run_lm_vae_pretraining_phdist_beta.py

# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The 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.
"""
Fine-tuning the library models for language modeling on a text file (GPT, GPT-2, BERT, RoBERTa).
GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT and RoBERTa are fine-tuned
using a masked language modeling (MLM) loss.
"""

from __future__ import absolute_import, division, print_function


import pdb
import argparse
import glob
import logging

import os, sys
import pickle
import random
from pathlib import Path
import os.path as op
import time, json
from io import open
import re

import numpy as np
import torch
from torch.utils.data import DataLoader, Dataset, SequentialSampler, RandomSampler, TensorDataset
from torch.utils.data.distributed import DistributedSampler
from tensorboardX import SummaryWriter
from tqdm import tqdm, trange
from collections import defaultdict
import subprocess

import torch.nn.init as init

# from azure.cosmosdb.table.tableservice import TableService
# from azure.cosmosdb.table.models import Entity
from datetime import datetime

try:
    this_file = __file__
except NameError:
    this_file = sys.argv[0]
this_file = op.abspath(this_file)
print('current path: {}'.format(os.path.abspath(__file__)))
print('current folder: {}'.format(op.dirname(this_file)))
sys.path.insert(0, op.join(op.dirname(this_file), '../..'))


from pytorch_transformers import (WEIGHTS_NAME, AdamW, WarmupLinearSchedule,
                                  BertConfig, BertForLatentConnector, BertTokenizer,
                                  GPT2Config, GPT2ForLatentConnector, GPT2Tokenizer,
                                  OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer,
                                  RobertaConfig, RobertaForMaskedLM, RobertaTokenizer)

from utils import (calc_iwnll, calc_mi, calc_au, BucketingDataLoader, BucketingMultipleFiles_DataLoader, frange_cycle_linear, frange_cycle_zero_linear)

from modules import VAE


# logging.getLogger("azure").setLevel(logging.WARNING)
# logging.getLogger("TableService").setLevel(logging.WARNING)

logger = logging.getLogger(__name__)

MODEL_CLASSES = {
    'gpt2': (GPT2Config, GPT2ForLatentConnector, GPT2Tokenizer),
    'openai-gpt': (OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
    'bert': (BertConfig, BertForLatentConnector, BertTokenizer),
    'roberta': (RobertaConfig, RobertaForMaskedLM, RobertaTokenizer)
}

# ts = TableService(account_name=storage_name, account_key=key)


def ompi_rank():
    """Find OMPI world rank without calling mpi functions
    :rtype: int
    """
    return int(os.environ.get('OMPI_COMM_WORLD_RANK') or 0)

def ompi_size():
    """Find OMPI world size without calling mpi functions
    :rtype: int
    """
    return int(os.environ.get('OMPI_COMM_WORLD_SIZE') or 1)

def ompi_local_rank():
    """Find OMPI local rank without calling mpi functions
    :rtype: int
    """
    return int(os.environ.get('OMPI_COMM_WORLD_LOCAL_RANK') or 0)

def ompi_local_size():
    """Find OMPI local size without calling mpi functions
    :rtype: int
    """
    return int(os.environ.get('OMPI_COMM_WORLD_LOCAL_SIZE') or 1)

def get_master_machine():
    mpi_host_file = op.expanduser('~/mpi-hosts')
    with open(mpi_host_file, 'r') as f:
        master_name = f.readline().strip()
    return master_name

def get_master_ip(master_name=None):
    if master_name is None:
        master_name = get_master_machine()
    #etc_host_file = '/etc/hosts'
    etc_host_file = op.expanduser('~/etc-hosts')
    with open(etc_host_file, 'r') as f:
        name_ip_pairs = f.readlines()
    name2ip = {}
    for name_ip_pair in name_ip_pairs:
        pair_list = name_ip_pair.split(' ')
        key = pair_list[1].strip()
        value = pair_list[0]
        name2ip[key] = value
    return name2ip[master_name]

def get_gpus_nocache():
    """List of NVIDIA GPUs
    """
    cmds = 'nvidia-smi --query-gpu=name --format=csv,noheader'.split(' ')

    p = subprocess.Popen(cmds, stdout=subprocess.PIPE)
    ret = p.communicate()
    gpus_str = ret[0].decode("utf-8")
    gpus_arr = [gpu.strip() for gpu in gpus_str.strip().split('\n')]
    return gpus_arr

