def create_config(data):
config = trf.Config(data)
config.norm_config = 'linear'
config.batch_size = 100
config.noise_factor = 1
config.noise_sampler = None # '2gram'
config.data_factor = 1
config.data_sampler = config.noise_sampler
# config.lr_feat = lr.LearningRateTime(1e-4)
config.lr_net = lr.LearningRateTime(1e-3) #lr.LearningRateTime(1, 0.5, tc=1e3)
config.lr_logz = lr.LearningRateTime(0.01)
config.opt_feat_method = 'adam'
config.opt_net_method = 'adam'
config.opt_logz_method = 'adam'
config.max_epoch = 1000
config.init_logz = config.get_initial_logz()
config.init_global_logz = 0
# config.prior_model_path = 'lstm/lstm_e32_h32x1_BNCE_SGD/model.ckpt'
# feat config
# config.feat_config.feat_type_file = '../../tfcode/feat/g4.fs'
# config.feat_config.feat_cluster = None
config.feat_config = None
# net config
config.net_config.update(task.get_config_rnn(config.vocab_size))
# config.net_config.l2_reg = 1e-4
# wb.mkdir('word_emb')
# config.net_config.load_embedding_path = 'word_emb/ptb_d{}.emb'.format(config.net_config.embedding_dim)
config.write_dbg = False
return config
def create_config(data):
config = trf.Config(data)
config.write_dbg = False
config.max_epoch = 1000
config.batch_size = 100
config.noise_factor = 1
config.norm_config = 'multiple'
config.init_logz = config.get_initial_logz()
config.lr_feat = lr.LearningRateTime(1, 0.2, tc=1e3)
config.lr_net = lr.LearningRateTime(
1, 1., tc=1e3) #lr.LearningRateTime(1, 0.5, tc=1e3)
config.lr_logz = lr.LearningRateTime(1, 1., tc=1e3)
config.opt_feat_method = 'adam'
config.opt_net_method = 'adam'
config.opt_logz_method = 'adam'
# config.prior_model_path = 'lstm/lstm_e32_h32x1_BNCE_SGD/model.ckpt'
# feat config
# config.feat_config.feat_type_file = '../../tfcode/feat/g4.fs'
# config.feat_config.feat_cluster = None
config.feat_config = None
# net config
config.net_config.update(get_config_cnn(config.vocab_size))
config.write_dbg = True
# for sampler
config.sampler_config.learning_rate = 0.1
return config
def get_config(data):
config = trf.Config(data)
# config.pi_0 = data.get_pi0(config.pi_true)
# config.pi_true = config.pi_0
config.norm_config = 'linear'
config.batch_size = 100
config.noise_factor = 4
config.data_factor = 0
config.train_add_noise = False
# config.noise_sampler = '2gram'
# config.lr_feat = lr.LearningRateTime(1e-4)
config.lr_net = lr.LearningRateTime(1e-3) # lr.LearningRateTime(1, 0.5, tc=1e3)
config.lr_logz = lr.LearningRateTime(1e-2)
config.lr_sampler = lr.LearningRateEpochDelay(0.1)
config.opt_feat_method = 'adam'
config.opt_net_method = 'adam'
config.opt_logz_method = 'adam'
config.max_epoch = 1000
# sampler
# config.sampler_config.hidden_layers = 2
# config.load_sampler = 'sampler/lstm_e200_h200x2/sampler.ckpt'
# config.fix_sampler = True
config.init_logz = config.get_initial_logz()
config.init_global_logz = 0
config.feat_config = None
# net config
config.net_config.update(get_config_rnn(config.vocab_size))
# config.net_config.l2_reg = 1e-4
# wb.mkdir('word_emb')
# config.net_config.load_embedding_path = 'word_emb/ptb_d{}.emb'.format(config.net_config.embedding_dim)
config.write_dbg = False
config.add_sampler_as_prior = False
return config