def __init__(self, opt):
self.opt = opt
self.tt = torch.cuda if opt.cuda else torch
self.beam_accum = None
self.beta = opt.beta
self.alpha = opt.alpha
self.start_with_bos = opt.start_with_bos
self.fp16 = opt.fp16
self.attributes = opt.attributes # attributes split by |. for example: de|domain1
self.bos_token = opt.bos_token
self.sampling = opt.sampling
if self.attributes:
self.attributes = self.attributes.split("|")
self.models = list()
self.model_types = list()
# models are string with | as delimiter
models = opt.model.split("|")
print(models)
self.n_models = len(models)
self._type = 'text'
for i, model in enumerate(models):
if opt.verbose:
print('Loading model from %s' % model)
checkpoint = torch.load(model,
map_location=lambda storage, loc: storage)
model_opt = checkpoint['opt']
if i == 0:
if "src" in checkpoint['dicts']:
self.src_dict = checkpoint['dicts']['src']
else:
self._type = "audio"
self.tgt_dict = checkpoint['dicts']['tgt']
if "atb" in checkpoint["dicts"]:
self.atb_dict = checkpoint['dicts']['atb']
else:
self.atb_dict = None
self.bos_id = self.tgt_dict.labelToIdx[self.bos_token]
# Build model from the saved option
# if hasattr(model_opt, 'fusion') and model_opt.fusion == True:
# print("* Loading a FUSION model")
# model = build_fusion(model_opt, checkpoint['dicts'])
# else:
# model = build_model(model_opt, checkpoint['dicts'])
model = build_model(model_opt, checkpoint['dicts'])
model.load_state_dict(checkpoint['model'])
if model_opt.model in model_list:
# if model.decoder.positional_encoder.len_max < self.opt.max_sent_length:
# print("Not enough len to decode. Renewing .. ")
# model.decoder.renew_buffer(self.opt.max_sent_length)
model.renew_buffer(self.opt.max_sent_length)
if opt.fp16:
model = model.half()
if opt.cuda:
model = model.cuda()
else:
model = model.cpu()
model.eval()
self.models.append(model)
self.model_types.append(model_opt.model)
# language model
if opt.lm is not None:
if opt.verbose:
print('Loading language model from %s' % opt.lm)
lm_chkpoint = torch.load(opt.lm, map_location=lambda storage, loc: storage)
lm_opt = lm_chkpoint['opt']
lm_model = build_language_model(lm_opt, checkpoint['dicts'])
if opt.fp16:
lm_model = lm_model.half()
if opt.cuda:
lm_model = lm_model.cuda()
else:
lm_model = lm_model.cpu()
self.lm_model = lm_model
self.cuda = opt.cuda
self.ensemble_op = opt.ensemble_op
if opt.autoencoder is not None:
if opt.verbose:
print('Loading autoencoder from %s' % opt.autoencoder)
checkpoint = torch.load(opt.autoencoder,
map_location=lambda storage, loc: storage)
model_opt = checkpoint['opt']
# posSize= checkpoint['autoencoder']['nmt.decoder.positional_encoder.pos_emb'].size(0)
# self.models[0].decoder.renew_buffer(posSize)
# self.models[0].decoder.renew_buffer(posSize)
# Build model from the saved option
self.autoencoder = Autoencoder(self.models[0], model_opt)
self.autoencoder.load_state_dict(checkpoint['autoencoder'])
if opt.cuda:
self.autoencoder = self.autoencoder.cuda()
self.models[0] = self.models[0].cuda()
else:
self.autoencoder = self.autoencoder.cpu()
self.models[0] = self.models[0].cpu()
self.models[0].autoencoder = self.autoencoder
if opt.verbose:
print('Done')
def __init__(self, opt):
self.opt = opt
self.tt = torch.cuda if opt.cuda else torch
self.start_with_bos = opt.start_with_bos
self.fp16 = opt.fp16
self.models = list()
self.model_types = list()
# models are string with | as delimiter
models = opt.model.split("|")
print(models)
self.n_models = len(models)
self._type = 'text'
check_m = None
for i, model in enumerate(models):
if opt.verbose:
print('Loading model from %s' % model)
checkpoint = torch.load(model,
map_location=lambda storage, loc: storage)
model_opt = checkpoint['opt']
if i == 0:
if ("src" in checkpoint['dicts']):
self.src_dict = checkpoint['dicts']['src']
else:
self._type = "audio"
self.tgt_dict = checkpoint['dicts']['tgt']
# Build model from the saved option
model = build_model(model_opt, checkpoint['dicts'])
model.load_state_dict(checkpoint['model'])
check_m = checkpoint['model']
if opt.cuda:
model = model.cuda()
else:
