multitask.py

import argparse
import time
import math
import torch
import torch.nn as nn
from torch.autograd import Variable
from progress.bar import Bar
import data
import model
import sys
import numpy as np
import random
#import Decimal #TODO: need to install Decimal

sys.stderr.write('Libraries loaded\n')

## Parallelization notes:
##   Does not currently operate across multiple nodes
##   Single GPU is better for default: tied,emsize:200,nhid:200,nlayers:2,dropout:0.2
##
##   Multiple GPUs are better for tied,emsize:1500,nhid:1500,nlayers:2,dropout:0.65
##      4 GPUs train on wikitext-2 in 1/2 - 2/3 the time of 1 GPU

parser = argparse.ArgumentParser(description='PyTorch RNN/LSTM language modeling and CCG tagging multitask model')

parser.add_argument('--lm_data', type=str, default='./data/penn',
                    help='location of the language modeling corpus')
parser.add_argument('--ccg_data', type=str, default='./data/ccg.02-21.common',
                    help='location of the CCG corpus')
parser.add_argument('--model', type=str, default='LSTM',
                    help='type of recurrent net (RNN_TANH, RNN_RELU, LSTM, GRU)')
parser.add_argument('--emsize', type=int, default=200,
                    help='size of word embeddings')
parser.add_argument('--nhid', type=int, default=200,
                    help='number of hidden units per layer')
parser.add_argument('--nlayers', type=int, default=2,
                    help='number of layers')
parser.add_argument('--lr', type=float, default=20,
                    help='initial learning rate')
parser.add_argument('--clip', type=float, default=0.25,
                    help='gradient clipping')
parser.add_argument('--epochs', type=int, default=40,
                    help='upper epoch limit')
parser.add_argument('--batch_size', type=int, default=20, metavar='N',
                    help='batch size')
parser.add_argument('--bptt', type=int, default=35,
                    help='sequence length')
parser.add_argument('--dropout', type=float, default=0.2,
                    help='dropout applied to layers (0 = no dropout)')
parser.add_argument('--tied', action='store_true',
                    help='tie the word embedding and softmax weights')
parser.add_argument('--seed', type=int, default=1111,
                    help='random seed')
parser.add_argument('--cuda', action='store_true',
                    help='use CUDA')
parser.add_argument('--log-interval', type=int, default=200, metavar='N',
                    help='report interval')
parser.add_argument('--save', type=str,  default='model.pt',
                    help='path to save the final model')
parser.add_argument('--single', action='store_true',
                    help='use only a single GPU (even if more are available)')
parser.add_argument('--save_lm_data', type=str, default='lm_data.bin',
                    help='path to save the LM data')
parser.add_argument('--test', action='store_true',
                    help='test a trained LM')
parser.add_argument('--guess', action='store_true',
                    help='display best guesses at each time step')
parser.add_argument('--guessscores', action='store_true',
                    help='display guess scores along with guesses')
parser.add_argument('--guessratios', action='store_true',
                    help='display guess ratios normalized by best guess')
parser.add_argument('--guessprobs', action='store_true',
                    help='display guess probs along with guesses')
parser.add_argument('--guessn', type=int, default=1,
                    help='output top n guesses')
parser.add_argument('--words', action='store_true',
                    help='evaluate word-level complexities (instead of sentence-level loss)')
parser.add_argument('--trainfname', type=str, default='train.txt',
                    help='name of the training file')
parser.add_argument('--validfname', type=str, default='valid.txt',
                    help='name of the validation file')
parser.add_argument('--testfname', type=str, default='test.txt',
                    help='name of the test file')
args = parser.parse_args()

# Set the random seed manually for reproducibility.
torch.manual_seed(args.seed)
if torch.cuda.is_available():
    if not args.cuda:
        print("WARNING: You have a CUDA device, so you should probably run with --cuda")
    else:
        torch.cuda.manual_seed(args.seed)

###############################################################################
# Load data
###############################################################################

# Starting from sequential data, batchify arranges the dataset into columns.
# For instance, with the alphabet as the sequence and batch size 4, we'd get
#  a g m s 
#  b h n t 
#  c i o u 
#  d j p v 
#  e k q w 
#  f l r x 
# These columns are treated as independent by the model, which means that the
# dependence of e. g. 'g' on 'f' can not be learned, but allows more efficient
# batch processing.

def batchify(data, bsz):
    # Work out how cleanly we can divide the dataset into bsz parts.
    if isinstance(data, tuple):
        nbatch = data[0].size(0) // bsz
        # Trim off any extra elements that wouldn't cleanly fit (remainders).
        tag_data = data[1].narrow(0, 0, nbatch * bsz)
        data = data[0].narrow(0, 0, nbatch * bsz)
        # Evenly divide the data across the bsz batches.
        tag_data = tag_data.view(bsz, -1).t().contiguous()
    else:
        nbatch = data.size(0) // bsz
        # Trim off any extra elements that wouldn't cleanly fit (remainders).
        data = data.narrow(0, 0, nbatch * bsz)
    
