def encode_dataset(cfg):
out_dir = Path(utils.to_absolute_path(cfg.out_dir))
out_dir.mkdir(exist_ok=True, parents=True)
root_path = Path(utils.to_absolute_path("datasets")) / cfg.dataset.path
with open(root_path / "test.json") as file:
metadata = json.load(file)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
encoder = Encoder(**cfg.model.encoder)
encoder.to(device)
print("Load checkpoint from: {}:".format(cfg.checkpoint))
checkpoint_path = utils.to_absolute_path(cfg.checkpoint)
checkpoint = torch.load(checkpoint_path,
map_location=lambda storage, loc: storage)
encoder.load_state_dict(checkpoint["encoder"])
encoder.eval()
if cfg.save_auxiliary:
auxiliary = []
def hook(module, input, output):
auxiliary.append(output.clone())
encoder.encoder[-1].register_forward_hook(hook)
for _, _, _, path in tqdm(metadata):
path = root_path.parent / path
mel = torch.from_numpy(np.load(
path.with_suffix(".mel.npy"))).unsqueeze(0).to(device)
with torch.no_grad():
z, c, indices = encoder.encode(mel)
z = z.squeeze().cpu().numpy()
out_path = out_dir / path.stem
with open(out_path.with_suffix(".txt"), "w") as file:
np.savetxt(file, z, fmt="%.16f")
if cfg.save_auxiliary:
aux_path = out_dir.parent / "auxiliary_embedding1"
aux_path.mkdir(exist_ok=True, parents=True)
out_path = aux_path / path.stem
c = c.squeeze().cpu().numpy()
with open(out_path.with_suffix(".txt"), "w") as file:
np.savetxt(file, c, fmt="%.16f")
aux_path = out_dir.parent / "auxiliary_embedding2"
aux_path.mkdir(exist_ok=True, parents=True)
out_path = aux_path / path.stem
aux = auxiliary.pop().squeeze().cpu().numpy()
with open(out_path.with_suffix(".txt"), "w") as file:
np.savetxt(file, aux, fmt="%.16f")
def convert(cfg):
dataset_path = Path(utils.to_absolute_path("datasets")) / cfg.dataset.path
with open(dataset_path / "speakers.json") as file:
speakers = sorted(json.load(file))
synthesis_list_path = Path(utils.to_absolute_path(cfg.synthesis_list))
with open(synthesis_list_path) as file:
synthesis_list = json.load(file)
in_dir = Path(utils.to_absolute_path(cfg.in_dir))
out_dir = Path(utils.to_absolute_path(cfg.out_dir))
out_dir.mkdir(exist_ok=True, parents=True)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
encoder = Encoder(**cfg.model.encoder)
decoder = Decoder(**cfg.model.decoder)
encoder.to(device)
decoder.to(device)
print("Load checkpoint from: {}:".format(cfg.checkpoint))
checkpoint_path = utils.to_absolute_path(cfg.checkpoint)
checkpoint = torch.load(checkpoint_path, map_location=lambda storage, loc: storage)
encoder.load_state_dict(checkpoint["encoder"])
decoder.load_state_dict(checkpoint["decoder"])
encoder.eval()
decoder.eval()
for wav_path, speaker_id, out_filename in tqdm(synthesis_list):
wav_path = in_dir / wav_path
wav, _ = librosa.load(
wav_path.with_suffix(".wav"),
sr=cfg.preprocessing.sr)
wav = wav / np.abs(wav).max() * 0.999
mel = librosa.feature.melspectrogram(
preemphasis(wav, cfg.preprocessing.preemph),
sr=cfg.preprocessing.sr,
n_fft=cfg.preprocessing.n_fft,
n_mels=cfg.preprocessing.n_mels,
hop_length=cfg.preprocessing.hop_length,
win_length=cfg.preprocessing.win_length,
fmin=cfg.preprocessing.fmin,
power=1)
logmel = librosa.amplitude_to_db(mel, top_db=cfg.preprocessing.top_db)
logmel = logmel / cfg.preprocessing.top_db + 1
mel = torch.FloatTensor(logmel).unsqueeze(0).to(device)
speaker = torch.LongTensor([speakers.index(speaker_id)]).to(device)
with torch.no_grad():
z, _ = encoder.encode(mel)
output = decoder.generate(z, speaker)
path = out_dir / out_filename
librosa.output.write_wav(path.with_suffix(".wav"), output.astype(np.float32), sr=cfg.preprocessing.sr)
def get_encoder(latent_dim, fckpt=''):
E = Encoder(latent_dim)
if fckpt and os.path.exists(fckpt):
ckpt = torch.load(fckpt)
loaded_sd = ckpt['E']
try:
E.load_state_dict(loaded_sd)
except:
curr_params = E.state_dict()
curr_keys = list(curr_params.keys())
updated_params = {}
for k, v in loaded_sd.items():
if 'bn7' in k:
newk = k.replace('bn7', 'conv7')
else:
newk = k
if newk in curr_keys and loaded_sd[k].shape == curr_params[
newk].shape:
updated_params[newk] = v
else:
print('Failed to load:', k)
curr_params.update(updated_params)
E.load_state_dict(curr_params)
return E.to(device)
class Model:
def __init__(self, chpt_enc_path, chpt_dec_path, chpt_stat_path):
historyLength = 10
encoder_dim = hiddenDimension
lstm_input_dim = historyLength + 1
decoder_dim = hiddenDimension
attention_dim = hiddenDimension
output_dim = 1
self.decodeLength = 20
self.encoder = Encoder()
self.decoder = DecoderWithAttention(encoder_dim, lstm_input_dim, decoder_dim, attention_dim, output_dim)
self.encoder.load_state_dict(torch.load(chpt_enc_path))
self.decoder.load_state_dict(torch.load(chpt_dec_path))
self.encoder = self.encoder.to(device)
self.decoder = self.decoder.to(device)
self.encoder.eval()
self.decoder.eval()
with open(chpt_stat_path, 'rb') as f:
chpt_stat = pickle.load(f)
self.cMean = chpt_stat['cMean_tr']
self.cStd = chpt_stat['cStd_tr']
self.vMean = chpt_stat['vMean_tr']
self.vStd = chpt_stat['vStd_tr']
self.aMean = chpt_stat['aMean_tr']
self.aStd = chpt_stat['aStd_tr']
self.mean = torch.Tensor([self.vMean, self.aMean]).to(device)
self.std = torch.Tensor([self.vStd, self.aStd]).to(device)
def predict(self, curvatures, currentSpeed, histSpeeds, currentAccelX, histAccelXs):
curvatures = torch.FloatTensor(curvatures).to(device)
currentSpeed = torch.FloatTensor([currentSpeed]).to(device)
histSpeeds = torch.FloatTensor(histSpeeds).to(device)
currentAccelX = torch.FloatTensor([currentAccelX]).to(device)
histAccelXs = torch.FloatTensor(histAccelXs).to(device)
curvatures = (curvatures - self.cMean) / self.cStd
currentSpeed = (currentSpeed - self.vMean) / self.vStd
histSpeeds = (histSpeeds - self.vMean) / self.vStd
currentAccelX = (currentAccelX - self.aMean) / self.aStd
histAccelXs = (histAccelXs - self.aMean) / self.aStd
curvatures = self.encoder(curvatures.unsqueeze(dim=0).unsqueeze(dim=0))
predictions, alphas, alphas_target = self.decoder(curvatures, currentSpeed, histSpeeds.unsqueeze(dim=0), currentAccelX, histAccelXs.unsqueeze(dim=0),
self.decodeLength, self.vMean, self.vStd, self.aMean, self.aStd)
return (predictions.squeeze()*self.aStd + self.aMean).cpu().detach().numpy(), alphas.squeeze().cpu().detach().numpy()
def initialize_for_test(params):
data_loader = get_loader(params, mode='test')
encoder_file = os.path.join(params.encoder_save,
'epoch-%d.pkl' % params.num_epochs)
decoder_file = os.path.join(params.decoder_save,
'epoch-%d.pkl' % params.num_epochs)
vocab_size = len(data_loader.dataset.vocab)
# Initialize the encoder and decoder, and set each to inference mode.
encoder = Encoder(params)
decoder = Decoder(params, vocab_size)
encoder.eval()
decoder.eval()
