import sys
sys.path.insert(0, ".")
from adversarial_net.models import LanguageModel
from adversarial_net import arguments as flags
from adversarial_net.preprocessing import WordCounter
from adversarial_net import osp
flags.add_argument(
name="save_model_dir",
argtype=str,
default=
"E:/kaggle/avito/imdb_testset/adversarial_net/model/lm_model/lm_model.ckpt"
)
if __name__ == "__main__":
vocab_freqs = WordCounter().load(
osp.join(flags["lm_inputs"]["datapath"],
"imdb_word_freqs.pickle")).most_common_freqs(
flags["lm_sequence"]["vocab_size"])
flags.add_variable(name="vocab_freqs", value=vocab_freqs)
lm_model = LanguageModel()
lm_model.build()
lm_model.fit(save_model_path=flags["save_model_dir"])
def configure():
flags.register_variable(name="vocab_freqs")
flags.add_argument(scope="inputs", name="datapath", argtype=str)
flags.add_argument(scope="inputs", name="dataset", argtype=str)
flags.add_argument(scope="inputs",
name="eval_count_examples",
argtype=int,
default=-1)
flags.add_argument(scope="inputs",
name="vocab_size",
argtype=int,
default=50000)
flags.add_association(scope="inputs",
name="eval_max_words",
assoc_scope="inputs",
assoc_name="vocab_size")
flags.add_argument(scope="inputs",
name="batch_size",
argtype=int,
default=256)
flags.add_argument(scope="inputs",
name="unroll_steps",
argtype=int,
default=200)
flags.add_association(scope="inputs",
name="lstm_num_layers",
assoc_scope="lm_sequence",
assoc_name="rnn_num_layers")
flags.add_association(scope="inputs",
name="state_size",
assoc_scope="lm_sequence",
assoc_name="rnn_cell_size")
flags.add_argument(scope="inputs",
name="bidrec",
argtype=bool,
default=False)
flags.add_association(scope="lm_sequence",
name="vocab_size",
assoc_scope="inputs",
assoc_name="vocab_size")
flags.add_argument(scope="lm_sequence",
name="embedding_dim",
argtype=int,
default=256)
# should be same with gan rnn_cell_size
flags.add_argument(scope="lm_sequence",
name="rnn_cell_size",
argtype=int,
default=1024)
flags.add_argument(scope="lm_sequence",
name="normalize",
argtype="bool",
default=True)
flags.add_argument(scope="lm_sequence",
name="keep_embed_prob",
argtype=float,
default=1.0)
flags.add_argument(scope="lm_sequence",
name="lstm_keep_pro_out",
argtype=float,
default=1.0)
flags.add_argument(scope="lm_sequence",
name="rnn_num_layers",
argtype=int,
default=1)
flags.add_association(scope="lm_sequence",
name="vocab_freqs",
assoc_name="vocab_freqs")
flags.add_scope_association(scope="lm_inputs", assoc_scope="inputs")
flags.add_association(scope="lm_loss",
name="vocab_size",
assoc_scope="inputs",
assoc_name="vocab_size")
flags.add_argument(scope="lm_loss",
name="num_candidate_samples",
argtype=int,
default=1024)
flags.add_association(scope="lm_loss",
name="vocab_freqs",
assoc_name="vocab_freqs")
flags.add_scope_association(scope="ae_sequence", assoc_scope="lm_sequence")
flags.add_scope_association(scope="ae_inputs", assoc_scope="inputs")
flags.add_scope_association(scope="ae_loss", assoc_scope="lm_loss")
flags.add_scope_association(scope="adv_cl_inputs", assoc_scope="inputs")
flags.add_association(scope="adv_cl_inputs",
name="phase",
assoc_name="phase")
flags.add_argument(scope="adv_cl_sequence",
name="hidden_size",
argtype=int,
default=30)
flags.add_argument(scope="adv_cl_sequence",
name="num_layers",
argtype=int,
default=1)
flags.add_argument(scope="adv_cl_sequence",
name="num_classes",
argtype=int,
default=2)
flags.add_argument(scope="adv_cl_sequence",
name="keep_prob",
argtype=int,
default=1.0)
flags.add_association(scope="adv_cl_sequence",
name="input_size",
assoc_scope="lm_sequence",
assoc_name="rnn_cell_size")