    # Evenly divide the data across the bsz batches.
    data = data.view(bsz, -1).t().contiguous()
    # Turning the data over to CUDA at this point may lead to more OOM errors
    #if args.cuda:
     #    data = data.cuda()
    if isinstance(data,tuple):
        return data, tag_data
    return data
    
eval_batch_size = 10

corpus = data.SentenceCorpus(args.lm_data, args.ccg_data, args.save_lm_data, args.test,
                             trainfname=args.trainfname,
                             validfname=args.validfname,
                             testfname=args.testfname)

if args.test:
    test_lm_sentences, test_lm_data = corpus.test_lm
    test_ccg_sentences, test_ccg_data = corpus.test_ccg
else:
    train_lm_data = batchify(corpus.train_lm, args.batch_size)
    train_ccg_data = batchify(corpus.train_ccg, args.batch_size)
    val_lm_data = batchify(corpus.valid_lm, eval_batch_size)
    val_ccg_data = batchify(corpus.valid_ccg, eval_batch_size)

###############################################################################
# Build/load the model
###############################################################################

if not args.test:
    ntokens = len(corpus.dictionary)
    model = model.RNNModel(args.model, ntokens, args.emsize, args.nhid, args.nlayers, args.dropout, args.tied)
    if args.cuda:
        if (not args.single) and (torch.cuda.device_count() > 1):
            # Scatters minibatches (in dim=1) across available GPUs
            model = nn.DataParallel(model,dim=1)
        model.cuda()

criterion = nn.CrossEntropyLoss()

###############################################################################
# Complexity measures
###############################################################################

def get_entropy(o):
    ## o should be a vector scoring possible classes
    probs = nn.functional.softmax(o,dim=0)
    logprobs = nn.functional.log_softmax(o,dim=0) #numerically more stable than two separate operations
    return -1 * torch.sum(probs * logprobs)

def get_surps(o):
    ## o should be a vector scoring possible classes
    logprobs = nn.functional.log_softmax(o,dim=0)
    return -1 * logprobs

def get_guesses(o,scores=False):
    ## o should be a vector scoring possible classes
    guessvals, guessixes = torch.topk(o,args.guessn,0)
    # guessvals are the scores of each input cell
    # guessixes are the indices of the max cells
    if scores:
        return guessvals
    else:
        return guessixes

def get_guessscores(o):
    return get_guesses(o,True)

def get_complexity_iter(o,t):
    for corpuspos,targ in enumerate(t):
        word = corpus.dictionary.idx2word[targ]
        surp = get_surps(o[corpuspos])
        H = get_entropy(o[corpuspos])
        print(str(word)+' '+str(surp)+' '+str(H))

def get_complexity_apply(o,t,sentid,tags=False):
    ## Use apply() method
    Hs = torch.squeeze(apply(get_entropy,o))
    surps = apply(get_surps,o)
    
    if args.guess:
        guesses = apply(get_guesses, o)
        guessscores = apply(get_guessscores, o)
    ## Use dimensional indexing method
    ## NOTE: For some reason, this doesn't work.
    ##       May marginally speed things if we can determine why
    ##       Currently 'probs' ends up equivalent to o after the softmax
    #probs = nn.functional.softmax(o,dim=0)
    #logprobs = nn.functional.log_softmax(o,dim=0)
    #Hs = -1 * torch.sum(probs * logprobs),dim=1)
    #surps = -1 * logprobs
    ## Move along
    for corpuspos,targ in enumerate(t):
        if tags:
            word = corpus.dictionary.idx2tag[int(targ)]
        else:
            word = corpus.dictionary.idx2word[int(targ)]
        if word == '<eos>' or word == '<EOS>':
            #don't output the complexity of EOS
            continue
        surp = surps[corpuspos][int(targ)]
        if args.guess:
            outputguesses = []
            for g in range(args.guessn):
                if tags:
                    outputguesses.append(corpus.dictionary.idx2tag[int(guesses[corpuspos][g])])
                else:
                    outputguesses.append(corpus.dictionary.idx2word[int(guesses[corpuspos][g])])
                if args.guessscores:
                    ##output raw scores
                    outputguesses.append("{:.3f}".format(float(guessscores[corpuspos][g])))
                elif args.guessratios:
                    ##output scores (ratio of score(x)/score(best guess)
                    outputguesses.append("{:.3f}".format(float(guessscores[corpuspos][g])/float(guessscores[corpuspos][0])))
                elif args.guessprobs:
                  ##output probabilities ## Currently normalizes probs over N-best list; ideally it'd normalize to probs before getting the N-best
                  outputguesses.append("{:.3f}".format(math.exp(float(nn.functional.log_softmax(guessscores[corpuspos],dim=0)[g]))))
            outputguesses = ' '.join(outputguesses)
            print(str(word)+' '+str(sentid)+' '+str(corpuspos)+' '+str(len(word))+' '+str(float(surp))+' '+str(float(Hs[corpuspos]))+' '+str(outputguesses))
        else:
            print(str(word)+' '+str(sentid)+' '+str(corpuspos)+' '+str(len(word))+' '+str(float(surp))+' '+str(float(Hs[corpuspos])))