# Load the trained weights.
encoder.load_state_dict(torch.load(encoder_file))
decoder.load_state_dict(torch.load(decoder_file))
encoder.to(params.device)
decoder.to(params.device)
return data_loader, encoder, decoder
def load_encoder(data_root, weight_path, device):
encoder = Encoder()
if weight_path:
weight = torch.load(weight_path)
else:
weight = torch.load(get_best_weight(data_root))
encoder.load_state_dict(weight)
if device >= 0:
encoder = encoder.to(f"cuda:{device}")
encoder.eval()
return encoder
def main(data_name):
dataset = MyDataSet(data_name=data_name, reset=False)
vocab_size = dataset.vocab_size
corpus = dataset.corpus
id2word = {v: k for k, v in corpus.items()}
train_loader, val_loader = _get_data_loader(dataset, 0.5, batch_size)
embedding, embed_dim = load_embedding(basic_settings['word2vec'], corpus)
encoder = Encoder(dataset.feature_dim, output_dim=100)
decoder = DecoderWithAttention(encoder.get_output_dim(),
decoder_dim=100,
attn_dim=100,
embed_dim=embed_dim,
vocab_size=vocab_size)
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=lr)
encoder = encoder.to(device)
decoder = decoder.to(device)
criterion = torch.nn.CrossEntropyLoss().to(device)
best_bleu4 = 0
best_hypos = []
best_refs = []
for epoch in range(1, epoches + 1):
# One epoch's training
train_epoch(train_loader=train_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
optimizer=decoder_optimizer,
epoch=epoch)
# One epoch's validation
bleu4_score, refs, hypos = validate(val_loader=val_loader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
word2id=corpus)
if bleu4_score > best_bleu4:
best_bleu4 = bleu4_score
best_refs = refs
best_hypos = hypos
name = data_name + '_' + str(best_bleu4) + '.xlsx'
save_result(name, best_refs, best_hypos, id2word)
def main(args):
# Image preprocessing
# In generation phase, we need should random crop, just resize
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
# Load vocabulary wraper
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f)
# Build model
encoder = Encoder(embed_size=args.embed_size).eval()
decoder = Decoder(stateful=False,
embed_size=args.embed_size,
hidden_size=args.hidden_size,
vocab_size=len(vocab),
num_layers=args.num_layers).to(device)
encoder = encoder.to(device)
decoder = decoder.to(device)
# load the trained model parameters
encoder.load_state_dict(torch.load(args.encoder_path, map_location=device))
decoder.load_state_dict(torch.load(args.decoder_path, map_location=device))
# Prepare an image
image = load_image(args.image, transform)
image_tensor = image.to(device)
# Generate an caption from the image
feature = encoder(image_tensor)
sampled_ids = decoder.sample(feature)
sampled_ids = sampled_ids[0].cpu().numpy()
sampled_caption = []
for word_id in sampled_ids:
word = vocab.idx2word[word_id]
sampled_caption.append(word)
if word == '<<end>>':
break
sentence = ' '.join(sampled_caption)
print(sentence)
def load_model(
encoder_path,
decoder_path,
vocab_size,
layer_type='gru',
embed_size=256,
hidden_size=512,
num_layers=2,
):
if layer_type == 'lstm':
from model import Encoder, Decoder
else:
from model_gru import Encoder, Decoder
# eval mode (batchnorm uses moving mean/variance)
encoder = Encoder(embed_size).eval()
decoder = Decoder(embed_size, hidden_size, vocab_size, num_layers)
encoder = encoder.to(device)
decoder = decoder.to(device)
# Load the trained model parameters
encoder.load_state_dict(torch.load(encoder_path))
decoder.load_state_dict(torch.load(decoder_path))
return encoder, decoder
def train_model(cfg):
tensorboard_path = Path(
utils.to_absolute_path("tensorboard")) / cfg.checkpoint_dir
checkpoint_dir = Path(utils.to_absolute_path(cfg.checkpoint_dir))
writer = SummaryWriter(tensorboard_path)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
encoder = Encoder(**cfg.model.encoder)
cpc = CPCLoss(**cfg.model.cpc)
encoder.to(device)
cpc.to(device)
optimizer = optim.Adam(chain(encoder.parameters(), cpc.parameters()),
lr=cfg.training.scheduler.initial_lr)
scheduler = WarmupScheduler(
optimizer,
warmup_epochs=cfg.training.scheduler.warmup_epochs,
initial_lr=cfg.training.scheduler.initial_lr,
max_lr=cfg.training.scheduler.max_lr,
milestones=cfg.training.scheduler.milestones,
gamma=cfg.training.scheduler.gamma)
if cfg.resume:
print("Resume checkpoint from: {}:".format(cfg.resume))
resume_path = utils.to_absolute_path(cfg.resume)
checkpoint = torch.load(resume_path,
map_location=lambda storage, loc: storage)
encoder.load_state_dict(checkpoint["encoder"])
cpc.load_state_dict(checkpoint["cpc"])
optimizer.load_state_dict(checkpoint["optimizer"])
scheduler.load_state_dict(checkpoint["scheduler"])
start_epoch = checkpoint["epoch"]
else:
start_epoch = 1
root_path = Path(utils.to_absolute_path("datasets")) / cfg.dataset.path
dataset = CPCDataset(
root=root_path,
n_sample_frames=cfg.training.sample_frames +
cfg.training.n_prediction_steps,
n_utterances_per_speaker=cfg.training.n_utterances_per_speaker,
hop_length=cfg.preprocessing.hop_length,
sr=cfg.preprocessing.sr)
dataloader = DataLoader(dataset,
batch_size=cfg.training.n_speakers_per_batch,
shuffle=True,
num_workers=cfg.training.n_workers,
pin_memory=True,
drop_last=True)
for epoch in range(start_epoch, cfg.training.n_epochs + 1):
if epoch % cfg.training.log_interval == 0 or epoch == start_epoch:
average_cpc_loss = average_vq_loss = average_perplexity = 0
average_accuracies = np.zeros(cfg.training.n_prediction_steps // 2)
for i, (mels, _) in enumerate(tqdm(dataloader), 1):
mels = mels.to(device)
mels = mels.view(
cfg.training.n_speakers_per_batch *
cfg.training.n_utterances_per_speaker,
cfg.preprocessing.n_mels, -1)
optimizer.zero_grad()
z, c, vq_loss, perplexity = encoder(mels)
cpc_loss, accuracy = cpc(z, c)
loss = cpc_loss + vq_loss
loss.backward()
optimizer.step()
average_cpc_loss += (cpc_loss.item() - average_cpc_loss) / i
average_vq_loss += (vq_loss.item() - average_vq_loss) / i
average_perplexity += (perplexity.item() - average_perplexity) / i
average_accuracies += (np.array(accuracy) - average_accuracies) / i
scheduler.step()
if epoch % cfg.training.log_interval == 0 and epoch != start_epoch:
writer.add_scalar("cpc_loss/train", average_cpc_loss, epoch)
writer.add_scalar("vq_loss/train", average_vq_loss, epoch)
writer.add_scalar("perplexity/train", average_perplexity, epoch)
print(
"epoch:{}, cpc loss:{:.2E}, vq loss:{:.2E}, perpexlity:{:.3f}".
format(epoch, cpc_loss, average_vq_loss, average_perplexity))
print(100 * average_accuracies)
if epoch % cfg.training.checkpoint_interval == 0 and epoch != start_epoch:
save_checkpoint(encoder, cpc, optimizer, scheduler, epoch,
checkpoint_dir)
def main(epoch_num, batch_size, verbose, UNSEEN, SEEN, MODE):
[
hownet_file, sememe_file, word_index_file, word_vector_file,
dictionary_file, word_cilinClass_file
] = [
'hownet.json', 'sememe.json', 'word_index.json', 'word_vector.npy',
'dictionary_sense.json', 'word_cilinClass.json'
]
word2index, index2word, word2vec, sememe_num, label_size, label_size_chara, word_defi_idx_all = load_data(
hownet_file, sememe_file, word_index_file, word_vector_file,
dictionary_file, word_cilinClass_file)
(word_defi_idx_TrainDev, word_defi_idx_seen, word_defi_idx_test2000,
word_defi_idx_test200, word_defi_idx_test272) = word_defi_idx_all
index2word = np.array(index2word)
length = len(word_defi_idx_TrainDev)
valid_dataset = MyDataset(word_defi_idx_TrainDev[int(0.9 * length):])
test_dataset = MyDataset(word_defi_idx_test2000 + word_defi_idx_test200 +
word_defi_idx_test272)
if SEEN:
mode = 'S_' + MODE
print('*METHOD: Seen defi.')
print('*TRAIN: [Train + allSeen(2000+200+272)]')
print('*TEST: [2000rand1 + 200desc + 272desc]')
train_dataset = MyDataset(word_defi_idx_TrainDev[:int(0.9 * length)] +
word_defi_idx_seen)
elif UNSEEN:
mode = 'U_' + MODE
print('*METHOD: Unseen All words and defi.')
print('*TRAIN: [Train]')
print('*TEST: [2000rand1 + 200desc + 272desc]')
train_dataset = MyDataset(word_defi_idx_TrainDev[:int(0.9 * length)])
print('*MODE: [%s]' % mode)
train_dataloader = torch.utils.data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
collate_fn=my_collate_fn)
valid_dataloader = torch.utils.data.DataLoader(valid_dataset,
batch_size=batch_size,
shuffle=True,
collate_fn=my_collate_fn)
test_dataloader = torch.utils.data.DataLoader(
test_dataset,
batch_size=batch_size,
shuffle=False,
collate_fn=my_collate_fn_test)
print('Train dataset: ', len(train_dataset))
print('Valid dataset: ', len(valid_dataset))
print('Test dataset: ', len(test_dataset))
word_defi_idx = word_defi_idx_TrainDev + word_defi_idx_seen
wd2sem = word2sememe(word_defi_idx, len(word2index), sememe_num)
wd_sems = label_multihot(wd2sem, sememe_num)
wd_sems = torch.from_numpy(np.array(wd_sems[:label_size])).to(device)
wd_POSs = label_multihot(word2POS(word_defi_idx, len(word2index), 13), 13)
wd_POSs = torch.from_numpy(np.array(wd_POSs[:label_size])).to(device)
wd_charas = label_multihot(
word2chara(word_defi_idx, len(word2index), label_size_chara),
label_size_chara)
wd_charas = torch.from_numpy(np.array(wd_charas[:label_size])).to(device)
wd2Cilin1 = word2Cn(word_defi_idx, len(word2index), 'C1', 13)
wd_C1 = label_multihot(wd2Cilin1, 13) #13 96 1426 4098
wd_C1 = torch.from_numpy(np.array(wd_C1[:label_size])).to(device)
wd_C2 = label_multihot(word2Cn(word_defi_idx, len(word2index), 'C2', 96),
96)
wd_C2 = torch.from_numpy(np.array(wd_C2[:label_size])).to(device)
wd_C3 = label_multihot(word2Cn(word_defi_idx, len(word2index), 'C3', 1426),
1426)
wd_C3 = torch.from_numpy(np.array(wd_C3[:label_size])).to(device)
wd_C4 = label_multihot(word2Cn(word_defi_idx, len(word2index), 'C4', 4098),
4098)
wd_C4 = torch.from_numpy(np.array(wd_C4[:label_size])).to(device)
'''wd2Cilin = word2Cn(word_defi_idx, len(word2index), 'C', 5633)
wd_C0 = label_multihot(wd2Cilin, 5633)
wd_C0 = torch.from_numpy(np.array(wd_C0[:label_size])).to(device)
wd_C = [wd_C1, wd_C2, wd_C3, wd_C4, wd_C0]
'''
wd_C = [wd_C1, wd_C2, wd_C3, wd_C4]
#----------mask of no sememes
print('calculating mask of no sememes...')
mask_s = torch.zeros(label_size, dtype=torch.float32, device=device)
for i in range(label_size):
sems = set(wd2sem[i].detach().cpu().numpy().tolist()) - set(
[sememe_num])
if len(sems) == 0:
mask_s[i] = 1
mask_c = torch.zeros(label_size, dtype=torch.float32, device=device)
for i in range(label_size):
cc = set(wd2Cilin1[i].detach().cpu().numpy().tolist()) - set([13])
if len(cc) == 0:
mask_c[i] = 1
model = Encoder(vocab_size=len(word2index),
embed_dim=word2vec.shape[1],
hidden_dim=200,
layers=1,
class_num=label_size,
sememe_num=sememe_num,
chara_num=label_size_chara)
model.embedding.weight.data = torch.from_numpy(word2vec)
model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=0.001) # Adam
best_valid_accu = 0
DEF_UPDATE = True
for epoch in range(epoch_num):
print('epoch: ', epoch)
model.train()
train_loss = 0
label_list = list()
pred_list = list()
for words_t, sememes_t, definition_words_t, POS_t, sememes, POSs, charas_t, C, C_t in tqdm(
train_dataloader, disable=verbose):
optimizer.zero_grad()
loss, _, indices = model('train',
x=definition_words_t,
w=words_t,
ws=wd_sems,
wP=wd_POSs,
wc=wd_charas,
wC=wd_C,
msk_s=mask_s,
msk_c=mask_c,
mode=MODE)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1)
optimizer.step()
predicted = indices[:, :100].detach().cpu().numpy().tolist()
train_loss += loss.item()
label_list.extend(words_t.detach().cpu().numpy())
pred_list.extend(predicted)
train_accu_1, train_accu_10, train_accu_100 = evaluate(
label_list, pred_list)
del label_list
del pred_list
gc.collect()
print('train_loss: ', train_loss / len(train_dataset))
print('train_accu(1/10/100): %.2f %.2F %.2f' %
(train_accu_1, train_accu_10, train_accu_100))
model.eval()
with torch.no_grad():
valid_loss = 0
label_list = []
pred_list = []
for words_t, sememes_t, definition_words_t, POS_t, sememes, POSs, charas_t, C, C_t in tqdm(
valid_dataloader, disable=verbose):
loss, _, indices = model('train',
x=definition_words_t,
w=words_t,
ws=wd_sems,
wP=wd_POSs,
wc=wd_charas,
wC=wd_C,
msk_s=mask_s,
msk_c=mask_c,
mode=MODE)
predicted = indices[:, :100].detach().cpu().numpy().tolist()
valid_loss += loss.item()
label_list.extend(words_t.detach().cpu().numpy())
pred_list.extend(predicted)
valid_accu_1, valid_accu_10, valid_accu_100 = evaluate(
label_list, pred_list)
print('valid_loss: ', valid_loss / len(valid_dataset))
print('valid_accu(1/10/100): %.2f %.2F %.2f' %
(valid_accu_1, valid_accu_10, valid_accu_100))
del label_list
del pred_list
gc.collect()
if valid_accu_10 > best_valid_accu:
best_valid_accu = valid_accu_10
print('-----best_valid_accu-----')
#torch.save(model, 'saved.model')
label_list = []
pred_list = []
for words_t, definition_words_t in tqdm(test_dataloader,
disable=verbose):
indices = model('test',
x=definition_words_t,
w=words_t,
ws=wd_sems,
wP=wd_POSs,
wc=wd_charas,
wC=wd_C,
msk_s=mask_s,
msk_c=mask_c,
mode=MODE)
predicted = indices[:, :1000].detach().cpu().numpy(
).tolist()
label_list.extend(words_t.detach().cpu().numpy())
pred_list.extend(predicted)
test_accu_1, test_accu_10, test_accu_100, median, variance = evaluate_test(
label_list, pred_list)
print('test_accu(1/10/100): %.2f %.2F %.2f %.1f %.2f' %
(test_accu_1, test_accu_10, test_accu_100, median,
variance))
if epoch > 10:
json.dump((index2word[label_list]).tolist(),
open(mode + '_label_list.json', 'w'))
json.dump((index2word[np.array(pred_list)]).tolist(),
open(mode + '_pred_list.json', 'w'))
del label_list
del pred_list
gc.collect()
assert args.snapshot is not None
else:
if args.sort_by_freq is False:
assert args.order_free in ["pla", "mla"]
else:
if args.order_free:
raise ValueError(
'Sort by freq and order_free are mutually exclusive.')