_GPUS = get_gpus_nocache()
print('_GPUs: {}'.format(_GPUS))

def get_gpus():
    """List of NVIDIA GPUs
    """
    return _GPUS

def gpu_indices(divisible=True):
    """Get the GPU device indices for this process/rank
    :param divisible: if GPU count of all ranks must be the same
    :rtype: list[int]
    """
    local_size = ompi_local_size()
    local_rank = ompi_local_rank()
    assert 0 <= local_rank < local_size, "Invalid local_rank: {} local_size: {}".format(local_rank, local_size)
    gpu_count = len(get_gpus())
    assert gpu_count >= local_size > 0, "GPU count: {} must be >= LOCAL_SIZE: {} > 0".format(gpu_count, local_size)
    if divisible:
        ngpu = gpu_count / local_size
        gpus = np.arange(local_rank * ngpu, (local_rank + 1) * ngpu)
        if gpu_count % local_size != 0:
            logger.warning("gpu_count: {} not divisible by local_size: {}; some GPUs may be unused".format(gpu_count, local_size))
    else:
        gpus = np.array_split(range(gpu_count), local_size)[local_rank]

    ret_gpus = [int(g) for g in gpus]
    return ret_gpus


def build_dataload_and_cache_examples(args, tokenizer, evaluate=False):
    if isinstance(tokenizer, list):
        args.batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
        file_path=args.train_data_file
        dataloader = BucketingMultipleFiles_DataLoader(file_path, args.batch_size, args.max_seq_length, tokenizer, args, bucket=100, shuffle=True)
    else:
        pass 
    return dataloader


def set_seed(args):
    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)


def weights_init_rondom(model):
    model = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
    model_state_dict = model.state_dict()
    for key in model_state_dict:
        init.normal_(model_state_dict[key].data)  


def mask_tokens(inputs, tokenizer, args):
    """ Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original. """
    labels = inputs.clone()
    # We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
    
    masked_indices = torch.bernoulli(torch.full(labels.shape, args.mlm_probability)).to(torch.uint8)
    labels[masked_indices==1] = -1  # We only compute loss on masked tokens

    # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
    indices_replaced = torch.bernoulli(torch.full(labels.shape, 0.8)).to(torch.uint8) & masked_indices
    inputs[indices_replaced] = tokenizer.convert_tokens_to_ids(tokenizer.mask_token)

    # 10% of the time, we replace masked input tokens with random word
    indices_random = torch.bernoulli(torch.full(labels.shape, 0.5)).to(torch.uint8) & masked_indices & ~indices_replaced
    indices_random = indices_random
    random_words = torch.randint(len(tokenizer), labels.shape, dtype=torch.long)
    inputs[indices_random] = random_words[indices_random]

    # The rest of the time (10% of the time) we keep the masked input tokens unchanged
    return inputs, labels



def save_checkpoint(model_vae, optimizer, global_step, args):

    # Create output directory if needed
    # Save model checkpoint
    output_encoder_dir = os.path.join(args.output_dir, 'checkpoint-encoder-{}'.format(global_step))
    output_decoder_dir = os.path.join(args.output_dir, 'checkpoint-decoder-{}'.format(global_step))
    if not os.path.exists(output_encoder_dir) and args.local_rank in [-1, 0]:
        os.makedirs(output_encoder_dir)
    if not os.path.exists(output_decoder_dir) and args.local_rank in [-1, 0]:
        os.makedirs(output_decoder_dir)

    logger.info("Saving encoder model checkpoint to %s", output_encoder_dir)
    logger.info("Saving decoder model checkpoint to %s", output_decoder_dir)
    # Save a trained model, configuration and tokenizer using `save_pretrained()`.
    # They can then be reloaded using `from_pretrained()`

    model_encoder_to_save = model_vae.module.encoder if hasattr(model_vae, 'module') else model_vae.encoder  # Take care of distributed/parallel training
    model_decoder_to_save = model_vae.module.decoder if hasattr(model_vae, 'module') else model_vae.decoder  # Take care of distributed/parallel training

