mnist.py

import argparse
import gzip
import io
import json
import logging
import os
import struct
from collections import namedtuple

import mxnet as mx
import numpy as np


def load_data(path):
    with gzip.open(find_file(path, "labels.gz")) as flbl:
        struct.unpack(">II", flbl.read(8))
        labels = np.fromstring(flbl.read(), dtype=np.int8)
    with gzip.open(find_file(path, "images.gz")) as fimg:
        _, _, rows, cols = struct.unpack(">IIII", fimg.read(16))
        images = np.fromstring(fimg.read(), dtype=np.uint8).reshape(len(labels), rows, cols)
        images = images.reshape(images.shape[0], 1, 28, 28).astype(np.float32) / 255
    return labels, images


def find_file(root_path, file_name):
    for root, dirs, files in os.walk(root_path):
        if file_name in files:
            return os.path.join(root, file_name)


def build_graph():
    data = mx.sym.var("data")
    data = mx.sym.flatten(data=data)
    fc1 = mx.sym.FullyConnected(data=data, num_hidden=128)
    act1 = mx.sym.Activation(data=fc1, act_type="relu")
    fc2 = mx.sym.FullyConnected(data=act1, num_hidden=64)
    act2 = mx.sym.Activation(data=fc2, act_type="relu")
    fc3 = mx.sym.FullyConnected(data=act2, num_hidden=10)
    return mx.sym.SoftmaxOutput(data=fc3, name="softmax")


def get_training_context(num_gpus):
    if num_gpus:
        return [mx.gpu(i) for i in range(num_gpus)]
    else:
        return mx.cpu()


def train(
    batch_size,
    epochs,
    learning_rate,
    num_gpus,
    training_channel,
    testing_channel,
    hosts,
    current_host,
    model_dir,
):
    (train_labels, train_images) = load_data(training_channel)
    (test_labels, test_images) = load_data(testing_channel)

    # Data parallel training - shard the data so each host
    # only trains on a subset of the total data.
    shard_size = len(train_images) // len(hosts)
    for i, host in enumerate(hosts):
        if host == current_host:
            start = shard_size * i
            end = start + shard_size
            break

    train_iter = mx.io.NDArrayIter(
        train_images[start:end], train_labels[start:end], batch_size, shuffle=True
    )
    val_iter = mx.io.NDArrayIter(test_images, test_labels, batch_size)

    logging.getLogger().setLevel(logging.DEBUG)

    kvstore = "local" if len(hosts) == 1 else "dist_sync"

    mlp_model = mx.mod.Module(symbol=build_graph(), context=get_training_context(num_gpus))
    mlp_model.fit(
        train_iter,
        eval_data=val_iter,
        kvstore=kvstore,
        optimizer="sgd",
        optimizer_params={"learning_rate": learning_rate},
        eval_metric="acc",
        batch_end_callback=mx.callback.Speedometer(batch_size, 100),
        num_epoch=epochs,
    )

    if current_host == hosts[0]:
        save(model_dir, mlp_model)


def save(model_dir, model):
    model.symbol.save(os.path.join(model_dir, "model-symbol.json"))
    model.save_params(os.path.join(model_dir, "model-0000.params"))

    signature = [
        {"name": data_desc.name, "shape": [dim for dim in data_desc.shape]}
        for data_desc in model.data_shapes
    ]
    with open(os.path.join(model_dir, "model-shapes.json"), "w") as f:
        json.dump(signature, f)


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

    parser.add_argument("--batch-size", type=int, default=100)
    parser.add_argument("--epochs", type=int, default=10)
    parser.add_argument("--learning-rate", type=float, default=0.1)

    parser.add_argument("--model-dir", type=str, default=os.environ["SM_MODEL_DIR"])
    parser.add_argument("--train", type=str, default=os.environ["SM_CHANNEL_TRAIN"])
    parser.add_argument("--test", type=str, default=os.environ["SM_CHANNEL_TEST"])

    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()


### NOTE: model_fn and transform_fn are used to load the model and serve inference
def model_fn(model_dir):
    import neomx  # noqa: F401

    logging.info("Invoking user-defined model_fn")

    # change context to mx.gpu() when optimizing and deploying with Neo for GPU endpoints
    ctx = mx.cpu()

    Batch = namedtuple("Batch", ["data"])
    sym, arg_params, aux_params = mx.model.load_checkpoint(os.path.join(model_dir, "compiled"), 0)
    mod = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
    exe = mod.bind(
        for_training=False, data_shapes=[("data", (1, 28, 28))], label_shapes=mod._label_shapes
    )
    mod.set_params(arg_params, aux_params, allow_missing=True)
    # run warm-up inference on empty data
    data = mx.nd.empty((1, 28, 28), ctx=ctx)
    mod.forward(Batch([data]))
    return mod


def transform_fn(mod, payload, input_content_type, output_content_type):

    logging.info("Invoking user-defined transform_fn")
    Batch = namedtuple("Batch", ["data"])

    # change context to mx.gpu() when optimizing and deploying with Neo for GPU endpoints
    ctx = mx.cpu()

    if input_content_type != "application/x-npy":
        raise RuntimeError("Input content type must be application/x-npy")

    # pre-processing
    io_bytes_obj = io.BytesIO(payload)
    npy_payload = np.load(io_bytes_obj)
    mx_ndarray = mx.nd.array(npy_payload)
    mx_ndarray = mx_ndarray.expand_dims(axis=0)
    inference_payload = mx_ndarray.as_in_context(ctx)

    # prediction/inference
    mod.forward(Batch([inference_payload]))

    # post-processing
    result = mod.get_outputs()[0].asnumpy()
    result = np.squeeze(result)
    result_exp = np.exp(result - np.max(result))
    result = result_exp / np.sum(result_exp)
    output_json = json.dumps(result.tolist())
    output_content_type = "application/json"
    return output_json, output_content_type


if __name__ == "__main__":
    args = parse_args()
    num_gpus = int(os.environ["SM_NUM_GPUS"])
    train(
        args.batch_size,
        args.epochs,
        args.learning_rate,
        num_gpus,
        args.train,
        args.test,
        args.hosts,
        args.current_host,
        args.model_dir,
    )