resume = 0
highest_f1 = 0
epochs_without_imp = 0
iterations = 0
encoder = Encoder(encoder_weights=args.encoder_weights)
decoder = Decoder(args.hidden_size, args.embed_size, args.attention_size,
args.dropout)
encoder = encoder.to('cuda')
decoder = decoder.to('cuda')
snapshot = args.snapshot
test_model = args.test_model
train_from_scratch = args.train_from_scratch
swa_params = eval(args.swa_params)
finetune_encoder = args.finetune_encoder
if not test_model:
if finetune_encoder:
encoder_optimizer = torch.optim.Adam(encoder.parameters(),
lr=args.encoder_lr)
decoder_optimizer = torch.optim.Adam(decoder.parameters(),
lr=args.decoder_lr)
else:
embedding_dimension=opt.embedding_dim,
hidden_size=opt.rnn_hidden,
num_layer=opt.num_layers)
optimizer1 = torch.optim.Adam(encoder.parameters(), lr=opt.lr)
decoder = BahdanauAttnDecoderRNN(opt.rnn_hidden,
opt.embedding_dim,
len(en_config.word2ix),
n_layers=2,
dropout_p=0.1)
# decoder =
optimizer2 = torch.optim.Adam(decoder.parameters(), lr=opt.lr)
if opt.save_path:
encoder.load_state_dict(torch.load(opt.save_path + 'encoder.pth'))
decoder.load_state_dict(torch.load(opt.save_path + 'decoder.pth'))
print('load update model')
encoder.to(device)
decoder.to(device)
loss_meter = AverageValueMeter()
'''
for epoch in range(200):
loss_meter.reset()
for ii, ((in_lang,in_lengths),(out_lang,out_lengths)) in tqdm(enumerate(train_dataloader)):
in_lang = in_lang.to(device)
out_lang = out_lang.to(device)
optimizer1.zero_grad()
optimizer2.zero_grad()
encoder_outputs, encoder_hidden = encoder(in_lang,in_lengths) # MAX_LENGTH, BATCH_SIZE, EMBEDDING DIMENSION // n_layer, BATCH_SIZE, EMBEDDING DIMENSION
# Prepare input and output variables
decoder_input = torch.LongTensor([fr_config.word2ix[opt.start]] * in_lang.shape[1]).to(device)
class Solver(object):
"""Solver for training and testing"""
def __init__(self, data_loader, config):
"""Initialize configurations."""
# Data loader.
self.data_loader = data_loader
# Model configurations.
self.a_dim = config.a_dim
self.id_dim = config.id_dim
# Training configurations.
self.batch_size = config.batch_size
self.num_iters = config.num_iters
self.num_iters_decay = config.num_iters_decay
self.lr = config.lr
self.n_critic = config.n_critic
self.beta1 = config.beta1
self.beta2 = config.beta2
self.resume_iters = config.resume_iters
# Test configurations.
self.test_iters = config.test_iters
# Miscellaneous.
self.use_tensorboard = config.use_tensorboard
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Directories.
self.log_dir = config.log_dir
self.sample_dir = config.sample_dir
self.model_save_dir = config.model_save_dir
self.result_dir = config.result_dir
# Step size.
self.log_step = config.log_step
self.sample_step = config.sample_step
self.model_save_step = config.model_save_step
self.lr_update_step = config.lr_update_step
# Build the model and tensorboard.
self.build_model()
if self.use_tensorboard:
self.build_tensorboard()
def build_model(self):
"""Create a generator and a discriminator."""
self.G = Generator()
self.D = Discriminator()
self.I = Encoder()
self.C = Encoder()
self.A = Attribute()
self.g_optimizer = torch.optim.Adam(self.G.parameters(), self.lr, [self.beta1, self.beta2])
self.d_optimizer = torch.optim.Adam(self.D.parameters(), self.lr, [self.beta1, self.beta2])
self.i_optimizer = torch.optim.Adam(self.I.parameters(), self.lr, [self.beta1, self.beta2])
self.c_optimizer = torch.optim.Adam(self.C.parameters(), self.lr, [self.beta1, self.beta2])
self.a_optimizer = torch.optim.Adam(self.A.parameters(), self.lr, [self.beta1, self.beta2])
self.G.to(self.device)
self.D.to(self.device)
self.A.to(self.device)
self.I.to(self.device)
self.C.to(self.device)
def print_network(self, model, name):
"""Print out the network information."""
num_params = 0
for p in model.parameters():
num_params += p.numel()
print(model)
print(name)
print("The number of parameters: {}".format(num_params))
def restore_model(self, resume_iters):
"""Restore the trained generator and discriminator."""
print('Loading the trained models from step {}...'.format(resume_iters))
G_path = os.path.join(self.model_save_dir, '{}-G.ckpt'.format(resume_iters))
D_path = os.path.join(self.model_save_dir, '{}-D.ckpt'.format(resume_iters))
A_path = os.path.join(self.model_save_dir, '{}-A.ckpt'.format(resume_iters))
I_path = os.path.join(self.model_save_dir, '{}-I.ckpt'.format(resume_iters))
C_path = os.path.join(self.model_save_dir, '{}-C.ckpt'.format(resume_iters))
self.A.load_state_dict(torch.load(A_path, map_location=lambda storage, loc: storage))
self.I.load_state_dict(torch.load(I_path, map_location=lambda storage, loc: storage))
self.C.load_state_dict(torch.load(C_path, map_location=lambda storage, loc: storage))
self.G.load_state_dict(torch.load(G_path, map_location=lambda storage, loc: storage))
self.D.load_state_dict(torch.load(D_path, map_location=lambda storage, loc: storage))
def build_tensorboard(self):
"""Build a tensorboard logger."""
from logger import Logger
self.logger = Logger(self.log_dir)
def update_lr(self, lr):
"""Decay learning rates of the generator and discriminator."""
for param_group in self.g_optimizer.param_groups:
param_group['lr'] = lr
for param_group in self.d_optimizer.param_groups:
param_group['lr'] = lr
for param_group in self.i_optimizer.param_groups:
param_group['lr'] = lr
for param_group in self.a_optimizer.param_groups:
param_group['lr'] = lr
for param_group in self.c_optimizer.param_groups:
param_group['lr'] = lr
def reset_grad(self):
"""Reset the gradient buffers."""
self.g_optimizer.zero_grad()
self.d_optimizer.zero_grad()
self.i_optimizer.zero_grad()
self.a_optimizer.zero_grad()
self.c_optimizer.zero_grad()
def denorm(self, x):
"""Convert the range from [-1, 1] to [0, 1]."""
out = (x + 1) / 2
return out.clamp_(0, 1)
def classification_loss(self, logit, target):
"""Compute binary or softmax cross entropy loss."""
return F.cross_entropy(logit, target)
def mse_loss(self, out, gt):
"""Computes the MSE between model output and scalar gt"""
loss = 0.5 * torch.mean(torch.abs(out - gt)**2)
return loss
def L1_loss(self, pred, target):
"""
Calculate L1 loss
"""
return torch.mean(torch.abs(pred - target))
def reparameterization(self, mu, logvar):
std = torch.exp(logvar / 2)
sampled_z = torch.FloatTensor(np.random.normal(0, 1, (mu.size(0), 8))).to(self.device)
z = sampled_z * std + mu
return z
def train(self):
"""Train StarGAN within a single dataset."""
# Set data loader.
data_loader = self.data_loader
# Fetch fixed inputs for debugging.
data_iter = iter(data_loader)
batch_fixed = next(data_iter)
for k in batch_fixed:
batch_fixed[k] = batch_fixed[k].to(self.device)
# Learning rate cache for decaying.
lr = self.lr
# Start training from scratch or resume training.
start_iters = 0
if self.resume_iters:
start_iters = self.resume_iters
self.restore_model(self.resume_iters)
# Start training.
print('Start training...')