model = model.cpu()
if opt.fp16:
model = model.half()
model.eval()
self.models.append(model)
self.model_types.append(model_opt.model)
self.cuda = opt.cuda
## Autoencoder
if opt.verbose:
print('Loading autoencoder from %s' % opt.autoencoder)
checkpoint = torch.load(opt.autoencoder,
map_location=lambda storage, loc: storage)
model_opt = checkpoint['opt']
posSize = checkpoint['autoencoder'][
'nmt.decoder.positional_encoder.pos_emb'].size(0)
self.models[0].decoder.renew_buffer(posSize)
# Build model from the saved option
self.autoencoder = Autoencoder(self.models[0], model_opt)
self.autoencoder.load_state_dict(checkpoint['autoencoder'])
for k in checkpoint['autoencoder']:
if (k.startswith("nmt") and k[4:] in check_m):
n = checkpoint['autoencoder'][k]
o = check_m[k[4:]]
if (o.size() != n.size()):
print("Different size:", k[4:])
elif ((n - o).sum() != 0):
print("Different weight:", k[4:])
if self.autoencoder.nmt.decoder.positional_encoder.len_max < self.opt.max_sent_length:
self.autoencoder.nmt.decoder.renew_buffer(self.opt.max_sent_length)
if opt.cuda:
self.autoencoder = self.autoencoder.cuda()
else:
self.autoencoder = self.autoencoder.cpu()
if opt.fp16:
self.autoencoder = self.autoencoder.half()
self.autoencoder.eval()
if opt.verbose:
print('Done')
def main():
if opt.data_format == 'raw':
start = time.time()
print("Loading data from '%s'" % opt.data)
if opt.data.endswith(".train.pt"):
print("Loading data from '%s'" % opt.data)
dataset = torch.load(opt.data)
else:
print("Loading data from %s" % opt.data + ".train.pt")
dataset = torch.load(opt.data + ".train.pt")
elapse = str(datetime.timedelta(seconds=int(time.time() - start)))
print("Done after %s" % elapse)
trainData = onmt.Dataset(dataset['train']['src'],
dataset['train']['tgt'], opt.batch_size_words,
data_type=dataset.get("type", "text"),
batch_size_sents=opt.batch_size_sents,
multiplier=opt.batch_size_multiplier)
validData = onmt.Dataset(dataset['valid']['src'],
dataset['valid']['tgt'], opt.batch_size_words,
data_type=dataset.get("type", "text"),
batch_size_sents=opt.batch_size_sents)
dicts = dataset['dicts']
if ("src" in dicts):
print(' * vocabulary size. source = %d; target = %d' %
(dicts['src'].size(), dicts['tgt'].size()))
else:
print(' * vocabulary size. target = %d' %
(dicts['tgt'].size()))
print(' * number of training sentences. %d' %
len(dataset['train']['src']))
print(' * maximum batch size (words per batch). %d' % opt.batch_size_words)
elif opt.data_format == 'bin':
from onmt.data_utils.IndexedDataset import IndexedInMemoryDataset
dicts = torch.load(opt.data + ".dict.pt")
# ~ train = {}
train_path = opt.data + '.train'
train_src = IndexedInMemoryDataset(train_path + '.src')
train_tgt = IndexedInMemoryDataset(train_path + '.tgt')
trainData = onmt.Dataset(train_src,
train_tgt, opt.batch_size_words,
batch_size_sents=opt.batch_size_sents,
multiplier=opt.batch_size_multiplier)
valid_path = opt.data + '.valid'
valid_src = IndexedInMemoryDataset(valid_path + '.src')
valid_tgt = IndexedInMemoryDataset(valid_path + '.tgt')
validData = onmt.Dataset(valid_src,
valid_tgt, opt.batch_size_words,
batch_size_sents=opt.batch_size_sents)
else:
raise NotImplementedError
print('Building model...')
model = build_model(opt, dicts)
autoencoder = Autoencoder(model,opt)
""" Building the loss function """
loss_function = nn.MSELoss(size_average=False)
nParams = sum([p.nelement() for p in autoencoder.parameters()])
print('* number of parameters: %d' % nParams)
# load nmt model
checkpoint = None
if opt.load_from:
checkpoint = torch.load(opt.load_from, map_location=lambda storage, loc: storage)
else:
raise NotImplementedError
if checkpoint is not None:
print('Loading model from checkpoint at %s' % opt.load_from)
model.load_state_dict(checkpoint['model'])
del checkpoint['model']
del checkpoint['optim']
del checkpoint
if len(opt.gpus) > 1 or opt.virtual_gpu > 1:
# ~ trainer = MultiGPUXETrainer(model, loss_function, trainData, validData, dataset, opt)
raise NotImplementedError("Warning! Multi-GPU training is not fully tested and potential bugs can happen.")
