init

README.md

+# Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context
+
+This repository contains the code in both **PyTorch** and **TensorFlow** for our paper
+>[Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context]()
+
+>Zihang Dai\*, Zhilin Yang\*, Yiming Yang, William W. Cohen, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov (*: equal contribution)
+
+>Preprint 2018
+
+#### TensorFlow
+
+- The source code is in the `tf/` folder, supporting (1) single-node multi-gpu training, and (2) multi-host TPU training.
+- Besides the source code, we also provide pretrained "TensorFlow" models with state-of-the-art (SoTA) performances reported in the paper.
+- Please refer to `tf/README.md` for details.
+
+#### PyTorch
+
+- The source code is in the `pytorch/` folder, supporting single-node multi-gpu training via the module `nn.DataPrallel`.
+- Please refer to `pytorch/README.md` for details.
+
+#### Results
+
+Transformer-XL achieves new state-of-the-art results on multipole language modeling benchmarks. Transformer-XL is also the first to break through the 1.0 barrier on char-level language modeling. Below is a summary.
+
+Method | enwiki8 | text8 | One Billion Word | WT-103 | PTB (w/o finetuning)
+-- | -- | -- | -- | -- | -- 
+Previous Best | 1.06 | 1.13 | 23.7 | 20.5 | 55.5
+Transformer-XL | **0.99** | **1.08** | **21.8** | **18.3** | **54.5**

getdata.sh

+echo "=== Acquiring datasets ==="
+echo "---"
+
+mkdir -p data
+cd data
+
+if [[ ! -d 'wikitext-2' ]]; then
+    echo "- Downloading WikiText-2 (WT2)"
+    wget --quiet --continue https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-v1.zip
+    unzip -q wikitext-2-v1.zip
+    cd wikitext-2
+    mv wiki.train.tokens train.txt
+    mv wiki.valid.tokens valid.txt
+    mv wiki.test.tokens test.txt
+    cd ..
+fi
+
+echo "- Downloading WikiText-103 (WT2)"
+if [[ ! -d 'wikitext-103' ]]; then
+    wget --continue https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-103-v1.zip
+    unzip -q wikitext-103-v1.zip
+    cd wikitext-103
+    mv wiki.train.tokens train.txt
+    mv wiki.valid.tokens valid.txt
+    mv wiki.test.tokens test.txt
+    cd ..
+fi
+
+echo "- Downloading enwik8 (Character)"
+if [[ ! -d 'enwik8' ]]; then
+    mkdir -p enwik8
+    cd enwik8
+    wget --continue http://mattmahoney.net/dc/enwik8.zip
+    wget https://raw.githubusercontent.com/salesforce/awd-lstm-lm/master/data/enwik8/prep_enwik8.py
+    python3 prep_enwik8.py
+    cd ..
+fi
+
+echo "- Downloading text8 (Character)"
+if [[ ! -d 'text8' ]]; then
+    mkdir -p text8
+    cd text8
+    wget --continue http://mattmahoney.net/dc/text8.zip
+    python ../../prep_text8.py
+    cd ..
+fi
+
+echo "- Downloading Penn Treebank (PTB)"
+if [[ ! -d 'penn' ]]; then
+    wget --quiet --continue http://www.fit.vutbr.cz/~imikolov/rnnlm/simple-examples.tgz
+    tar -xzf simple-examples.tgz
+
+    mkdir -p penn
+    cd penn
+    mv ../simple-examples/data/ptb.train.txt train.txt
+    mv ../simple-examples/data/ptb.test.txt test.txt
+    mv ../simple-examples/data/ptb.valid.txt valid.txt
+    cd ..
+
+    echo "- Downloading Penn Treebank (Character)"
+    mkdir -p pennchar
+    cd pennchar
+    mv ../simple-examples/data/ptb.char.train.txt train.txt
+    mv ../simple-examples/data/ptb.char.test.txt test.txt
+    mv ../simple-examples/data/ptb.char.valid.txt valid.txt
+    cd ..
+
+    rm -rf simple-examples/
+fi
+
+echo "- Downloading 1B words"
+
+if [[ ! -d 'one-billion-words' ]]; then
+    mkdir -p one-billion-words
+    cd one-billion-words
+
+    wget --no-proxy http://www.statmt.org/lm-benchmark/1-billion-word-language-modeling-benchmark-r13output.tar.gz
+    tar xzvf 1-billion-word-language-modeling-benchmark-r13output.tar.gz
+
+    path="1-billion-word-language-modeling-benchmark-r13output/heldout-monolingual.tokenized.shuffled/"
+    cat ${path}/news.en.heldout-00000-of-00050 > valid.txt
+    cat ${path}/news.en.heldout-00000-of-00050 > test.txt
+
+    wget https://github.com/rafaljozefowicz/lm/raw/master/1b_word_vocab.txt
+
+    cd ..
+fi
+
+echo "---"
+echo "Happy language modeling :)"

prep_text8.py

+#!/usr/bin/env python
+# coding=utf-8
+
+import os
+import sys
+import zipfile
+
+from io import open
+
+if os.path.exists('train.txt'):
+    print('Tokenized text8 already exists - skipping processing')
+    sys.exit()
+
+data = zipfile.ZipFile('text8.zip').extractall()
+data = open('text8', 'r', encoding='utf-8').read()
+
+print('Length of text8: {}'.format(len(data)))
+
+num_test_chars = 5000000
+
+train_data = data[: -2 * num_test_chars]
+valid_data = data[-2 * num_test_chars: -num_test_chars]
+test_data = data[-num_test_chars:]
+
+for fn, part in [('train.txt', train_data), ('valid.txt', valid_data), ('test.txt', test_data)]:
+    print('{} will have {} bytes'.format(fn, len(part)))
+    print('- Tokenizing...')
+    # Change space ' ' to underscore '_'
+    part_str = ' '.join(['_' if c == ' ' else c for c in part.strip()])
+    print('- Writing...')
+    f = open(fn, 'w').write(part_str)
+    f = open(fn + '.raw', 'w', encoding='utf-8').write(part)

pytorch/README.md

+#### Prerequisite
+
+- Pytorch 0.4: `conda install pytorch torchvision -c pytorch`
+
+
+
+#### Data Prepration
+
+`bash getdata.sh`
+
+
+
+#### Replicate the "bpc = 1.06" result on `enwik8` with a 12-layer Transformer-XL
+
+- Make sure the machine have **4 GPUs**, each with **at least 11G memory**
+
+- Training
+
+  `bash run_enwik8.sh train --work_dir PATH_TO_WORK_DIR`
+
+- Testing
+
+  `bash run_enwik8.sh eval --work_dir PATH_TO_WORK_DIR`
+
+
+
+#### Replicate the "PPL = 24.03" result on `wikitext-103` with Transformer-XL
+
+- Make sure the machine have **4 GPUs**, each with **at least 11G memory**
+
+- Training
+
+  `bash run_wt103.sh train --work_dir PATH_TO_WORK_DIR`
+
+- Testing
+
+  `bash run_wt103.sh eval --work_dir PATH_TO_WORK_DIR`
+
+
+
+#### Other options:
+
+- `--batch_chunk`: this option allows one to trade speed for memory. For `batch_chunk > 1`, the program will split each training batch into `batch_chunk` sub-batches and perform forward and backward on each sub-batch sequentially, with the gradient accumulated and divided by `batch_chunk`. Hence, the memory usage will propertionally lower while the computation time will inversely higher. 
+- `--div_val`: when using adaptive softmax and embedding, the embedding dimension is divided by `div_val` from bin $i$ to bin $i+1$. This saves both GPU memory and the parameter budget.
+- `--fp16` and `--dynamic-loss-scale`: Run in pseudo-fp16 mode (fp16 storage fp32 math) with dynamic loss scaling. 
+  - Note: to explore the `--fp16` option, please make sure the `apex` package is installed (https://github.com/NVIDIA/apex/).
+- `--attn_type`: set `attn_type` to 2 to use standard Transformer without any recurrence.
+
+
+
+#### Other datasets:
+
+- `Text8` character-level language modeling: check out `run_text8.sh`
+- `lm1b` word-level language modeling: check out `run_lm1b.sh`