start_time = time.time()
for i in range(start_iters, self.num_iters):
# =================================================================================== #
# 1. Preprocess input data #
# =================================================================================== #
# Fetch real images and labels.
try:
batch = next(data_iter)
except:
data_iter = iter(data_loader)
batch = next(data_iter)
for k in batch:
batch[k] = batch[k].to(self.device)
# =================================================================================== #
# 2. Train the discriminator #
# =================================================================================== #
loss = {}
# get identity z
id_z, _ = self.I(batch['img_profile'])
# get attribute z
mu, logvar = self.A(batch['img_frontal'])
a_z = self.reparameterization(mu, logvar)
# get x'
x = torch.cat([id_z, a_z], 1)
x_fake = self.G(x)
# Get the predicted identity
id_pred, _ = self.C(batch['img_profile'])
# distinguish the true and the false
d_real, _ = self.D(batch['img_frontal'])
d_fake, _ = self.D(x_fake.detach())
# train I
loss_Li = self.classification_loss(id_z, batch['label'])
# train A
loss_KL = torch.sum(0.5 * (mu**2 + torch.exp(logvar) - logvar - 1))
loss_GR = self.mse_loss(batch['img_frontal'], x_fake)
# triain C
loss_C = self.classification_loss(id_pred, batch['label'])
# train D
loss_D = - torch.mean(d_real) + torch.mean(d_fake)
d_loss = loss_D + loss_C + loss_GR + loss_KL + loss_Li
self.reset_grad()
d_loss.backward()
self.d_optimizer.step()
self.c_optimizer.step()
self.a_optimizer.step()
self.i_optimizer.step()
loss['C/loss_C'] = loss_C.item()
loss['A/loss_GR'] = loss_GR.item()
loss['I/loss_Li'] = loss_Li.item()
loss['D/loss_D'] = loss_D.item()
# =================================================================================== #
# 3. Train the generator #
# =================================================================================== #
if (i + 1) % self.n_critic == 0:
id_z, _ = self.I(batch['img_profile'])
# get attribute z
mu, logvar = self.A(batch['img_frontal'])
a_z = self.reparameterization(mu, logvar)
# get x'
x = torch.cat([id_z, a_z], 1)
x_fake = self.G(x)
# Get the predicted identity
_, c_f_s = self.C(batch['img_profile'])
_, c_f_x = self.C(x_fake)
# distinguish the true and the false
d_real, d_f_a = self.D(batch['img_frontal'])
d_fake, d_f_x = self.D(x_fake)
loss_GR = self.mse_loss(batch['img_frontal'], x_fake)
# triain C
loss_GC = self.mse_loss(c_f_x, c_f_s)
loss_GD = self.mse_loss(d_f_x, d_f_a)
loss_g = - torch.mean(d_fake)
g_loss = loss_g + loss_GC + loss_GR + loss_GD
self.reset_grad()
g_loss.backward()
self.g_optimizer.step()
# Logging.
loss['G/loss_GR'] = loss_GR.item()
loss['G/loss_GC'] = loss_GC.item()
loss['G/loss_GD'] = loss_GD.item()
loss['G/loss_g'] = loss_g.item()
# =================================================================================== #
# 4. Miscellaneous #
# =================================================================================== #
# Print out training information.
if (i + 1) % self.log_step == 0:
et = time.time() - start_time
et = str(datetime.timedelta(seconds=et))[:-7]
log = "Elapsed [{}], Iteration [{}/{}]".format(et, i + 1, self.num_iters)
for tag, value in loss.items():
log += ", {}: {:.4f}".format(tag, value)
print(log)
if self.use_tensorboard:
for tag, value in loss.items():
self.logger.scalar_summary(tag, value, i + 1)
# Translate fixed images for debugging.
if (i + 1) % self.sample_step == 0:
for k in batch_fixed:
batch_fixed[k] = batch_fixed[k].to(self.device)
with torch.no_grad():
x_fake_list = [batch_fixed['img_profile']]
id_z, _ = self.I(batch_fixed['img_profile'])
# get attribute z
mu, logvar = self.A(batch_fixed['img_frontal'])
a_z = self.reparameterization(mu, logvar)
# get x'
x = torch.cat([id_z, a_z], 1)
x_fake = self.G(x)
x_fake_list.append(x_fake)
x_concat = torch.cat(x_fake_list, dim=3)
sample_path = os.path.join(self.sample_dir, '{}-images.jpg'.format(i + 1))
save_image(self.denorm(x_concat.data.cpu()), sample_path, nrow=2, padding=5)
print('Saved real and fake images into {}...'.format(sample_path))
# Save model checkpoints.
if (i + 1) % self.model_save_step == 0:
G_path = os.path.join(self.model_save_dir, '{}-G.ckpt'.format(i + 1))
D_path = os.path.join(self.model_save_dir, '{}-D.ckpt'.format(i + 1))
torch.save(self.G.state_dict(), G_path)
torch.save(self.D.state_dict(), D_path)
print('Saved model checkpoints into {}...'.format(self.model_save_dir))
# Decay learning rates.
if (i + 1) % self.lr_update_step == 0 and (i + 1) > (self.num_iters - self.num_iters_decay):
lr -= (self.lr / float(self.num_iters_decay))
self.update_lr(lr)
print('Decayed learning rates, lr: {}'.format(lr))
def main():
epoch = 1000
batch_size = 64
hidden_dim = 256
encoder = Encoder(num_words, hidden_dim)
if args.attn:
attn_model = 'dot'
decoder = LuongAttnDecoderLength(attn_model, hidden_dim, num_words,
MAX_TGT_LEN)
else:
decoder = DecoderTask1(hidden_dim, num_words)
if args.train:
weight = torch.ones(num_words)
weight[word2idx_mapping[PAD_TOKEN]] = 0
encoder = encoder.to(device)
decoder = decoder.to(device)
weight = weight.to(device)
encoder_optimizer = Adam(encoder.parameters(), lr=0.001)
decoder_optimizer = Adam(decoder.parameters(), lr=0.001)
criterion = nn.NLLLoss(ignore_index=int(word2idx_mapping[PAD_TOKEN]),
size_average=True)
#criterion = nn.CrossEntropyLoss(weight=weight)
np.random.seed(1124)
order = np.arange(len(train_data))
best_loss = 1e5
best_epoch = 0
for e in range(epoch):
#if e - best_epoch > 20: break
#np.random.shuffle(order)
choice = np.random.choice(order, 10000, replace=False)
shuffled_train_data = train_data[choice]
train_loss = 0
valid_loss = 0
for b in tqdm(range(int(len(choice) // batch_size))):
batch_x = torch.LongTensor(
shuffled_train_data[b * batch_size:(b + 1) *
batch_size][:, 0].tolist()).t()
batch_y = torch.LongTensor(
shuffled_train_data[b * batch_size:(b + 1) *
batch_size][:, 1].tolist()).t()
batch_x, batch_y = batch_x.to(device), batch_y.to(device)
train_loss += train(batch_x, batch_y, encoder, decoder,
encoder_optimizer, decoder_optimizer,
criterion, False)
train_loss /= b
'''
for b in range(len(valid_data) // batch_size):
batch_x = torch.LongTensor(valid_data[b*batch_size: (b+1)*batch_size][:, 0].tolist()).t()
batch_y = torch.LongTensor(valid_data[b*batch_size: (b+1)*batch_size][:, 1].tolist()).t()
batch_x, batch_y = batch_x.to(device), batch_y.to(device)
valid_loss += train(batch_x, batch_y, max_seqlen, encoder, decoder, encoder_optimizer, decoder_optimizer, criterion, True)
valid_loss /= b
'''
print(
"epoch {}, train_loss {:.4f}, valid_loss {:.4f}, best_epoch {}, best_loss {:.4f}"
.format(e, train_loss, valid_loss, best_epoch, best_loss))
'''
if valid_loss < best_loss:
best_loss = valid_loss
best_epoch = e
torch.save(encoder.state_dict(), args.encoder_path + '.best')
torch.save(decoder.state_dict(), args.decoder_path + '.best')
'''
torch.save(encoder.state_dict(), args.encoder_path)
torch.save(decoder.state_dict(), args.decoder_path)
print(encoder)
print(decoder)
else:
encoder.load_state_dict(
torch.load(args.encoder_path, map_location=torch.device(device)))
decoder.load_state_dict(
torch.load(args.decoder_path, map_location=torch.device(device)))
print(encoder)
print(decoder)
print("==========================================================")
predict(encoder, decoder)
image, padding=2, normalize=False))
netE = Encoder(ngpu=1, nz=nz, nc=3)
netD = Res_Discriminator(channel=6)
netG = ResnetGenerator32(z_dim=nz)
netD2 = ResnetDiscriminator32(stack=6, ch=opt.ch)
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
netE = nn.DataParallel(netE)
netD = nn.DataParallel(netD)
netG = nn.DataParallel(netG)
netD2 = nn.DataParallel(netD2)
netE.to(device)
netD2.to(device)
netG.to(device)
netD.to(device)
optimizerD = optim.Adam(netD.parameters(), lr=opt.lr, betas=(0, .9))
optimizerD2 = optim.Adam(netD2.parameters(), lr=opt.lr, betas=(0, .9))
optimizerG = optim.Adam(netG.parameters(), lr=opt.lr, betas=(0, .9))
optimizerE = optim.Adam(netE.parameters(), lr=opt.lr, betas=(.5, .9))
start = opt.start
def adjust_learning_rate(optimizer, epoch, num_epochs):
"""Sets the learning rate to the initial LR decayed by 10 every 30 epochs"""
lr = opt.lr - opt.lr * (epoch - 45) / (num_epochs - 45)
def train_model(resume):
with open(Path("./cfg/cfg.json").absolute()) as file:
para = json.load(file)
tensorboard_path = Path("./tensorboard/writer").absolute()
checkpoint_dir = Path("./checkpoint").absolute()
writer = SummaryWriter(tensorboard_path)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
encoder = Encoder(in_channels=para['encoder']['in_channels'], channels=para['encoder']['channels'],
n_embeddings=para['encoder']['n_embeddings'], embedding_dim=para['encoder']['embedding_dim'], jitter=para['encoder']['jitter'])
decoder = Decoder(in_channels=para['decoder']['in_channels'], conditioning_channels=para['decoder']['conditioning_channels'],
n_speakers = para['decoder']['n_speakers'], speaker_embedding_dim=para['decoder']['speaker_embedding_dim'],
