def load_test_model(opt, model_path=None):
if model_path is None:
model_path = opt.models[0]
checkpoint = torch.load(model_path,
map_location=lambda storage, loc: storage)
model_opt = ArgumentParser.ckpt_model_opts(checkpoint['opt'])
ArgumentParser.update_model_opts(model_opt)
ArgumentParser.validate_model_opts(model_opt)
vocab = checkpoint['vocab']
if inputters.old_style_vocab(vocab):
fields = inputters.load_old_vocab(vocab,
opt.data_type,
dynamic_dict=model_opt.copy_attn)
else:
fields = vocab
model = build_base_model(model_opt, fields, use_gpu(opt), checkpoint,
opt.gpu)
if opt.fp32:
model.float()
model.eval()
model.generator.eval()
# TODO(yida)
if model_opt.pos_gen:
model.pos_generator.eval()
return fields, model, model_opt
def load_test_multitask_model(opt, model_path=None):
if model_path is None:
model_path = opt.models[0]
checkpoint = torch.load(model_path,
map_location=lambda storage, loc: storage)
model = checkpoint['whole_model']
vocab = checkpoint['vocab']
src_tgtpair = opt.src_lang + '-' + opt.tgt_lang
vocab = vocab if type(vocab) is dict else vocab[src_tgtpair]
if inputters.old_style_vocab(vocab):
fields = inputters.load_old_vocab(vocab,
opt.data_type,
dynamic_dict=model_opt.copy_attn)
else:
fields = vocab
if opt.data_type == 'audio' and not (isinstance(
checkpoint.get('vocab')[src_tgtpair]['src'], AudioSeqField)):
vocab_path = "/home/vazquezj/Documents/iwslt2019/_ready_to_train/onmt_ready/ENaudio_DEtext/data"
fields = torch.load(vocab_path + '.vocab.pt')
model_opt = ArgumentParser.ckpt_model_opts(checkpoint['opt'])
device = torch.device("cuda" if use_gpu(opt) else "cpu")
model.to(device)
model.eval()
return fields, model, model_opt
def load_test_model(opt, model_path=None):
if model_path is None:
model_path = opt.models[0]
checkpoint = torch.load(model_path,
map_location=lambda storage, loc: storage)
model_opt = ArgumentParser.ckpt_model_opts(checkpoint['opt'])
ArgumentParser.update_model_opts(model_opt)
ArgumentParser.validate_model_opts(model_opt)
vocab = checkpoint['vocab']
if inputters.old_style_vocab(vocab):
fields = inputters.load_old_vocab(
vocab, opt.data_type, dynamic_dict=model_opt.copy_attn
)
else:
fields = vocab
# @memray, to make tgt_field be aware of format of targets (multiple phrases)
if opt.data_type == "keyphrase":
fields["tgt"].type = opt.tgt_type
model = build_base_model(model_opt, fields, use_gpu(opt), checkpoint,
opt.gpu)
if opt.fp32:
model.float()
model.eval()
model.generator.eval()
return fields, model, model_opt
def load_test_model(opt, dummy_opt, model_path=None):
if model_path is None:
model_path = opt.models[0]
checkpoint = torch.load(model_path,
map_location=lambda storage, loc: storage)
vocab = checkpoint['vocab']
if inputters.old_style_vocab(vocab):
fields = inputters.load_fields_from_vocab(vocab, opt.data_type)
else:
fields = vocab
model_opt = checkpoint['opt']
for arg in dummy_opt:
if arg not in model_opt:
model_opt.__dict__[arg] = dummy_opt[arg]
# changed to my_build_base_model by wchen
if 'hr' in model_opt.encoder_type or 'hr' in model_opt.decoder_type or 'CatSeqD' in model_opt.decoder_type:
model = my_build_base_model(model_opt, fields, use_gpu(opt),
checkpoint)
else:
model = build_base_model(model_opt, fields, use_gpu(opt), checkpoint)
model.eval()
model.generator.eval()
return fields, model, model_opt
def load_test_model(opt, args):
model_path = opt.models[0]
checkpoint = torch.load(model_path,
map_location=lambda storage, loc: storage)
