sentiment.py

from __future__ import print_function

import argparse
import bisect
import json
import logging
import os
import random
import time
from collections import Counter
from itertools import chain, islice

import mxnet as mx
import numpy as np
from mxnet import gluon, autograd, nd
from mxnet.io import DataIter, DataBatch, DataDesc

logging.basicConfig(level=logging.DEBUG)

# ------------------------------------------------------------ #
# Training methods                                             #
# ------------------------------------------------------------ #


def train(current_host, hosts, num_cpus, num_gpus, training_dir, model_dir,
          batch_size, epochs, learning_rate, log_interval, embedding_size):
    if len(hosts) == 1:
        kvstore = 'device' if num_gpus > 0 else 'local'
    else:
        kvstore = 'dist_device_sync' if num_gpus > 0 else 'dist_sync'

    ctx = mx.gpu() if num_gpus > 0 else mx.cpu()

    checkpoints_dir = '/opt/ml/checkpoints'
    checkpoints_enabled = os.path.exists(checkpoints_dir)

    train_sentences, train_labels, _ = get_dataset(training_dir + '/train')
    val_sentences, val_labels, _ = get_dataset(training_dir + '/test')

    num_classes = len(set(train_labels))
    vocab = create_vocab(train_sentences)
    vocab_size = len(vocab)

    train_sentences = [[vocab.get(token, 1) for token in line if len(line) > 0] for line in train_sentences]
    val_sentences = [[vocab.get(token, 1) for token in line if len(line) > 0] for line in val_sentences]

    # Alternatively to splitting in memory, the data could be pre-split in S3 and use ShardedByS3Key
    # to do parallel training.
    shard_size = len(train_sentences) // len(hosts)
    for i, host in enumerate(hosts):
        if host == current_host:
            start = shard_size * i
            end = start + shard_size
            break

    train_iterator = BucketSentenceIter(train_sentences[start:end], train_labels[start:end], batch_size)
    val_iterator = BucketSentenceIter(val_sentences, val_labels, batch_size)

    # define the network
    net = TextClassifier(vocab_size, embedding_size, num_classes)

    # Collect all parameters from net and its children, then initialize them.
    net.initialize(mx.init.Xavier(magnitude=2.24), ctx=ctx)
    # Trainer is for updating parameters with gradient.
    trainer = gluon.Trainer(net.collect_params(), 'adam',
                            {'learning_rate': learning_rate},
                            kvstore=kvstore)
    metric = mx.metric.Accuracy()
    loss = gluon.loss.SoftmaxCrossEntropyLoss()
    net.hybridize()

    best_acc_score = 0.0
    for epoch in range(epochs):
        # reset data iterator and metric at begining of epoch.
        metric.reset()
        btic = time.time()
        i = 0
        for batch in train_iterator:
            # Copy data to ctx if necessary
            data = batch.data[0].as_in_context(ctx)
            label = batch.label[0].as_in_context(ctx)

            # Start recording computation graph with record() section.
            # Recorded graphs can then be differentiated with backward.
            with autograd.record():
                output = net(data)
                L = loss(output, label)
                L.backward()
            # take a gradient step with batch_size equal to data.shape[0]
            trainer.step(data.shape[0])
            # update metric at last.
            metric.update([label], [output])

            if i % log_interval == 0 and i > 0:
                name, acc = metric.get()
                print('[Epoch %d Batch %d] Training: %s=%f, %f samples/s' %
                      (epoch, i, name, acc, batch_size / (time.time() - btic)))

            btic = time.time()
            i += 1

        name, acc = metric.get()
        print('[Epoch %d] Training: %s=%f' % (epoch, name, acc))

        name, val_acc = test(ctx, net, val_iterator)
        print('[Epoch %d] Validation: %s=%f' % (epoch, name, val_acc))
        if checkpoints_enabled and val_acc > best_acc_score:
            best_acc_score = val_acc
            logging.info('Saving the model, params and optimizer state.')
            net.export(checkpoints_dir + '/%.4f-gluon_sentiment' % (best_acc_score), epoch)
            trainer.save_states(checkpoints_dir + '/%.4f-gluon_sentiment-%d.states' % (best_acc_score, epoch))
        train_iterator.reset()
    return net, vocab


class BucketSentenceIter(DataIter):
    """Simple bucketing iterator for text classification model.