pytorch/data_utils.py

+import os, sys
+import glob
+
+from collections import Counter, OrderedDict
+import numpy as np
+import torch
+
+from utils.vocabulary import Vocab
+
+class LMOrderedIterator(object):
+    def __init__(self, data, bsz, bptt, device='cpu', ext_len=None):
+        """
+            data -- LongTensor -- the LongTensor is strictly ordered
+        """
+        self.bsz = bsz
+        self.bptt = bptt
+        self.ext_len = ext_len if ext_len is not None else 0
+
+        self.device = device
+
+        # Work out how cleanly we can divide the dataset into bsz parts.
+        self.n_step = data.size(0) // bsz
+
+        # Trim off any extra elements that wouldn't cleanly fit (remainders).
+        data = data.narrow(0, 0, self.n_step * bsz)
+
+        # Evenly divide the data across the bsz batches.
+        self.data = data.view(bsz, -1).t().contiguous().to(device)
+
+        # Number of mini-batches
+        self.n_batch = (self.n_step + self.bptt - 1) // self.bptt
+
+    def get_batch(self, i, bptt=None):
+        if bptt is None: bptt = self.bptt
+        seq_len = min(bptt, self.data.size(0) - 1 - i)
+
+        end_idx = i + seq_len
+        beg_idx = max(0, i - self.ext_len)
+
+        data = self.data[beg_idx:end_idx]
+        target = self.data[i+1:i+1+seq_len]
+
+        return data, target, seq_len
+
+    def get_fixlen_iter(self, start=0):
+        for i in range(start, self.data.size(0) - 1, self.bptt):
+            yield self.get_batch(i)
+
+    def get_varlen_iter(self, start=0, std=5, min_len=5, max_deviation=3):
+        max_len = self.bptt + max_deviation * std
+        i = start
+        while True:
+            bptt = self.bptt if np.random.random() < 0.95 else self.bptt / 2.
+            bptt = min(max_len, max(min_len, int(np.random.normal(bptt, std))))
+            data, target, seq_len = self.get_batch(i, bptt)
+            i += seq_len
+            yield data, target, seq_len
+            if i >= self.data.size(0) - 2:
+                break
+
+    def __iter__(self):
+        return self.get_fixlen_iter()
+
+
+class LMShuffledIterator(object):
+    def __init__(self, data, bsz, bptt, device='cpu', ext_len=None, shuffle=False):
+        """
+            data -- list[LongTensor] -- there is no order among the LongTensors
+        """
+        self.data = data
+
+        self.bsz = bsz
+        self.bptt = bptt
+        self.ext_len = ext_len if ext_len is not None else 0
+
+        self.device = device
+        self.shuffle = shuffle
+
+    def get_sent_stream(self):
+        # index iterator
+        epoch_indices = np.random.permutation(len(self.data)) if self.shuffle \
+            else np.array(range(len(self.data)))
+
+        # sentence iterator
+        for idx in epoch_indices:
+            yield self.data[idx]
+
+    def stream_iterator(self, sent_stream):
+        # streams for each data in the batch
+        streams = [None] * self.bsz
+
+        data = torch.LongTensor(self.bptt, self.bsz)
+        target = torch.LongTensor(self.bptt, self.bsz)
+
+        n_retain = 0
+
+        while True:
+            # data   : [n_retain+bptt x bsz]
+            # target : [bptt x bsz]
+            data[n_retain:].fill_(-1)
+            target.fill_(-1)
+
+            valid_batch = True
+
+            for i in range(self.bsz):
+                n_filled = 0
+                try:
+                    while n_filled < self.bptt:
+                        if streams[i] is None or len(streams[i]) <= 1:
+                            streams[i] = next(sent_stream)
+                        # number of new tokens to fill in
+                        n_new = min(len(streams[i]) - 1, self.bptt - n_filled)
+                        # first n_retain tokens are retained from last batch
+                        data[n_retain+n_filled:n_retain+n_filled+n_new, i] = \
+                            streams[i][:n_new]
+                        target[n_filled:n_filled+n_new, i] = \
+                            streams[i][1:n_new+1]
+                        streams[i] = streams[i][n_new:]
+                        n_filled += n_new
+                except StopIteration:
+                    valid_batch = False
+                    break
+
+            if not valid_batch:
+                return
+
+            data = data.to(self.device)
+            target = target.to(self.device)
+
+            yield data, target, self.bptt
+
+            n_retain = min(data.size(0), self.ext_len)
+            if n_retain > 0:
+                data[:n_retain] = data[-n_retain:]
+            data.resize_(n_retain + self.bptt, data.size(1))
+
+    def __iter__(self):
+        # sent_stream is an iterator
+        sent_stream = self.get_sent_stream()
+
+        for batch in self.stream_iterator(sent_stream):
+            yield batch
+
+
+class LMMultiFileIterator(LMShuffledIterator):
+    def __init__(self, paths, vocab, bsz, bptt, device='cpu', ext_len=None,
+        shuffle=False):
+
+        self.paths = paths
+        self.vocab = vocab
+
+        self.bsz = bsz
+        self.bptt = bptt
+        self.ext_len = ext_len if ext_len is not None else 0
+
+        self.device = device
+        self.shuffle = shuffle
+
+    def get_sent_stream(self, path):
+        sents = self.vocab.encode_file(path, add_double_eos=True)
+        if self.shuffle:
+            np.random.shuffle(sents)
+        sent_stream = iter(sents)
+
+        return sent_stream
+
+    def __iter__(self):
+        if self.shuffle:
+            np.random.shuffle(self.paths)
+
+        for path in self.paths:
+            # sent_stream is an iterator
+            sent_stream = self.get_sent_stream(path)
+            for batch in self.stream_iterator(sent_stream):
+                yield batch
+
+
+class Corpus(object):
+    def __init__(self, path, dataset, *args, **kwargs):
+        self.dataset = dataset
+        self.vocab = Vocab(*args, **kwargs)
+
+        if self.dataset in ['ptb', 'wt2', 'enwik8', 'text8']:
+            self.vocab.count_file(os.path.join(path, 'train.txt'))
+            self.vocab.count_file(os.path.join(path, 'valid.txt'))
+            self.vocab.count_file(os.path.join(path, 'test.txt'))
+        elif self.dataset == 'wt103':
+            self.vocab.count_file(os.path.join(path, 'train.txt'))
+        elif self.dataset == 'lm1b':
+            train_path_pattern = os.path.join(
+                path, '1-billion-word-language-modeling-benchmark-r13output',
+                'training-monolingual.tokenized.shuffled', 'news.en-*')
+            train_paths = glob.glob(train_path_pattern)
+            # the vocab will load from file when build_vocab() is called
+
+        self.vocab.build_vocab()
+
+        if self.dataset in ['ptb', 'wt2', 'wt103']:
+            self.train = self.vocab.encode_file(
+                os.path.join(path, 'train.txt'), ordered=True)
+            self.valid = self.vocab.encode_file(
+                os.path.join(path, 'valid.txt'), ordered=True)
+            self.test  = self.vocab.encode_file(
+                os.path.join(path, 'test.txt'), ordered=True)
+        elif self.dataset in ['enwik8', 'text8']:
+            self.train = self.vocab.encode_file(
+                os.path.join(path, 'train.txt'), ordered=True, add_eos=False)
+            self.valid = self.vocab.encode_file(
+                os.path.join(path, 'valid.txt'), ordered=True, add_eos=False)
+            self.test  = self.vocab.encode_file(
+                os.path.join(path, 'test.txt'), ordered=True, add_eos=False)
+        elif self.dataset == 'lm1b':
+            self.train = train_paths
+            self.valid = self.vocab.encode_file(
+                os.path.join(path, 'valid.txt'), ordered=False, add_double_eos=True)
+            self.test  = self.vocab.encode_file(
+                os.path.join(path, 'test.txt'), ordered=False, add_double_eos=True)
+
+    def get_iterator(self, split, *args, **kwargs):
+        if split == 'train':
+            if self.dataset in ['ptb', 'wt2', 'wt103', 'enwik8', 'text8']:
+                data_iter = LMOrderedIterator(self.train, *args, **kwargs)
+            elif self.dataset == 'lm1b':
+                kwargs['shuffle'] = True
+                data_iter = LMMultiFileIterator(self.train, self.vocab, *args, **kwargs)
+        elif split in ['valid', 'test']:
+            data = self.valid if split == 'valid' else self.test
+            if self.dataset in ['ptb', 'wt2', 'wt103', 'enwik8', 'text8']:
+                data_iter = LMOrderedIterator(data, *args, **kwargs)
+            elif self.dataset == 'lm1b':
+                data_iter = LMShuffledIterator(data, *args, **kwargs)
+
+        return data_iter
+
+
+def get_lm_corpus(datadir, dataset):
+    fn = os.path.join(datadir, 'cache.pt')
+    if os.path.exists(fn):
+        print('Loading cached dataset...')
+        corpus = torch.load(fn)
+    else:
+        print('Producing dataset {}...'.format(dataset))
+        kwargs = {}
+        if dataset in ['wt103', 'wt2']:
+            kwargs['special'] = ['<eos>']
+            kwargs['lower_case'] = False
+        elif dataset == 'ptb':
+            kwargs['special'] = ['<eos>']
+            kwargs['lower_case'] = True
+        elif dataset == 'lm1b':
+            kwargs['special'] = []
+            kwargs['lower_case'] = False
+            kwargs['vocab_file'] = os.path.join(datadir, '1b_word_vocab.txt')
+        elif dataset in ['enwik8', 'text8']:
+            pass
+
+        corpus = Corpus(datadir, dataset, **kwargs)
+        torch.save(corpus, fn)
+
+    return corpus
+
+if __name__ == '__main__':
+    import argparse
+    parser = argparse.ArgumentParser(description='unit test')
+    parser.add_argument('--datadir', type=str, default='../data/text8',
+                        help='location of the data corpus')
+    parser.add_argument('--dataset', type=str, default='text8',
+                        choices=['ptb', 'wt2', 'wt103', 'lm1b', 'enwik8', 'text8'],
+                        help='dataset name')
+    args = parser.parse_args()
+
+    corpus = get_lm_corpus(args.datadir, args.dataset)
+    print('Vocab size : {}'.format(len(corpus.vocab.idx2sym)))

