diff --git a/dataset.py b/dataset.py
new file mode 100644
index 0000000..4ac4682
--- /dev/null
+++ b/dataset.py
@@ -0,0 +1,109 @@
+from torch.utils.data import Dataset
+import torch
+import tqdm
+import json
+import logging
+logger = logging.getLogger(__name__)
+
+class VAEDataset(Dataset):
+    def __init__(self, source_path, tokenizer, device=torch.device('cuda:0')):
+        self.data = []
+        self.tokenizer = tokenizer
+        self.device = device
+        with open(source_path) as f:
+            for line in tqdm.tqdm(f, desc='Loading data...'):
+                line = line.strip()
+                if line == '':
+                    continue
+                line = line.split('\t')[-1]
+                self.data.append(line)
+
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, idx):
+        return self.tokenizer.encode(self.data[idx])
+
+    @staticmethod
+    def create_mask(num_tokens, max_len):
+        base_position_matrix = torch.arange(
+            0, max_len, dtype=num_tokens.dtype, device=num_tokens.device).view(1, -1)
+        mask = (base_position_matrix < num_tokens.view(-1, 1)).type_as(num_tokens)
+        return mask
+
+    def collate_fn(self, samples):
+        samples = [[self.tokenizer.bos_id] + s + [self.tokenizer.eos_id] for s in samples]
+        length_list = [len(s) for s in samples]
+        max_t = max(length_list)
+        new_samples = [s + [self.tokenizer.pad_id] * (max_t - len(s)) for s in samples]
+        new_samples = torch.LongTensor(new_samples)
+        attention_mask = self.create_mask(torch.LongTensor(length_list), max_t)
+        return {
+            'input_ids': new_samples.to(self.device),
+            'attention_mask': attention_mask.byte().to(self.device),
+        }
+
+class WPDataset(Dataset):
+    def __init__(self, source_path, tokenizer, device=torch.device('cuda:0'), max_length=700, add_prefix=False, add_special_token=False):
+        self.source = []
+        self.target = []
+        self.tokenizer = tokenizer
+        self.device = device
+        self.max_length = max_length
+        self.add_special_token = add_special_token
+        self.add_prefix = add_prefix
+        with open(source_path) as f:
+            for line in tqdm.tqdm(f, desc='Loading data...'):
+                line = json.loads(line.strip())
+                source = line['source'].replace('<newline>', '\n')
+                target = line['target'].replace('<newline>', '\n')
+                if len(source.split()) + len(target.split()) < self.max_length:
+                    self.source.append(source)
+                    self.target.append(target)
+
+    def __len__(self):
+        return len(self.source)
+
+    def __getitem__(self, idx):
+        source = self.tokenizer.encode(self.source[idx])
+        target = self.tokenizer.encode(self.target[idx])
+        return source, target
+
+    @staticmethod
+    def create_mask(num_tokens, max_len):
+        base_position_matrix = torch.arange(
+            0, max_len, dtype=num_tokens.dtype, device=num_tokens.device).view(1, -1)
+        mask = (base_position_matrix < num_tokens.view(-1, 1)).type_as(num_tokens)
+        return mask
+
+    def collate_fn(self, samples):
+        source_initial = [item[0] for item in samples]
+        target_initial = [item[1] for item in samples]
+        source = [s + [self.tokenizer.eos_id] for s in source_initial]
+        target = [s + t + [self.tokenizer.eos_id] for s, t in zip(source, target_initial)]
+        labels = [[self.tokenizer.pad_id] * len(s) + t + [self.tokenizer.eos_id] for s, t in zip(source, target_initial)]
+        source = [item[:self.max_length] for item in source]
+        target = [item[:self.max_length] for item in target]
+        labels = [item[:self.max_length] for item in labels]
+
+        source_length_list = [len(s) for s in source]
+        source_max_t = max(source_length_list)
+        new_source = [s + [self.tokenizer.pad_id] * (source_max_t - len(s)) for s in source]
+        new_source = torch.LongTensor(new_source)
+        source_attention_mask = self.create_mask(torch.LongTensor(source_length_list), source_max_t)
+
+        target_length_list = [len(s) for s in target]
+        target_max_t = max(target_length_list)
+        new_target = [s + [self.tokenizer.pad_id] * (target_max_t - len(s)) for s in target]
+        new_labels = [s + [self.tokenizer.pad_id] * (target_max_t - len(s)) for s in labels]
+        new_target = torch.LongTensor(new_target)
+        new_labels = torch.LongTensor(new_labels)
+        target_attention_mask = self.create_mask(torch.LongTensor(target_length_list), target_max_t)
+
+        return {
+            'input_ids': new_target.to(self.device),
+            'attention_mask': target_attention_mask.byte().to(self.device),
+            'labels': new_labels.to(self.device),
+            'condition': new_source.to(self.device),
+            'condition_mask': source_attention_mask.byte().to(self.device),
+        }
\ No newline at end of file
diff --git a/dist.py b/dist.py
new file mode 100644
index 0000000..49d73e5
--- /dev/null
+++ b/dist.py
@@ -0,0 +1,43 @@
+import torch
+import numpy as np
+
+@torch.jit.script
+def soft_clamp5(x: torch.Tensor):
+    return x.div_(5.).tanh_().mul(5.)
+
+class Normal:
+    def __init__(self, mu, log_sigma):
+        self.mu = torch.clamp(mu, -5, 5)
+        log_sigma = torch.clamp(log_sigma, -5, 5)
+        self.std = log_sigma.mul(0.5).exp()
+
+    def sample(self):
+        eps = self.mu.mul(0).normal_()
+        z = eps.mul_(self.std).add_(self.mu)
+        return z, eps
+
+    @staticmethod
+    def get_standard(bs, nz, device):
+        zeros = torch.zeros(bs, nz).to(device)
+        return Normal(zeros, zeros)
+
+    def sample_given_eps(self, eps):
+        return eps * self.std + self.mu
+
+    def log_p(self, samples):
+        normalized_samples = (samples - self.mu) / self.std
+        log_p = - 0.5 * normalized_samples * normalized_samples - 0.5 * np.log(2 * np.pi) - torch.log(self.std)
+        log_p = torch.sum(log_p, dim=-1)
+        return log_p
+
+    def kl(self, normal_dist):
+        assert normal_dist.mu.shape == self.mu.shape
+        term1 = (self.mu - normal_dist.mu) / normal_dist.std
+        term2 = self.std / normal_dist.std
+        loss = 0.5 * (term1 * term1 + term2 * term2) - 0.5 - torch.log(term2)
+        loss = torch.sum(loss, dim=-1)
+        return loss
+
+    def set_device(self, cuda_id):
+        self.mu = self.mu.to(cuda_id)
+        self.std = self.std.to(cuda_id)
diff --git a/main.py b/main.py
new file mode 100644
index 0000000..ec4c2f4
--- /dev/null
+++ b/main.py
@@ -0,0 +1,236 @@
+import argparse
+import logging
+import os
+import json
+import torch
+import random
+import numpy as np
+import time
+
+from torch.nn.parallel import DataParallel
+from torch.utils.data import DataLoader
+
+from dataset import VAEDataset, WPDataset
+from train import train, valid, generate
+
+from model import Della
+
+from transformers import AutoConfig, AutoModel, AutoTokenizer
+
+logger = logging.getLogger(__name__)
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--train_file", default='./data/yelp/yelp.train.txt', type=str,
+                        help="Data path for training.")
+    parser.add_argument("--valid_file", default='./data/yelp/yelp.train.txt', type=str,
+                        help="Data path for valid")
+    parser.add_argument("--test_file", default='./data/yelp/yelp.train.txt', type=str,
+                        help="Data path for test")
+    parser.add_argument("--pretrained_model", type=str, default='gpt2', 
+                        help="Pretrained model to be loaded")
+    parser.add_argument("--dataset_type", type=str, default='vae', choices=['vae', 'wp'], 
+                        help="Dataset type")
+    parser.add_argument("--output_dir", default='./checkpoints', type=str,
+                        help="The output directory where the model checkpoints and predictions will be written.")
+    parser.add_argument("--model_name", default='della', type=str,
+                        help="The model name")
+    parser.add_argument("--generation_output_dir", default='./generation_output', type=str,
+                        help="The output directory where the log will be written.")
+    # Other parameters\
+    parser.add_argument("--load_epoch", default=None, type=int, help="the epochs of trained model to load")
+    parser.add_argument("--epochs", default=40, type=int, help="total epochs")
+    parser.add_argument("--per_gpu_train_batch_size", default=4, type=int,help="Batch size per GPU for training.")
+    parser.add_argument("--no_cuda", action='store_true',
+                        help="Whether not to use CUDA when available")
+    parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
+    parser.add_argument('--gradient_accumulation_steps', type=int, default=8,
+                        help="Number of updates steps to accumulate before performing a backward/update pass.")
+    parser.add_argument("--weight_decay", default=0.01, type=float,
+                        help="Weight decay if we apply some.")
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
+                        help="Epsilon for Adam optimizer.")
