.upper(), epoch, args.epochs, torch.mean(tracker['ELBO']), correct/total, close/total))
if args.tensorboard_logging:
writer.add_scalar("%s-Epoch/ELBO"%split.upper(), torch.mean(tracker['ELBO']), epoch)
# save a dump of all sentences and the encoded latent space
if split == 'valid':
dump = {'target_sents':tracker['target_sents'], 'z':tracker['z'].tolist()}
if not os.path.exists(os.path.join('dumps', ts)):
os.makedirs('dumps/'+ts)
with open(os.path.join('dumps/'+ts+'/valid_E%i.json'%epoch), 'w') as dump_file:
json.dump(dump,dump_file)
# save checkpoint
if split == 'train':
checkpoint_path = os.path.join(save_model_path, "E%i.pytorch"%(epoch))
torch.save(model.state_dict(), checkpoint_path)
print("Model saved at %s"%checkpoint_path)
def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
splits = ['train', 'valid']
datasets = OrderedDict()
for split in splits:
datasets[split] = PoetryDataset(
data_dir=args.data_dir,
split=split,
create_data=args.create_data,
max_sequence_length=args.max_sequence_length,
min_occ=args.min_occ)
model = SentenceVAE(vocab_size=datasets['train'].vocab_size,
sos_idx=datasets['train'].sos_idx,
eos_idx=datasets['train'].eos_idx,
pad_idx=datasets['train'].pad_idx,
unk_idx=datasets['train'].unk_idx,
max_sequence_length=args.max_sequence_length,
embedding_size=args.embedding_size,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
word_dropout=args.word_dropout,
embedding_dropout=args.embedding_dropout,
latent_size=args.latent_size,
num_layers=args.num_layers,
bidirectional=args.bidirectional,
condition_size=7)
if torch.cuda.is_available():
model = model.cuda()
if args.tensorboard_logging:
writer = SummaryWriter(
os.path.join(args.logdir, expierment_name(args, ts)))
writer.add_text("model", str(model))
writer.add_text("args", str(args))
writer.add_text("ts", ts)
save_model_path = os.path.join(args.save_model_path, ts)
os.makedirs(save_model_path)
def kl_anneal_function(anneal_function, step, k, x0):
if anneal_function == 'logistic':
return float(1 / (1 + np.exp(-k * (step - x0))))
elif anneal_function == 'linear':
return min(1, step / x0)
NLL = torch.nn.NLLLoss(size_average=False,
ignore_index=datasets['train'].pad_idx)
def calculate_bleu_scores(original, decoded):
reference = original.split(' ')
hypothesis = decoded.split(' ')
return nltk.translate.bleu_score.sentence_bleu([reference], hypothesis)
def loss_fn(logp, target, length, mean, logv, anneal_function, step, k,
x0):
# cut-off unnecessary padding from target, and flatten
target = target[:, :torch.max(length)].contiguous().view(-1)
logp = logp.view(-1, logp.size(2))
# Negative Log Likelihood
NLL_loss = NLL(logp, target)
# KL Divergence
KL_loss = -0.5 * torch.sum(1 + logv - mean.pow(2) - logv.exp())
KL_weight = kl_anneal_function(anneal_function, step, k, x0)
return NLL_loss, KL_loss, KL_weight
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
tensor = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.Tensor
step = 0
for epoch in range(args.epochs):
total_BLEU_score = 0
for split in splits:
data_loader = DataLoader(dataset=datasets[split],
batch_size=args.batch_size,
shuffle=split == 'train',
num_workers=0,
pin_memory=torch.cuda.is_available())
tracker = defaultdict(tensor)
# Enable/Disable Dropout
if split == 'train':
model.train()
else:
model.eval()
for iteration, batch in enumerate(data_loader):
batch_size = batch['input'].size(0)
for k, v in batch.items():
if torch.is_tensor(v):
batch[k] = to_var(v)
# Forward pass
logp, mean, logv, z = model(
batch['input'],
batch['length'],
condition=batch['category'].float())
# logp, mean, logv, z = model(batch['input'], batch['length'], condition=None)
# loss calculation
NLL_loss, KL_loss, KL_weight = loss_fn(logp, batch['target'],
batch['length'], mean,
logv,
args.anneal_function,
step, args.k, args.x0)
loss = (NLL_loss + KL_weight * KL_loss) / batch_size
# backward + optimization
if split == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
# bookkeepeing
tracker['ELBO'] = torch.cat(
(tracker['ELBO'], loss.data.unsqueeze(0)))
if args.tensorboard_logging:
writer.add_scalar("%s/ELBO" % split.upper(), loss.data[0],
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/NLL Loss" % split.upper(),
NLL_loss.data[0] / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL Loss" % split.upper(),
KL_loss.data[0] / batch_size,
epoch * len(data_loader) + iteration)
writer.add_scalar("%s/KL Weight" % split.upper(),
KL_weight,
epoch * len(data_loader) + iteration)
if iteration % args.print_every == 0 or iteration + 1 == len(
data_loader):
print(
"%s Batch %04d/%i, Loss %9.4f, NLL-Loss %9.4f, KL-Loss %9.4f, KL-Weight %6.3f"
% (split.upper(), iteration, len(data_loader) - 1,
loss.data.item(), NLL_loss.data.item() / batch_size,
KL_loss.data.item() / batch_size, KL_weight))
if split == 'valid':
if 'target_sents' not in tracker:
tracker['target_sents'] = list()
tracker['target_sents'] += idx2word(
batch['target'].data,
i2w=datasets['train'].get_i2w(),
pad_idx=datasets['train'].pad_idx)
tracker['z'] = torch.cat((tracker['z'], z.data), dim=0)
# Calculate BLEU score
decoded = torch.argmax(logp, dim=-1)
for i in range(decoded.shape[0]):
decoded_poem = idx2word(
[decoded[i]],
i2w=datasets['train'].get_i2w(),
pad_idx=datasets['train'].pad_idx)[0]
original_poem = idx2word(
[batch['target'].data[i]],
i2w=datasets['train'].get_i2w(),
pad_idx=datasets['train'].pad_idx)[0]
total_BLEU_score += calculate_bleu_scores(
original_poem, decoded_poem)
print("%s Epoch %02d/%i, Mean ELBO %9.4f" %
(split.upper(), epoch, args.epochs,
torch.mean(tracker['ELBO'])))
if split == 'valid':
print("Average BLEU {}".format(total_BLEU_score /
decoded.shape[0]))
if args.tensorboard_logging:
writer.add_scalar("%s-Epoch/ELBO" % split.upper(),
torch.mean(tracker['ELBO']), epoch)
# save a dump of all sentences and the encoded latent space
if split == 'valid':
dump = {
'target_sents': tracker['target_sents'],
'z': tracker['z'].tolist()
}
if not os.path.exists(os.path.join('dumps', ts)):
os.makedirs('dumps/' + ts)
with open(
os.path.join('dumps/' + ts +
'/valid_E%i.json' % epoch),
'w') as dump_file:
json.dump(dump, dump_file)
# save checkpoint
if split == 'train':
checkpoint_path = os.path.join(save_model_path,
"E%i.pytorch" % (epoch))
torch.save(model.state_dict(), checkpoint_path)
print("Model saved at %s" % checkpoint_path)