else:
trainer = AETrainer(autoencoder,model, loss_function, trainData, validData, dicts, opt)
trainer.run(save_file=False)
def __init__(self, opt):
self.opt = opt
self.tt = torch.cuda if opt.cuda else torch
self.beam_accum = None
self.beta = opt.beta
self.alpha = opt.alpha
self.start_with_bos = opt.start_with_bos
self.fp16 = opt.fp16
self.attributes = opt.attributes # attributes split by |. for example: de|domain1
# self.bos_token = opt.bos_token
self.sampling = opt.sampling
self.src_lang = opt.src_lang
self.tgt_lang = opt.tgt_lang
if self.attributes:
self.attributes = self.attributes.split("|")
self.models = list()
self.model_types = list()
# models are string with | as delimiter
models = opt.model.split("|")
print(models)
self.n_models = len(models)
self._type = 'text'
for i, model_path in enumerate(models):
checkpoint = torch.load(model_path,
map_location=lambda storage, loc: storage)
model_opt = checkpoint['opt']
model_opt = backward_compatible(model_opt)
if hasattr(model_opt, "enc_state_dict"):
model_opt.enc_state_dict = None
model_opt.dec_state_dict = None
self.main_model_opt = model_opt
dicts = checkpoint['dicts']
# update special tokens
onmt.constants = add_tokenidx(model_opt, onmt.constants, dicts)
self.bos_token = model_opt.tgt_bos_word
if i == 0:
if "src" in checkpoint['dicts']:
self.src_dict = checkpoint['dicts']['src']
else:
self._type = "audio"
# self.src_dict = self.tgt_dict
self.tgt_dict = checkpoint['dicts']['tgt']
if "langs" in checkpoint["dicts"]:
self.lang_dict = checkpoint['dicts']['langs']
else:
self.lang_dict = {'src': 0, 'tgt': 1}
self.bos_id = self.tgt_dict.labelToIdx[self.bos_token]
model = build_model(model_opt, checkpoint['dicts'])
optimize_model(model)
if opt.verbose:
print('Loading model from %s' % model_path)
model.load_state_dict(checkpoint['model'])
if model_opt.model in model_list:
# if model.decoder.positional_encoder.len_max < self.opt.max_sent_length:
# print("Not enough len to decode. Renewing .. ")
# model.decoder.renew_buffer(self.opt.max_sent_length)
model.renew_buffer(self.opt.max_sent_length)
# model.convert_autograd()
if opt.fp16:
model = model.half()
if opt.cuda:
model = model.cuda()
else:
model = model.cpu()
if opt.dynamic_quantile == 1:
engines = torch.backends.quantized.supported_engines
if 'fbgemm' in engines:
torch.backends.quantized.engine = 'fbgemm'
else:
print(
"[INFO] fbgemm is not found in the available engines. Possibly the CPU does not support AVX2."
" It is recommended to disable Quantization (set to 0)."
)
torch.backends.quantized.engine = 'qnnpack'
# convert the custom functions to their autograd equivalent first
model.convert_autograd()
model = torch.quantization.quantize_dynamic(
model, {torch.nn.LSTM, torch.nn.Linear}, dtype=torch.qint8)
model.eval()
self.models.append(model)
self.model_types.append(model_opt.model)
# language model
if opt.lm is not None:
if opt.verbose:
print('Loading language model from %s' % opt.lm)
lm_chkpoint = torch.load(opt.lm,
map_location=lambda storage, loc: storage)
lm_opt = lm_chkpoint['opt']
lm_model = build_language_model(lm_opt, checkpoint['dicts'])
if opt.fp16:
lm_model = lm_model.half()
if opt.cuda:
lm_model = lm_model.cuda()
else:
lm_model = lm_model.cpu()
self.lm_model = lm_model
self.cuda = opt.cuda
self.ensemble_op = opt.ensemble_op
if opt.autoencoder is not None:
if opt.verbose:
print('Loading autoencoder from %s' % opt.autoencoder)
checkpoint = torch.load(opt.autoencoder,
map_location=lambda storage, loc: storage)
model_opt = checkpoint['opt']
# posSize= checkpoint['autoencoder']['nmt.decoder.positional_encoder.pos_emb'].size(0)
# self.models[0].decoder.renew_buffer(posSize)
# self.models[0].decoder.renew_buffer(posSize)
# Build model from the saved option
self.autoencoder = Autoencoder(self.models[0], model_opt)
self.autoencoder.load_state_dict(checkpoint['autoencoder'])
if opt.cuda:
self.autoencoder = self.autoencoder.cuda()
self.models[0] = self.models[0].cuda()
else:
self.autoencoder = self.autoencoder.cpu()
self.models[0] = self.models[0].cpu()
self.models[0].autoencoder = self.autoencoder
if opt.verbose:
print('Done')