    Args:
        sentences (list[list[int]]): Encoded sentences.
        labels (list[int]): Corresponding labels.
        batch_size (int): Batch size of the data.
        buckets (list[int]): Optional. Size of the data buckets. Automatically generated if None.
        invalid_label (int): Optional. Key for invalid label, e.g. <unk. The default is 0.
        dtype (str): Optional. Data type of the encoding. The default data type is 'float32'.
        data_name (str): Optional. Name of the data. The default name is 'data'.
        label_name (str): Optional. Name of the label. The default name is 'softmax_label'.
        layout (str): Optional. Format of data and label. 'NT' means (batch_size, length)
            and 'TN' means (length, batch_size).
    """
    def __init__(self, sentences, labels, batch_size, buckets=None, invalid_label=0,
                 data_name='data', label_name='softmax_label', dtype='float32',
                 layout='NT'):
        super(BucketSentenceIter, self).__init__()
        if not buckets:
            buckets = [i for i, j in enumerate(np.bincount([len(s) for s in sentences]))
                       if j >= batch_size]
        buckets.sort()

        ndiscard = 0
        self.data = [[] for _ in buckets]
        self.labels = [[] for _ in buckets]
        for i, sent in enumerate(sentences):
            buck = bisect.bisect_left(buckets, len(sent))
            if buck == len(buckets):
                ndiscard += 1
                continue
            buff = np.full((buckets[buck],), invalid_label, dtype=dtype)
            buff[:len(sent)] = sent
            self.data[buck].append(buff)
            self.labels[buck].append(labels[i])

        self.data = [np.asarray(i, dtype=dtype) for i in self.data]
        self.labels = [np.asarray(i, dtype=dtype) for i in self.labels]

        print('WARNING: discarded %d sentences longer than the largest bucket.' % ndiscard)

        self.batch_size = batch_size
        self.buckets = buckets
        self.data_name = data_name
        self.label_name = label_name
        self.dtype = dtype
        self.invalid_label = invalid_label
        self.nddata = []
        self.ndlabel = []
        self.major_axis = layout.find('N')
        self.layout = layout
        self.default_bucket_key = max(buckets)

        if self.major_axis == 0:
            self.provide_data = [DataDesc(
                name=self.data_name, shape=(batch_size, self.default_bucket_key),
                layout=self.layout)]
            self.provide_label = [DataDesc(
                name=self.label_name, shape=(batch_size,),
                layout=self.layout)]
        elif self.major_axis == 1:
            self.provide_data = [DataDesc(
                name=self.data_name, shape=(self.default_bucket_key, batch_size),
                layout=self.layout)]
            self.provide_label = [DataDesc(
                name=self.label_name, shape=(self.default_bucket_key, batch_size),
                layout=self.layout)]
        else:
            raise ValueError('Invalid layout %s: Must by NT (batch major) or TN (time major)')

        self.idx = []
        for i, buck in enumerate(self.data):
            self.idx.extend([(i, j) for j in range(0, len(buck) - batch_size + 1, batch_size)])
        self.curr_idx = 0
        self.reset()

    def reset(self):
        """Resets the iterator to the beginning of the data."""
        self.curr_idx = 0
        random.shuffle(self.idx)
        for i in range(len(self.data)):
            data, labels = self.data[i], self.labels[i]
            p = np.random.permutation(len(data))
            self.data[i], self.labels[i] = data[p], labels[p]

        self.nddata = []
        self.ndlabel = []
        for buck, label_buck in zip(self.data, self.labels):
            self.nddata.append(nd.array(buck, dtype=self.dtype))
            self.ndlabel.append(nd.array(label_buck, dtype=self.dtype))

    def next(self):
        """Returns the next batch of data."""
        if self.curr_idx == len(self.idx):
            raise StopIteration
        i, j = self.idx[self.curr_idx]
        self.curr_idx += 1

        if self.major_axis == 1:
            data = self.nddata[i][j:j+self.batch_size].T
            label = self.ndlabel[i][j:j+self.batch_size].T
        else:
            data = self.nddata[i][j:j+self.batch_size]
            label = self.ndlabel[i][j:j+self.batch_size]

        return DataBatch([data], [label], pad=0,
                         bucket_key=self.buckets[i],
                         provide_data=[DataDesc(
                             name=self.data_name, shape=data.shape,
                             layout=self.layout)],
                         provide_label=[DataDesc(
                             name=self.label_name, shape=label.shape,
                             layout=self.layout)])


class TextClassifier(gluon.HybridBlock):
    def __init__(self, vocab_size, embedding_size, classes, **kwargs):
        super(TextClassifier, self).__init__(**kwargs)
        with self.name_scope():
            self.dense = gluon.nn.Dense(classes)
            self.embedding = gluon.nn.Embedding(input_dim=vocab_size, output_dim=embedding_size)

    def hybrid_forward(self, F, x):
        x = self.embedding(x)
        x = F.mean(x, axis=1)
        x = self.dense(x)
        return x


def get_dataset(filename):
    labels = []
    sentences = []
    max_length = -1
    with open(filename) as f:
        for line in f:
            tokens = line.split()
            label = int(tokens[0])
            words = tokens[1:]
            max_length = max(max_length, len(words))
            labels.append(label)
            sentences.append(words)
    return sentences, labels, max_length