+    parser.add_argument("--kl_threshold", default=0, type=float,
+                        help="The threshold of the minimum KL value, default as 0")
+    parser.add_argument("--latent_size", default=32, type=int,
+                        help="The dimension of latent space")
+    parser.add_argument("--latent_lmf_rank", default=4, type=int,
+                        help="latent size")
+    parser.add_argument("--max_length", default=200, type=int,
+                        help="Max length for generation")
+    parser.add_argument('--seed', type=int, default=42,
+                        help="Random seed for initialization")
+    parser.add_argument('--log_step', type=int, default=100,
+                        help="Steps for logging")
+    parser.add_argument('--num_beams', type=int, default=10,
+                        help="Beam size for searching")
+    parser.add_argument('--greedy_decoding', action='store_true',
+                        help="Choose to use greedy decoding")
+    parser.add_argument('--top_k', type=int, default=-1, help='Set top k')
+    parser.add_argument('--top_p', type=float, default=0.9, help='Set top p')
+    parser.add_argument('--repetition_penalty', type=float, default=1.2)
+    parser.add_argument('--model_parallel', action='store_true', 
+                        help="Choose to use model parallel, mapping the layers to different devices")
+    parser.add_argument('--eval', action='store_true', help='Choose to eval the model')
+    parser.add_argument('--eval_metrics', action='store_true',
+                        help="Choose to eval the metrics for representation learning")
+    parser.add_argument('--generation', action='store_true', help='Choose to generate')
+    parser.add_argument('--use_scheduler', action='store_true',
+                        help="Choose to use lr scheduler")
+    parser.add_argument('--cycle_annealing', action='store_true',
+                        help="Choose to use cycle annealing")
+    parser.add_argument('--cycle_iters', type=int, default=2,
+                        help="Set the iters for cycle annealing")
+    parser.add_argument('--sample_times', type=int, default=30,
+                        help="The total times of sample when computing PPL with importance weighted sampling")
+    parser.add_argument('--use_bow', action='store_true',
+                        help="Choose to use bow loss")
+    parser.add_argument('--bow_weight',type=float, default=0.2,
+                        help="Set the weight of bow loss term")
+    parser.add_argument("--begin_layer", default=None, type=int,
+                        help="The beginning layer to consider the latent vector, default as the first layer of model")
+    parser.add_argument("--end_layer", default=None, type=int,
+                        help="The end layer to consider the latent vector, default as the last layer of model")
+    args = parser.parse_args()
+    return args
+
+def prepare(args):
+    torch.set_num_threads(3)
+
+    if not args.eval and not args.generation:
+        os.makedirs(os.path.join(args.output_dir, args.model_name), exist_ok=True)
+        json.dump(args.__dict__, open(os.path.join(
+            args.output_dir, args.model_name, 'train_opt.json'), 'w'), sort_keys=True, indent=2)
+
+    if args.no_cuda:
+        args.n_gpu = 1
+    else:
+        args.n_gpu = torch.cuda.device_count()
+    args.batch_size = args.per_gpu_train_batch_size * args.n_gpu
+    
+    # Setup logging
+    logger = logging.getLogger()
+    logger.setLevel(logging.INFO)
+    formatter = logging.Formatter("%(asctime)s - %(filename)s[line:%(lineno)d] - %(levelname)s: %(message)s")
+    sh = logging.StreamHandler()
+    sh.setFormatter(formatter)
+    logger.addHandler(sh)
+
+    # Set seed
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+
+    if args.n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+
+    logger.info("Training/evaluation parameters %s", args)
+
+    if args.no_cuda:
+        args.device = torch.device('cpu')
+    else:
+        args.device = torch.device('cuda:0')
+
+def init_para_frompretrained(model, gpt2):
+    logger.info('load gpt2 pretrained model parameters')
+    model = model.encoder
+    model.wte.weight = gpt2.wte.weight
+    model.wpe.weight = gpt2.wpe.weight
+
+    for i in range(len(gpt2.h)):
+        model.h[i].ln_1.weight = gpt2.h[i].ln_1.weight
+        model.h[i].ln_1.bias = gpt2.h[i].ln_1.bias
+        model.h[i].attn.c_attn.weight = gpt2.h[i].attn.c_attn.weight
+        model.h[i].attn.c_attn.bias = gpt2.h[i].attn.c_attn.bias
+        model.h[i].attn.c_proj.weight = gpt2.h[i].attn.c_proj.weight
+        model.h[i].attn.c_proj.bias = gpt2.h[i].attn.c_proj.bias
+        model.h[i].ln_2.weight = gpt2.h[i].ln_2.weight
+        model.h[i].ln_2.bias = gpt2.h[i].ln_2.bias
+        model.h[i].mlp.c_fc.weight = gpt2.h[i].mlp.c_fc.weight
+        model.h[i].mlp.c_fc.bias = gpt2.h[i].mlp.c_fc.bias
+        model.h[i].mlp.c_proj.weight = gpt2.h[i].mlp.c_proj.weight
+        model.h[i].mlp.c_proj.bias = gpt2.h[i].mlp.c_proj.bias
+
+    model.ln_f.weight = gpt2.ln_f.weight
+    model.ln_f.bias = gpt2.ln_f.bias
+
+def prepare_model(args):
+    tokenizer = AutoTokenizer.from_pretrained(args.pretrained_model)
+    if '<s>' not in tokenizer.vocab:
+        tokenizer._add_tokens(['<s>'])
+    if '</s>' not in tokenizer.vocab:
+        tokenizer._add_tokens(['</s>'])
+    tokenizer.pad_id = 50256
+    
+    tokenizer.bos_id = tokenizer.convert_tokens_to_ids('<s>')
+    tokenizer.eos_id = tokenizer.convert_tokens_to_ids('</s>')
+
+    model_config = AutoConfig.from_pretrained(args.pretrained_model)
+    model_config.vocab_size = len(tokenizer)
+    model_config.pad_token_id = tokenizer.pad_id
+    model_config.kl_threshold = args.kl_threshold
+    model_config.is_cvae = (args.dataset_type == 'wp')
+    model_config.use_bow = args.use_bow
+    model_config.begin_layer = args.begin_layer
+    model_config.end_layer = args.end_layer
+
+    for arg in vars(args):
+        if arg.startswith('latent'):
+            setattr(model_config, arg, getattr(args, arg))
+    
+    model = Della(model_config)
+    pretrained_model = AutoModel.from_pretrained(args.pretrained_model)
+    logging.info('loading pretrained model parameters...')
+    init_para_frompretrained(model, pretrained_model)
+    model.encoder.resize_token_embeddings(len(tokenizer))
+    model.decoder.wte = model.encoder.wte
+    if args.load_epoch is not None:
+        model_path = os.path.join(args.output_dir, args.model_name, 'model_epoch_{}.pt'.format(args.load_epoch))
+        model_state_dict = torch.load(model_path, map_location=args.device)
+        model.load_state_dict(model_state_dict)
+        logging.info('load model_epoch_{}.pt finish'.format(args.load_epoch))
+    else:
+        args.load_epoch = -1
+
+    if args.model_parallel and torch.cuda.device_count() > 1:  
+        logging.info('model paralleize...')
+        model.parallelize()
+    else:
+        model = model.to(args.device)
+        if torch.cuda.device_count() > 1:
+            model = DataParallel(model)
+    return model, tokenizer
+
+def prepare_data(tokenizer, args):
+    dataset_class = {'vae': VAEDataset, 'wp': WPDataset}
+    if args.eval or args.generation:
+        logging.info("eval model: the epoch {} of {}".format(args.load_epoch, args.model_name))
+        test_dataset = dataset_class[args.dataset_type](args.test_file, tokenizer, args.device)
+        test_iter = DataLoader(test_dataset, batch_size=args.batch_size, collate_fn=test_dataset.collate_fn)
+        return test_iter
+    else:
+        train_dataset = dataset_class[args.dataset_type](args.train_file, tokenizer, args.device)
+        valid_dataset = dataset_class[args.dataset_type](args.valid_file, tokenizer, args.device)
+        train_iter = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, collate_fn=train_dataset.collate_fn)
+        valid_iter = DataLoader(valid_dataset, batch_size=args.batch_size, shuffle=True, collate_fn=valid_dataset.collate_fn)
+        logging.info('training with {} samples...'.format(len(train_dataset)))
+        return train_iter, valid_iter
+
+def main():
+    args = get_args()
+    prepare(args)
+    model, tokenizer = prepare_model(args)
+    total_params = sum(p.numel() for p in model.parameters())
+    logging.info('total parameters: {}'.format(total_params))
+    if args.eval or args.generation:
+        test_iter = prepare_data(tokenizer, args)
+        if args.eval:
+            valid(model, test_iter, args.load_epoch, args)
+        if args.generation:
+            generate(model, test_iter, tokenizer, args)
+    else:
+        train_iter, valid_iter = prepare_data(tokenizer, args)
+        train(model, train_iter, valid_iter, args)
+
+if __name__ == "__main__":
+    main()
diff --git a/model.py b/model.py
new file mode 100644
index 0000000..3f12046
--- /dev/null
+++ b/model.py
@@ -0,0 +1,1202 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch OpenAI GPT-2 model."""