pytorch/eval.py

+# coding: utf-8
+import argparse
+import time
+import math
+import os, sys
+
+import torch
+
+from data_utils import get_lm_corpus
+from mem_transformer import MemTransformerLM
+from utils.exp_utils import get_logger
+
+parser = argparse.ArgumentParser(description='PyTorch Transformer Language Model')
+parser.add_argument('--data', type=str, default='../data/wikitext-103',
+                    help='location of the data corpus')
+parser.add_argument('--dataset', type=str, default='wt103',
+                    choices=['wt103', 'lm1b', 'enwik8', 'text8'],
+                    help='dataset name')
+parser.add_argument('--split', type=str, default='all',
+                    choices=['all', 'valid', 'test'],
+                    help='which split to evaluate')
+parser.add_argument('--batch_size', type=int, default=10,
+                    help='batch size')
+parser.add_argument('--tgt_len', type=int, default=5,
+                    help='number of tokens to predict')
+parser.add_argument('--ext_len', type=int, default=0,
+                    help='length of the extended context')
+parser.add_argument('--mem_len', type=int, default=0,
+                    help='length of the retained previous heads')
+parser.add_argument('--clamp_len', type=int, default=-1,
+                    help='max positional embedding index')
+parser.add_argument('--cuda', action='store_true',
+                    help='use CUDA')
+parser.add_argument('--work_dir', type=str, required=True,
+                    help='path to the work_dir')
+parser.add_argument('--no_log', action='store_true',
+                    help='do not log the eval result')
+parser.add_argument('--same_length', action='store_true',
+                    help='set same length attention with masking')
+args = parser.parse_args()
+assert args.ext_len >= 0, 'extended context length must be non-negative'
+
+device = torch.device("cuda" if args.cuda else "cpu")
+
+# Get logger
+logging = get_logger(os.path.join(args.work_dir, 'log.txt'),
+                     log_=not args.no_log)
+
+# Load dataset
+corpus = get_lm_corpus(args.data, args.dataset)
+ntokens = len(corpus.vocab)
+
+va_iter = corpus.get_iterator('valid', args.batch_size, args.tgt_len,
+    device=device, ext_len=args.ext_len)
+te_iter = corpus.get_iterator('test', args.batch_size, args.tgt_len,
+    device=device, ext_len=args.ext_len)
+
+# Load the best saved model.
+with open(os.path.join(args.work_dir, 'model.pt'), 'rb') as f:
+    model = torch.load(f)
+model.backward_compatible()
+model = model.to(device)
+
+logging('Evaluating with bsz {} tgt_len {} ext_len {} mem_len {} clamp_len {}'.format(
+       args.batch_size, args.tgt_len, args.ext_len, args.mem_len, args.clamp_len))
+
+model.reset_length(args.tgt_len, args.ext_len, args.mem_len)
+if args.clamp_len > 0:
+    model.clamp_len = args.clamp_len
+if args.same_length:
+    model.same_length = True
+
+###############################################################################
+# Evaluation code
+###############################################################################
+def evaluate(eval_iter):
+    # Turn on evaluation mode which disables dropout.
+    model.eval()
+    total_len, total_loss = 0, 0.
+    start_time = time.time()
+    with torch.no_grad():
+        mems = tuple()
+        for idx, (data, target, seq_len) in enumerate(eval_iter):
+            ret = model(data, target, *mems)
+            loss, mems = ret[0], ret[1:]
+            loss = loss.mean()
+            total_loss += seq_len * loss.item()
+            total_len += seq_len
+        total_time = time.time() - start_time
+    logging('Time : {:.2f}s, {:.2f}ms/segment'.format(
+            total_time, 1000 * total_time / (idx+1)))
+    return total_loss / total_len
+
+# Run on test data.
+if args.split == 'all':
+    test_loss = evaluate(te_iter)
+    valid_loss = evaluate(va_iter)
+elif args.split == 'valid':
+    valid_loss = evaluate(va_iter)
+    test_loss = None
+elif args.split == 'test':
+    test_loss = evaluate(te_iter)
+    valid_loss = None
+
+def format_log(loss, split):
+    if args.dataset in ['enwik8', 'text8']:
+        log_str = '| {0} loss {1:5.2f} | {0} bpc {2:9.5f} '.format(
+            split, loss, loss / math.log(2))
+    else:
+        log_str = '| {0} loss {1:5.2f} | {0} ppl {2:9.3f} '.format(
+            split, loss, math.exp(loss))
+    return log_str
+
+log_str = ''
+if valid_loss is not None:
+    log_str += format_log(valid_loss, 'valid')
+if test_loss is not None:
+    log_str += format_log(test_loss, 'test')
+
+logging('=' * 100)
+logging(log_str)
+logging('=' * 100)