mu_embedding_dim=para['decoder']['mu_embedding_dim'], rnn_channels=para['decoder']['rnn_channels'], fc_channels=para['decoder']['fc_channels'],
bits=para['decoder']['bits'], hop_length=para['decoder']['hop_length'])
encoder.to(device)
decoder.to(device)
if resume:
resume_path = Path("./checkpoint/model.pt").absolute()
print("Resume checkpoint from: {}:".format(str(resume_path)))
checkpoint = torch.load(resume_path, map_location=lambda storage, loc: storage)
print(checkpoint.keys())
encoder.load_state_dict(checkpoint["encoder"])
decoder.load_state_dict(checkpoint["decoder"])
optimizer = optim.Adam(
chain(encoder.parameters(), decoder.parameters()),
lr=1e-5)
# [encoder, decoder], optimizer = amp.initialize([encoder, decoder], optimizer, opt_level="O1")
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer, milestones=[300000, 400000],
gamma=0.5)
optimizer.load_state_dict(checkpoint["optimizer"])
#amp.load_state_dict(checkpoint["amp"])
scheduler.load_state_dict(checkpoint["scheduler"])
global_step = checkpoint["step"]
else:
global_step = 0
optimizer = optim.Adam(
chain(encoder.parameters(), decoder.parameters()),
lr=1e-5)
# [encoder, decoder], optimizer = amp.initialize([encoder, decoder], optimizer, opt_level="O1")
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer, milestones=[300000, 400000],
gamma=0.5)
sdataset = SpeechDataset(
root='./preprocessed_file/train',
hop_length=para['preprocess']['hop_length'],
sr=para['preprocess']['sr'],
sample_frames=para['preprocess']['sample_frames'])
print(len(sdataset))
dataloader = DataLoader(
dataset=sdataset,
batch_size=16,
shuffle=True,
num_workers=1,
pin_memory=True,
drop_last=True)
print(len(dataloader))
n_epochs = 1
# start_epoch = global_step // len(dataloader) + 1
for epoch in range(global_step, global_step+n_epochs):
average_recon_loss = average_vq_loss = average_perplexity = 0
for i, (audio, mels, speakers) in enumerate(tqdm(dataloader), 1):
#audio, mels, speakers = audio.to(device), mels.to(device), speakers.to(device)
#print(speakers)
optimizer.zero_grad()
z, vq_loss, perplexity = encoder(mels)
output = decoder(audio[:, :-1], z, speakers)
recon_loss = F.cross_entropy(output.transpose(1, 2), audio[:, 1:])
loss = recon_loss + vq_loss
loss.backward()
#with amp.scale_loss(loss, optimizer) as scaled_loss:
# scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), 1)
optimizer.step()
scheduler.step()
average_recon_loss += (recon_loss.item() - average_recon_loss) / i
average_vq_loss += (vq_loss.item() - average_vq_loss) / i
average_perplexity += (perplexity.item() - average_perplexity) / i
global_step += 1
save_checkpoint(
encoder, decoder, optimizer, amp,
scheduler, global_step, checkpoint_dir)
writer.add_scalar("recon_loss/train", average_recon_loss, global_step)
writer.add_scalar("vq_loss/train", average_vq_loss, global_step)
writer.add_scalar("average_perplexity", average_perplexity, global_step)
print("epoch:{}, recon loss:{:.2E}, vq loss:{:.2E}, perpexlity:{:.3f}"
.format(epoch, average_recon_loss, average_vq_loss, average_perplexity))
def main():
input_lang, output_lang, pairs, data1, data2 = read_langs("eng", "fra", True)
input_tensor = [[input_lang.word2index[s] for s in es.split(' ')] for es in data1]
target_tensor = [[output_lang.word2index[s] for s in es.split(' ')] for es in data2]
max_length_inp, max_length_tar = max_length(input_tensor), max_length(target_tensor)
input_tensor = [pad_sequences(x, max_length_inp) for x in input_tensor]
target_tensor = [pad_sequences(x, max_length_tar) for x in target_tensor]
print(len(target_tensor))
input_tensor_train, input_tensor_val, target_tensor_train, target_tensor_val = train_test_split(input_tensor,
target_tensor,
test_size=0.2)
# Show length
print(len(input_tensor_train), len(target_tensor_train), len(input_tensor_val), len(target_tensor_val))
BUFFER_SIZE = len(input_tensor_train)
BATCH_SIZE = 64
N_BATCH = BUFFER_SIZE // BATCH_SIZE
embedding_dim = 256
units = 1024
vocab_inp_size = len(input_lang.word2index)
vocab_tar_size = len(output_lang.word2index)
train_dataset = MyData(input_tensor_train, target_tensor_train)
val_dataset = MyData(input_tensor_val, target_tensor_val)
dataset = DataLoader(train_dataset, batch_size=BATCH_SIZE,
drop_last=True,
shuffle=True)
device = torch.device("cpu")
encoder = Encoder(vocab_inp_size, embedding_dim, units, BATCH_SIZE)
decoder = Decoder(vocab_tar_size, embedding_dim, units, units, BATCH_SIZE)
encoder.to(device)
decoder.to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(list(encoder.parameters()) + list(decoder.parameters()),
lr=0.001)
EPOCHS = 10
for epoch in range(EPOCHS):
start = time()
encoder.train()
decoder.train()
total_loss = 0
for (batch, (inp, targ, inp_len)) in enumerate(dataset):
loss = 0
xs, ys, lens = sort_batch(inp, targ, inp_len)
enc_output, enc_hidden = encoder(xs.to(device), lens, device)
dec_hidden = enc_hidden
dec_input = torch.tensor([[output_lang.word2index['<sos>']]] * BATCH_SIZE)
for t in range(1, ys.size(1)):
predictions, dec_hidden, _ = decoder(dec_input.to(device),
dec_hidden.to(device),
enc_output.to(device))
loss += loss_function(criterion, ys[:, t].to(device), predictions.to(device))
# loss += loss_
dec_input = ys[:, t].unsqueeze(1)
batch_loss = (loss / int(ys.size(1)))
total_loss += batch_loss
optimizer.zero_grad()
loss.backward()
### UPDATE MODEL PARAMETERS
optimizer.step()
if batch % 100 == 0:
print('Epoch {} Batch {} Loss {:.4f}'.format(epoch + 1,
batch,
batch_loss.detach().item()))
### TODO: Save checkpoint for model
print('Epoch {} Loss {:.4f}'.format(epoch + 1,
total_loss / N_BATCH))
print('Time taken for 1 epoch {} sec\n'.format(time.time() - start))
valid_loader = torch.utils.data.DataLoader(CaptionDataset(DATA_FOLDER, 'VAL'),
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=1,
pin_memory=True)
# Optimizer
optimizer = torch.optim.Adam(decoder.parameters(), lr=LEARNING_RATE)
# Parameters check
model_parameters = filter(lambda p: p.requires_grad, decoder.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
print('\n>> {} parameters\n'.format(params))
encoder = encoder.to(DEVICE)
decoder = decoder.to(DEVICE)
#=========================================================================================================
#=========================================================================================================
#================================ 3. TRAINING
for epoch in range(START_EPOCH, START_EPOCH + N_EPOCHS):
decoder.train()
encoder.train()
epoch_loss = 0.
time = datetime.now()
for i, (image, caption, length) in enumerate(tqdm(train_loader)):
def main(args):
# Image preprocessing
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
# Load vocabulary wrapper
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f)
# Build models
if args.model_type == 'no_attention':
encoder = Encoder(args.embed_size).to(device)
decoder = Decoder(args.embed_size, args.hidden_size, len(vocab),
args.num_layers).to(device)
elif args.model_type == 'attention':
encoder = EncoderAtt(encoded_image_size=9).to(device)
decoder = DecoderAtt(vocab, args.encoder_dim, args.hidden_size,
args.attention_dim, args.embed_size,
args.dropout_ratio, args.alpha_c).to(device)
elif args.model_type == 'transformer':
model = Transformer(len(vocab), args.embed_size,
args.transformer_layers, 8,
args.dropout_ratio).eval()
else:
print('Select model_type attention or no_attention')
if args.model_type != 'transformer':
encoder = encoder.to(device)
decoder = decoder.to(device)
# Load the trained model parameters
encoder.load_state_dict(
torch.load(args.encoder_path, map_location=torch.device('cpu')))
decoder.load_state_dict(
torch.load(args.decoder_path, map_location=torch.device('cpu')))
else:
model = model.to(device)
model.load_state_dict(
torch.load(args.model_path, map_location=torch.device('cpu')))
filenames = os.listdir(args.image_dir)
predicted = {}
for file in tqdm(filenames):
if file == '.DS_Store':
continue
# Prepare an image
image = load_image(os.path.join(args.image_dir, file), transform)
image_tensor = image.to(device)
if args.model_type == 'attention':
features = encoder(image_tensor)
sampled_ids, _ = decoder.sample(features)
sampled_ids = sampled_ids[0].cpu().numpy()
sampled_caption = ['<start>']
elif args.model_type == 'no_attention':
features = encoder(image_tensor)
sampled_ids = decoder.sample(features)
sampled_ids = sampled_ids[0].cpu().numpy()
sampled_caption = ['<start>']
elif args.model_type == 'transformer':
e_outputs = model.encoder(image_tensor)
max_seq_length = 20
sampled_ids = torch.zeros(max_seq_length, dtype=torch.long)
sampled_ids[0] = torch.LongTensor([[vocab.word2idx['<start>']]
]).to(device)
for i in range(1, max_seq_length):
trg_mask = np.triu(np.ones((1, i, i)), k=1).astype('uint8')
trg_mask = Variable(torch.from_numpy(trg_mask) == 0).to(device)
out = model.decoder(sampled_ids[:i].unsqueeze(0), e_outputs,
trg_mask)
out = model.out(out)
out = F.softmax(out, dim=-1)
val, ix = out[:, -1].data.topk(1)
sampled_ids[i] = ix[0][0]
sampled_ids = sampled_ids.cpu().numpy()
sampled_caption = []
# Convert word_ids to words
#sampled_caption = []
for word_id in sampled_ids:
word = vocab.idx2word[word_id]