    # Good practice: save your training arguments together with the trained model
    if args.use_philly:
        save_solid = False
        while not save_solid:
            try:
                model_encoder_to_save.save_pretrained(output_encoder_dir)
                torch.save(args, os.path.join(output_encoder_dir, 'training_encoder_args.bin'))
                save_solid = True
            except:
                pass
    else:
        model_encoder_to_save.save_pretrained(output_encoder_dir)
        torch.save(args, os.path.join(output_encoder_dir, 'training_encoder_args.bin'))


    if args.use_philly:
        save_solid = False
        while not save_solid:
            try:
                model_decoder_to_save.save_pretrained(output_decoder_dir)
                torch.save(args, os.path.join(output_decoder_dir, 'training_decoder_args.bin'))
                save_solid = True
            except:
                pass
    else:
        model_decoder_to_save.save_pretrained(output_decoder_dir)
        torch.save(args, os.path.join(output_decoder_dir, 'training_encoder_args.bin'))


    # save the full model and optmizer into a checkpoint
    model_to_save = model_vae.module if hasattr(model_vae, 'module') else model_vae  # Take care of distributed/parallel training

    checkpoint = {
    'iter': global_step,
    'model_state_dict': model_to_save.state_dict(),
    'optimizer_state_dict': optimizer.state_dict(),
    'beta': model_to_save.args.beta,
    'args': args
    }

    output_full_dir = os.path.join(args.output_dir, 'checkpoint-full-{}'.format(global_step))
    if not os.path.exists(output_full_dir) and args.local_rank in [-1, 0]:
        os.makedirs(output_full_dir)

    logger.info("Start saving full model checkpoint to %s", output_full_dir)
    if args.use_philly:
        save_solid = False
        n_save_attempts = 0
        while not save_solid:
            try:
                n_save_attempts += 1
                logger.info(f"Saving full checkpoint: {n_save_attempts} attempts made")
                torch.save(checkpoint, os.path.join(output_full_dir, 'training.bin'))
                logger.info("Saving full checkpoint to %s,", output_full_dir)
                save_solid = True
            except:
                pass
    else:
        torch.save(checkpoint, os.path.join(output_full_dir, 'training.bin'))
        logger.info("Saving full checkpoint to %s", output_full_dir)


def train(args, train_dataloader, model_vae, encoder_tokenizer, decoder_tokenizer, table_name):
    """ Train the model """
    #gpus = list(gpu_indices())

    if args.local_rank in [-1, 0]: tb_writer = SummaryWriter()


    args.n_gpu = (torch.distributed.get_world_size() if args.local_rank != -1 else 1)
    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
    # train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
    # train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)

    if args.max_steps > 0:
        t_total = args.max_steps
        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
    else:
        t_total = len(train_dataloader) // args.gradient_accumulation_steps #* args.num_train_epochs

    if args.distributed:
        t_total = t_total // ompi_size()

    # Prepare optimizer and schedule (linear warmup and decay)


    # model_encoder, model_decoder, model_connector = model_vae.encoder,  model_vae.decoder, model_vae.linear
    no_decay = ['bias', 'LayerNorm.weight']
    optimizer_grouped_parameters = [
        {'params': [p for n, p in model_vae.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
        {'params': [p for n, p in model_vae.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
        ]
    
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)


    if args.fp16:
        try:
            from apex import amp
        except ImportError:
            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
        model_vae, optimizer = amp.initialize(model_vae, optimizer, opt_level=args.fp16_opt_level)

    # multi-gpu training (should be after apex fp16 initialization)
    #if args.n_gpu > 1:
    #    model_vae = torch.nn.DataParallel(model_vae, device_ids=range(args.n_gpu)).to(args.device)