pytorch/run_enwik8.sh

+#!/bin/bash
+
+if [[ $1 == 'train' ]]; then
+    echo 'Run training...'
+    python train.py \
+        --cuda \
+        --data ../data/enwik8/ \
+        --dataset enwik8 \
+        --n_layer 12 \
+        --d_model 512 \
+        --n_head 8 \
+        --d_head 64 \
+        --d_inner 2048 \
+        --dropout 0.1 \
+        --dropatt 0.0 \
+        --optim adam \
+        --lr 0.00025 \
+        --warmup_step 0 \
+        --max_step 400000 \
+        --tgt_len 512 \
+        --mem_len 512 \
+        --eval_tgt_len 128 \
+        --batch_size 22 \
+        --multi_gpu \
+        --gpu0_bsz 4 \
+        ${@:2}
+elif [[ $1 == 'eval' ]]; then
+    echo 'Run evaluation...'
+    python eval.py \
+        --cuda \
+        --data ../data/enwik8/ \
+        --dataset enwik8 \
+        --tgt_len 80 \
+        --mem_len 2100 \
+        --clamp_len 820 \
+        --same_length \
+        --split test \
+        ${@:2}
+else
+    echo 'unknown argment 1'
+fi

pytorch/run_lm1b.sh

+#!/bin/bash
+
+if [[ $1 == 'train' ]]; then
+    echo 'Run training...'
+    python train.py \
+        --cuda \
+        --data ../data/one-billion-words/ \
+        --dataset lm1b \
+        --adaptive \
+        --n_layer 12 \
+        --d_model 1024 \
+        --div_val 4 \
+        --n_head 8 \
+        --d_head 128 \
+        --d_inner 2048 \
+        --dropout 0.0 \
+        --dropatt 0.0 \
+        --optim adam \
+        --warmup_step 4000 \
+        --max_step 500000 \
+        --lr 0.00025 \
+        --tgt_len 32 \
+        --mem_len 32 \
+        --eval_tgt_len 32 \
+        --batch_size 224 \
+        --multi_gpu \
+        --gpu0_bsz 32 \
+        ${@:2}
+elif [[ $1 == 'eval' ]]; then
+    echo 'Run evaluation...'
+    python eval.py \
+        --cuda \
+        --data ../data/one-billion-words/ \
+        --dataset lm1b \
+        --batch_size 64 \
+        --tgt_len 32 \
+        --mem_len 64 \
+        --split test \
+        --same_length \
+        ${@:2}
+else
+    echo 'unknown argment 1'
+fi

pytorch/run_text8.sh

+#!/bin/bash
+
+if [[ $1 == 'train' ]]; then
+    echo 'Run training...'
+    python train.py \
+        --cuda \
+        --data ../data/text8/ \
+        --dataset text8 \
+        --n_layer 12 \
+        --d_model 512 \
+        --n_head 8 \
+        --d_head 64 \
+        --d_inner 2048 \
+        --dropout 0.2 \
+        --optim adam \
+        --lr 0.00025 \
+        --tgt_len 256 \
+        --mem_len 256 \
+        --eval_tgt_len 256 \
+        --batch_size 16 \
+        --max_step 1000000 \
+        ${@:2}
+elif [[ $1 == 'eval' ]]; then
+    echo 'Run evaluation...'
+    python eval.py \
+        --cuda \
+        --data ../data/text8/ \
+        --dataset text8 \
+        --tgt_len 256 \
+        --same_length \
+        --split test \
+        ${@:2}
+else
+    echo 'unknown argment 1'
+fi

pytorch/run_wt103.sh

+#!/bin/bash
+
+if [[ $1 == 'train' ]]; then
+    echo 'Run training...'
+    python train.py \
+        --cuda \
+        --data ../data/wikitext-103/ \
+        --dataset wt103 \
+        --adaptive \
+        --n_layer 16 \
+        --d_model 410 \
+        --n_head 10 \
+        --d_head 41 \
+        --d_inner 2100 \
+        --dropout 0.1 \
+        --dropatt 0.0 \
+        --optim adam \
+        --lr 0.00025 \
+        --warmup_step 0 \
+        --max_step 200000 \
+        --tgt_len 150 \
+        --mem_len 150 \
+        --eval_tgt_len 150 \
+        --batch_size 60 \
+        --multi_gpu \
+        --gpu0_bsz 4 \
+        ${@:2}
+elif [[ $1 == 'eval' ]]; then
+    echo 'Run evaluation...'
+    python eval.py \
+        --cuda \
+        --data ../data/wikitext-103/ \
+        --dataset wt103 \
+        --tgt_len 64 \
+        --mem_len 640 \
+        --clamp_len 400 \
+        --same_length \
+        --split test \
+        ${@:2}
+else
+    echo 'unknown argment 1'
+fi