flags.add_argument(scope="adv_cl_loss",
name="adv_reg_coeff",
argtype=float,
default=1.0)
flags.add_argument(scope="adv_cl_loss",
name="perturb_norm_length",
argtype=float,
default=5.0)
flags.add_argument(scope="vir_adv_loss",
name="num_power_iteration",
argtype=int,
default=1)
flags.add_argument(scope="vir_adv_loss",
name="small_constant_for_finite_diff",
argtype=float,
default=1e-1)
flags.add_argument(scope="gan",
name="critic_iters",
argtype=int,
default=5)
flags.add_argument(scope="gan",
name="rnn_cell_size",
argtype=int,
default=1024)
flags.add_argument(scope="summary",
name="rnn_cell_size",
argtype=int,
default=512)
flags.add_argument(scope="summary",
name="rnn_keep_prob_out",
argtype=float,
default=1.0)
flags.add_argument(scope="summary",
name="maximum_iterations",
argtype=int,
default=50)
flags.add_argument(scope="summary",
name="beam_width",
argtype=int,
default=3)
flags.add_argument(name="phase", argtype=str, default="train")
flags.add_argument(name="max_grad_norm", argtype=float, default=1.0)
flags.add_argument(name="lr", argtype=float, default=1e-3)
flags.add_argument(name="lr_decay", argtype=float, default=0.9999)
flags.add_argument(name="max_steps", argtype=int, default=100000)
flags.add_argument(name="save_steps", argtype=int, default=100)
flags.add_argument(name="save_best", argtype="bool", default=True)
flags.add_argument(name="save_best_check_steps", argtype=int, default=100)
flags.add_argument(name="eval_acc", argtype=bool, default=False)
flags.add_argument(name="eval_steps", argtype=int, default=100)
flags.add_argument(name="should_restore_if_could",
argtype="bool",
default=True)
flags.add_argument(name="tf_debug_trace", argtype=bool, default=False)
flags.add_argument(name="tf_timeline_dir", argtype=str, default=None)
flags.add_argument(name="no_need_clip_grads", argtype=bool, default=False)
flags.add_argument(name="with_lr_decay", argtype="bool", default=True)
flags.add_argument(name="lr_decay_step", argtype=int, default=1)
# continue from break point
flags.add_argument(name="best_loss_val", argtype=float, default=99999999.0)
flags.add_argument(name="extra_save_dir", argtype=str, default=None)
flags.add_argument(name="tfdebug_root", argtype=str, default=None)
flags.add_argument(name="use_exp_mov_avg_loss",
argtype=bool,
default=False)
flags.add_argument(name="use_exp_mov_avg_loss_decay",
argtype=float,
default=0.9)
flags.add_argument(name="logging_file", argtype=str, default=None)
def eval_from(value):
eval_from_vals = [
"generator", "topic_generator", "pretrain_cl", "final_cl"
]
assert value in eval_from_vals, "step is one of %s" % eval_from_vals
return value
def adv_type(value):
adv_types = ["adv", "vir_adv"]
assert value in adv_types, "adv_type is one of %s" % adv_types
return value
flags.add_argument(name="step", argtype=training_step)
flags.add_argument(name="save_model_dir", argtype=str)
flags.add_argument(name="pretrain_model_dir", argtype=str, default=None)
flags.add_argument(name="eval_from", argtype=eval_from, default="generator")
flags.add_argument(name="eval_batch_size", argtype=int, default=2)
flags.add_argument(name="eval_topic_count", argtype=int, default=2)
flags.add_argument(name="eval_seq_length", argtype=int, default=200)
# lm/ae model args
flags.add_argument(name="no_loss_sampler", argtype=bool, default=False)
flags.add_argument(name="hard_mode", argtype=bool, default=False)
flags.add_argument(name="forget_bias", argtype=float, default=0.0)
# model prefix
flags.add_argument(name="model_prefix", argtype=str, default=None)
# adversarial training type
flags.add_argument(name="adv_type", argtype=adv_type, default="adv")