    # Distributed training (should be after apex fp16 initialization)
    #if args.local_rank != -1:
        #model_vae = torch.nn.parallel.DistributedDataParallel(model_vae, device_ids=gpus, output_device=args.local_rank, find_unused_parameters=True)
        #model_vae = torch.nn.parallel.DistributedDataParallel(model_vae, device_ids=gpus)


    files = Path(args.train_data_file)
    num_files = len(list(files.glob('*seq64*.json')))


    # Train!
    logger.info("***** Running training *****")
    logger.info("  Num files = %d", num_files)
    logger.info("  Num examples of first file = %d", train_dataloader.num_examples)
    logger.info("  Num Epochs = %d", args.num_train_epochs)
    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
    logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
                   args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
    logger.info("  Total optimization steps = %d", t_total)


    global_step = 0
    tr_loss, logging_loss = 0.0, 0.0

    model_vae.zero_grad()
    num_train_epochs_iterator = trange(int(args.num_train_epochs), desc="Epoch") #, disable=args.local_rank not in [-1, 0])

    #n_iter = int(args.num_train_epochs) * len(train_dataloader)
    n_iter_per_file = train_dataloader.num_examples / args.train_batch_size
    n_iter = int(args.num_train_epochs * n_iter_per_file * num_files)
    beta_t_list = frange_cycle_zero_linear(n_iter, start=0.0, stop=args.beta, n_cycle=10,  ratio_increase=args.ratio_increase, ratio_zero=args.ratio_zero)
    logger.info(f"Total iters (estimated): {n_iter}; Length of beta schedule: {len(beta_t_list)}; #Iter per file {n_iter_per_file}")

    beta_t = 0.0
    tmp_list = []
    dict_token_length = defaultdict(int)

    set_seed(args)  # Added here for reproducibility (even between python 2 and 3)
    for epoch in range(int(args.num_train_epochs)): # num_train_epochs_iterator:
        train_dataloader.reset()
        for idx_file in range(num_files-1):

            logger.info(f"Rank {ompi_rank()}, Epoch {epoch}, File idx {train_dataloader.file_idx}")
            #epoch_iterator = tqdm(train_dataloader, desc="Iteration") #disable=disable=args.local_rank not in [-1, 0])
            for step, batch in enumerate(train_dataloader):
                tokenized_text0, tokenized_text1, tokenized_text_lengths = batch
                
                #dict_token_length[tokenized_text_lengths[0,0].item()] += 1
                # continue
                # tokenized_text0 = tokenized_text0.to(args.device)
                # tokenized_text1 = tokenized_text1.to(args.device)
                # prepare input-output data for reconstruction

                inputs, labels = mask_tokens(tokenized_text0, encoder_tokenizer, args) if args.mlm else (tokenized_text0, tokenized_text1)
                labels = tokenized_text1

                tokenized_text1 = tokenized_text1.to(args.device)
                inputs = inputs.to(args.device)
                labels = labels.to(args.device)

                model_vae.train()

                if args.use_beta_schedule:
                    if global_step >= len(beta_t_list):
                        beta_t = 1.0
                    else:
                        beta_t = beta_t_list[global_step]

                    #try:
                    #    beta_t = beta_t_list[global_step] #[step + idx_file* n_iter_per_file]
                    #except:
                    #    beta_t = 0.0

                #beta_t = 0.0 # beta_t_list[step +  epoch*len(epoch_iterator)]
                model_vae.module.args.beta = beta_t

                if beta_t == 0.0:
                    model_vae.module.args.fb_mode = 0
                else:
                    model_vae.module.args.fb_mode = 1
                
                if args.use_deterministic_connect:
                    model_vae.module.args.fb_mode = 2

                loss_rec, loss_kl, loss = model_vae(inputs, labels)

                loss_rec = loss_rec.mean()  # mean() to average on multi-gpu parallel training
                loss_kl = loss_kl.mean()
                loss = loss.mean()