pytorch/utils/adaptive_softmax.py

+from collections import defaultdict
+
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class AdaptiveLogSoftmax(nn.Module):
+    def __init__(self, in_features, n_classes, cutoffs, keep_order=False):
+        super(AdaptiveLogSoftmax, self).__init__()
+
+        cutoffs = list(cutoffs)
+
+        if (cutoffs != sorted(cutoffs)) \
+                or (min(cutoffs) <= 0) \
+                or (max(cutoffs) >= (n_classes - 1)) \
+                or (len(set(cutoffs)) != len(cutoffs)) \
+                or any([int(c) != c for c in cutoffs]):
+
+            raise ValueError("cutoffs should be a sequence of unique, positive "
+                             "integers sorted in an increasing order, where "
+                             "each value is between 1 and n_classes-1")
+
+        self.in_features = in_features
+        self.n_classes = n_classes
+        self.cutoffs = cutoffs + [n_classes]
+
+        self.shortlist_size = self.cutoffs[0]
+        self.n_clusters = len(self.cutoffs) - 1
+        self.head_size = self.shortlist_size + self.n_clusters
+
+        self.cluster_weight = nn.Parameter(torch.zeros(self.n_clusters, self.in_features))
+        self.cluster_bias = nn.Parameter(torch.zeros(self.n_clusters))
+
+        self.keep_order = keep_order
+
+
+    def forward(self, hidden, target, weight, bias, keep_order=False):
+        if hidden.size(0) != target.size(0):
+            raise RuntimeError('Input and target should have the same size '
+                               'in the batch dimension.')
+
+        head_weight = torch.cat(
+            [weight[:self.shortlist_size], self.cluster_weight], dim=0)
+        head_bias = torch.cat(
+            [bias[:self.shortlist_size], self.cluster_bias], dim=0)
+
+        head_logit = F.linear(hidden, head_weight, bias=head_bias)
+        head_logprob = F.log_softmax(head_logit, dim=1)
+
+        nll = torch.zeros_like(target,
+                dtype=hidden.dtype, device=hidden.device)
+
+        offset = 0
+        cutoff_values = [0] + self.cutoffs
+        for i in range(len(cutoff_values) - 1):
+            l_idx, h_idx = cutoff_values[i], cutoff_values[i + 1]
+
+            mask_i = (target >= l_idx) & (target < h_idx)
+            indices_i = mask_i.nonzero().squeeze()
+
+            if indices_i.numel() == 0:
+                continue
+
+            target_i = target.index_select(0, indices_i) - l_idx
+            head_logprob_i = head_logprob.index_select(0, indices_i)
+
+            if i == 0:
+                logprob_i = head_logprob_i.gather(1, target_i[:,None]).squeeze(1)
+            else:
+                weight_i = weight[l_idx:h_idx]
+                bias_i = bias[l_idx:h_idx]
+
+                hidden_i = hidden.index_select(0, indices_i)
+
+                tail_logit_i = F.linear(hidden_i, weight_i, bias=bias_i)
+                tail_logprob_i = F.log_softmax(tail_logit_i, dim=1)
+
+                logprob_i = head_logprob_i[:, -i] \
+                          + tail_logprob_i.gather(1, target_i[:,None]).squeeze(1)
+
+            if (hasattr(self, 'keep_order') and self.keep_order) or keep_order:
+                nll.index_copy_(0, indices_i, -logprob_i)
+            else:
+                nll[offset:offset+logprob_i.size(0)].copy_(-logprob_i)
+
+            offset += logprob_i.size(0)
+
+        return nll

pytorch/utils/data_parallel.py

+
+from torch.nn.parallel import DataParallel
+import torch
+from torch.nn.parallel._functions import Scatter
+from torch.nn.parallel.parallel_apply import parallel_apply
+
+def scatter(inputs, target_gpus, chunk_sizes, dim=0):
+    r"""
+    Slices tensors into approximately equal chunks and
+    distributes them across given GPUs. Duplicates
+    references to objects that are not tensors.
+    """
+    def scatter_map(obj):
+        if isinstance(obj, torch.Tensor):
+            try:
+                return Scatter.apply(target_gpus, chunk_sizes, dim, obj)
+            except:
+                print('obj', obj.size())
+                print('dim', dim)
+                print('chunk_sizes', chunk_sizes)
+                quit()
+        if isinstance(obj, tuple) and len(obj) > 0:
+            return list(zip(*map(scatter_map, obj)))
+        if isinstance(obj, list) and len(obj) > 0:
+            return list(map(list, zip(*map(scatter_map, obj))))
+        if isinstance(obj, dict) and len(obj) > 0:
+            return list(map(type(obj), zip(*map(scatter_map, obj.items()))))
+        return [obj for targets in target_gpus]
+
+    # After scatter_map is called, a scatter_map cell will exist. This cell
+    # has a reference to the actual function scatter_map, which has references
+    # to a closure that has a reference to the scatter_map cell (because the
+    # fn is recursive). To avoid this reference cycle, we set the function to
+    # None, clearing the cell
+    try:
+        return scatter_map(inputs)
+    finally:
+        scatter_map = None
+
+def scatter_kwargs(inputs, kwargs, target_gpus, chunk_sizes, dim=0):
+    r"""Scatter with support for kwargs dictionary"""
+    inputs = scatter(inputs, target_gpus, chunk_sizes, dim) if inputs else []
+    kwargs = scatter(kwargs, target_gpus, chunk_sizes, dim) if kwargs else []
+    if len(inputs) < len(kwargs):
+        inputs.extend([() for _ in range(len(kwargs) - len(inputs))])
+    elif len(kwargs) < len(inputs):
+        kwargs.extend([{} for _ in range(len(inputs) - len(kwargs))])
+    inputs = tuple(inputs)
+    kwargs = tuple(kwargs)
+    return inputs, kwargs
+
+class BalancedDataParallel(DataParallel):
+    def __init__(self, gpu0_bsz, *args, **kwargs):
+        self.gpu0_bsz = gpu0_bsz
+        super().__init__(*args, **kwargs)
+
+    def forward(self, *inputs, **kwargs):
+        if not self.device_ids:
+            return self.module(*inputs, **kwargs)
+        if self.gpu0_bsz == 0:
+            device_ids = self.device_ids[1:]
+        else:
+            device_ids = self.device_ids
+        inputs, kwargs = self.scatter(inputs, kwargs, device_ids)
+        if len(self.device_ids) == 1:
+            return self.module(*inputs[0], **kwargs[0])
+        replicas = self.replicate(self.module, self.device_ids)
+        if self.gpu0_bsz == 0:
+            replicas = replicas[1:]
+        outputs = self.parallel_apply(replicas, device_ids, inputs, kwargs)
+        return self.gather(outputs, self.output_device)
+
+    def parallel_apply(self, replicas, device_ids, inputs, kwargs):
+        return parallel_apply(replicas, inputs, kwargs, device_ids)
+
+    def scatter(self, inputs, kwargs, device_ids):
+        bsz = inputs[0].size(self.dim)
+        num_dev = len(self.device_ids)
+        gpu0_bsz = self.gpu0_bsz
+        bsz_unit = (bsz - gpu0_bsz) // (num_dev - 1)
+        if gpu0_bsz < bsz_unit:
+            chunk_sizes = [gpu0_bsz] + [bsz_unit] * (num_dev - 1)
+            delta = bsz - sum(chunk_sizes)
+            for i in range(delta):
+                chunk_sizes[i + 1] += 1
+            if gpu0_bsz == 0:
+                chunk_sizes = chunk_sizes[1:]
+        else:
+            return super().scatter(inputs, kwargs, device_ids)
+        return scatter_kwargs(inputs, kwargs, device_ids, chunk_sizes, dim=self.dim)
+

pytorch/utils/exp_utils.py

+import functools
+import os, shutil
+
+import numpy as np
+
+import torch
+
+
+def logging(s, log_path, print_=True, log_=True):
+    if print_:
+        print(s)
+    if log_:
+        with open(log_path, 'a+') as f_log:
+            f_log.write(s + '\n')
+
+def get_logger(log_path, **kwargs):
+    return functools.partial(logging, log_path=log_path, **kwargs)
+
+def create_exp_dir(dir_path, scripts_to_save=None, debug=False):
+    if debug:
+        print('Debug Mode : no experiment dir created')
+        return functools.partial(logging, log_path=None, log_=False)
+
+    if not os.path.exists(dir_path):
+        os.makedirs(dir_path)
+
+    print('Experiment dir : {}'.format(dir_path))
+    if scripts_to_save is not None:
+        script_path = os.path.join(dir_path, 'scripts')
+        if not os.path.exists(script_path):
+            os.makedirs(script_path)
+        for script in scripts_to_save:
+            dst_file = os.path.join(dir_path, 'scripts', os.path.basename(script))
+            shutil.copyfile(script, dst_file)
+
+    return get_logger(log_path=os.path.join(dir_path, 'log.txt'))
+
+def save_checkpoint(model, optimizer, path, epoch):
+    torch.save(model, os.path.join(path, 'model_{}.pt'.format(epoch)))
+    torch.save(optimizer.state_dict(), os.path.join(path, 'optimizer_{}.pt'.format(epoch)))