+
+import os
+from typing import Optional, Tuple
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss, MSELoss, RNNCell
+
+from torch.nn.parameter import Parameter
+from torch.nn.init import xavier_normal_
+
+from transformers.activations import ACT2FN
+from transformers.file_utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    SequenceClassifierOutputWithPast,
+)
+from transformers.modeling_utils import (
+    Conv1D,
+    PreTrainedModel,
+    SequenceSummary,
+    find_pruneable_heads_and_indices,
+    prune_conv1d_layer,
+)
+from transformers.utils import logging
+from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
+from transformers import GPT2Config, EncoderDecoderModel, EncoderDecoderConfig
+
+from dist import Normal
+from modeloutput import EncoderOutput, DecoderOutput
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "gpt2"
+_CONFIG_FOR_DOC = "GPT2Config"
+_TOKENIZER_FOR_DOC = "GPT2Tokenizer"
+
+GPT2_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "gpt2",
+    "gpt2-medium",
+    "gpt2-large",
+    "gpt2-xl",
+    "distilgpt2",
+    # See all GPT-2 models at https://huggingface.co/models?filter=gpt2
+]
+
+class LMFLayer(nn.Module):
+    def __init__(self, config):
+        super(LMFLayer, self).__init__()
+        self.rank = config.latent_lmf_rank
+        self.hidden_size = config.hidden_size
+        self.latent_size = config.latent_size
+        
+        self.text_factor = nn.ModuleList([nn.Linear(self.hidden_size + 1, self.hidden_size) for _ in range(self.rank)])
+        self.latent_factor = nn.ModuleList([nn.Linear(self.latent_size + 1, self.hidden_size) for _ in range(self.rank)])
+
+    def forward(self, hidden_states, latent):
+        '''
+        Args:
+            hidden_states: tensor of shape (batch_size, sequence_len, text_in)
+            latent: tensor of shape (batch_size, latent_size)
+        '''
+        batch_size, seq_len, _ = hidden_states.size()
+        device = hidden_states.device
+        text_cat = torch.ones(batch_size, seq_len, 1, dtype=torch.float, device=device)
+        latent_cat = torch.ones(batch_size, 1, dtype=torch.float, device=device)
+        hidden_states = torch.cat((hidden_states, text_cat), dim=-1)
+        latent = torch.cat((latent, latent_cat), dim=-1)
+        text_fusion_output = []
+        latent_fusion_output = []
+        for text_factor, latent_factor in zip(self.text_factor, self.latent_factor):
+            text_fusion = text_factor(hidden_states)
+            latent_fusion = latent_factor(latent)
+            text_fusion_output.append(text_fusion)
+            latent_fusion_output.append(latent_fusion)
+        text_fusion = torch.stack(text_fusion_output).sum(0)
+        latent_fusion = torch.stack(latent_fusion_output).sum(0)
+        text_fusion = text_fusion.transpose(1, 0)
+        output = text_fusion * latent_fusion
+        output = output.transpose(1, 0)
+        return output
+
+class AverageSelfAttention(nn.Module):
+    def __init__(self, config):
+        super(AverageSelfAttention, self).__init__()
+        w = torch.empty(config.hidden_size)
+        nn.init.normal_(w, std=0.02)
+        self.attention_weights = nn.Parameter(w)
+        self.softmax = nn.Softmax(dim=-1)
+        self.non_linearity = ACT2FN[config.activation_function]
+
+    def forward(self, inputs, attention_mask=None):
+        scores = self.non_linearity(inputs.matmul(self.attention_weights))
+        if attention_mask is not None:
+            scores = scores + attention_mask
+
+        scores = self.softmax(scores)
+        weighted = torch.mul(inputs, scores.unsqueeze(-1).expand_as(inputs))
+        representations = weighted.sum(1).squeeze(1)
+
+        return representations, scores
+
+class GPT2Attention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+
+        max_positions = config.max_position_embeddings
+        self.register_buffer(
+            "bias",
+            torch.tril(torch.ones((max_positions, max_positions), dtype=torch.uint8)).view(
+                1, 1, max_positions, max_positions
+            ),
+        )
+        self.register_buffer("masked_bias", torch.tensor(-1e4))
+
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        self.split_size = self.embed_dim
+        self.lmf = LMFLayer(config)
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"`embed_dim` must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {self.num_heads})."
+            )
+
+        self.scale_attn_weights = config.scale_attn_weights
+        
+        self.c_attn = Conv1D(3 * self.embed_dim, self.embed_dim)
+        self.c_proj = Conv1D(self.embed_dim, self.embed_dim)
+
+        self.attn_dropout = nn.Dropout(config.attn_pdrop)
+        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+
+        self.pruned_heads = set()
+
+        self.config = config
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(heads, self.num_heads, self.head_dim, self.pruned_heads)
+        index_attn = torch.cat([index, index + self.split_size, index + (2 * self.split_size)])
+
+        # Prune conv1d layers
+        self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
+        self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)
+
+        # Update hyper params
+        self.split_size = (self.split_size // self.num_heads) * (self.num_heads - len(heads))
+        self.num_heads = self.num_heads - len(heads)
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def _attn(self, query, key, value, attention_mask=None, head_mask=None, use_causal_mask=True):
+        attn_weights = torch.matmul(query, key.transpose(-1, -2))
+
+        if self.scale_attn_weights:
+            attn_weights = attn_weights / (float(value.size(-1)) ** 0.5)
+
+        if use_causal_mask:
+            # if only "normal" attention layer implements causal mask
+            query_length, key_length = query.size(-2), key.size(-2)
+            causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length].bool()
+            attn_weights = torch.where(causal_mask, attn_weights, self.masked_bias.to(attn_weights.dtype))
+
+        if attention_mask is not None:
+            # Apply the attention mask
+            attn_weights = attn_weights + attention_mask
+        attn_weights = nn.Softmax(dim=-1)(attn_weights)
+        attn_weights = self.attn_dropout(attn_weights)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attn_weights = attn_weights * head_mask
+
+        attn_output = torch.matmul(attn_weights, value)
+
+        return attn_output, attn_weights
+
+    def _split_heads(self, tensor, num_heads, attn_head_size):
+        """
+        Splits hidden_size dim into attn_head_size and num_heads
+        """
+        new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
+        tensor = tensor.view(*new_shape)
+        return tensor.permute(0, 2, 1, 3)  # (batch, head, seq_length, head_features)
+
+    def _merge_heads(self, tensor, num_heads, attn_head_size):
+        """
+        Merges attn_head_size dim and num_attn_heads dim into hidden_size
+        """
+        tensor = tensor.permute(0, 2, 1, 3).contiguous()
+        new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
+        return tensor.view(new_shape)
+
+    def forward(
+        self,
+        hidden_states,
+        layer_past=None,
+        attention_mask=None,
+        head_mask=None,
+        use_cache=False,
+        output_attentions=False,
+        use_causal_mask=False,
+        latent=None,
+    ):
+        if latent is not None:
+            query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)
+            value = self.lmf(value, latent)
+        else:
+            query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)
+        
+        query = self._split_heads(query, self.num_heads, self.head_dim)
+        key = self._split_heads(key, self.num_heads, self.head_dim)
+        value = self._split_heads(value, self.num_heads, self.head_dim)
+
+        if layer_past is not None:
+            past_key, past_value = layer_past
+            key = torch.cat((past_key, key), dim=-2)
+            value = torch.cat((past_value, value), dim=-2)
+
+        if use_cache is True:
+            present = (key, value)
+        else:
+            present = None
+
+        attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask, use_causal_mask=use_causal_mask)
+        attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim)
+        attn_output = self.c_proj(attn_output)
+        attn_output = self.resid_dropout(attn_output)
+
+        outputs = (attn_output, present)
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs  # a, present, (attentions)
+
+
+class GPT2MLP(nn.Module):
+    def __init__(self, intermediate_size, config):
+        super().__init__()
+        embed_dim = config.hidden_size
+        self.c_fc = Conv1D(intermediate_size, embed_dim)
+        self.c_proj = Conv1D(embed_dim, intermediate_size)
+        self.act = ACT2FN[config.activation_function]
+        self.dropout = nn.Dropout(config.resid_pdrop)
+
+    def forward(self, hidden_states):
+        hidden_states = self.c_fc(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.c_proj(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        return hidden_states
+
+
+class GPT2Block(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        hidden_size = config.hidden_size
+        inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
+
+        self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.attn = GPT2Attention(config)
+        self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+
+        self.mlp = GPT2MLP(inner_dim, config)
+
+    def forward(
+        self,
+        hidden_states,
+        layer_past=None,
+        attention_mask=None,
+        head_mask=None,
+        latent=None,
+        use_cache=False,
+        output_attentions=False,
+        use_causal_mask=False,
+        **latent_inject_kwargs,
+    ):
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+        
+        attn_outputs = self.attn(
+            hidden_states,
+            layer_past=layer_past,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            use_causal_mask=use_causal_mask,
+            latent=latent,
+            **latent_inject_kwargs,
+        )
+        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
+        outputs = attn_outputs[1:]
+        # residual connection
+        hidden_states = attn_output + residual
+
+        residual = hidden_states
+        hidden_states = self.ln_2(hidden_states)
+        feed_forward_hidden_states = self.mlp(hidden_states)
+        # residual connection
+        hidden_states = residual + feed_forward_hidden_states
+
+        if use_cache:
+            outputs = (hidden_states,) + outputs
+        else:
+            outputs = (hidden_states,) + outputs[1:]
+
+        return outputs  # hidden_states, present, (attentions, cross_attentions)
+
+
+class GPT2PreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = GPT2Config
+    base_model_prefix = "transformer"
+    is_parallelizable = True
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, (nn.Linear, Conv1D)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+GPT2_START_DOCSTRING = r"""
+
+    This model inherits from :class:`~transformers.PreTrainedModel`. Check the superclass documentation for the generic
+    methods the library implements for all its model (such as downloading or saving, resizing the input embeddings,
+    pruning heads etc.)