                if args.use_philly:
                    #if args.local_rank in [-1, 0]:
                    if args.logging_steps > 0 and global_step % args.logging_steps == 0:
                        logger.info("Steps {}, Rank {}, File {}, Epoch: [{}/{}][{}/{}], Beta: {}, Loss: {}".format(global_step, ompi_rank(), train_dataloader.file_idx,
                                    epoch, args.num_train_epochs, step, n_iter_per_file, model_vae.module.args.beta, loss_rec))
                        logger.info("PROGRESS: {}%".format(round(100 * global_step /n_iter, 4)))
                        logger.info("EVALERR: {}%".format(loss_rec))

                if args.gradient_accumulation_steps > 1:
                    loss = loss / args.gradient_accumulation_steps

                if args.fp16:
                    with amp.scale_loss(loss, optimizer) as scaled_loss:
                        scaled_loss.backward()                                   
                else:
                    loss.backward()

                tr_loss += loss.item()
                if (step + 1) % args.gradient_accumulation_steps == 0:
                    if args.fp16:
                        torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
                    else:
                        torch.nn.utils.clip_grad_norm_(model_vae.parameters(), args.max_grad_norm)

                    optimizer.step()
                    scheduler.step()  # Update learning rate schedule
                    model_vae.zero_grad()

                    global_step += 1

                    if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
                        # Log metrics
                        if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
                            results = evaluate(args, model_vae, encoder_tokenizer, decoder_tokenizer)
                            for key, value in results.items():
                                tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
                        tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
                        tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
                        logging_loss = tr_loss

                    if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
                        save_checkpoint(model_vae, optimizer, global_step, args)
       

                if args.max_steps > 0 and global_step > args.max_steps:
                    #epoch_iterator.close()
                    break
                


    # print(dict_token_length)
    # with open('wikipedia_stats.json', 'w') as fp:
    #     json.dump(dict_token_length, fp)

    return global_step, tr_loss / global_step, optimizer


def main():
    parser = argparse.ArgumentParser()

    ## Required parameters
    parser.add_argument("--train_data_file", default=None, type=str, required=True,
                        help="The input training data file (a text file).")
    parser.add_argument("--output_dir", default=None, type=str, required=True,
                        help="The output directory where the model predictions and checkpoints will be written.")
    parser.add_argument("--dataset", default=None, type=str, help="The dataset.")

    ## Other parameters
    parser.add_argument("--eval_data_file", default=None, type=str,
                        help="An optional input evaluation data file to evaluate the perplexity on (a text file).")
    parser.add_argument("--ExpName", default="", type=str,
                        help="The experiment name used in Azure Table.")

    ## Encoder options
    parser.add_argument("--encoder_model_type", default="bert", type=str,
                        help="The encoder model architecture to be fine-tuned.")
    parser.add_argument("--encoder_model_name_or_path", default="bert-base-cased", type=str,
                        help="The encoder model checkpoint for weights initialization.")
    parser.add_argument("--encoder_config_name", default="", type=str,
                        help="Optional pretrained config name or path if not the same as model_name_or_path")
    parser.add_argument("--encoder_tokenizer_name", default="", type=str,
                        help="Optional pretrained tokenizer name or path if not the same as model_name_or_path")

    ## Decoder options
    parser.add_argument("--decoder_model_type", default="gpt2", type=str,
                        help="The decoder model architecture to be fine-tuned.")
    parser.add_argument("--decoder_model_name_or_path", default="bert-base-cased", type=str,
                        help="The decoder model checkpoint for weights initialization.")
    parser.add_argument("--decoder_config_name", default="", type=str,
                        help="Optional pretrained config name or path if not the same as model_name_or_path")
    parser.add_argument("--decoder_tokenizer_name", default="", type=str,
                        help="Optional pretrained tokenizer name or path if not the same as model_name_or_path")

    ## Variational auto-encoder
    parser.add_argument("--latent_size", default=32, type=int, help="Latent space dimension.")
    parser.add_argument("--use_deterministic_connect", action='store_true',
                        help="Use deterministic inference to generate latent codes, i.e., standard auto-encoders.")
    parser.add_argument("--use_beta_schedule", action='store_true', help="Use cyclical beta schedule for auto-encoders.")