sampled_caption.append(word)
if word == '<end>':
break
sentence = ' '.join(sampled_caption)
#print(sentence)
predicted[file] = sentence
#print(file, sentence)
json.dump(predicted, open(args.predict_json, 'w'))
s_testloader = DataLoader(s_testset, batch_size=batch_size, shuffle=True)
t_trainset, t_testset = load_usps(data_per_class) #transformの指定は禁止
t_trainloader = DataLoader(t_trainset, batch_size=batch_size, shuffle=True)
t_testloader = DataLoader(t_testset, batch_size=64, shuffle=True)
net_g = Encoder()
net_h = classifier()
net_DCD = DCD()
loss_func = torch.nn.CrossEntropyLoss() #損失関数は共通
#ソースにおいてgとhを訓練
print("part 1 : initial training for g and h")
optimizer = torch.optim.Adam(list(net_g.parameters()) +
list(net_h.parameters()),
lr=0.001) #optimizerが両者を更新
net_g = net_g.to(device)
net_h = net_h.to(device)
net_DCD = net_DCD.to(device)
if not device == "cpu":
net_g = nn.DataParallel(net_g)
net_h = nn.DataParallel(net_h)
net_DCD = nn.DataParallel(net_DCD)
for epoch in range(num_ep_init_gh):
for data, label in s_trainloader:
data, label = data.to(device), label.to(device)
optimizer.zero_grad()
pred = net_h(net_g(data))
loss = loss_func(pred, label)
loss.backward()
optimizer.step()
def main():
device = torch.device('cuda:0')
n_features = 256
n_epochs = 40
batch_size = 64
skip_training = False
# Create the transformer model
encoder = Encoder(src_vocab_size=trainset.input_lang.n_words,
n_blocks=3,
n_features=n_features,
n_heads=16,
n_hidden=1024)
decoder = Decoder(tgt_vocab_size=trainset.output_lang.n_words,
n_blocks=3,
n_features=n_features,
n_heads=16,
n_hidden=1024)
encoder.to(device)
decoder.to(device)
# define training loop parameters
parameters = list(encoder.parameters()) + list(decoder.parameters())
adam = torch.optim.Adam(parameters, lr=0, betas=(0.9, 0.98), eps=1e-9)
optimizer = NoamOptimizer(n_features, 2, 10000, adam)
loss_method = nn.NLLLoss(ignore_index=0, reduction='mean')
# prepare data
data_dir = tools.select_data_dir()
trainset = TranslationDataset(data_dir, train=True)
trainloader = DataLoader(dataset=trainset,
batch_size=64,
shuffle=True,
collate_fn=collate,
pin_memory=True)
# training
if not skip_training:
for epoch in range(n_epochs):
loss = training_loop(encoder, decoder, optimizer, loss_method,
trainloader)
print(f'Train Epoch {epoch+1}: Loss: {loss}')
# save and load trained model
tools.save_model(encoder, 'tr_encoder.pth')
tools.save_model(decoder, 'tr_decoder.pth')
else:
encoder = Encoder(src_vocab_size=trainset.input_lang.n_words,
n_blocks=3,
n_features=256,
n_heads=16,
n_hidden=1024)
tools.load_model(encoder, 'tr_encoder.pth', device)
decoder = Decoder(tgt_vocab_size=trainset.output_lang.n_words,
n_blocks=3,
n_features=256,
n_heads=16,
n_hidden=1024)
tools.load_model(decoder, 'tr_decoder.pth', device)
# Generate translations with the trained model
# translate sentences from the training set
print('Translate training data:')
print('-----------------------------')
for i in range(5):
src_sentence, tgt_sentence = trainset[np.random.choice(len(trainset))]
print(
'>', ' '.join(trainset.input_lang.index2word[i.item()]
for i in src_sentence))
print(
'=', ' '.join(trainset.output_lang.index2word[i.item()]
for i in tgt_sentence))
out_sentence = translate(encoder, decoder, src_sentence)
print(
'<', ' '.join(trainset.output_lang.index2word[i.item()]
for i in out_sentence), '\n')
# translate sentences from the test set
testset = TranslationDataset(data_dir, train=False)
print('Translate test data:')
print('-----------------------------')
for i in range(5):
input_sentence, target_sentence = testset[np.random.choice(
len(testset))]
print(
'>', ' '.join(testset.input_lang.index2word[i.item()]
for i in input_sentence))
print(
'=', ' '.join(testset.output_lang.index2word[i.item()]
for i in target_sentence))
output_sentence = translate(encoder, decoder, input_sentence)
print(
'<', ' '.join(testset.output_lang.index2word[i.item()]
for i in output_sentence), '\n')
def DDF(cfg):
filter_list_path = Path(utils.to_absolute_path(cfg.filter_list))
with open(filter_list_path) as file:
filter_list = json.load(file)
in_dir = Path(utils.to_absolute_path(cfg.in_dir))
out_dir = Path(utils.to_absolute_path(cfg.out_dir))
out_dir.mkdir(exist_ok=True, parents=True)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
encoder = Encoder(**cfg.model.encoder)
decoder = Decoder(**cfg.model.decoder)
encoder.to(device)
decoder.to(device)
print("Load checkpoint from: {}:".format(cfg.checkpoint))
checkpoint_path = utils.to_absolute_path(cfg.checkpoint)
checkpoint = torch.load(checkpoint_path,
map_location=lambda storage, loc: storage)
encoder.load_state_dict(checkpoint["encoder"])
decoder.load_state_dict(checkpoint["decoder"])
encoder.eval()
decoder.eval()
meter = pyloudnorm.Meter(cfg.preprocessing.sr)
#---------------------------------------
if cfg.privacy_preference == "Low":
for wav_path, speaker_id, out_filename in tqdm(filter_list):
wav_path = in_dir / wav_path
# librosa.load (it will return audio time series, and its sampling rate)
wav, _ = librosa.load(wav_path.with_suffix(".wav"),
sr=cfg.preprocessing.sr)
ref_loudness = meter.integrated_loudness(wav)
wav = wav / np.abs(wav).max() * 0.999
path = out_dir / out_filename
# to return raw recording in mel-spectrogram without any filtering
if cfg.output_type == "Embedding":
mel = librosa.feature.melspectrogram(
preemphasis(wav, cfg.preprocessing.preemph),
sr=cfg.preprocessing.sr,
n_fft=cfg.preprocessing.n_fft,
n_mels=cfg.preprocessing.n_mels,
hop_length=cfg.preprocessing.hop_length,
win_length=cfg.preprocessing.win_length,
fmin=cfg.preprocessing.fmin,
power=1)
logmel = librosa.amplitude_to_db(
mel, top_db=cfg.preprocessing.top_db)
logmel = logmel / cfg.preprocessing.top_db + 1
mel = torch.FloatTensor(logmel).squeeze().to(device).numpy()
np.savetxt(path.with_suffix(".mel.txt"), mel)
# to return raw recording in waveform without any filtering
if cfg.output_type == "Recording":
librosa.output.write_wav(path.with_suffix(".wav"),
wav.astype(np.float32),
sr=cfg.preprocessing.sr)
#---------------------------------------
if cfg.privacy_preference == "Moderate":
dataset_path = Path(
utils.to_absolute_path("Training/Datasets")) / cfg.dataset.path
with open(dataset_path / "speakers.json") as file:
speakers = sorted(json.load(file))
for wav_path, speaker_id, out_filename in tqdm(filter_list):
wav_path = in_dir / wav_path
wav, _ = librosa.load(wav_path.with_suffix(".wav"),
sr=cfg.preprocessing.sr)
ref_loudness = meter.integrated_loudness(wav)
wav = wav / np.abs(wav).max() * 0.999
mel = librosa.feature.melspectrogram(
preemphasis(wav, cfg.preprocessing.preemph),
sr=cfg.preprocessing.sr,
n_fft=cfg.preprocessing.n_fft,
n_mels=cfg.preprocessing.n_mels,
hop_length=cfg.preprocessing.hop_length,
win_length=cfg.preprocessing.win_length,
fmin=cfg.preprocessing.fmin,
power=1)
logmel = librosa.amplitude_to_db(mel,
top_db=cfg.preprocessing.top_db)
logmel = logmel / cfg.preprocessing.top_db + 1
mel = torch.FloatTensor(logmel).unsqueeze(0).to(device)
speaker = torch.LongTensor([speakers.index(speaker_id)]).to(device)
path = out_dir / out_filename
if cfg.output_type == "Recording":
with torch.no_grad():
vq, _ = encoder.encode(mel)
output = decoder.generate(vq, speaker)
output_loudness = meter.integrated_loudness(output)
output = pyloudnorm.normalize.loudness(output, output_loudness,
ref_loudness)
librosa.output.write_wav(path.with_suffix(".wav"),
output.astype(np.float32),
sr=cfg.preprocessing.sr)
if cfg.output_type == "Embedding":
with torch.no_grad():
vq, _ = encoder.encode(mel)
speaker = decoder.speaker(speaker)
vq = vq.squeeze().to(device).numpy()
speaker = speaker.squeeze().to(device).numpy()
np.savetxt(path.with_suffix(".vq.txt"), vq)
np.savetxt(path.with_suffix(".speaker.txt"), speaker)
#---------------------------------------
if cfg.privacy_preference == "High":
dataset_path = Path(
utils.to_absolute_path("Training/Datasets")) / cfg.dataset.path
with open(dataset_path / "speakers.json") as file:
speakers = sorted(json.load(file))
for wav_path, speaker_id, out_filename in tqdm(filter_list):
wav_path = in_dir / wav_path
wav, _ = librosa.load(wav_path.with_suffix(".wav"),
sr=cfg.preprocessing.sr)
ref_loudness = meter.integrated_loudness(wav)
wav = wav / np.abs(wav).max() * 0.999
mel = librosa.feature.melspectrogram(
preemphasis(wav, cfg.preprocessing.preemph),
sr=cfg.preprocessing.sr,
n_fft=cfg.preprocessing.n_fft,
n_mels=cfg.preprocessing.n_mels,
hop_length=cfg.preprocessing.hop_length,
win_length=cfg.preprocessing.win_length,
fmin=cfg.preprocessing.fmin,
power=1)
logmel = librosa.amplitude_to_db(mel,
top_db=cfg.preprocessing.top_db)
logmel = logmel / cfg.preprocessing.top_db + 1
mel = torch.FloatTensor(logmel).unsqueeze(0).to(device)
speaker = torch.LongTensor([speakers.index(speaker_id)]).to(device)
path = out_dir / out_filename
if cfg.output_type == "Recording":
with torch.no_grad():
vq, _ = encoder.encode(mel)
output = decoder.generate(vq, speaker)
output_loudness = meter.integrated_loudness(output)
output = pyloudnorm.normalize.loudness(output, output_loudness,
ref_loudness)