+
+    This model is also a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__
+    subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to
+    general usage and behavior.
+
+    Parameters:
+        config (:class:`~transformers.GPT2Config`): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
+            weights.
+"""
+
+GPT2_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length`` if :obj:`past_key_values` is ``None`` else
+            ``past_key_values[0][0].shape[-2]`` (``sequence_length`` of input past key value states). Indices of input
+            sequence tokens in the vocabulary.
+
+            If :obj:`past_key_values` is used, only ``input_ids`` that do not have their past calculated should be
+            passed as ``input_ids``.
+
+            Indices can be obtained using :class:`~transformers.GPT2Tokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers`):
+            Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
+            :obj:`past_key_values` output below). Can be used to speed up sequential decoding. The ``input_ids`` which
+            have their past given to this model should not be passed as ``input_ids`` as they have already been
+            computed.
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`, `optional`):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in ``[0,
+            1]``:
+
+            - 0 corresponds to a `sentence A` token,
+            - 1 corresponds to a `sentence B` token.
+
+            `What are token type IDs? <../glossary.html#token-type-ids>`_
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
+            config.max_position_embeddings - 1]``.
+
+            `What are position IDs? <../glossary.html#position-ids>`_
+        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
+            vectors than the model's internal embedding lookup matrix.
+
+            If :obj:`past_key_values` is used, optionally only the last :obj:`inputs_embeds` have to be input (see
+            :obj:`past_key_values`).
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        output_attentions (:obj:`bool`, `optional`):
+            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
+            tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
+"""
+PARALLELIZE_DOCSTRING = r"""
+    This is an experimental feature and is a subject to change at a moment's notice.
+
+    Uses a device map to distribute attention modules of the model across several devices. If no device map is given,
+    it will evenly distribute blocks across all devices.
+
+    Args:
+        device_map (:obj:`Dict[int, list]`, optional, defaults to None):
+            A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always
+            automatically mapped to the first device (for esoteric reasons). That means that the first device should
+            have fewer attention modules mapped to it than other devices. For reference, the gpt2 models have the
+            following number of attention modules:
+
+                - gpt2: 12
+                - gpt2-medium: 24
+                - gpt2-large: 36
+                - gpt2-xl: 48
+
+    Example::
+
+            # Here is an example of a device map on a machine with 4 GPUs using gpt2-xl, which has a total of 48 attention modules:
+            model = GPT2LMHeadModel.from_pretrained('gpt2-xl')
+            device_map = {0: [0, 1, 2, 3, 4, 5, 6, 7, 8],
+
+                          1: [9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21],
+                          2: [22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34],
+                          3: [35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47]}
+            model.parallelize(device_map)
+"""
+DEPARALLELIZE_DOCSTRING = r"""
+    Moves the model to cpu from a model parallel state.
+
+    Example::
+
+        # On a 4 GPU machine with gpt2-large:
+        model = GPT2LMHeadModel.from_pretrained('gpt2-large')
+        device_map = {0: [0, 1, 2, 3, 4, 5, 6, 7],
+
+                    1: [8, 9, 10, 11, 12, 13, 14, 15],
+                    2: [16, 17, 18, 19, 20, 21, 22, 23],
+                    3: [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35]}
+        model.parallelize(device_map) # Splits the model across several devices
+        model.deparallelize() # Put the model back on cpu and cleans memory by calling torch.cuda.empty_cache()
+"""
+
+class GPT2VAEEncoder(GPT2PreTrainedModel):
+    _keys_to_ignore_on_load_missing = ["attn.masked_bias"]
+
+    def __init__(self, config, wte, wpe, h, ln_f, dropout):
+        super().__init__(config)
+
+        self.wte = wte
+        self.wpe = wpe
+        self.drop = dropout
+        self.h = h
+        self.ln_f = ln_f
+
+        self.num_layers = config.num_hidden_layers
+        self.latent_size = config.latent_size
+        self.kl_threshold = config.kl_threshold
+        self.is_cvae = config.is_cvae
+
+        self.begin_layer = config.begin_layer if config.begin_layer else 0
+        self.end_layer = config.end_layer if config.end_layer else len(self.h) - 1
+        assert 0 <= self.begin_layer <= self.end_layer <= len(self.h) - 1
+
+        self.recog_network = nn.ModuleList([nn.Linear(config.hidden_size + self.latent_size, 2 * self.latent_size) for _ in range(self.num_layers)])
+        if self.is_cvae:
+            self.prior_network = nn.ModuleList([nn.Linear(self.latent_size + config.hidden_size, 2 * self.latent_size) for _ in range(self.num_layers)])
+        else:
+            self.prior_network = nn.ModuleList([nn.Linear(self.latent_size, 2 * self.latent_size) for _ in range(self.num_layers)])
+        self.reccurnt_cell_weight_hh = nn.ModuleList([nn.Linear(self.latent_size, self.latent_size) for _ in range(self.num_layers)])
+        self.reccurnt_cell_weight_ih = nn.ModuleList([nn.Linear(self.latent_size, self.latent_size) for _ in range(self.num_layers)])
+        self.pooling = nn.ModuleList([AverageSelfAttention(config) for _ in range(len(h))])
+        self.tanh_activate = nn.Tanh()
+
+        self.init_weights()
+        
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        # Check validity of device_map
+        self.device_map = (
+            get_device_map(len(self.h), range(torch.cuda.device_count())) if device_map is None else device_map
+        )
+        assert_device_map(self.device_map, len(self.h))
+        self.model_parallel = True
+        self.first_device = "cpu" if "cpu" in self.device_map.keys() else "cuda:" + str(min(self.device_map.keys()))
+        self.last_device = "cuda:" + str(max(self.device_map.keys()))
+        self.wte = self.wte.to(self.first_device)
+        self.wpe = self.wpe.to(self.first_device)
+        # Load onto devices
+        for k, v in self.device_map.items():
+            for block in v:
+                cuda_device = "cuda:" + str(k)
+                self.h[block] = self.h[block].to(cuda_device)
+                self.pooling[block] = self.pooling[block].to(cuda_device)
+                self.recog_network[block] = self.recog_network[block].to(cuda_device)
+                self.prior_network[block] = self.prior_network[block].to(cuda_device)
+                self.reccurnt_cell_weight_hh[block] = self.reccurnt_cell_weight_hh[block].to(cuda_device)
+                self.reccurnt_cell_weight_ih[block] = self.reccurnt_cell_weight_ih[block].to(cuda_device)
+        # ln_f to last
+        self.ln_f = self.ln_f.to(self.last_device)
+
+    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
+    def deparallelize(self):
+        self.model_parallel = False
+        self.device_map = None
+        self.first_device = "cpu"
+        self.last_device = "cpu"
+        self.wte = self.wte.to("cpu")
+        self.wpe = self.wpe.to("cpu")
+        for index in range(len(self.h)):
+            self.h[index] = self.h[index].to("cpu")
+            self.pooling[index] = self.pooling[index].to("cpu")
+            self.recog_network[index] = self.recog_network[index].to("cpu")
+            self.prior_network[index] = self.prior_network[index].to("cpu")
+            self.reccurnt_cell_weight_hh[index] = self.reccurnt_cell_weight_hh[index].to("cpu")
+            self.reccurnt_cell_weight_ih[index] = self.reccurnt_cell_weight_ih[index].to("cpu")
+        self.ln_f = self.ln_f.to("cpu")
+        torch.cuda.empty_cache()
+
+    def get_input_embeddings(self):
+        return self.wte
+
+    def set_input_embeddings(self, new_embeddings):
+        self.wte = new_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+        """
+        for layer, heads in heads_to_prune.items():
+            self.h[layer].attn.prune_heads(heads)
+
+    def _recurrent_latent(self, post_latent, prior_latent_representaion, weight_ih, weight_hh):
+        h = self.tanh_activate(weight_hh(prior_latent_representaion) + weight_ih(post_latent))
+        return h
+
+    def get_prior(self, batch_size, device, condition=None, condition_mask=None):
+        if condition is not None:
+            condition_output = self(condition, attention_mask=condition_mask, compute_kl=False)
+            condition_hidden_states = condition_output.hidden_states
+        else:
+            condition_hidden_states = tuple([None] * self.num_layers)
+        prior_dist = Normal.get_standard(batch_size, self.latent_size, device)
+        prior_latent_representaion = torch.zeros(batch_size, self.latent_size).to(device)
+        prior_latent = torch.zeros_like(prior_latent_representaion)
+        all_prior_latent = ()
+
+        for idx in range(self.begin_layer, self.end_layer + 1):
+            cond = condition_hidden_states[idx - self.begin_layer]
+            prior_network = self.prior_network[idx - self.begin_layer]
+            weight_ih = self.reccurnt_cell_weight_ih[idx - self.begin_layer]
+            weight_hh = self.reccurnt_cell_weight_hh[idx - self.begin_layer]
+            prior_latent_representaion = self._recurrent_latent(prior_latent, prior_latent_representaion, weight_ih, weight_hh)
+            if cond is not None:
+                prior_representaion = torch.cat((prior_latent_representaion, cond), dim=-1)
+            else:
+                prior_representaion = prior_latent_representaion
+            prior_mu, prior_sigma = torch.chunk(prior_network(prior_representaion), 2, dim=-1)
+            prior_dist = Normal(prior_mu, prior_sigma)
+            prior_latent, _ = prior_dist.sample()
+            all_prior_latent = all_prior_latent + (prior_latent, )
+        return all_prior_latent
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPastAndCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        return_dict=None,
+        condition_hidden_states=None,
+        compute_kl=True,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+            batch_size = input_ids.shape[0]
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size = inputs_embeds.shape[0]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.view(-1, input_shape[-1])
+        if position_ids is not None:
+            position_ids = position_ids.view(-1, input_shape[-1])
+
+        if past_key_values is None:
+            past_length = 0
+            past_key_values = tuple([None] * len(self.h))
+        else:
+            past_length = past_key_values[0][0].size(-2)
+        if position_ids is None:
+            position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
+
+        # GPT2Attention mask.