model_opt = ArgumentParser.ckpt_model_opts(checkpoint['opt'])
ArgumentParser.update_model_opts(model_opt)
ArgumentParser.validate_model_opts(model_opt)
vocab = checkpoint['vocab']
if inputters.old_style_vocab(vocab):
fields = inputters.load_old_vocab(vocab,
opt.data_type,
dynamic_dict=model_opt.copy_attn)
else:
fields = vocab
model = build_base_model(model_opt, fields, use_gpu(opt), args, checkpoint,
opt.gpu)
if args.data_type == 'fp32':
model.float()
elif args.data_type == 'fp16':
model.half()
else:
raise ValueError('wrong data_type argument {}'.format(args.data_type))
model.eval()
model.generator.eval()
return fields, model, model_opt
def ltm(opt, model_path=None):
if model_path is None:
model_path = opt.models[0]
checkpoint = torch.load(model_path,
map_location=lambda storage, loc: storage)
model_opt = ArgumentParser.ckpt_model_opts(checkpoint["opt"])
ArgumentParser.update_model_opts(model_opt)
ArgumentParser.validate_model_opts(model_opt)
vocab = checkpoint["vocab"]
if inputters.old_style_vocab(vocab):
fields = inputters.load_old_vocab(vocab,
opt.data_type,
dynamic_dict=model_opt.copy_attn)
else:
fields = vocab
# This will randomly initialize
if settings.RANDOM_WEIGHTS:
checkpoint = None
model = build_base_model(model_opt, fields, use_gpu(opt), checkpoint,
opt.gpu)
if opt.fp32:
model.float()
model.eval()
model.generator.eval()
return fields, model, model_opt
def load_test_model(opt, model_path=None):
if model_path is None:
model_path = opt.models[0]
checkpoint = torch.load(model_path,
map_location=lambda storage, loc: storage)
model_opt = ArgumentParser.ckpt_model_opts(checkpoint['opt'])
ArgumentParser.update_model_opts(model_opt)
ArgumentParser.validate_model_opts(model_opt)
vocab = checkpoint['vocab']
if inputters.old_style_vocab(vocab):
fields = inputters.load_old_vocab(
vocab, opt.data_type, dynamic_dict=model_opt.copy_attn
)
else:
fields = vocab
arae_model_path = opt.model_arae if opt.arae and checkpoint else None
model = build_base_model(model_opt, fields, use_gpu(opt), checkpoint,
opt.gpu, arae_setting=opt.arae, arae_model_path=arae_model_path)
if opt.arae:
model, gan_g, gan_d = model
gan_g.eval()
gan_d.eval()
if opt.fp32:
model.float()
model.eval()
model.generator.eval()
if opt.arae:
model = model, gan_g, gan_d
return fields, model, model_opt
def load_test_model(opt, model_path=None):
if model_path is None:
model_path = opt.models
checkpoint = torch.load(model_path[0],
map_location=lambda storage, loc: storage)
model_opt = ArgumentParser.ckpt_model_opts(checkpoint['opt'])
ArgumentParser.update_model_opts(model_opt)
ArgumentParser.validate_model_opts(model_opt)
vocab = checkpoint['vocab']
teacher_vocab = checkpoint['teacher_vocab']
if inputters.old_style_vocab(vocab):
fields = inputters.load_old_vocab(vocab,
opt.data_type,
dynamic_dict=model_opt.copy_attn)
else:
fields = vocab
teacher_fields = teacher_vocab
fields_opt = {'original': fields, 'teacher': teacher_fields}
# setattr(fields,"true_tgt_vocab",true_tgt_field.vocab)
model = build_base_model(model_opt, fields_opt, use_gpu(opt), checkpoint,
opt.gpu)
if opt.fp32:
model.float()
model.eval()
model.generator.eval()
return fields, model, model_opt
def main():
dummy_parser = argparse.ArgumentParser(description='train.py')
onmt.opts.model_opts(dummy_parser)
dummy_opt = dummy_parser.parse_known_args([])[0]
opt = parser.parse_args()
opt.cuda = opt.gpu > -1
if opt.cuda:
torch.cuda.set_device(opt.gpu)
