main_fp16_optimizer.py

# coding: utf-8
import argparse
import time
import math
import torch
import torch.nn as nn
import data
import model

try:
    from apex.fp16_utils import *
except ImportError:
    raise ImportError("Please install apex from https://www.github.com/nvidia/apex to run this example.")

parser = argparse.ArgumentParser(description='PyTorch Wikitext-2 RNN/LSTM Language Model')
parser.add_argument('--data', type=str, default='./data/wikitext-2',
                    help='location of the data 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=1504,
                    help='size of word embeddings')
parser.add_argument('--nhid', type=int, default=1504,
                    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=24, 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('--fp16', action='store_true',
                    help='Run model in pseudo-fp16 mode (fp16 storage fp32 math).')
parser.add_argument('--static-loss-scale', type=float, default=128.0,
                    help='Static loss scale, positive power of 2 values can improve fp16 convergence.')
parser.add_argument('--dynamic-loss-scale', action='store_true',
                    help='Use dynamic loss scaling.  If supplied, this argument supersedes ' +
                    '--static-loss-scale.')

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")
if args.fp16 and not args.cuda:
    print("WARNING: --fp16 requires --cuda, ignoring --fp16 option")

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

# Ensure that the dictionary length is a multiple of 8,
# so that the decoder's GEMMs will use Tensor Cores.
corpus = data.Corpus(args.data, pad_to_multiple_of=8)

# 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.
    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()
    if args.cuda:
        data = data.cuda()
    return data

eval_batch_size = 10
train_data = batchify(corpus.train, args.batch_size)
val_data = batchify(corpus.valid, eval_batch_size)
test_data = batchify(corpus.test, eval_batch_size)

###############################################################################
# Build the model
###############################################################################

ntokens = len(corpus.dictionary)

if args.fp16 and args.cuda:
    if ntokens%8 != 0:
        print("Warning: the dictionary size (ntokens = {}) should be a multiple of 8 to ensure "
              "Tensor Core use for the decoder's GEMMs.".format(ntokens))
    if args.emsize%8 != 0 or args.nhid%8 != 0 or args.batch_size%8 != 0:
        print("Warning: emsize = {}, nhid = {}, batch_size = {} should all be multiples of 8 "
              "to ensure Tensor Core use for the RNN's GEMMs.".format(
              args.emsize, args.nhid, args.batch_size))

model = model.RNNModel(args.model, ntokens, args.emsize, args.nhid, args.nlayers, args.dropout, args.tied)

if args.cuda and args.fp16:
    model.type(torch.cuda.HalfTensor)
elif args.cuda:
    model.cuda()
criterion = nn.CrossEntropyLoss()

optimizer = torch.optim.SGD(model.parameters(), lr=args.lr)

###############################################################################
# Create the FP16_Optimizer instance
###############################################################################

if args.cuda and args.fp16:
    # If args.dynamic_loss_scale is False, static_loss_scale will be used.
    # If args.dynamic_loss_scale is True, it will take precedence over static_loss_scale.
    optimizer = FP16_Optimizer(optimizer,
                               static_loss_scale = args.static_loss_scale,
                               dynamic_loss_scale = args.dynamic_loss_scale)

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


def repackage_hidden(h):
    """Detaches hidden states from their history."""
    if torch.is_tensor(h):
        return h.detach()
    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 get_batch(source, i):
    seq_len = min(args.bptt, len(source) - 1 - i)
    data = source[i:i+seq_len]
    target = source[i+1:i+1+seq_len].view(-1)
    return data, target


def evaluate(data_source):
    # Turn on evaluation mode which disables dropout.
    model.eval()
    total_loss = 0
    ntokens = len(corpus.dictionary)
    hidden = model.init_hidden(eval_batch_size)
    with torch.no_grad():
        for i in range(0, data_source.size(0) - 1, args.bptt):
            data, targets = get_batch(data_source, i)
            output, hidden = model(data, hidden)
            output_flat = output.view(-1, ntokens)
            #total loss can overflow if accumulated in fp16.
            total_loss += len(data) * criterion(output_flat, targets).data.float()
            hidden = repackage_hidden(hidden)
    return to_python_float(total_loss) / len(data_source)


def train():
    # Turn on training mode which enables dropout.
    model.train()
    total_loss = 0
    start_time = time.time()
    ntokens = len(corpus.dictionary)
    hidden = model.init_hidden(args.batch_size)
    for batch, i in enumerate(range(0, train_data.size(0) - 1, args.bptt)):
        data, targets = get_batch(train_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)

        # Clipping gradients helps prevent the exploding gradient problem in RNNs / LSTMs.
        if args.cuda and args.fp16:
            optimizer.backward(loss)
            optimizer.clip_master_grads(args.clip)
        else:
            loss.backward()
            # `clip_grad_norm` helps prevent the exploding gradient problem in RNNs / LSTMs.
            # apex.fp16_utils.clip_grad_norm selects between "torch.nn.utils.clip_grad_norm" 
            # and "torch.nn.utils.clip_grad_norm_" based on Pytorch version.  
            # It's not FP16-specific, just a small fix to avoid deprecation warnings.
            clip_grad_norm(model.parameters(), args.clip)

        optimizer.step()

        total_loss += loss.data

        if batch % args.log_interval == 0 and batch > 0:
            cur_loss = to_python_float(total_loss) / 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_data) // args.bptt, lr,
                      elapsed * 1000 / args.log_interval, cur_loss, math.exp(min(cur_loss, 20))))
            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.
try:
    for epoch in range(1, args.epochs+1):
        epoch_start_time = time.time()
        train()
        val_loss = evaluate(val_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(min(val_loss, 20))))
        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
            for param_group in optimizer.param_groups:
                param_group['lr'] = lr
except KeyboardInterrupt:
    print('-' * 89)
    print('Exiting from training early')

# Load the best saved model.
with open(args.save, 'rb') as f:
    model = torch.load(f)

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