+        if attention_mask is not None:
+            assert batch_size > 0, "batch_size has to be defined and > 0"
+            attention_mask = attention_mask.view(batch_size, -1)
+            attention_mask = attention_mask[:, None, None, :]
+            attention_mask = attention_mask.to(dtype=self.dtype)
+            attention_mask = (1.0 - attention_mask) * -10000.0
+
+        if condition_hidden_states is None:
+            condition_hidden_states = tuple([None] * len(self.h))
+        else:
+            assert self.is_cvae
+        head_mask = self.get_head_mask(head_mask, self.config.n_layer)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
+        position_embeds = self.wpe(position_ids)
+        hidden_states = inputs_embeds + position_embeds
+
+        if token_type_ids is not None:
+            token_type_embeds = self.wte(token_type_ids)
+            hidden_states = hidden_states + token_type_embeds
+
+        hidden_states = self.drop(hidden_states)
+        output_shape = input_shape + (hidden_states.size(-1),)
+
+        prior_dist = Normal.get_standard(batch_size, self.latent_size, device)
+        prior_latent_representaion = torch.zeros(batch_size, self.latent_size).to(device)
+        post_latent = torch.zeros_like(prior_latent_representaion)
+        all_post_latent = ()
+        all_kl_loss = ()
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = ()
+        all_log_prior = ()
+        all_log_post = ()
+        all_post_mu = ()
+        all_post_sigma = ()
+        loops = zip(self.h, past_key_values, self.pooling, condition_hidden_states, self.recog_network, self.prior_network, self.reccurnt_cell_weight_ih, self.reccurnt_cell_weight_hh)
+        for i, (block, layer_past, pooling_layer, condition, recog_network, prior_network, weight_ih, weight_hh) in enumerate(loops):
+
+            # Model parallel
+            if self.model_parallel:
+                torch.cuda.set_device(hidden_states.device)
+                # Ensure layer_past is on same device as hidden_states (might not be correct)
+                if layer_past is not None:
+                    layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
+                # Ensure that attention_mask is always on the same device as hidden_states
+                if attention_mask is not None:
+                    attention_mask = attention_mask.to(hidden_states.device)
+                if isinstance(head_mask, torch.Tensor):
+                    head_mask = head_mask.to(hidden_states.device)
+
+            outputs = block(
+                hidden_states,
+                layer_past=layer_past,
+                attention_mask=attention_mask,
+                head_mask=head_mask[i],
+                use_cache=use_cache,
+                use_causal_mask=False,
+                output_attentions=output_attentions,
+            )
+
+            hidden_states = outputs[0]
+            representation, _ = pooling_layer(hidden_states)
+            all_hidden_states = all_hidden_states + (representation, )
+            if compute_kl and self.begin_layer <= i <= self.end_layer:
+                # update prior latent
+                prior_latent_representaion = self._recurrent_latent(post_latent, prior_latent_representaion, weight_ih, weight_hh)
+                if condition is not None:
+                    prior_representaion = torch.cat((prior_latent_representaion, condition), dim=-1)
+                else:
+                    prior_representaion = prior_latent_representaion
+                prior_mu, prior_sigma = torch.chunk(prior_network(prior_representaion), 2, dim=-1)
+                prior_dist = Normal(prior_mu, prior_sigma)
+
+                post_latent_representation = recog_network(torch.cat((representation, prior_latent_representaion), dim=-1))
+                post_mu, post_sigma = torch.chunk(post_latent_representation, 2, dim=-1)
+                post_dist = Normal(post_mu, post_sigma)
+                post_latent, _ = post_dist.sample()
+                all_post_latent = all_post_latent + (post_latent, )
+                kl_loss = post_dist.kl(prior_dist)
+                kl_threshold = torch.Tensor([self.kl_threshold]).type_as(kl_loss)
+                kl_loss = torch.max(kl_loss, kl_threshold)
+
+                log_prior_z = prior_dist.log_p(post_latent)
+                log_post_z = post_dist.log_p(post_latent)
+                if self.model_parallel:
+                    kl_loss = kl_loss.to(self.first_device)
+                    log_prior_z = log_prior_z.to(self.first_device)
+                    log_post_z = log_post_z.to(self.first_device)
+                all_kl_loss = all_kl_loss + (kl_loss, )
+                all_log_prior = all_log_prior + (log_prior_z, )
+                all_log_post = all_log_post + (log_post_z, )
+                all_post_mu = all_post_mu + (post_mu, )
+                all_post_sigma = all_post_sigma + (post_sigma, )
+                
+
+            if use_cache is True:
+                presents = presents + (outputs[1],)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 2],)
+
+            # Model Parallel: If it's the last layer for that device, put things on the next device
+            if self.model_parallel:
+                for k, v in self.device_map.items():
+                    if i == v[-1] and "cuda:" + str(k) != self.last_device:
+                        hidden_states = hidden_states.to("cuda:" + str(k + 1))
+
+        hidden_states = self.ln_f(hidden_states)
+        hidden_states = hidden_states.view(*output_shape)
+        
+        if compute_kl:
+            kl_loss = torch.stack(all_kl_loss).mean(0)
+            log_prior = torch.stack(all_log_prior).mean(0)
+            log_post = torch.stack(all_log_post).mean(0)
+
+            mu = torch.cat(all_post_mu, dim=-1)
+            sigma = torch.cat(all_post_sigma, dim=-1)
+        else:
+            kl_loss = None
+            log_prior = None
+            log_post = None
+            mu, sigma = None, None
+
+        if self.model_parallel:
+            torch.cuda.set_device(self.first_device)
+            hidden_states = hidden_states.to(self.first_device)
+
+        return EncoderOutput(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            latent=all_post_latent,
+            dist_parameter=(mu, sigma),
+            kl_loss=kl_loss,
+            log_prior=log_prior,
+            log_post=log_post,
+        )
+
+class GPT2VAEDecoderWithLMHead(GPT2PreTrainedModel):
+    _keys_to_ignore_on_load_missing = ["attn.masked_bias"]
+
+    def __init__(self, config, wte, wpe, h, ln_f, dropout):
+        super().__init__(config)
+
+        self.wte = wte
+        self.wpe = wpe
+        self.drop = dropout
+        self.h = h
+        self.ln_f = ln_f
+
+        self.init_weights()
+        
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+
+        self.begin_layer = config.begin_layer if config.begin_layer else 0
+        self.end_layer = config.end_layer if config.end_layer else len(self.h) - 1
+        assert 0 <= self.begin_layer <= self.end_layer <= len(self.h) - 1
+
+        self.use_bow = config.use_bow
+        if self.use_bow:
+            self.mlp_bow = nn.Linear(config.latent_size, config.n_embd)
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        # Check validity of device_map
+        self.device_map = (
+            get_device_map(len(self.h), range(torch.cuda.device_count())) if device_map is None else device_map
+        )
+        assert_device_map(self.device_map, len(self.h))
+        self.model_parallel = True
+        self.first_device = "cpu" if "cpu" in self.device_map.keys() else "cuda:" + str(min(self.device_map.keys()))
+        self.last_device = "cuda:" + str(max(self.device_map.keys()))
+        self.lm_head = self.lm_head.to(self.first_device)
+
+    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
+    def deparallelize(self):
+        self.model_parallel = False
+        self.device_map = None
+        self.first_device = "cpu"
+        self.last_device = "cpu"
+        self.lm_head = self.lm_head.to("cpu")
+        torch.cuda.empty_cache()
+
+    def get_input_embeddings(self):
+        return self.wte
+
+    def set_input_embeddings(self, new_embeddings):
+        self.wte = new_embeddings
+
+    def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
+        token_type_ids = kwargs.get("token_type_ids", None)
+        # only last token for inputs_ids if past is defined in kwargs
+        if past:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
+
+        attention_mask = kwargs.get("attention_mask", None)
+        position_ids = kwargs.get("position_ids", None)
+        latent = kwargs.get("latent", None)
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+        else:
+            position_ids = None
+        
+        return {
+            "input_ids": input_ids,
+            "past_key_values": past,
+            "use_cache": kwargs.get("use_cache"),
+            "position_ids": position_ids,
+            "attention_mask": attention_mask,
+            "token_type_ids": token_type_ids,
+            "all_latent": latent,
+        }
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+        """
+        for layer, heads in heads_to_prune.items():
+            self.h[layer].attn.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPastAndCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        all_latent=None,
+        return_dict=None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+            batch_size = input_ids.shape[0]
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size = inputs_embeds.shape[0]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if position_ids is not None:
+            position_ids = position_ids.view(-1, input_shape[-1])
+
+        if past_key_values is None:
+            past_length = 0
+            past_key_values = tuple([None] * len(self.h))
+        else:
+            past_length = past_key_values[0][0].size(-2)
+
+        if position_ids is None:
+            position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
+
+        # GPT2Attention mask.