pytorch/utils/log_uniform_sampler.py

+import torch
+from torch import nn
+import numpy as np
+
+class LogUniformSampler(object):
+    def __init__(self, range_max, n_sample):
+        """
+        Reference : https://github.com/tensorflow/tensorflow/blob/r1.10/tensorflow/python/ops/candidate_sampling_ops.py
+            `P(class) = (log(class + 2) - log(class + 1)) / log(range_max + 1)`
+
+        expected count can be approximated by 1 - (1 - p)^n
+        and we use a numerically stable version -expm1(num_tries * log1p(-p))
+
+        Our implementation fixes num_tries at 2 * n_sample, and the actual #samples will vary from run to run
+        """
+        with torch.no_grad():
+            self.range_max = range_max
+            log_indices = torch.arange(1., range_max+2., 1.).log_()
+            self.dist = (log_indices[1:] - log_indices[:-1]) / log_indices[-1]
+            # print('P', self.dist.numpy().tolist()[-30:])
+
+            self.log_q = (- (-self.dist.double().log1p_() * 2 * n_sample).expm1_()).log_().float()
+
+        self.n_sample = n_sample
+
+    def sample(self, labels):
+        """
+            labels: [b1, b2]
+        Return
+            true_log_probs: [b1, b2]
+            samp_log_probs: [n_sample]
+            neg_samples: [n_sample]
+        """
+
+        # neg_samples = torch.empty(0).long()
+        n_sample = self.n_sample
+        n_tries = 2 * n_sample
+
+        with torch.no_grad():
+            neg_samples = torch.multinomial(self.dist, n_tries, replacement=True).unique()
+            device = labels.device
+            neg_samples = neg_samples.to(device)
+            true_log_probs = self.log_q[labels].to(device)
+            samp_log_probs = self.log_q[neg_samples].to(device)
+            return true_log_probs, samp_log_probs, neg_samples
+
+def sample_logits(embedding, bias, labels, inputs, sampler):
+    """
+        embedding: an nn.Embedding layer
+        bias: [n_vocab]
+        labels: [b1, b2]
+        inputs: [b1, b2, n_emb]
+        sampler: you may use a LogUniformSampler
+    Return
+        logits: [b1, b2, 1 + n_sample]
+    """
+    true_log_probs, samp_log_probs, neg_samples = sampler.sample(labels)
+    n_sample = neg_samples.size(0)
+    b1, b2 = labels.size(0), labels.size(1)
+    all_ids = torch.cat([labels.view(-1), neg_samples])
+    all_w = embedding(all_ids)
+    true_w = all_w[: -n_sample].view(b1, b2, -1)
+    sample_w = all_w[- n_sample:].view(n_sample, -1)
+
+    all_b = bias[all_ids]
+    true_b = all_b[: -n_sample].view(b1, b2)
+    sample_b = all_b[- n_sample:]
+
+    hit = (labels[:, :, None] == neg_samples).detach()
+
+    true_logits = torch.einsum('ijk,ijk->ij',
+        [true_w, inputs]) + true_b - true_log_probs
+    sample_logits = torch.einsum('lk,ijk->ijl',
+        [sample_w, inputs]) + sample_b - samp_log_probs
+    sample_logits.masked_fill_(hit, -1e30)
+    logits = torch.cat([true_logits[:, :, None], sample_logits], -1)
+
+    return logits
+
+
+# class LogUniformSampler(object):
+#     def __init__(self, range_max, unique=False):
+#         """
+#         Reference : https://github.com/tensorflow/tensorflow/blob/r1.10/tensorflow/python/ops/candidate_sampling_ops.py
+#             `P(class) = (log(class + 2) - log(class + 1)) / log(range_max + 1)`
+#         """
+#         self.range_max = range_max
+#         log_indices = torch.arange(1., range_max+2., 1.).log_()
+#         self.dist = (log_indices[1:] - log_indices[:-1]) / log_indices[-1]
+
+#         self.unique = unique
+
+#         if self.unique:
+#             self.exclude_mask = torch.ByteTensor(range_max).fill_(0)
+
+#     def sample(self, n_sample, labels):
+#         pos_sample, new_labels = labels.unique(return_inverse=True)
+#         n_pos_sample = pos_sample.size(0)
+#         n_neg_sample = n_sample - n_pos_sample
+
+#         if self.unique:
+#             self.exclude_mask.index_fill_(0, pos_sample, 1)
+#             sample_dist = self.dist.clone().masked_fill_(self.exclude_mask, 0)
+#             self.exclude_mask.index_fill_(0, pos_sample, 0)
+#         else:
+#             sample_dist = self.dist
+
+#         neg_sample = torch.multinomial(sample_dist, n_neg_sample)
+
+#         sample = torch.cat([pos_sample, neg_sample])
+#         sample_prob = self.dist[sample]
+
+#         return new_labels, sample, sample_prob
+
+
+if __name__ == '__main__':
+    S, B = 3, 4
+    n_vocab = 10000
+    n_sample = 5
+    H = 32
+
+    labels = torch.LongTensor(S, B).random_(0, n_vocab)
+
+    # sampler = LogUniformSampler(n_vocab, unique=False)
+    # new_labels, sample, sample_prob = sampler.sample(n_sample, labels)
+
+    sampler = LogUniformSampler(n_vocab, unique=True)
+    # true_probs, samp_probs, neg_samples = sampler.sample(n_sample, labels)
+
+    # print('true_probs', true_probs.numpy().tolist())
+    # print('samp_probs', samp_probs.numpy().tolist())
+    # print('neg_samples', neg_samples.numpy().tolist())
+
+    # print('sum', torch.sum(sampler.dist).item())
+
+    # assert torch.all(torch.sort(sample.unique())[0].eq(torch.sort(sample)[0])).item()
+
+    embedding = nn.Embedding(n_vocab, H)
+    bias = torch.zeros(n_vocab)
+    inputs = torch.Tensor(S, B, H).normal_()
+
+    logits, out_labels = sample_logits(embedding, bias, labels, inputs, sampler, n_sample)
+    print('logits', logits.detach().numpy().tolist())
+    print('logits shape', logits.size())
+    print('out_labels', out_labels.detach().numpy().tolist())
+    print('out_labels shape', out_labels.size())
+