    ## Objective functions
    parser.add_argument("--mlm", action='store_true',
                        help="Train with masked-language modeling loss instead of language modeling.")
    parser.add_argument("--mlm_probability", type=float, default=0.15,
                        help="Ratio of tokens to mask for masked language modeling loss")
    parser.add_argument("--beta", type=float, default=1.0,
                        help="The weighting hyper-parameter of the KL term in VAE")


    parser.add_argument("--cache_dir", default="", type=str,
                        help="Optional directory to store the pre-trained models downloaded from s3 (instread of the default one)")
    parser.add_argument("--max_seq_length", default=512, type=int,
                        help="Optional input sequence length before tokenization. The sequence will be dropped if it is longer the max_seq_length")
    parser.add_argument("--block_size", default=-1, type=int,
                        help="Optional input sequence length after tokenization."
                             "The training dataset will be truncated in block of this size for training."
                             "Default to the model max input length for single sentence inputs (take into account special tokens).")
    parser.add_argument("--do_train", action='store_true',
                        help="Whether to run training.")
    parser.add_argument("--do_eval", action='store_true',
                        help="Whether to run eval on the dev set.")
    parser.add_argument("--evaluate_during_training", action='store_true',
                        help="Run evaluation during training at each logging step.")
    parser.add_argument("--do_lower_case", action='store_true',
                        help="Set this flag if you are using an uncased model.")
    parser.add_argument("--use_random_weight", action='store_true',
                        help="Use random weights as initialization")


    # Training Schedules
    parser.add_argument("--ratio_increase", default=0.25, type=float,
                        help="Learning schedule, the percentage for the annealing stage.") 
    parser.add_argument("--ratio_zero", default=0.25, type=float,
                        help="Learning schedule, the percentage for the pure auto-encoding stage.")     
    parser.add_argument("--fb_mode", default=0, type=int,
                        help="free bit training mode.")   
    parser.add_argument("--dim_target_kl", default=3.0, type=float,
                        help="dim_target_kl free bit training mode.")                            
    parser.add_argument("--per_gpu_train_batch_size", default=4, type=int,
                        help="Batch size per GPU/CPU for training.")
    parser.add_argument("--per_gpu_eval_batch_size", default=1, type=int,
                        help="Batch size per GPU/CPU for evaluation.")
    parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
                        help="Number of updates steps to accumulate before performing a backward/update pass.")
    parser.add_argument("--learning_rate", default=5e-5, type=float,
                        help="The initial learning rate for Adam.")
    parser.add_argument("--weight_decay", default=0.0, type=float,
                        help="Weight deay if we apply some.")
    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
                        help="Epsilon for Adam optimizer.")
    parser.add_argument("--max_grad_norm", default=1.0, type=float,
                        help="Max gradient norm.")
    parser.add_argument("--num_train_epochs", default=1.0, type=float,
                        help="Total number of training epochs to perform.")
    parser.add_argument("--max_steps", default=-1, type=int,
                        help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
    parser.add_argument("--warmup_steps", default=0, type=int,
                        help="Linear warmup over warmup_steps.")
    parser.add_argument("--use_philly", action='store_true',
                        help="Use Philly for computing.")

    ## IO: Logging and Saving
    parser.add_argument('--logging_steps', type=int, default=50,
                        help="Log every X updates steps.")
    parser.add_argument('--save_steps', type=int, default=50,
                        help="Save checkpoint every X updates steps.")
    parser.add_argument("--eval_all_checkpoints", action='store_true',
                        help="Evaluate all checkpoints starting with the same prefix as model_name_or_path ending and ending with step number")
    parser.add_argument("--no_cuda", action='store_true',
                        help="Avoid using CUDA when available")
    parser.add_argument('--overwrite_output_dir', action='store_true',
                        help="Overwrite the content of the output directory")
    parser.add_argument('--overwrite_cache', action='store_true',
                        help="Overwrite the cached training and evaluation sets")
    parser.add_argument('--seed', type=int, default=42,
                        help="random seed for initialization")
    parser.add_argument('--gloabl_step_eval', type=int, default=661,
                        help="Evaluate the results at the given global step")