+        if attention_mask is not None:
+            assert batch_size > 0, "batch_size has to be defined and > 0"
+            attention_mask = attention_mask.view(batch_size, -1)
+            # We create a 3D attention mask from a 2D tensor mask.
+            # Sizes are [batch_size, 1, 1, to_seq_length]
+            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+            # this attention mask is more simple than the triangular masking of causal attention
+            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+            attention_mask = attention_mask[:, None, None, :]
+
+            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+            # masked positions, this operation will create a tensor which is 0.0 for
+            # positions we want to attend and -10000.0 for masked positions.
+            # Since we are adding it to the raw scores before the softmax, this is
+            # effectively the same as removing these entirely.
+            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+            attention_mask = (1.0 - attention_mask) * -10000.0
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # head_mask has shape n_layer x batch x n_heads x N x N
+        head_mask = self.get_head_mask(head_mask, self.config.n_layer)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
+        position_embeds = self.wpe(position_ids)
+        hidden_states = inputs_embeds + position_embeds
+
+        hidden_states = self.drop(hidden_states)
+        output_shape = input_shape + (hidden_states.size(-1),)
+
+        all_hidden_states = ()
+        presents = ()
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
+        
+        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
+            
+            if self.begin_layer <= i <= self.end_layer:
+                latent = all_latent[i - self.begin_layer]
+            else:
+                latent = None
+            # Model parallel
+            if self.model_parallel:
+                torch.cuda.set_device(hidden_states.device)
+                # Ensure layer_past is on same device as hidden_states (might not be correct)
+                if layer_past is not None:
+                    layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
+                # Ensure that attention_mask is always on the same device as hidden_states
+                if attention_mask is not None:
+                    attention_mask = attention_mask.to(hidden_states.device)
+                if isinstance(head_mask, torch.Tensor):
+                    head_mask = head_mask.to(hidden_states.device)
+                if latent is not None:
+                    latent = latent.to(hidden_states.device)
+            
+            outputs = block(
+                hidden_states,
+                layer_past=layer_past,
+                attention_mask=attention_mask,
+                head_mask=head_mask[i],
+                latent=latent,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+                use_causal_mask=True,
+            )
+            hidden_states = outputs[0]
+            if use_cache is True:
+                presents = presents + (outputs[1],)
+            all_hidden_states = all_hidden_states + (hidden_states,)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 2],)
+
+            # Model Parallel: If it's the last layer for that device, put things on the next device
+            if self.model_parallel:
+                for k, v in self.device_map.items():
+                    if i == v[-1] and "cuda:" + str(k) != self.last_device:
+                        hidden_states = hidden_states.to("cuda:" + str(k + 1))
+
+        hidden_states = self.ln_f(hidden_states)
+        hidden_states = hidden_states.view(*output_shape)
+
+        if self.model_parallel:
+            torch.cuda.set_device(self.first_device)
+            hidden_states = hidden_states.to(self.lm_head.weight.device)
+        lm_logits = self.lm_head(hidden_states)
+        if self.use_bow:
+            bow_logits = self.lm_head(self.mlp_bow(all_latent[-1]))
+        else:
+            bow_logits = None
+        return DecoderOutput(
+            logits=lm_logits,
+            bow_logits=bow_logits,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor) -> Tuple[Tuple[torch.Tensor]]:
+        """
+        This function is used to re-order the :obj:`past_key_values` cache if
+        :meth:`~transformers.PreTrainedModel.beam_search` or :meth:`~transformers.PreTrainedModel.beam_sample` is
+        called. This is required to match :obj:`past_key_values` with the correct beam_idx at every generation step.
+        """
+        return tuple(
+            tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
+            for layer_past in past
+        )
+
+class Della(GPT2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        wte = nn.Embedding(config.vocab_size, config.hidden_size)
+        wpe = nn.Embedding(config.max_position_embeddings, config.hidden_size)
+
+        h = nn.ModuleList([GPT2Block(config) for _ in range(config.num_hidden_layers)])
+        ln_f = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+        drop = nn.Dropout(config.embd_pdrop)
+
+        self.encoder = GPT2VAEEncoder(config, wte, wpe, h, ln_f, drop)
+        self.decoder = GPT2VAEDecoderWithLMHead(config, wte, wpe, h, ln_f, drop)
+        self.loss_fn = nn.CrossEntropyLoss(ignore_index=config.pad_token_id, reduce=False)
+        self.loss_fn_reduced = nn.CrossEntropyLoss(ignore_index=config.pad_token_id)
+
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        self.device_map = (
+            get_device_map(len(self.encoder.h), range(torch.cuda.device_count()))
+            if device_map is None
+            else device_map
+        )
+        assert_device_map(self.device_map, len(self.encoder.h))
+        self.encoder.parallelize(self.device_map)
+        self.decoder.parallelize(self.device_map)
+        self.first_device = self.encoder.first_device
+        self.model_parallel = True
+    
+    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
+    def deparallelize(self):
+        self.encoder.deparallelize()
+        self.decoder.deparallelize()
+        torch.cuda.empty_cache()
+
+    def get_encode_states(self, input_ids, attention_mask):
+        encoder_outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
+        mu, sigma = encoder_outputs.dist_parameter
+        return mu, sigma
+
+    def get_reduced_celoss(self, input_ids, attention_mask):
+        encoder_outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
+        latent = encoder_outputs.latent
+        decoder_outputs = self.decoder(input_ids, attention_mask=attention_mask, all_latent=latent)
+        lm_logits = decoder_outputs.logits
+        
+        shift_logits = lm_logits[..., :-1, :].contiguous()
+        shift_labels = input_ids[..., 1:].contiguous()
+    
+        ce_loss = self.loss_fn_reduced(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+        return ce_loss
+
+    def get_celoss(self, input_ids, attention_mask, latent):
+        batch_size = input_ids.size(0)
+        decoder_outputs = self.decoder(input_ids, attention_mask=attention_mask, all_latent=latent)
+        lm_logits = decoder_outputs.logits
+        
+        shift_logits = lm_logits[..., :-1, :].contiguous()
+        shift_labels = input_ids[..., 1:].contiguous()
+    
+        ce_loss = self.loss_fn(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+        ce_loss = ce_loss.view(batch_size, -1).sum(-1)
+        return ce_loss
+
+    def get_klloss(self, input_ids, attention_mask):
+        encoder_outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
+        return encoder_outputs.kl_loss
+
+    def get_neg_entropy(self, input_ids, attention_mask, ns=30):
+        log_post_list = []
+        for _ in range(ns):
+            encoder_outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
+            log_post_list.append(encoder_outputs.log_post)
+        neg_entropy = torch.stack(log_post_list).mean(0).sum()
+        latentoutput = encoder_outputs.latent
+        latent = []
+        for l in latentoutput:
+            if self.model_parallel:
+                temp = l.to(self.first_device)
+                latent.append(temp)
+            else:
+                latent.append(l)
+        latent = torch.cat(latent, dim=-1)
+        return neg_entropy, latent
+
+    def iw_sample(self, input_ids, attention_mask):
+        encoder_outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
+        latent = encoder_outputs.latent
+        log_prior = encoder_outputs.log_prior
+        log_post = encoder_outputs.log_post
+        log_gen = -self.get_celoss(input_ids=input_ids, attention_mask=attention_mask, latent=latent)
+        log_likelihood = log_gen + log_prior - log_post
+        return log_gen, log_likelihood
+
+    def get_prior(self, batch_size, device, condition=None, condition_mask=None):