pytorch/utils/proj_adaptive_softmax.py

+from collections import defaultdict
+
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+CUDA_MAJOR = int(torch.version.cuda.split('.')[0])
+CUDA_MINOR = int(torch.version.cuda.split('.')[1])
+
+class ProjectedAdaptiveLogSoftmax(nn.Module):
+    def __init__(self, n_token, d_embed, d_proj, cutoffs, div_val=1,
+                 keep_order=False):
+        super(ProjectedAdaptiveLogSoftmax, self).__init__()
+
+        self.n_token = n_token
+        self.d_embed = d_embed
+        self.d_proj = d_proj
+
+        self.cutoffs = cutoffs + [n_token]
+        self.cutoff_ends = [0] + self.cutoffs
+        self.div_val = div_val
+
+        self.shortlist_size = self.cutoffs[0]
+        self.n_clusters = len(self.cutoffs) - 1
+        self.head_size = self.shortlist_size + self.n_clusters
+
+        if self.n_clusters > 0:
+            self.cluster_weight = nn.Parameter(torch.zeros(self.n_clusters, self.d_embed))
+            self.cluster_bias = nn.Parameter(torch.zeros(self.n_clusters))
+
+        self.out_layers = nn.ModuleList()
+        self.out_projs = nn.ParameterList()
+
+        if div_val == 1:
+            for i in range(len(self.cutoffs)):
+                if d_proj != d_embed:
+                    self.out_projs.append(
+                        nn.Parameter(torch.Tensor(d_proj, d_embed))
+                    )
+                else:
+                    self.out_projs.append(None)
+
+            self.out_layers.append(nn.Linear(d_embed, n_token))
+        else:
+            for i in range(len(self.cutoffs)):
+                l_idx, r_idx = self.cutoff_ends[i], self.cutoff_ends[i+1]
+                d_emb_i = d_embed // (div_val ** i)
+
+                self.out_projs.append(
+                    nn.Parameter(torch.Tensor(d_proj, d_emb_i))
+                )
+
+                self.out_layers.append(nn.Linear(d_emb_i, r_idx-l_idx))
+
+        self.keep_order = keep_order
+
+    def _compute_logit(self, hidden, weight, bias, proj):
+        if proj is None:
+            logit = F.linear(hidden, weight, bias=bias)
+        else:
+            # if CUDA_MAJOR <= 9 and CUDA_MINOR <= 1:
+            proj_hid = F.linear(hidden, proj.t().contiguous())
+            logit = F.linear(proj_hid, weight, bias=bias)
+            # else:
+            #     logit = torch.einsum('bd,de,ev->bv', (hidden, proj, weight.t()))
+            #     if bias is not None:
+            #         logit = logit + bias
+
+        return logit
+
+    def forward(self, hidden, target, keep_order=False):
+        '''
+            hidden :: [len*bsz x d_proj]
+            target :: [len*bsz]
+        '''
+
+        if hidden.size(0) != target.size(0):
+            raise RuntimeError('Input and target should have the same size '
+                               'in the batch dimension.')
+
+        if self.n_clusters == 0:
+            logit = self._compute_logit(hidden, self.out_layers[0].weight,
+                                        self.out_layers[0].bias, self.out_projs[0])
+            nll = -F.log_softmax(logit, dim=-1) \
+                    .gather(1, target.unsqueeze(1)).squeeze(1)
+        else:
+            # construct weights and biases
+            weights, biases = [], []
+            for i in range(len(self.cutoffs)):
+                if self.div_val == 1:
+                    l_idx, r_idx = self.cutoff_ends[i], self.cutoff_ends[i + 1]
+                    weight_i = self.out_layers[0].weight[l_idx:r_idx]
+                    bias_i = self.out_layers[0].bias[l_idx:r_idx]
+                else:
+                    weight_i = self.out_layers[i].weight
+                    bias_i = self.out_layers[i].bias
+
+                if i == 0:
+                    weight_i = torch.cat(
+                        [weight_i, self.cluster_weight], dim=0)
+                    bias_i = torch.cat(
+                        [bias_i, self.cluster_bias], dim=0)
+
+                weights.append(weight_i)
+                biases.append(bias_i)
+
+            head_weight, head_bias, head_proj = weights[0], biases[0], self.out_projs[0]
+
+            head_logit = self._compute_logit(hidden, head_weight, head_bias, head_proj)
+            head_logprob = F.log_softmax(head_logit, dim=1)
+
+            nll = torch.zeros_like(target,
+                    dtype=hidden.dtype, device=hidden.device)
+
+            offset = 0
+            cutoff_values = [0] + self.cutoffs
+            for i in range(len(cutoff_values) - 1):
+                l_idx, r_idx = cutoff_values[i], cutoff_values[i + 1]
+
+                mask_i = (target >= l_idx) & (target < r_idx)
+                indices_i = mask_i.nonzero().squeeze()
+
+                if indices_i.numel() == 0:
+                    continue
+
+                target_i = target.index_select(0, indices_i) - l_idx
+                head_logprob_i = head_logprob.index_select(0, indices_i)
+
+                if i == 0:
+                    logprob_i = head_logprob_i.gather(1, target_i[:,None]).squeeze(1)
+                else:
+                    weight_i, bias_i, proj_i = weights[i], biases[i], self.out_projs[i]
+
+                    hidden_i = hidden.index_select(0, indices_i)
+
+                    tail_logit_i = self._compute_logit(hidden_i, weight_i, bias_i, proj_i)
+                    tail_logprob_i = F.log_softmax(tail_logit_i, dim=1)
+
+                    logprob_i = head_logprob_i[:, -i] \
+                              + tail_logprob_i.gather(1, target_i[:,None]).squeeze(1)
+
+                if (hasattr(self, 'keep_order') and self.keep_order) or keep_order:
+                    nll.index_copy_(0, indices_i, -logprob_i)
+                else:
+                    nll[offset:offset+logprob_i.size(0)].copy_(-logprob_i)
+
+                offset += logprob_i.size(0)
+
+        return nll