    # Precision & Distributed Training 
    parser.add_argument('--fp16', action='store_true',
                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
    parser.add_argument('--fp16_opt_level', type=str, default='O1',
                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
                             "See details at https://nvidia.github.io/apex/amp.html")
    parser.add_argument("--local_rank", type=int, default=-1,
                        help="For distributed training: local_rank")
    parser.add_argument('--server_ip', type=str, default='', help="For distant debugging.")
    parser.add_argument('--server_port', type=str, default='', help="For distant debugging.")

    parser.add_argument('--world-size', default=ompi_size(), type=int, help='number of distributed processes')
    parser.add_argument('--dist-url', default='tcp://' + get_master_ip() + ':23456', type=str,
                        help='url used to set up distributed training')
    parser.add_argument('--dist-backend', default='nccl', type=str, help='distributed backend')
    parser.add_argument('--port', type=str, default='51115', help="Port")

    args = parser.parse_args()

    args.dist_url = 'tcp://' + get_master_ip() + ':' + args.port

    # Setup logging
    logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
                        datefmt='%m/%d/%Y %H:%M:%S',
                        level=logging.INFO)
    logger = logging.getLogger(__name__)

    rank_node = ompi_rank()
    args.distributed = args.world_size > 1
    logger.info("Rank {} distributed: {}".format(rank_node, args.distributed))

    if args.decoder_model_type in ["bert", "roberta"] and not args.mlm:
        raise ValueError("BERT and RoBERTa do not have LM heads but masked LM heads. They must be run using the --mlm "
                         "flag (masked language modeling).")
    if args.eval_data_file is None and args.do_eval:
        raise ValueError("Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file "
                         "or remove the --do_eval argument.")

    if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
        raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))

    # Setup distant debugging if needed
    if args.server_ip and args.server_port:
        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
        import ptvsd
        print("Waiting for debugger attach")
        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
        ptvsd.wait_for_attach()

    if args.distributed:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
        torch.distributed.init_process_group(
            backend=args.dist_backend,
            init_method=args.dist_url,
            world_size=args.world_size,
            rank=ompi_rank(),
            group_name='mtorch')
        logger.info("World Size is {}, Backend is {}, Init Method is {}, rank is {}".format(args.world_size, args.dist_backend, args.dist_url, ompi_rank()))

    gpus = list(gpu_indices())
    args.n_gpu = len(gpus)
    args.local_rank = ompi_rank() #gpus[0]
    torch.cuda.set_device(gpus[0])
    device = torch.device("cuda", gpus[0])

    args.device = device
    logger.info('Rank {}, gpus: {}, get_rank: {}'.format(rank_node, gpus, torch.distributed.get_rank()))
    logger.info(f'Local rank is {args.local_rank}, {rank_node}')

    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s", args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)

    args.ExpName = 'Vae_' + args.dataset + '_Nz_' + str(args.latent_size) + '_Beta_'  + str(args.beta) + '_Dkl_' + str(args.dim_target_kl) + '_Ra_' + str(args.ratio_increase) + '_R0_' + str(args.ratio_zero)
    table_name = 'Vae' + args.dataset + 'Nz' + str(args.latent_size) 
    if ompi_rank() == 0:
        try:
            ts.create_table(table_name)
        except:
            pass


    # Set seed
    set_seed(args)

    # Load pretrained model and tokenizer
    #if args.local_rank not in [-1, 0]: torch.distributed.barrier()  # Barrier to make sure only the first process in distributed training download model & vocab

