usage: run.py [-h] [--name NAME] [--restore RESTORE] [--mode MODE] [--lr LR]
                  [--eh_size EH_SIZE] [--dh_size DH_SIZE] [--droprate DROPRATE]
                  [--batchnorm] [--decode_mode {max,sample,gumbel,samplek,beam}]
                  [--top_k TOP_K] [--length_penalty LENGTH_PENALTY]
                  [--dataset DATASET] [--dataid DATAID] [--epoch EPOCH]
                  [--batch_per_epoch BATCH_PER_EPOCH] [--wvclass WVCLASS]
                  [--wvid WVID] [--out_dir OUT_DIR] [--log_dir LOG_DIR]
                  [--model_dir MODEL_DIR] [--cache_dir CACHE_DIR] [--cpu]
                  [--debug] [--cache] [--seed SEED]

    A seq2seq model with GRU encoder and decoder. Attention, beamsearch, dropout
    and batchnorm is supported.

    optional arguments:
      -h, --help            show this help message and exit
      --name NAME           The name of your model, used for tensorboard, etc.
                            Default: runXXXXXX_XXXXXX (initialized by current
                            time)
      --restore RESTORE     Checkpoints name to load. "NAME_last" for the last
                            checkpoint of model named NAME. "NAME_best" means the
                            best checkpoint. You can also use "last" and "best",
                            by default use last model you run. Attention:
                            "NAME_last" and "NAME_best" are not guaranteed to work
                            when 2 models with same name run in the same time.
                            "last" and "best" are not guaranteed to work when 2
                            models run in the same time. Default: None (don't load
                            anything)
      --mode MODE           "train" or "test". Default: train
      --lr LR               Learning rate. Default: 0.001
      --eh_size EH_SIZE     Size of encoder GRU
      --dh_size DH_SIZE     Size of decoder GRU
      --droprate DROPRATE   The probability to be zerod in dropout. 0 indicates
                            for don't use dropout
      --batchnorm           Use bathnorm
      --decode_mode {max,sample,gumbel,samplek,beam}
                            The decode strategy when freerun. Choices: max,
                            sample, gumbel(=sample), samplek(sample from topk),
                            beam(beamsearch). Default: beam
      --top_k TOP_K         The top_k when decode_mode == "beam" or "samplek"
      --length_penalty LENGTH_PENALTY
                            The beamsearch penalty for short sentences. The
                            penalty will get larger when this becomes smaller.
      --dataset DATASET     Dataloader class. Default: OpenSubtitles
      --dataid DATAID       Resource id for data set. It can be a resource name or
                            a local path. Default: resources://OpenSubtitles
      --epoch EPOCH         Epoch for training. Default: 100
      --batch_per_epoch BATCH_PER_EPOCH
                            Batches per epoch. Default: 1500
      --wvclass WVCLASS     Wordvector class, none for not using pretrained
                            wordvec. Default: Glove
      --wvid WVID           Resource id for pretrained wordvector. Default:
                            resources://Glove300d
      --out_dir OUT_DIR     Output directory for test output. Default: ./output
      --log_dir LOG_DIR     Log directory for tensorboard. Default: ./tensorboard
      --model_dir MODEL_DIR
                            Checkpoints directory for model. Default: ./model
      --cache_dir CACHE_DIR
                            Checkpoints directory for cache. Default: ./cache
      --cpu                 Use cpu.
      --debug               Enter debug mode (using ptvsd).
      --cache               Use cache for speeding up load data and wordvec. (It
                            may cause problems when you switch dataset.)
      --seed SEED           Specify random seed. Default: 0

perplexity:     48.194050
bleu:    0.320098
post:   my name is josie .
resp:   <unk> <unk> , pennsylvania , the <unk> state .
gen:    i' m a teacher .
post:   i put premium gasoline in her .
resp:   josie , i told you .
gen:    i don' t know .
post:   josie , dont hang up
resp:   they do it to aii the new kids .
gen:    aii right , you guys , you know what ?
post:   about playing a part .
resp:   and thats the theme of as you like it .
gen:    i don' t know .
......