# Add in default model arguments, possibly added since training.
checkpoint = torch.load(opt.model,
map_location=lambda storage, loc: storage)
model_opt = checkpoint['opt']
vocab = checkpoint['vocab']
if inputters.old_style_vocab(vocab):
fields = onmt.inputters.load_old_vocab(vocab)
else:
fields = vocab
src_dict = fields['src'].base_field.vocab # assumes src is text
tgt_dict = fields['tgt'].base_field.vocab
model_opt = checkpoint['opt']
for arg in dummy_opt.__dict__:
if arg not in model_opt:
model_opt.__dict__[arg] = dummy_opt.__dict__[arg]
model = onmt.model_builder.build_base_model(model_opt, fields,
use_gpu(opt), opt.length_model,
checkpoint)
encoder = model.encoder
decoder = model.decoder
encoder_embeddings = encoder.embeddings.word_lut.weight.data.tolist()
decoder_embeddings = decoder.embeddings.word_lut.weight.data.tolist()
logger.info("Writing source embeddings")
write_embeddings(opt.output_dir + "/src_embeddings.txt", src_dict,
encoder_embeddings)
logger.info("Writing target embeddings")
write_embeddings(opt.output_dir + "/tgt_embeddings.txt", tgt_dict,
decoder_embeddings)
logger.info('... done.')
logger.info('Converting model...')
def load_test_model(opt, dummy_opt, model_path=None):
if model_path is None:
model_path = opt.models[0]
checkpoint = torch.load(model_path,
map_location=lambda storage, loc: storage)
model_opt = checkpoint['opt']
vocab = checkpoint['vocab']
if inputters.old_style_vocab(vocab):
fields = inputters.load_old_vocab(vocab,
opt.data_type,
dynamic_dict=model_opt.copy_attn)
else:
fields = vocab
for arg in dummy_opt:
if arg not in model_opt:
model_opt.__dict__[arg] = dummy_opt[arg]
model = build_base_model(model_opt, fields, use_gpu(opt), checkpoint)
model.eval()
model.generator.eval()
return fields, model, model_opt
def load_test_model(opt, model_path=None):
if model_path is None:
model_path = opt.models[0]
checkpoint = torch.load(model_path,
map_location=lambda storage, loc: storage)
model_opt = ArgumentParser.ckpt_model_opts(checkpoint['opt'])
ArgumentParser.update_model_opts(model_opt)
ArgumentParser.validate_model_opts(model_opt)
vocab = checkpoint['vocab']
if inputters.old_style_vocab(vocab):
fields = inputters.load_old_vocab(
vocab, opt.data_type, dynamic_dict=model_opt.copy_attn
)
else:
fields = vocab
model = build_base_model(model_opt, fields, use_gpu(opt), checkpoint,
opt.gpu)
if opt.fp32:
model.float()
model.eval()
model.generator.eval()
return fields, model, model_opt
def __init__(self, model_dir):
# Model dir
self._model_dir = os.path.abspath(model_dir)
if not os.path.isdir(self._model_dir):
msg = f"{model_dir} doesn't exists'"
raise ValueError(msg)
# Extended model
self._extended_model = ExtendedModel(model_dir)
# Config
self._config = self._extended_model.config
# Options
self._opts = self._config.opts
# Get the model options
model_path = self._opts.models[0]
checkpoint = torch.load(
model_path, map_location=lambda storage, loc: storage
)
self._model_opts = ArgumentParser.ckpt_model_opts(checkpoint["opt"])
ArgumentParser.update_model_opts(self._model_opts)
ArgumentParser.validate_model_opts(self._model_opts)
# Extract vocabulary
vocab = checkpoint["vocab"]
if inputters.old_style_vocab(vocab):
self._fields = inputters.load_old_vocab(
vocab, "text", dynamic_dict=False
)
else:
self._fields = vocab
# Train_steps
self._train_steps = self._model_opts.train_steps
# Build openmmt model
self._opennmt_model = build_base_model(
self._model_opts,
self._fields,
use_gpu(self._opts),
checkpoint,
self._opts.gpu,
)
# Translator
try:
min_length = self._opts.min_length
except:
min_length = 0
try:
max_length = self._opts.max_length
except:
max_length = 100
try:
beam_size = self._opts.beam_size
except:
beam_size = 5
try:
replace_unk = self._opts.replace_unk
except:
replace_unk = 0
self._translator = Translator(
self._opennmt_model,
self._fields,
TextDataReader(),
TextDataReader(),
gpu=self._opts.gpu,
min_length=min_length,
max_length=max_length,
beam_size=beam_size,
replace_unk=replace_unk,
copy_attn=self._model_opts.copy_attn,
global_scorer=GNMTGlobalScorer(0.0, -0.0, "none", "none"),
seed=self.SEED,
)
online_learning = self._config.online_learning
if online_learning:
# Optim
optimizer_opt = type("", (), {})()
optimizer_opt.optim = "sgd"
optimizer_opt.learning_rate = self._opts.learning_rate
optimizer_opt.train_from = ""
optimizer_opt.adam_beta1 = 0
optimizer_opt.adam_beta2 = 0
optimizer_opt.model_dtype = "fp32"
optimizer_opt.decay_method = "none"
optimizer_opt.start_decay_steps = 100000
optimizer_opt.learning_rate_decay = 1.0
optimizer_opt.decay_steps = 100000
optimizer_opt.max_grad_norm = 5
self._optim = Optimizer.from_opt(
self._opennmt_model, optimizer_opt, checkpoint=None
)
trainer_opt = type("", (), {})()
trainer_opt.lambda_coverage = 0.0
trainer_opt.copy_attn = False
trainer_opt.label_smoothing = 0.0
trainer_opt.truncated_decoder = 0
trainer_opt.model_dtype = "fp32"
trainer_opt.max_generator_batches = 32
trainer_opt.normalization = "sents"
trainer_opt.accum_count = [1]
trainer_opt.accum_steps = [0]
trainer_opt.world_size = 1
trainer_opt.average_decay = 0
trainer_opt.average_every = 1
trainer_opt.dropout = 0
trainer_opt.dropout_steps = (0,)
trainer_opt.gpu_verbose_level = 0
trainer_opt.early_stopping = 0
trainer_opt.early_stopping_criteria = (None,)
trainer_opt.tensorboard = False
trainer_opt.report_every = 50
trainer_opt.gpu_ranks = []
if self._opts.gpu != -1:
trainer_opt.gpu_ranks = [self._opts.gpu]
self._trainer = build_trainer(
trainer_opt,
self._opts.gpu,
self._opennmt_model,
self._fields,
self._optim,
)
else:
self._trainer = None
def __init__(self, model_dir):
# Model dir
self._model_dir = os.path.abspath(model_dir)
if not os.path.isdir(self._model_dir):
msg = f"{model_dir} doesn't exists'"
raise ValueError(msg)
# Extended model
self._extended_model = ExtendedModel(model_dir)
# Config
self._config = self._extended_model.config
# Options
self._opts = self._config.opts
# Get the model options
model_path = self._opts.models[0]
checkpoint = torch.load(model_path,
map_location=lambda storage, loc: storage)
self._model_opts = ArgumentParser.ckpt_model_opts(checkpoint['opt'])
ArgumentParser.update_model_opts(self._model_opts)
ArgumentParser.validate_model_opts(self._model_opts)
# Train_steps
self._train_steps = self._model_opts.train_steps
# Extract vocabulary
vocab = checkpoint['vocab']
if inputters.old_style_vocab(vocab):
self._fields = inputters.load_old_vocab(
vocab,
self._opts.data_type,
dynamic_dict=self._model_opts.copy_attn)
else:
self._fields = vocab
# Build model
self._model = build_base_model(self._model_opts, self._fields,
use_gpu(self._opts), checkpoint,
self._opts.gpu)
if self._opts.fp32:
self._model.float()
#Translator
scorer = GNMTGlobalScorer.from_opt(self._opts)
self.translator = OnmtxTranslator.from_opt(
self._model,
self._fields,
self._opts,
self._model_opts,
global_scorer=scorer,
out_file=None,
report_score=False,
logger=None,
)
# Create trainer
self._optim = Optimizer.from_opt(self._model,
self._opts,
checkpoint=checkpoint)
device_id = -1 # TODO Handle GPU
self.trainer = build_trainer(self._opts, device_id, self._model,
self._fields, self._optim)
