def __init__(self, args=None, vocab=None, pretrain=None, model_file=None, use_cuda=False): self.use_cuda = use_cuda if model_file is not None: # load everything from file self.load(pretrain, model_file) else: assert all(var is not None for var in [args, vocab, pretrain]) # build model from scratch self.args = args self.vocab = vocab self.model = Parser(args, vocab, emb_matrix=pretrain.emb) self.parameters = [ p for p in self.model.parameters() if p.requires_grad ] if self.use_cuda: self.model.cuda() else: self.model.cpu() self.optimizer = utils.get_optimizer(self.args['optim'], self.parameters, self.args['lr'], betas=(0.9, self.args['beta2']), eps=1e-6)
def __init__(self, args=None, vocab=None, emb_matrix=None, model_file=None, use_cuda=False): self.use_cuda = use_cuda if model_file is not None: # load everything from file self.load(model_file, use_cuda) else: # build model from scratch self.args = args self.model = None if args['dict_only'] else Seq2SeqModel(args, emb_matrix=emb_matrix, use_cuda=use_cuda) self.vocab = vocab # dict-based components self.word_dict = dict() self.composite_dict = dict() if not self.args['dict_only']: if self.args.get('edit', False): self.crit = loss.MixLoss(self.vocab['char'].size, self.args['alpha']) print("[Running seq2seq lemmatizer with edit classifier]") else: self.crit = loss.SequenceLoss(self.vocab['char'].size) self.parameters = [p for p in self.model.parameters() if p.requires_grad] if use_cuda: self.model.cuda() self.crit.cuda() else: self.model.cpu() self.crit.cpu() self.optimizer = utils.get_optimizer(self.args['optim'], self.parameters, self.args['lr'])
def __init__(self, args=None, vocab=None, emb_matrix=None, model_file=None, use_cuda=False): self.use_cuda = use_cuda if model_file is not None: # load from file self.load(model_file, use_cuda) else: self.args = args self.model = None if args['dict_only'] else Seq2SeqModel( args, emb_matrix=emb_matrix) self.vocab = vocab self.expansion_dict = dict() if not self.args['dict_only']: self.crit = loss.SequenceLoss(self.vocab.size) self.parameters = [ p for p in self.model.parameters() if p.requires_grad ] if use_cuda: self.model.cuda() self.crit.cuda() else: self.model.cpu() self.crit.cpu() self.optimizer = utils.get_optimizer(self.args['optim'], self.parameters, self.args['lr'])
def init_from_lm(self, lm_model, freeze: bool = True, m_names=[ 'word_emb', 'lemma_emb', 'upos_emb', 'xpos_emb', 'ufeats_emb', 'charmodel', 'trans_char', 'trans_char', 'trans_pretrained', 'lstm_forward', 'lstm_backward' ]): """ Initialize the paramters from a pretrained langauge model. lm_model: LSTMBiLM object freeze: bool (optional) if True, the initialized paramters are freezed and will be from the optimizer's param group m_names: List[str] (optional) a list of module names to initialize """ for m in m_names: if hasattr(self.model, m): print('Initilizing {}'.format(m)) if not hasattr(lm_model, m): raise ValueError( 'pretrained language model does not have attribute {}'. format(m)) module = getattr(self.model, m) copy_weights(getattr(lm_model, m), module) if freeze: freeze_net(module) else: print('Skipping {}'.format(m)) self.parameters = [ p for p in self.model.parameters() if p.requires_grad ] self.optimizer = utils.get_optimizer(self.args['optim'], self.parameters, self.args['lr'], betas=(self.args['beta1'], self.args['beta2']), weight_decay=self.args['wdecay'])