pytorch/utils/vocabulary.py

+import os
+from collections import Counter, OrderedDict
+
+import torch
+
+class Vocab(object):
+    def __init__(self, special=[], min_freq=0, max_size=None, lower_case=True,
+                 delimiter=None, vocab_file=None):
+        self.counter = Counter()
+        self.special = special
+        self.min_freq = min_freq
+        self.max_size = max_size
+        self.lower_case = lower_case
+        self.delimiter = delimiter
+        self.vocab_file = vocab_file
+
+    def tokenize(self, line, add_eos=False, add_double_eos=False):
+        line = line.strip()
+        # convert to lower case
+        if self.lower_case:
+            line = line.lower()
+
+        # empty delimiter '' will evaluate False
+        if self.delimiter == '':
+            symbols = line
+        else:
+            symbols = line.split(self.delimiter)
+
+        if add_double_eos: # lm1b
+            return ['<S>'] + symbols + ['<S>']
+        elif add_eos:
+            return symbols + ['<eos>']
+        else:
+            return symbols
+
+    def count_file(self, path, verbose=False, add_eos=False):
+        if verbose: print('counting file {} ...'.format(path))
+        assert os.path.exists(path)
+
+        sents = []
+        with open(path, 'r', encoding='utf-8') as f:
+            for idx, line in enumerate(f):
+                if verbose and idx > 0 and idx % 500000 == 0:
+                    print('    line {}'.format(idx))
+                symbols = self.tokenize(line, add_eos=add_eos)
+                self.counter.update(symbols)
+                sents.append(symbols)
+
+        return sents
+
+    def count_sents(self, sents, verbose=False):
+        """
+            sents : a list of sentences, each a list of tokenized symbols
+        """
+        if verbose: print('counting {} sents ...'.format(len(sents)))
+        for idx, symbols in enumerate(sents):
+            if verbose and idx > 0 and idx % 500000 == 0:
+                print('    line {}'.format(idx))
+            self.counter.update(symbols)
+
+    def _build_from_file(self, vocab_file):
+        self.idx2sym = []
+        self.sym2idx = OrderedDict()
+
+        with open(vocab_file, 'r', encoding='utf-8') as f:
+            for line in f:
+                symb = line.strip().split()[0]
+                self.add_symbol(symb)
+        self.unk_idx = self.sym2idx['<UNK>']
+
+    def build_vocab(self):
+        if self.vocab_file:
+            print('building vocab from {}'.format(self.vocab_file))
+            self._build_from_file(self.vocab_file)
+            print('final vocab size {}'.format(len(self)))
+        else:
+            print('building vocab with min_freq={}, max_size={}'.format(
+                self.min_freq, self.max_size))
+            self.idx2sym = []
+            self.sym2idx = OrderedDict()
+
+            for sym in self.special:
+                self.add_special(sym)
+
+            for sym, cnt in self.counter.most_common(self.max_size):
+                if cnt < self.min_freq: break
+                self.add_symbol(sym)
+
+            print('final vocab size {} from {} unique tokens'.format(
+                len(self), len(self.counter)))
+
+    def encode_file(self, path, ordered=False, verbose=False, add_eos=True,
+            add_double_eos=False):
+        if verbose: print('encoding file {} ...'.format(path))
+        assert os.path.exists(path)
+        encoded = []
+        with open(path, 'r', encoding='utf-8') as f:
+            for idx, line in enumerate(f):
+                if verbose and idx > 0 and idx % 500000 == 0:
+                    print('    line {}'.format(idx))
+                symbols = self.tokenize(line, add_eos=add_eos,
+                    add_double_eos=add_double_eos)
+                encoded.append(self.convert_to_tensor(symbols))
+
+        if ordered:
+            encoded = torch.cat(encoded)
+
+        return encoded
+
+    def encode_sents(self, sents, ordered=False, verbose=False):
+        if verbose: print('encoding {} sents ...'.format(len(sents)))
+        encoded = []
+        for idx, symbols in enumerate(sents):
+            if verbose and idx > 0 and idx % 500000 == 0:
+                print('    line {}'.format(idx))
+            encoded.append(self.convert_to_tensor(symbols))
+
+        if ordered:
+            encoded = torch.cat(encoded)
+
+        return encoded
+
+    def add_special(self, sym):
+        if sym not in self.sym2idx:
+            self.idx2sym.append(sym)
+            self.sym2idx[sym] = len(self.idx2sym) - 1
+            setattr(self, '{}_idx'.format(sym.strip('<>')), self.sym2idx[sym])
+
+    def add_symbol(self, sym):
+        if sym not in self.sym2idx:
+            self.idx2sym.append(sym)
+            self.sym2idx[sym] = len(self.idx2sym) - 1
+
+    def get_sym(self, idx):
+        assert 0 <= idx < len(self), 'Index {} out of range'.format(idx)
+        return self.idx2sym[idx]
+
+    def get_idx(self, sym):
+        if sym in self.sym2idx:
+            return self.sym2idx[sym]
+        else:
+            # print('encounter unk {}'.format(sym))
+            assert '<eos>' not in sym
+            assert hasattr(self, 'unk_idx')
+            return self.sym2idx.get(sym, self.unk_idx)
+
+    def get_symbols(self, indices):
+        return [self.get_sym(idx) for idx in indices]
+
+    def get_indices(self, symbols):
+        return [self.get_idx(sym) for sym in symbols]
+
+    def convert_to_tensor(self, symbols):
+        return torch.LongTensor(self.get_indices(symbols))
+
+    def convert_to_sent(self, indices, exclude=None):
+        if exclude is None:
+            return ' '.join([self.get_sym(idx) for idx in indices])
+        else:
+            return ' '.join([self.get_sym(idx) for idx in indices if idx not in exclude])
+
+    def __len__(self):
+        return len(self.idx2sym)

tf/README.md

+## Prerequisite
+
+- Python 2.7
+- Tensorflow 0.1.12
+
+
+
+## Obtain and evaluate pretrained SoTA models
+
+#### 1. Download preprocessed data (vocab) & pretrained models
+
+(a) Set your own `DATA_ROOT` in `sota/download.sh` (default to `./`), which will be the root diretory of downloaded model.
+
+(b) Then, download the model & data by `bash sota/download.sh`. After downloading, the expected directory structure is as follows
+
+```markdown
+pretrained_xl
+  tf_enwik8/
+    data/
+      cache.pkl
+      corpus-info.json
+    model/
+      checkpoint
+      model.ckpt*
+  tf_wt103/
+  	...
+  ...
+```
+
+**Note**: we include preprocessed data in the download files to make sure the **same vocabulary** is used. Please see the code `tf/data_utils.py` to understand the data structure.
+
+
+
+#### 2. Run evaluation scripts to replicate SoTA results on GPUs
+
+- **enwik8**: modify the script `sota/enwik8.sh` accordingly (see below)
+  - set `DATA_ROOT` to the same folder used in the download step (default to `./`)
+  - set `TEST_NUM_CORE ` (number of GPUs to use): we recommend 2 GPUs => about 60 mins
+  - run the script: `bash sota/enwik8.sh`
+
+- **lm1b**: modify the script `sota/lm1b.sh` accordingly  (see below)
+  - set `DATA_ROOT` to the same folder used in the download step (default to `./`)
+  - set `TEST_NUM_CORE ` (number of GPUs to use): we recommend 1 GPUs => less than 5 mins
+  - run the script: `bash sota/lm1b.sh`
+
+- **wt103**:  modify the script `sota/wt103.sh` accordingly  (see below)
+  - set `DATA_ROOT` to the same folder used in the download step (default to `./`)
+  - set `TEST_NUM_CORE ` (number of GPUs to use): we recommend 1 GPUs => less than 5 mins
+  - run the script: `bash sota/wt103.sh`
+
+- **text8**:  modify the script `sota/text8.sh` accordingly  (see below)
+  - set `DATA_ROOT` to the same folder used in the download step (default to `./`)
+  - set `TEST_NUM_CORE ` (number of GPUs to use): we recommend 2 GPUs => about 60 mins
+  - run the script: `bash sota/text8.sh`
+
+
+
+## Train "Transformer-XL" from scratch with GPUs or TPUs
+
+### 1. Download raw data
+
+`bash getdata.sh`
+
+
+
+### 2. Preprocess, training and evaluation
+
+For `dataset` in `[enwik8, lm1b, wt103, text8]`:
+
+- check out `scripts/dataset_gpu.sh` for GPU training and evaluation
+- check out `scripts/dataset_tpu.sh` for TPU training and evaluation
+
+
+
+#### (1) Preprocess raw data and create tfrecords
+
+**NOTE**: The preprocessing for GPU and TPU are different. So, you have to run them separately.
+
+GPU:
+
+- create training and validation data: `bash scripts/dataset_gpu.sh train_data`
+- create test data: `bash scripts/dataset_gpu.sh test_data`
+
+TPU:
+
+- Set the Google storage URL  in `scripts/dataset_tpu.sh`:
+  - `GSDATA`: data URL
+  - `GSEXP`: experiment URL
+- create training and validation data: `bash scripts/dataset_tpu.sh train_data`
+- create test data: `bash scripts/dataset_tpu.sh test_data`
+
+
+
+#### (2) Run training
+
+GPU:
+
+- Modify the configurations in `scripts/dataset_gpu.sh`  according to your needs.
+- `bash scripts/dataset_gpu.sh train`
+
+TPU:
+
+- Modify the configurations in `scripts/dataset_tpu.sh`  according to your needs.
+- `bash scripts/dataset_tpu.sh train`
+
+
+
+#### (3) Run evaluation
+
+GPU:
+
+- `bash scripts/dataset_gpu.sh eval --eval_ckpt_path PATH_TO_CKPT`
+
+TPU:
+
+- `bash scripts/dataset_tpu.sh eval --eval_ckpt_path PATH_TO_CKPT`