+        return self.encoder.get_prior(batch_size, device, condition=condition, condition_mask=condition_mask)
+
+    def forward(self, input_ids, labels=None, attention_mask=None, condition=None, condition_mask=None):
+        if condition is not None:
+            condition_output = self.encoder(condition, attention_mask=condition_mask, compute_kl=False)
+            condition = condition_output.hidden_states
+        else: 
+            condition = None
+
+        encoder_output = self.encoder(input_ids, attention_mask=attention_mask, condition_hidden_states=condition)
+        batch_size = input_ids.size(0)
+        latent = encoder_output.latent
+        kl_loss = encoder_output.kl_loss.mean()
+        decoder_output = self.decoder(input_ids, attention_mask=attention_mask, all_latent=latent)
+        lm_logits = decoder_output.logits
+        bow_logits = decoder_output.bow_logits
+
+        shift_logits = lm_logits[..., :-1, :].contiguous()
+        if labels is None:
+            shift_labels = input_ids[..., 1:].contiguous()
+        else:
+            shift_labels = labels[..., 1:].contiguous()
+    
+        ce_loss = self.loss_fn(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+        ce_loss = ce_loss.view(batch_size, -1).sum(-1)
+        ce_loss = ce_loss.mean()
+
+        seq_len = input_ids.size(-1)
+        if bow_logits is not None:
+            bow_logits = bow_logits.unsqueeze(1).repeat(1, seq_len, 1)[..., :-1, :].contiguous()
+            bow_loss = self.loss_fn(bow_logits.view(-1, bow_logits.size(-1)), shift_labels.view(-1))
+            bow_loss = bow_loss.view(batch_size, -1).sum(-1)
+            bow_loss = bow_loss.mean()
+            return ce_loss, kl_loss, bow_loss, encoder_output, decoder_output
+        else:
+            return ce_loss, kl_loss, encoder_output, decoder_output
diff --git a/modeloutput.py b/modeloutput.py
new file mode 100644
index 0000000..1e7eca3
--- /dev/null
+++ b/modeloutput.py
@@ -0,0 +1,25 @@
+import torch
+from transformers.file_utils import ModelOutput
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+@dataclass
+class EncoderOutput(ModelOutput):
+    last_hidden_state: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    latent: Optional[Tuple[torch.FloatTensor]] = None
+    kl_loss: Optional[torch.FloatTensor] = None
+    log_prior: Optional[torch.FloatTensor] = None
+    log_post: Optional[torch.FloatTensor] = None
+    dist_parameter: Optional[Tuple[torch.FloatTensor]] = None
+
+@dataclass
+class DecoderOutput(ModelOutput):
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    bow_logits: Optional[torch.FloatTensor] = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
\ No newline at end of file
diff --git a/train.py b/train.py
new file mode 100755
index 0000000..ef5a4d4
--- /dev/null
+++ b/train.py
@@ -0,0 +1,267 @@
+import torch
+import os
+from transformers import AdamW, get_linear_schedule_with_warmup
+import torch.nn as nn
+import logging
+from tqdm import tqdm
+from train_utils import *
+import pandas as pd
+logger = logging.getLogger(__name__)
+
+def prepare_for_training(args, model, train_iter):
+    optimizer = AdamW(model.parameters(), lr=args.learning_rate, correct_bias=True)
+    t_total = len(train_iter) * args.epochs
+    if args.use_scheduler:
+        scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+    else:
+        scheduler = None
+
+    return model, optimizer, scheduler
+
+def compute_loss(logits, target_tokens, kl_loss=None, beta=None, ignore_index=50256):
+    loss_fn = nn.CrossEntropyLoss(ignore_index=ignore_index)
+    shift_logits = logits[..., :-1, :].contiguous()
+    shift_labels = target_tokens[..., 1:].contiguous()
+    
+    ce_loss = loss_fn(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+    if kl_loss is not None:
+        loss = ce_loss + beta * kl_loss
+    else:
+        loss = ce_loss
+    return loss, ce_loss, kl_loss
+
+def train(model, train_iter, valid_iter, args):
+    logging.info('begin trainging...')
+    model, optimizer, scheduler = prepare_for_training(args, model, train_iter)
+    if args.cycle_annealing:
+        beta = 1e-5
+        beta_0 = 1e-5
+    else:
+        beta = 1
+    global_step = 0
+    
+    one_epoch_step = len(train_iter) // args.gradient_accumulation_steps
+    beta_zero = beta_increase = args.cycle_iters // 2
+    running_loss = 0
+    running_ce_loss = 0
+    running_kl_loss = 0
+    running_bow_loss = 0
+    for epoch in range(1 + args.load_epoch, args.epochs + args.load_epoch + 1):
+        model.train()
+        for i, inputs in enumerate(train_iter):
+            model_output = model(**inputs)
+            if args.use_bow:
+                ce_loss, kl_loss, bow_loss, _, _ = model_output
+                loss = ce_loss + beta * kl_loss + args.bow_weight * bow_loss
+            else:
+                ce_loss, kl_loss, _, _ = model_output
+                loss = ce_loss + beta * kl_loss
+                    
+            if args.gradient_accumulation_steps > 1:
+                loss = loss / args.gradient_accumulation_steps
+            loss = loss.mean()
+            loss.backward()
+            
+            running_loss += loss.item()
+            running_ce_loss += ce_loss.mean().item() / args.gradient_accumulation_steps
+            running_kl_loss += kl_loss.mean().item() / args.gradient_accumulation_steps
+            if args.use_bow:
+                running_bow_loss += bow_loss.mean().item() / args.gradient_accumulation_steps
+            
+            if (i + 1) % args.gradient_accumulation_steps == 0:
+                optimizer.step()
+                optimizer.zero_grad()
+                if scheduler is not None:
+                    scheduler.step()
+
+                global_step += 1
+                if args.cycle_annealing:
+                    one_period = epoch % args.cycle_iters
+                    if one_period < beta_zero:
+                        beta = beta_0
+                    else:
+                        beta = min(1.0, beta + (1 - beta_0) / (beta_increase * one_epoch_step / 2))
+
+                if global_step % args.log_step == 0:
+                    logging.info('training loss: step [{}~{}], loss {}, ce_loss {}, kl_loss {}, bow_loss {}, lr {}, beta {}'.
+                        format(global_step - args.log_step, global_step, running_loss / args.log_step, running_ce_loss / args.log_step, 
+                                running_kl_loss / args.log_step, running_bow_loss / args.log_step, optimizer.param_groups[0]['lr'], beta))
+                    running_loss = 0
+                    running_kl_loss = 0
+                    running_ce_loss = 0
+                    running_bow_loss = 0
+
+        valid(model, valid_iter, epoch, args, beta)
+        save(model, args, epoch)
+    logging.info('training finished')
+
+def valid(model, valid_iter, epoch, args, beta=1):
+    model.eval()
+    with torch.no_grad():
+        valid_loss = 0
+        valid_kl_loss = 0
+        valid_ce_loss = 0
+        valid_bow_loss = 0
+        for inputs in tqdm(valid_iter, desc='valid epoch {}'.format(epoch)):
+            model_output = model(**inputs)
+            if args.use_bow:
+                ce_loss, kl_loss, bow_loss, _, _ = model_output
+                loss = ce_loss + beta * kl_loss + args.bow_weight * bow_loss
+            else:
+                ce_loss, kl_loss, _, _ = model_output
+                loss = ce_loss + beta * kl_loss
+            loss = loss.mean()
+            valid_loss += loss.item()
+            valid_ce_loss += ce_loss.mean().item()
+            valid_kl_loss += kl_loss.mean().item()
+            if args.use_bow:
+                valid_bow_loss += bow_loss.mean().item()
+        
+        valid_loss = valid_loss / len(valid_iter)
+        valid_ce_loss = valid_ce_loss / len(valid_iter)
+        valid_kl_loss = valid_kl_loss / len(valid_iter)
+        valid_bow_loss = valid_bow_loss / len(valid_iter)
+        logging.info('valid result: epoch {}, loss {}, ce_loss {}, kl {}, bow {}'.format(epoch, valid_loss, valid_ce_loss, valid_kl_loss, valid_bow_loss))
+        
+        if args.eval_metrics:
+            ppl, elbo, nll, kl = calc_iwnll(model, valid_iter, ns=args.sample_times)
+            au = calc_au(model, valid_iter)
+            logging.info('valid result: epoch {}, ppl {}, elbo {}, nll {}, kl {}'.format(epoch, ppl, elbo, nll, kl))
+            logging.info('valid result: epoch {}, au {}'.format(epoch, au))
+
+def save(model, args, epoch):
+    save_path = os.path.join(args.output_dir, args.model_name, 'model_epoch_{}.pt'.format(epoch))
+    if not os.path.exists(os.path.join(args.output_dir, args.model_name)):
+        os.makedirs(os.path.join(args.output_dir, args.model_name), exist_ok=True)
+    try:
+        model_to_save = model.module
+    except:
+        model_to_save = model
+    torch.save(model_to_save.state_dict(), save_path)
+
+def generate(model, test_iter, tokenizer, args): 
+    if args.dataset_type == 'wp':
+        has_condition = "conditional"