def create_vocab(sentences, min_count=5, num_words=100000):
    BOS_SYMBOL = '<s>'
    EOS_SYMBOL = '</s>'
    UNK_SYMBOL = '<unk>'
    PAD_SYMBOL = '<pad>'
    VOCAB_SYMBOLS = [PAD_SYMBOL, UNK_SYMBOL, BOS_SYMBOL, EOS_SYMBOL]
    raw_vocab = Counter(token for line in sentences for token in line)
    pruned_vocab = sorted(((c, w) for w, c in raw_vocab.items() if c >= min_count), reverse=True)
    vocab = islice((w for c, w in pruned_vocab), num_words)
    word_to_id = {word: idx for idx, word in enumerate(chain(VOCAB_SYMBOLS, vocab))}
    return word_to_id


def vocab_to_json(vocab, path):
    with open(path, 'w') as out:
        json.dump(vocab, out, indent=4, ensure_ascii=True)
        print('Vocabulary saved to "%s"', path)


def vocab_from_json(path):
    with open(path) as inp:
        vocab = json.load(inp)
        print('Vocabulary (%d words) loaded from "%s"', len(vocab), path)
        return vocab


def save(net, model_dir):
    net, vocab = net
    y = net(mx.sym.var('data'))
    y.save('%s/model.json' % model_dir)
    net.collect_params().save('%s/model.params' % model_dir)
    vocab_to_json(vocab, '%s/vocab.json' % model_dir)


def test(ctx, net, val_data):
    val_data.reset()
    metric = mx.metric.Accuracy()
    for batch in val_data:
        data = batch.data[0].as_in_context(ctx)
        label = batch.label[0].as_in_context(ctx)
        output = net(data)
        metric.update([label], [output])
    return metric.get()


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

    # retrieve the hyperparameters we set in notebook (with some defaults)
    parser.add_argument('--batch-size', type=int, default=8)
    parser.add_argument('--epochs', type=int, default=2)
    parser.add_argument('--learning-rate', type=float, default=0.01)
    parser.add_argument('--log-interval', type=int, default=1000)
    parser.add_argument('--embedding-size', type=int, default=50)

    parser.add_argument('--model-dir', type=str, default=os.environ['SM_MODEL_DIR'])
    parser.add_argument('--training_channel', type=str, default=os.environ['SM_CHANNEL_TRAINING'])

    parser.add_argument('--current-host', type=str, default=os.environ['SM_CURRENT_HOST'])
    parser.add_argument('--hosts', type=list, default=json.loads(os.environ['SM_HOSTS']))

    return parser.parse_args()


if __name__ == '__main__':
    args = parse_args()
    num_cpus = int(os.environ['SM_NUM_CPUS'])
    num_gpus = int(os.environ['SM_NUM_GPUS'])

    model = train(args.current_host, args.hosts, num_cpus, num_gpus, args.training_channel, args.model_dir,
                  args.batch_size, args.epochs, args.learning_rate, args.log_interval, args.embedding_size)

    if args.current_host == args.hosts[0]:
        save(model, args.model_dir)


# ------------------------------------------------------------ #
# Hosting methods                                              #
# ------------------------------------------------------------ #

def model_fn(model_dir):
    """Loads the Gluon model. Called once when hosting service starts.

    Args:
        model_dir (str): The directory where model files are stored.

    Returns:
        mxnet.gluon.block.Block: a Gluon network.
    """
    symbol = mx.sym.load('%s/model.json' % model_dir)
    vocab = vocab_from_json('%s/vocab.json' % model_dir)
    outputs = mx.symbol.softmax(data=symbol, name='softmax_label')
    inputs = mx.sym.var('data')
    param_dict = gluon.ParameterDict('model_')
    net = gluon.SymbolBlock(outputs, inputs, param_dict)
    net.load_params('%s/model.params' % model_dir, ctx=mx.cpu())
    return net, vocab


def transform_fn(net, data, input_content_type, output_content_type):
    """Transforms a request using the Gluon model. Called once per request.

    Args:
        net (mxnet.gluon.block.Block): The Gluon model.
        data (obj): The request payload.
        input_content_type (str): The request content type.
        output_content_type (str): The (desired) response content type.

    Returns:
        tuple[obj, str]: The response payload and content type.
    """
    # we can use content types to vary input/output handling, but
    # here we just assume json for both
    net, vocab = net
    parsed = json.loads(data)
    outputs = []
    for row in parsed:
        tokens = [vocab.get(token, 1) for token in row.split()]
        nda = mx.nd.array([tokens])
        output = net(nda)
        prediction = mx.nd.argmax(output, axis=1)
        outputs.append(int(prediction.asscalar()))
    response_body = json.dumps(outputs)
    return response_body, output_content_type