mnist_classifier_multi_ipus.py

# Copyright 2018 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""An MNIST example with single-program multiple-data (SPMD) data parallelism.

The aim here is to illustrate how to use JAX's `pmap` to express and execute
SPMD programs for data parallelism along a batch dimension, while also
minimizing dependencies by avoiding the use of higher-level layers and
optimizers libraries.
"""
from jax.config import config

# To be used for testing `pmap` with multiple CPU devices
# os.environ['XLA_FLAGS'] = '--xla_force_host_platform_device_count=4'
# Initialize parameters on CPU by default.
config.FLAGS.jax_platforms = "cpu,ipu"

from functools import partial
import time

import numpy as np
import numpy.random as npr

import jax
from jax import jit, grad, pmap
from jax.scipy.special import logsumexp
from jax.tree_util import tree_map
from jax import lax
import jax.numpy as jnp
import datasets


def init_random_params(scale, layer_sizes, rng=npr.RandomState(0)):
  return [(scale * rng.randn(m, n), scale * rng.randn(n))
          for m, n, in zip(layer_sizes[:-1], layer_sizes[1:])]

def predict(params, inputs):
  activations = inputs
  for w, b in params[:-1]:
    outputs = jnp.dot(activations, w) + b
    activations = jnp.tanh(outputs)

  final_w, final_b = params[-1]
  logits = jnp.dot(activations, final_w) + final_b
  return logits - logsumexp(logits, axis=1, keepdims=True)

def loss(params, batch):
  inputs, targets = batch
  preds = predict(params, inputs)
  return -jnp.mean(jnp.sum(preds * targets, axis=1))

@jit
def accuracy(params, batch):
  inputs, targets = batch
  target_class = jnp.argmax(targets, axis=1)
  predicted_class = jnp.argmax(predict(params, inputs), axis=1)
  return jnp.mean(predicted_class == target_class)


if __name__ == "__main__":
  # For this manual SPMD example, we get the number of devices (e.g. GPUs, IPUs or
  # TPU cores) that we're using, and use it to reshape data minibatches.
  num_devices = jax.device_count("ipu")

  batch_multiplier = num_devices
  layer_sizes = [784, 1024, 1024, 10]
  param_scale = 0.1
  step_size = 0.01 * batch_multiplier
  batch_size = 128 * batch_multiplier
  num_epochs = 10

  train_images, train_labels, test_images, test_labels = datasets.mnist()
  num_train = train_images.shape[0]
  num_complete_batches, leftover = divmod(num_train, batch_size)
  num_batches = num_complete_batches + bool(leftover)

  def data_stream():
    rng = npr.RandomState(0)
    while True:
      perm = rng.permutation(num_train)
      for i in range(num_batches):
        batch_idx = perm[i * batch_size:(i + 1) * batch_size]
        images, labels = train_images[batch_idx], train_labels[batch_idx]
        # For this SPMD example, we reshape the data batch dimension into two
        # batch dimensions, one of which is mapped over parallel devices.
        batch_size_per_device, ragged = divmod(images.shape[0], num_devices)
        if ragged:
          msg = "batch size must be divisible by device count, got {} and {}."
          raise ValueError(msg.format(batch_size, num_devices))

        # IPU: skip last incomplete batch with different static shape.
        if batch_size_per_device == batch_size // num_devices:
          shape_prefix = (num_devices, batch_size_per_device)
          images = images.reshape(shape_prefix + images.shape[1:])
          labels = labels.reshape(shape_prefix + labels.shape[1:])
          yield images, labels
  batches = data_stream()

  # IPU: specify backend + donate_argnums for model
  @partial(pmap, axis_name='batch', backend="ipu", donate_argnums=(0,))
  def spmd_update(params, batch):
    grads = grad(loss)(params, batch)
    # We compute the total gradients, summing across the device-mapped axis,
    # using the `lax.psum` SPMD primitive, which does a fast all-reduce-sum.
    grads = [(lax.psum(dw, 'batch'), lax.psum(db, 'batch')) for dw, db in grads]
    return [(w - step_size * dw, b - step_size * db)
            for (w, b), (dw, db) in zip(params, grads)]

  # We replicate the parameters so that the constituent arrays have a leading
  # dimension of size equal to the number of devices we're pmapping over.
  init_params = init_random_params(param_scale, layer_sizes)
  replicate_array = lambda x: np.broadcast_to(x, (num_devices,) + x.shape)
  replicated_params = tree_map(replicate_array, init_params)

  print("Number of IPU visible devices:", len(jax.devices("ipu")))

  for epoch in range(num_epochs):
    start_time = time.time()
    for _ in range(num_batches):
      replicated_params = spmd_update(replicated_params, next(batches))

    # Block to get accurate timing of the epoch!
    tree_map(lambda x: x.block_until_ready(), replicated_params)
    epoch_time = time.time() - start_time

    # We evaluate using the jitted `accuracy` function (not using pmap) by
    # grabbing just one of the replicated parameter values.
    params = tree_map(lambda x: jax.device_get(x[0]), replicated_params)
    train_acc = accuracy(params, (train_images, train_labels))
    test_acc = accuracy(params, (test_images, test_labels))
    print(f"Epoch {epoch} in {epoch_time:0.2f} sec")
    print(f"Training set accuracy {train_acc}")
    print(f"Test set accuracy {test_acc}")