    ## Encoder 
    encoder_config_class, encoder_model_class, encoder_tokenizer_class = MODEL_CLASSES[args.encoder_model_type]
    encoder_config = encoder_config_class.from_pretrained(args.encoder_config_name if args.encoder_config_name else args.encoder_model_name_or_path)
    tokenizer_encoder = encoder_tokenizer_class.from_pretrained(args.encoder_tokenizer_name if args.encoder_tokenizer_name else args.encoder_model_name_or_path, do_lower_case=args.do_lower_case)
    if args.block_size <= 0:
        args.block_size = tokenizer_encoder.max_len_single_sentence  # Our input block size will be the max possible for the model
    args.block_size = min(args.block_size, tokenizer_encoder.max_len_single_sentence)
    model_encoder = encoder_model_class.from_pretrained(args.encoder_model_name_or_path, from_tf=bool('.ckpt' in args.encoder_model_name_or_path), config=encoder_config, latent_size=args.latent_size)
    # model_encoder.to(args.device)

    ## Decoder 
    decoder_config_class, decoder_model_class, decoder_tokenizer_class = MODEL_CLASSES[args.decoder_model_type]
    decoder_config = decoder_config_class.from_pretrained(args.decoder_config_name if args.decoder_config_name else args.decoder_model_name_or_path)
    tokenizer_decoder = decoder_tokenizer_class.from_pretrained(args.decoder_tokenizer_name if args.decoder_tokenizer_name else args.decoder_model_name_or_path, do_lower_case=args.do_lower_case)
    if args.block_size <= 0:
        args.block_size = tokenizer_decoder.max_len_single_sentence  # Our input block size will be the max possible for the model
    args.block_size = min(args.block_size, tokenizer_decoder.max_len_single_sentence)
    setattr(decoder_config, "latent_size", args.latent_size)
    model_decoder = decoder_model_class.from_pretrained(args.decoder_model_name_or_path, from_tf=bool('.ckpt' in args.decoder_model_name_or_path), config=decoder_config, latent_size=args.latent_size)
    
    # Chunyuan: Add Padding token to GPT2
    special_tokens_dict = {'pad_token': '<PAD>', 'bos_token': '<BOS>', 'eos_token': '<EOS>'}
    num_added_toks = tokenizer_decoder.add_special_tokens(special_tokens_dict)
    print('We have added', num_added_toks, 'tokens to GPT2')
    model_decoder.resize_token_embeddings(len(tokenizer_decoder))  # Notice: resize_token_embeddings expect to receive the full size of the new vocabulary, i.e. the length of the tokenizer.
    assert tokenizer_decoder.pad_token == '<PAD>'

    #model_decoder.to(args.device)

    model_vae = VAE(model_encoder, model_decoder, tokenizer_encoder, tokenizer_decoder, args).to(args.device) #
    #model_vae.cuda()
    if args.use_random_weight:
        model_vae.apply(weights_init_rondom)

    # Distributed training (should be after apex fp16 initialization)
    if args.distributed:
        # model_vae = torch.nn.parallel.DistributedDataParallel(model_vae, device_ids=gpus, output_device=args.local_rank, find_unused_parameters=True)
        model_vae = torch.nn.parallel.DistributedDataParallel(model_vae, device_ids=gpus)
    elif args.n_gpu > 1:
        model_vae = torch.nn.DataParallel(model_vae)#.to(args.device)

    # on_gpu = next(model_vae.parameters()).is_cuda

    #if args.local_rank == 0: torch.distributed.barrier()  # End of barrier to make sure only the first process in distributed training download model & vocab

    logger.info("Training/evaluation parameters %s", args)

    global_step=0
    if args.do_train:
        #if args.local_rank not in [-1, 0]: torch.distributed.barrier()  # Barrier to make sure only the first process in distributed training process the dataset, and the others will use the cache

        train_dataloader = build_dataload_and_cache_examples(args, [tokenizer_encoder, tokenizer_decoder], evaluate=False)

        #if args.local_rank == 0: torch.distributed.barrier()

        global_step, tr_loss, optimizer = train(args, train_dataloader, model_vae, tokenizer_encoder, tokenizer_decoder, table_name)
        logger.info("Rank %d, global_step = %s, average loss = %s", ompi_rank(), global_step, tr_loss)


    # Saving best-practices: if you use save_pretrained for the model and tokenizer, you can reload them using from_pretrained()
    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
        save_checkpoint(model_vae, optimizer, global_step, args)



if __name__ == "__main__":
    main()