+    else:
+        has_condition = "unconditional"
+    if args.top_k > 0:
+        generate_param = "topk_{}".format(args.top_k)
+    elif args.greedy_decoding:
+        generate_param = "greedy_decoding"
+    else:
+        generate_param = "beamsearch_{}".format(args.num_beams)
+    
+    logging.info('{} generate with {}'.format(has_condition, generate_param))
+    def filter_sen(sen):
+        sen = sen.replace('<sep>', '')
+        sen = sen.replace('<s>', '')
+        sen = sen.replace('</s>', '')
+        sen = sen.replace('<pad>', '')
+        sen = sen.replace('<|endoftext|>', '')
+        sen = sen.replace('<eos>', '')
+        sen = ' '.join(sen.split())
+        return sen
+    model.eval()
+    model.decoder.config.is_encoder_decoder = False
+
+    output_list = []
+    target_list = []
+    source_list = []
+    
+    with torch.no_grad():
+        for inputs in tqdm(test_iter):
+            target = inputs['input_ids']
+            if args.dataset_type == 'wp':
+                source = inputs['condition']
+            
+            batch_size = target.size(0)
+            device = target.device
+            input_ids = target[:, 0].unsqueeze(1)
+            model_kwargs = {}
+            if args.dataset_type == 'wp':
+                prior_latent = model.get_prior(batch_size, device, condition=inputs['condition'], condition_mask=inputs['condition_mask'])
+                model_kwargs['attention_mask'] = inputs['condition_mask']
+                input_ids = inputs['condition']
+            else:
+                prior_latent = model.get_prior(batch_size, device)
+            
+            gen_model = model.decoder
+            if args.top_k > 0:
+                ans = gen_model.generate(
+                    input_ids, 
+                    latent=prior_latent, 
+                    bos_token_id=tokenizer.bos_id, 
+                    eos_token_id=tokenizer.eos_id, 
+                    pad_token_id=tokenizer.pad_id, 
+                    do_sample=True,
+                    top_k=args.top_k, 
+                    top_p=args.top_p, 
+                    min_length=input_ids.size(-1) + 3, 
+                    max_length=min(args.max_length, 1024),
+                    repetition_penalty=args.repetition_penalty, 
+                    **model_kwargs,
+                )
+            elif args.greedy_decoding:
+                ans = gen_model.generate(
+                    input_ids, 
+                    latent=prior_latent, 
+                    bos_token_id=tokenizer.bos_id, 
+                    eos_token_id=tokenizer.eos_id, 
+                    pad_token_id=tokenizer.pad_id, 
+                    min_length=input_ids.size(-1) + 3, 
+                    max_length=min(args.max_length, 1024),
+                    repetition_penalty=args.repetition_penalty, 
+                    **model_kwargs,
+                )
+            else:
+                if prior_latent is not None:
+                    if isinstance(prior_latent, tuple):
+                        latent = [item.repeat_interleave(args.num_beams, dim=0) for item in prior_latent]
+                    else:
+                        latent = prior_latent.repeat_interleave(args.num_beams, dim=0)
+                else:
+                    latent = None
+                ans = gen_model.generate(
+                    input_ids, 
+                    latent=latent, 
+                    bos_token_id=tokenizer.bos_id, 
+                    eos_token_id=tokenizer.eos_id, 
+                    pad_token_id=tokenizer.pad_id, 
+                    num_beams=args.num_beams, 
+                    min_length=input_ids.size(-1) + 3, 
+                    max_length=min(args.max_length, 1024), 
+                    repetition_penalty=args.repetition_penalty, 
+                    **model_kwargs,
+                )
+            ans = ans.cpu().numpy()
+
+            if args.dataset_type == 'wp':
+                target = target.cpu().numpy()
+                source = source.cpu().numpy()
+            for i in range(len(ans)):
+                text_ans = tokenizer.decode(ans[i], clean_up_tokenization_spaces=False)
+                text_ans = filter_sen(text_ans)
+                if len(text_ans) > 0:
+                    output_list.append(text_ans)
+                    if args.dataset_type in 'wp':
+                        target_text = tokenizer.decode(target[i], clean_up_tokenization_spaces=False)
+                        target_text = filter_sen(target_text)
+                        target_list.append(target_text)
+                        source_text = tokenizer.decode(source[i], clean_up_tokenization_spaces=False)
+                        source_text = filter_sen(source_text)
+                        source_list.append(source_text)
+
+    save_dir = os.path.join(args.generation_output_dir, args.model_name)
+    file_name = '{}_output_{}_epoch_{}_outputs.txt'.format(has_condition, generate_param, args.load_epoch)
+    logging.info('generation output save at {}'.format(save_dir))
+    if not os.path.exists(save_dir):
+        os.makedirs(save_dir, exist_ok=True)
+    with open(os.path.join(save_dir, file_name), 'w') as f:
+        f.write('\n'.join(output_list))
+    if args.dataset_type == 'wp':
+        file_name = '{}_output_{}_epoch_{}_targets.txt'.format(has_condition, generate_param, args.load_epoch)
+        with open(os.path.join(save_dir, file_name), 'w') as f:
+            f.write('\n'.join(target_list))
+        file_name = '{}_output_{}_epoch_{}_sources.txt'.format(has_condition, generate_param, args.load_epoch)
+        with open(os.path.join(save_dir, file_name), 'w') as f:
+            f.write('\n'.join(source_list))
diff --git a/train_utils.py b/train_utils.py
new file mode 100755
index 0000000..ede6f02
--- /dev/null
+++ b/train_utils.py
@@ -0,0 +1,84 @@
+import numpy as np
+from dist import Normal
+from tqdm import tqdm
+import torch
+import math
+
+def log_sum_exp(value, dim=None, keepdim=False):
+    """Numerically stable implementation of the operation
+    value.exp().sum(dim, keepdim).log()
+    """
+    if dim is not None:
+        m, _ = torch.max(value, dim=dim, keepdim=True)
+        value0 = value - m
+        if keepdim is False:
+            m = m.squeeze(dim)
+        return m + torch.log(torch.sum(torch.exp(value0), dim=dim, keepdim=keepdim))
+    else:
+        m = torch.max(value)
+        sum_exp = torch.sum(torch.exp(value - m))
+        return m + torch.log(sum_exp)
+
+def calc_iwnll(model, iters, ns=30):
+    report_kl_loss = report_ce_loss = report_loss = 0
+    report_num_words = report_num_sents = 0
+    for inputs in tqdm(iters, desc="Evaluating PPL"):
+        input_ids = inputs['input_ids']
+        attention_mask = inputs['attention_mask']
+        seq_len = attention_mask.sum(-1).long()
+        report_num_sents += input_ids.size(0)
+        report_num_words += seq_len.sum().item()
+        kl_loss = model.get_klloss(input_ids, attention_mask)
+        ll_tmp = []
+        ce_tmp = []
+        for _ in range(ns):
+            log_gen, log_likelihood = model.iw_sample(input_ids, attention_mask)
+            ce_tmp.append(log_gen)
+            ll_tmp.append(log_likelihood)
+
+        ll_tmp = torch.stack(ll_tmp, dim=0)
+        log_prob_iw = log_sum_exp(ll_tmp, dim=0) - math.log(ns)
+        log_ce_iw = torch.mean(torch.stack(ce_tmp), dim=0)
+        report_kl_loss += kl_loss.sum().item()
+        report_ce_loss += log_ce_iw.sum().item()
+        report_loss += log_prob_iw.sum().item()
+
+    elbo = (report_kl_loss - report_ce_loss) / report_num_sents
+    nll = - report_ce_loss / report_num_sents
+    kl = report_kl_loss / report_num_sents
+    ppl = np.exp(-report_loss / report_num_words)
+    return ppl, elbo, nll, kl
+
+def calc_au(model, iters, delta=0.2):
+    """compute the number of active units
+    """
+    cnt = 0
+    for inputs in tqdm(iters, desc="Evaluating Active Units, Stage 1"):
+        input_ids = inputs['input_ids']
+        attention_mask = inputs['attention_mask']
+        mean, _ = model.get_encode_states(input_ids=input_ids, attention_mask=attention_mask)
+        if cnt == 0:
+            means_sum = mean.sum(dim=0, keepdim=True)
+        else:
+            means_sum = means_sum + mean.sum(dim=0, keepdim=True)
+        cnt += mean.size(0)
+
+    # (1, nz)
+    mean_mean = means_sum / cnt
+
+    cnt = 0
+    for inputs in tqdm(iters, desc="Evaluating Active Units, Stage 2"):
+        input_ids = inputs['input_ids']
+        attention_mask = inputs['attention_mask']
+        mean, _ = model.get_encode_states(input_ids=input_ids, attention_mask=attention_mask)
+        if cnt == 0:
+            var_sum = ((mean - mean_mean) ** 2).sum(dim=0)
+        else:
+            var_sum = var_sum + ((mean - mean_mean) ** 2).sum(dim=0)
+        cnt += mean.size(0)
+
+    # (nz)
+    au_var = var_sum / (cnt - 1)
+    au = (au_var >= delta).sum().item()
+    au_prop = au / mean.size(-1)
+    return au_prop
\ No newline at end of file