def apply(func, M):
    ## applies a function along a given dimension
    tList = [func(m) for m in torch.unbind(M, dim=0) ]
    res = torch.stack(tList, dim=0)
    return res

###############################################################################
# Training code
###############################################################################

def repackage_hidden(h):
    """Wraps hidden states in new Variables, to detach them from their history."""
    if type(h) == Variable:
        return Variable(h.data)
    else:
        return tuple(repackage_hidden(v) for v in h)

# get_batch subdivides the source data into chunks of length args.bptt.
# If source is equal to the example output of the batchify function, with
# a bptt-limit of 2, we'd get the following two Variables for i = 0:
#  a g m s      b h n t 
#  b h n t      c i o u 
# Note that despite the name of the function, the subdivison of data is not
# done along the batch dimension (i.e. dimension 1), since that was handled
# by the batchify function. The chunks are along dimension 0, corresponding
# to the seq_len dimension in the LSTM.
    
def test_get_batch(source, evaluation=False):
    if isinstance(source, tuple):
        seq_len = len(source[0]) - 1
        data = Variable(source[0][:seq_len], volatile=evaluation)
        target = Variable(source[1][:seq_len], volatile=evaluation)
        
    else:
        seq_len = len(source) - 1
        data = Variable(source[:seq_len], volatile=evaluation)
        target = Variable(source[1:1+seq_len].view(-1))
    # This is where data should be CUDA-fied to lessen OOM errors
    if args.cuda:
        return data.cuda(), target.cuda()
    else:
        return data, target
    
def get_batch(source, i, evaluation=False):
    if isinstance(source, tuple):
        seq_len = min(args.bptt, len(source[0]) - 1 - i)
        data = Variable(source[0][i:i+seq_len], volatile=evaluation)
        target = Variable(source[1][i:i+seq_len].view(-1))
    else:
        seq_len = min(args.bptt, len(source) - 1 - i)
        data = Variable(source[i:i+seq_len], volatile=evaluation)
        target = Variable(source[i+1:i+1+seq_len].view(-1))
    #This is where data should be CUDA-fied to lessen OOM errors
    if args.cuda:
        return data.cuda(), target.cuda()
    else:
        return data, target


def test_evaluate(test_lm_sentences, test_ccg_sentences, lm_data_source, ccg_data_source):
    # Turn on evaluation mode which disables dropout.
    model.eval()
    total_loss = 0
    ntokens = len(corpus.dictionary)
    if args.words:
        print('word sentid sentpos wlen surp entropy')#,end='')
        if args.guess:
            for i in range(args.guessn):
                print(' guess'+str(i))#,end='')
                if args.guessscores:
                    print(' gscore'+str(i))#,end='')
        sys.stdout.write('\n')
    bar = Bar('Processing', max=len(lm_data_source)+len(ccg_data_source))
    for i in range(len(lm_data_source)+len(ccg_data_source)):
        if i >= len(lm_data_source):
            sent_ids = ccg_data_source[i-len(lm_data_source)]
            sent = test_ccg_sentences[i-len(lm_data_source)]
        else:
            sent_ids = lm_data_source[i]
            sent = test_lm_sentences[i]
        if args.cuda:
            sent_ids = sent_ids.cuda()
        if (not args.single) and (torch.cuda.device_count() > 1):
            # "module" is necessary when using DataParallel
            hidden = model.module.init_hidden(1) # number of parallel sentences being processed
        else:
            hidden = model.init_hidden(1) # number of parallel sentences being processed
        data, targets = test_get_batch(sent_ids, evaluation=True)
        data=data.unsqueeze(1) # only needed if there is just a single sentence being processed
        print data
        output, hidden = model(data, hidden)
        output_flat = output.view(-1, ntokens)
        curr_loss = criterion(output_flat, targets).data
        #curr_loss = len(data) * criterion(output_flat, targets).data # needed if there is more than a single sentence being processed
        total_loss += curr_loss
        if args.words:
            # output word-level complexity metrics
            if i >= len(lm_data_source):
                get_complexity_apply(output_flat,targets,i-len(lm_data_source),tags=True)
            else:
                get_complexity_apply(output_flat,targets,i)
        else:
            # output sentence-level loss
            print(str(sent)+":"+str(curr_loss[0]))
        hidden = repackage_hidden(hidden)
        bar.next()
    bar.finish()
    return total_loss[0] / (len(lm_data_source)+len(ccg_data_source))