librosa.output.write_wav(path.with_suffix(".wav"),
output.astype(np.float32),
sr=cfg.preprocessing.sr)
if cfg.output_type == "Embedding":
with torch.no_grad():
vq, _ = encoder.encode(mel)
vq = vq.squeeze().cpu().numpy()
np.savetxt(path.with_suffix(".vq.txt"), vq)
embedding_sd = checkpoint['embedding']
embedding = nn.Embedding(Config.vocab_size, Config.hidden_size)
embedding.load_state_dict(embedding_sd)
encoder = Encoder(embedding)
attn_model = 'dot'
decoder = Decoder(
attn_model,
embedding,
)
encoder.load_state_dict(encoder_sd)
decoder.load_state_dict(decoder_sd)
encoder = encoder.to(Config.device)
decoder = decoder.to(Config.device)
# Set dropout layers to eval mode
encoder.eval()
decoder.eval()
# Initialize search module
searcher = GreedySearchDecoder(encoder, decoder)
vocab2id = json.load(open('./data/vocab2id.json', 'r'))
id2vocab = json.load(open('./data/id2vocab.json', 'r'))
print(id2vocab)
# Begin chatting (uncomment and run the following line to begin)
evaluateInput(encoder, decoder, searcher, vocab2id, id2vocab)
def train(save_path, checkpoint, data_root, batch_size, dataset):
device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
transform = transforms.Compose(
[transforms.Resize((128, 128)),
transforms.ToTensor()])
target_transform = transforms.Compose(
[transforms.Resize((128, 128)),
ToTensor()])
if dataset == 'cityscapes':
train_data = Cityscapes(str(data_root),
split='train',
mode='fine',
target_type='semantic',
transform=transform,
target_transform=transform)
eG = 35
dG = [35, 35, 20, 14, 10, 4, 1]
eC = 8
dC = 280
n_classes = len(Cityscapes.classes)
update_lr = update_lr_default
epoch = 200
else:
train_data = Deepfashion(str(data_root),
split='train',
transform=transform,
target_transform=transform)
n_classes = len(Deepfashion.eclasses)
eG = 8
eC = 64
dG = [8, 8, 4, 4, 2, 2, 1]
dC = 160
update_lr = update_lr_deepfashion
epoch = 100
data_loader = torch.utils.data.DataLoader(train_data,
batch_size=batch_size,
num_workers=1)
os.makedirs(save_path, exist_ok=True)
n_channels = 3
encoder = Encoder(n_classes * n_channels, C=eC, G=eG)
decoder = Decoder(8 * eG, n_channels, n_classes, C=dC, Gs=dG)
discriminator = Discriminator(n_classes + n_channels)
vgg = Vgg19().eval()
encoder = torch.nn.DataParallel(encoder)
decoder = torch.nn.DataParallel(decoder)
discriminator = torch.nn.DataParallel(discriminator)
vgg = torch.nn.DataParallel(vgg)
gen_opt = optim.Adam(list(encoder.parameters()) +
list(decoder.parameters()),
lr=0.0001,
betas=(0, 0.9))
dis_opt = optim.Adam(discriminator.parameters(), lr=0.0004, betas=(0, 0.9))
gen_scheduler = optim.lr_scheduler.LambdaLR(gen_opt, update_lr)
dis_scheduler = optim.lr_scheduler.LambdaLR(gen_opt, update_lr)
params = [
'encoder', 'decoder', 'discriminator', 'gen_opt', 'dis_opt',
'gen_scheduler', 'dis_scheduler'
]
if os.path.exists(checkpoint):
cp = torch.load(checkpoint)
print(f'Load checkpoint: {checkpoint}')
for param in params:
eval(param).load_state_dict(cp[param])
# encoder.load_state_dict(cp['encoder'])
# decoder.load_state_dict(cp['decoder'])
# discriminator.load_state_dict(cp['discriminator'])
# gen_opt.load_state_dict(cp['gen_opt'])
# dis_opt.load_state_dict(cp['dis_opt'])
# gen_scheduler.load_state_dict(cp['gen_scheduler'])
# dis_scheduler.load_state_dict(cp['dis_scheduler'])
def to_device_optimizer(opt):
for state in opt.state.values():
for k, v in state.items():
if isinstance(v, torch.Tensor):
state[k] = v.to(device)
to_device_optimizer(gen_opt)
to_device_optimizer(dis_opt)
encoder = encoder.to(device)
decoder = decoder.to(device)
discriminator = discriminator.to(device)
vgg = vgg.to(device)
print(len(data_loader))
for epoch in range(epoch):
e_g_loss = []
e_d_loss = []
for i, batch in tqdm(enumerate(data_loader)):
x, sem = batch
x = x.to(device)
sem = sem.to(device)
sem = sem * 255.0
sem = sem.long()
s = split_class(x, sem, n_classes)
sem_target = sem.clone()
del sem
sem = torch.zeros(x.size()[0],
n_classes,
sem_target.size()[2],
sem_target.size()[3],
device=x.device)
sem.scatter_(1, sem_target, 1)
s = s.detach()
s = s.to(device)
mu, sigma = encoder(s)
z = mu + torch.exp(0.5 * sigma) * torch.rand(mu.size(),
device=mu.device)
gen = decoder(z, sem)
d_fake = discriminator(gen, sem)
d_real = discriminator(x, sem)
l1loss = nn.L1Loss()
gen_opt.zero_grad()
loss_gen = 0.5 * d_fake[0][-1].mean() + 0.5 * d_fake[1][-1].mean()
loss_fm = sum([
sum([l1loss(f, g) for f, g in zip(fs, rs)])
for fs, rs in zip(d_fake, d_real)
]).mean()
f_fake = vgg(gen)
f_real = vgg(x)
# loss_p = 1.0 / 32 * l1loss(f_fake.relu1_2, f_real.relu1_2) + \
# 1.0 / 16 * l1loss(f_fake.relu2_2, f_real.relu2_2) + \
# 1.0 / 8 * l1loss(f_fake.relu3_3, f_real.relu3_3) + \
# 1.0 / 4 * l1loss(f_fake.relu4_3, f_real.relu4_3) + \
# l1loss(f_fake.relu5_3, f_real.relu5_3)
loss_p = 1.0 / 32 * l1loss(f_fake[0], f_real[0]) + \
1.0 / 16 * l1loss(f_fake[1], f_real[1]) + \
1.0 / 8 * l1loss(f_fake[2], f_real[2]) + \
1.0 / 4 * l1loss(f_fake[3], f_real[3]) + \
l1loss(f_fake[4], f_real[4])
loss_kl = -0.5 * torch.sum(1 + sigma - mu * mu - torch.exp(sigma))
loss = loss_gen + 10.0 * loss_fm + 10.0 * loss_p + 0.05 * loss_kl
loss.backward(retain_graph=True)
gen_opt.step()
dis_opt.zero_grad()
loss_dis = torch.mean(-torch.mean(torch.min(d_real[0][-1] - 1, torch.zeros_like(d_real[0][-1]))) +
-torch.mean(torch.min(-d_fake[0][-1] - 1, torch.zeros_like(d_fake[0][-1])))) + \
torch.mean(-torch.mean(torch.min(d_real[1][-1] - 1, torch.zeros_like(d_real[1][-1]))) +
-torch.mean(torch.min(-d_fake[1][-1] - 1, torch.zeros_like(d_fake[1][-1]))))
loss_dis.backward()
dis_opt.step()
e_g_loss.append(loss.item())
e_d_loss.append(loss_dis.item())
#plt.imshow((gen.detach().cpu().numpy()[0]).transpose(1, 2, 0))
#plt.pause(.01)
#print(i, 'g_loss', e_g_loss[-1], 'd_loss', e_d_loss[-1])
os.makedirs(save_path / str(epoch), exist_ok=True)
Image.fromarray((gen.detach().cpu().numpy()[0].transpose(1, 2, 0) *
255.0).astype(np.uint8)).save(
save_path / str(epoch) / f'{i}.png')
print('g_loss', np.mean(e_g_loss), 'd_loss', np.mean(e_d_loss))
# save
cp = {}
for param in params:
cp[param] = eval(param).state_dict()
torch.save(cp, save_path / 'latest.pth'
) #{param:eval(param).state_dict() for param in params})
def convert(cfg):
dataset_path = Path(utils.to_absolute_path(
"datasets")) / cfg.dataset.path #zerospeech/datasets/2019/english
with open(dataset_path / "speakers.json") as file: # 말하는 사람들 이름 써있는 데이터
speakers = sorted(json.load(file)) # speakers라는 객체로 저장
synthesis_list_path = Path(utils.to_absolute_path(
cfg.synthesis_list)) # ???인걸 보니 우리가 파이썬에서 돌릴때 지정해줘야함
with open(synthesis_list_path) as file:
synthesis_list = json.load(
file) # datasets/2019/english에 있는 synthesis.json보면됨
in_dir = Path(utils.to_absolute_path(
cfg.in_dir)) # ???임. zerospeech 폴더로 경로따면 될듯. (./)
out_dir = Path(utils.to_absolute_path(
cfg.out_dir)) #???임. 목소리 변환된 결과를 저장할 경로
out_dir.mkdir(exist_ok=True, parents=True)
device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu") # gpu안되면 cpu로
encoder = Encoder(
**cfg.model.encoder) #ZeroSpeech/config/model/default에 있는 encoder
decoder = Decoder(
**cfg.model.decoder) #ZeroSpeech/config/model/default에 있는 decoder
encoder.to(device) # cpu or gpu
decoder.to(device) # cpu or gpu
print("Load checkpoint from: {}:".format(cfg.checkpoint)
) ### ???로 되어있는데 pretrained, 혹은 checkpoint까지 학습된 모델 있으면 그 모델의 위치로 지정
checkpoint_path = utils.to_absolute_path(cfg.checkpoint)
checkpoint = torch.load(checkpoint_path,
map_location=lambda storage, loc: storage
) # checkpoint에 지정된 weight들을 불러옵니다
encoder.load_state_dict(checkpoint["encoder"])
decoder.load_state_dict(checkpoint["decoder"])
encoder.eval()
decoder.eval()
meter = pyloudnorm.Meter(
cfg.preprocessing.sr
) #sr:16000으로 조정?? https://www.christiansteinmetz.com/projects-blog/pyloudnorm 소음 관련같습니다..
for wav_path, speaker_id, out_filename in tqdm(
synthesis_list
): #"english/test/S002_0379088085","V002","V002_0379088085"
wav_path = in_dir / wav_path # ./english/test/S002_0379088085
wav, _ = librosa.load(wav_path.with_suffix(".wav"),
sr=cfg.preprocessing.sr)
ref_loudness = meter.integrated_loudness(wav) #인풋의 음량을 측정인듯
wav = wav / np.abs(wav).max() * 0.999
mel = librosa.feature.melspectrogram(
preemphasis(wav, cfg.preprocessing.preemph),
sr=cfg.preprocessing.sr,
n_fft=cfg.preprocessing.n_fft,
n_mels=cfg.preprocessing.n_mels,
hop_length=cfg.preprocessing.hop_length,
win_length=cfg.preprocessing.win_length,
fmin=cfg.preprocessing.fmin,
power=1)
logmel = librosa.amplitude_to_db(mel, top_db=cfg.preprocessing.top_db)
logmel = logmel / cfg.preprocessing.top_db + 1
mel = torch.FloatTensor(logmel).unsqueeze(0).to(
device) #unsqueeze()함수는 인수로 받은 위치에 새로운 차원을 삽입
#https://subinium.github.io/pytorch-Tensor-Variable/#%EB%8D%94%EB%AF%B8-%EC%B0%A8%EC%9B%90-%EC%B6%94%EA%B0%80%EC%99%80-%EC%82%AD%EC%A0%9C--squeeze--unsqueeze
#https://datascienceschool.net/view-notebook/4f3606fd839f4320a4120a56eec1e228/
speaker = torch.LongTensor([speakers.index(speaker_id)
]).to(device) # 마찬가지로 텐서로 만드는데
#텐서에는 자료형이라는 것이 있습니다. 각 데이터형별로 정의되어져 있는데,
#예를 들어 32비트의 유동 소수점은 torch.FloatTensor를, 64비트의 부호 있는 정수는 torch.LongTensor를 사용합니다.
#GPU 연산을 위한 자료형도 있습니다. 예를 들어 torch.cuda.FloatTensor가 그 예입니다.
# 즉 mel은 소수점있고 speaker는 소숫점 없으니까!
with torch.no_grad(
): # 자동미분,벡터연산한 결과의 연산기록 추적못하게 https://bob3rdnewbie.tistory.com/315
z, _ = encoder.encode(mel)
output = decoder.generate(z, speaker)
output_loudness = meter.integrated_loudness(output) #아웃풋의 음량을 측정인듯
output = pyloudnorm.normalize.loudness(output, output_loudness,
ref_loudness)
# 아웃풋의 음량을 input에 넣은 wav의 음량과 동일하게 변경
path = out_dir / out_filename
librosa.output.write_wav(path.with_suffix(".wav"),
output.astype(np.float32),
sr=cfg.preprocessing.sr)
def test(args):
'''
compute bleu score on all images, and average its bleu score
'''
train_json_path = './data/annotations/captions_train2014.json'
test_json_path = './data/annotations/captions_val2014.json'
train_image_dir = './data/train2014'
test_image_dir = './data/val2014'
if args.eval == 'eval':
print('eval bleu')
jsonPath = test_json_path
image_dir = test_image_dir
else:
print('train bleu')
jsonPath = train_json_path
image_dir = train_image_dir
# Image preprocessing
# In generation phase, we need should not random crop, just resize
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
# Load vocabulary wraper
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f)
# Build model
encoder = Encoder(embed_size=args.embed_size).eval()
decoder = Decoder(stateful=False,
embed_size=args.embed_size,
hidden_size=args.hidden_size,
vocab_size=len(vocab),
num_layers=args.num_layers).eval()
encoder = encoder.to(device)
decoder = decoder.to(device)
# load the trained model parameters
encoder.load_state_dict(torch.load(args.encoder_path, map_location=device))
decoder.load_state_dict(torch.load(args.decoder_path, map_location=device))
name_caption_frame = get_image_name(jsonPath)
unique_image_names = pd.unique(name_caption_frame['file_name'])
# Add image directory path train2014 or val2014
unique_image_names = [
os.path.join(image_dir, image_name)
for image_name in unique_image_names
]
total_generated_score4 = []
total_theoratical_score4 = []
total_generated_score3 = []
total_theoratical_score3 = []
total_generated_score2 = []
total_theoratical_score2 = []
total_generated_score1 = []
total_theoratical_score1 = []
# Parallelize the process
def score_helper(image_path):
caption = generate_caption(image_path, vocab, encoder, decoder,
transform)
generated_score4, theoratical_score4 = bleu4_score(
image_path, caption, name_caption_frame)
total_generated_score4.append(generated_score4)
total_theoratical_score4.append(theoratical_score4)
generated_score3, theoratical_score3 = bleu3_score(
image_path, caption, name_caption_frame)
total_generated_score3.append(generated_score3)
total_theoratical_score3.append(theoratical_score3)
generated_score2, theoratical_score2 = bleu2_score(
image_path, caption, name_caption_frame)
total_generated_score2.append(generated_score2)
total_theoratical_score2.append(theoratical_score2)
generated_score1, theoratical_score1 = bleu1_score(
image_path, caption, name_caption_frame)
total_generated_score1.append(generated_score1)
total_theoratical_score1.append(theoratical_score1)
_ = pd.Series(unique_image_names).apply(score_helper)
print('Average bleu-4 score:',
sum(total_generated_score4) / len(total_generated_score4),
' | Average theoratical bleu-4 score:',
sum(total_theoratical_score4) / len(total_theoratical_score4))
print('Average bleu-3 score:',
sum(total_generated_score3) / len(total_generated_score3),
' | Average theoratical bleu-3 score:',
sum(total_theoratical_score3) / len(total_theoratical_score3))
print('Average bleu-2 score:',
sum(total_generated_score2) / len(total_generated_score2),
' | Average theoratical bleu-2 score:',
sum(total_theoratical_score2) / len(total_theoratical_score2))
print('Average bleu-1 score:',
sum(total_generated_score1) / len(total_generated_score1),
' | Average theoratical bleu-1 score:',
sum(total_theoratical_score1) / len(total_theoratical_score1))
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_bleu4 = checkpoint['bleu-4']
decoder = checkpoint['decoder']
decoder_optimizer = checkpoint['decoder_optimizer']
encoder = checkpoint['encoder']
encoder_optimizer = checkpoint['encoder_optimizer']
if fine_tune_encoder is True and encoder_optimizer is None:
encoder.fine_tune(fine_tune_encoder)
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
# Move to GPU, if available
decoder = decoder.to(device)
encoder = encoder.to(device)
# Loss function
criterion = nn.CrossEntropyLoss().to(device)
# Custom dataloaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_loader = torch.utils.data.DataLoader(
CaptionDataset(data_folder,
data_name,
'TRAIN',
transform=transforms.Compose([normalize])),
batch_size=batch_size,
shuffle=True,
def main():
global epochs_since_improvement, best_loss_tr
encoder = Encoder()
decoder = DecoderWithAttention(encoder_dim, lstm_input_dim, decoder_dim,
attention_dim, output_dim)
encoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, encoder.parameters()),
lr=encoder_lr)
decoder_optimizer = torch.optim.Adam(params=filter(
lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = encoder.to(device)
decoder = decoder.to(device)
trainLoader = torch.utils.data.DataLoader(Dataset(driver, circuit_tr,
curvatureLength,
historyLength,
predLength),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
cMean_tr = trainLoader.dataset.cMean
cStd_tr = trainLoader.dataset.cStd
vMean_tr = trainLoader.dataset.vMean
vStd_tr = trainLoader.dataset.vStd
aMean_tr = trainLoader.dataset.aMean
aStd_tr = trainLoader.dataset.aStd
validLoader = torch.utils.data.DataLoader(Dataset(driver,
circuit_vl,
curvatureLength,
historyLength,
predLength,
cMean=cMean_tr,
cStd=cStd_tr,
vMean=vMean_tr,
vStd=vStd_tr,
aMean=aMean_tr,
aStd=aStd_tr),
batch_size=batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
print('Training version.{} (A->V)'.format(vNumber))
print('Training data ({} - {})'.format(driver, circuit_tr))
print('Validation data ({} - {})'.format(driver, circuit_vl))
print('curvature len {}'.format(curvatureLength))
print('history len {}'.format(historyLength))
print('pred len {}'.format(predLength))
print('hiddenDimension {}'.format(hiddenDimension))
print('\nTraining...\n')
for epoch in tqdm(range(start_epoch, epochs)):
loss, vMape, vRmse, vCorr, aCorr = train(
trainLoader=trainLoader,
encoder=encoder,
decoder=decoder,
criterion=criterion,
encoder_optimizer=encoder_optimizer,
decoder_optimizer=decoder_optimizer,
epoch=epoch)
writer.add_scalars('Loss', {'tr': loss}, epoch)
writer.add_scalars('MAPE', {'tr': vMape}, epoch)
writer.add_scalars('RMSE', {'tr': vRmse}, epoch)
writer.add_scalars('vCorr', {'tr': vCorr}, epoch)
writer.add_scalars('aCorr', {'tr': aCorr}, epoch)
is_best = loss < best_loss_tr
best_loss_tr = min(loss, best_loss_tr)
if not is_best:
epochs_since_improvement += 1
print(
'\nEpoch {} Epoch Epochs since last improvement (unit: 100): {}\n'
.format(epoch, epochs_since_improvement))
else:
epochs_since_improvement = 0
if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
adjust_learning_rate(epoch, encoder_optimizer, 0.8)
adjust_learning_rate(epoch, decoder_optimizer, 0.8)
if epoch % 5 == 0:
loss_vl, vMape_vl, vRmse_vl, vCorr_vl, aCorr_vl = validate(
validLoader=validLoader,
encoder=encoder,
decoder=decoder,
criterion=criterion)
writer.add_scalars('Loss', {'vl': loss_vl}, epoch)
writer.add_scalars('MAPE', {'vl': vMape_vl}, epoch)
writer.add_scalars('RMSE', {'vl': vRmse_vl}, epoch)
writer.add_scalars('vCorr', {'vl': vCorr_vl}, epoch)
writer.add_scalars('aCorr', {'vl': aCorr_vl}, epoch)
if epoch % 10 == 0:
save_checkpoint(chptFolderPath, encoder, decoder, epoch, cMean_tr,
cStd_tr, vMean_tr, vStd_tr, aMean_tr, aStd_tr,
curvatureLength, historyLength)
writer.close()
def main():
epoch = 1000
batch_size = 256
hidden_dim = 128
encoder = Encoder(num_words,
hidden_dim,
n_layers=args.n_layers,
bidirectional=args.bidirectional).to(device)
if args.attn:
decoder = AttnDecoder(hidden_dim,
num_words,
max_seqlen,
n_layers=args.n_layers).to(device)
else:
decoder = Decoder(hidden_dim, num_words,
n_layers=args.n_layers).to(device)
if args.train:
weight = torch.ones(num_words)
weight[word2idx[PAD_TOKEN]] = 0
encoder = encoder.to(device)
decoder = decoder.to(device)
weight = weight.to(device)
encoder_optimizer = Adam(encoder.parameters(), lr=0.001)
decoder_optimizer = Adam(decoder.parameters(), lr=0.001)
criterion = nn.CrossEntropyLoss(ignore_index=word2idx[PAD_TOKEN])
np.random.seed(1124)
order = np.arange(len(train_X))
best_loss = 1e10
best_percentage = 0
best_percentage_epoch = 0
best_epoch = 0
start_epoch = 0
if args.resume:
start_epoch, best_loss = load_checkpoint(args.model_path, encoder,
encoder_optimizer,
decoder,
decoder_optimizer)
for e in range(start_epoch, start_epoch + epoch):
if e - best_percentage_epoch > 2: break
np.random.shuffle(order)
shuffled_train_X = train_X[order]
shuffled_train_Y = train_Y[order]
train_loss = 0
valid_loss = 0
for b in tqdm(range(int(len(order) // batch_size))):
batch_x = torch.LongTensor(
shuffled_train_X[b * batch_size:(b + 1) *
batch_size].tolist()).t()
batch_y = torch.LongTensor(
shuffled_train_Y[b * batch_size:(b + 1) *
batch_size].tolist()).t()
batch_x, batch_y = batch_x.to(device), batch_y.to(device)
train_loss += train(batch_x, batch_y, encoder, decoder,
encoder_optimizer, decoder_optimizer,
criterion)
train_loss /= b
all_control_cnt, all_hit_cnt = [], []
for b in range(len(valid_X) // batch_size):
batch_x = torch.LongTensor(valid_X[b * batch_size:(b + 1) *
batch_size].tolist()).t()
batch_y = torch.LongTensor(valid_Y[b * batch_size:(b + 1) *
batch_size].tolist()).t()
batch_x, batch_y = batch_x.to(device), batch_y.to(device)
val_loss, control_cnt, hit_cnt = valid(batch_x, batch_y,
encoder, decoder,
encoder_optimizer,
decoder_optimizer,
criterion)
valid_loss += val_loss
all_control_cnt.extend(control_cnt)
all_hit_cnt.extend(hit_cnt)
valid_loss /= b
all_control_cnt = np.array(all_control_cnt)
all_hit_cnt = np.array(all_hit_cnt)
nonzero = all_control_cnt != 0
all_control_cnt = all_control_cnt[nonzero]
all_hit_cnt = all_hit_cnt[nonzero]
percentage = np.mean(all_hit_cnt / all_control_cnt)
logger.info(
"epoch {}, train_loss {:.4f}, valid_loss {:.4f}, best_epoch {}, best_loss {:.4f}, control_cnt {}, hit_cnt {}, percentage {:.4f}"
.format(e, train_loss, valid_loss, best_epoch, best_loss,
np.sum(all_control_cnt), np.sum(all_hit_cnt),
percentage))
if percentage > best_percentage:
best_percentage = percentage
best_percentage_epoch = e
torch.save(
{
'encoder_state_dict':
encoder.state_dict(),
'encoder_optimizer_state_dict':
encoder_optimizer.state_dict(),
'decoder_state_dict':
decoder.state_dict(),
'decoder_optimizer_state_dict':
decoder_optimizer.state_dict(),
'epoch':
e,
'loss':
valid_loss,
'percentage':
best_percentage,
}, args.model_path)
if valid_loss < best_loss:
best_loss = valid_loss
best_epoch = e
torch.save(
{
'encoder_state_dict':
encoder.state_dict(),
'encoder_optimizer_state_dict':
encoder_optimizer.state_dict(),
'decoder_state_dict':
decoder.state_dict(),
'decoder_optimizer_state_dict':
decoder_optimizer.state_dict(),
'epoch':
e,
'loss':
valid_loss
}, args.model_path)
batch_x = torch.LongTensor(valid_X[:batch_size].tolist()).t()
batch_y = torch.LongTensor(valid_Y[:batch_size].tolist()).t()
batch_x, batch_y = batch_x.to(device), batch_y.to(device)
input_chinese, output_chinese = predict(batch_x, batch_y, encoder,
decoder, encoder_optimizer,
decoder_optimizer, criterion,
20)
logger.info('*** Results ***')
logger.info('Best Hit Accuracy: {}'.format(best_percentage))
logger.info(
'Best Hit Accuracy Epoch: {}'.format(best_percentage_epoch))
for inp, out in zip(input_chinese, output_chinese):
logger.info('{}\t||\t{}'.format(inp, out))
logger.info(encoder)
logger.info(decoder)
logger.info('\n\n' + '=' * 100 + '\n\n')
else:
print(encoder)
print(decoder)