def evaluate(lm_data_source, ccg_data_source):
    # Turn on evaluation mode which disables dropout.
    model.eval()
    total_loss = 0
    ntokens = len(corpus.dictionary)
    if (not args.single) and (torch.cuda.device_count() > 1):
        #"module" is necessary when using DataParallel
        hidden = model.module.init_hidden(eval_batch_size)
    else:
        hidden = model.init_hidden(eval_batch_size)
    for i in range(0, lm_data_source.size(0) + ccg_data_source.size(0) - 1, args.bptt):
        # TAG
        if i > lm_data_source.size(0):
            data, targets = get_batch(ccg_data_source, i - lm_data_source.size(0), evaluation=True)
        # LM
        else:
            data, targets = get_batch(lm_data_source, i, evaluation=True)
        output, hidden = model(data, hidden)
        output_flat = output.view(-1, ntokens)
        curr_loss = len(data) * criterion(output_flat, targets).data
        total_loss += curr_loss
        hidden = repackage_hidden(hidden)
    return total_loss[0] / (len(lm_data_source)+len(ccg_data_source))


def train():
    # Turn on training mode which enables dropout.
    model.train()
    total_loss = 0
    start_time = time.time()
    ntokens = len(corpus.dictionary)
    if (not args.single) and (torch.cuda.device_count() > 1):
        # "module" is necessary when using DataParallel
        hidden = model.module.init_hidden(args.batch_size)
    else:
        hidden = model.init_hidden(args.batch_size)
    # UNCOMMENT FOR DEBUGGING
    #random.seed(10)
    order = list(enumerate(range(0, train_lm_data.size(0) + train_ccg_data.size(0) - 1, args.bptt)))
    random.shuffle(order)
    for batch, i in order:#enumerate(range(0, train_lm_data.size(0) + train_ccg_data.size(0) - 1, args.bptt)):
        # TAG
        if i > train_lm_data.size(0):
            data, targets = get_batch(train_ccg_data, i - train_lm_data.size(0))
        # LM
        else:
            data, targets = get_batch(train_lm_data, i)
            
        # Starting each batch, we detach the hidden state from how it was previously produced.
        # If we didn't, the model would try backpropagating all the way to start of the dataset.
        hidden = repackage_hidden(hidden)
        model.zero_grad()
        output, hidden = model(data, hidden)
        loss = criterion(output.view(-1, ntokens), targets)
        loss.backward()

        # `clip_grad_norm` helps prevent the exploding gradient problem in RNNs / LSTMs.
        torch.nn.utils.clip_grad_norm(model.parameters(), args.clip)
        for p in model.parameters():
            p.data.add_(-lr, p.grad.data)

        total_loss += loss.data

        if batch % args.log_interval == 0 and batch > 0:
            cur_loss = total_loss[0] / args.log_interval
            elapsed = time.time() - start_time
            print('| epoch {:3d} | {:5d}/{:5d} batches | lr {:02.2f} | ms/batch {:5.2f} | '
                    'loss {:5.2f} | ppl {:8.2f}'.format(
                epoch, batch, len(train_lm_data)+len(train_ccg_data) // args.bptt, lr,
                elapsed * 1000 / args.log_interval, cur_loss, math.exp(cur_loss)))
            total_loss = 0
            start_time = time.time()

# Loop over epochs.
lr = args.lr
best_val_loss = None

# At any point you can hit Ctrl + C to break out of training early.
if not args.test:
    try:
        for epoch in range(1, args.epochs+1):
            epoch_start_time = time.time()
            train()
            val_loss = evaluate(val_lm_data, val_ccg_data)
            print('-' * 89)
            print('| end of epoch {:3d} | time: {:5.2f}s | valid loss {:5.2f} | '
                  'valid ppl {:8.2f}'.format(epoch, (time.time() - epoch_start_time),
                                             val_loss, math.exp(val_loss)))
            print('-' * 89)
            # Save the model if the validation loss is the best we've seen so far.
            if not best_val_loss or val_loss < best_val_loss:
                with open(args.save, 'wb') as f:
                    torch.save(model, f)
                    best_val_loss = val_loss
            else:
                # Anneal the learning rate if no improvement has been seen in the validation dataset.
                lr /= 4.0
    except KeyboardInterrupt:
        print('-' * 89)
        print('Exiting from training early')
else:
    # Load the best saved model.
    with open(args.save, 'rb') as f:
        model = torch.load(f)


    # Run on test data.
    test_loss = test_evaluate(test_lm_data, test_ccg_data)
    print('=' * 89)
    print('| End of training | test loss {:5.2f} | test ppl {:8.2f}'.format(
        test_loss, math.exp(test_loss)))
    print('=' * 89)