Non-deterministic outputs for book chapters (recognize_digits, etc) for a given random seed on GPU

[sidgoyal78]

For the below code (modified slightly from test_recognize_digits), the CPU version with the seed value 100 gives consistent and same results.

The random seed is set for 2 things:

• initial weights: fluid.default_startup_program().random_seed = 100
• shuffle for reader: random.seed(100)

However, when use_cuda = True, the results are non-deterministic for a given seed value. Note that this code uses the mlp version and doesn't employ any conv kernels (thus probably ruling out cuDNN issues). I have faced similar issue with other book chapters (like sentiment analysis) as well.

from __future__ import print_function
import argparse
import paddle.fluid as fluid
import paddle
import sys
import numpy
import unittest
import math
import sys
import os
import random

BATCH_SIZE = 64


def loss_net(hidden, label):
    prediction = fluid.layers.fc(input=hidden, size=10, act='softmax')
    loss = fluid.layers.cross_entropy(input=prediction, label=label)
    avg_loss = fluid.layers.mean(loss)
    acc = fluid.layers.accuracy(input=prediction, label=label)
    return prediction, avg_loss, acc


def mlp(img, label):
    hidden = fluid.layers.fc(input=img, size=200, act='tanh')
    hidden = fluid.layers.fc(input=hidden, size=200, act='tanh')
    return loss_net(hidden, label)


def conv_net(img, label):
    conv_pool_1 = fluid.nets.simple_img_conv_pool(
        input=img,
        filter_size=5,
        num_filters=20,
        pool_size=2,
        pool_stride=2,
        act="relu")
    conv_pool_1 = fluid.layers.batch_norm(conv_pool_1)
    conv_pool_2 = fluid.nets.simple_img_conv_pool(
        input=conv_pool_1,
        filter_size=5,
        num_filters=50,
        pool_size=2,
        pool_stride=2,
        act="relu")
    return loss_net(conv_pool_2, label)


def train(nn_type,
          use_cuda,
          parallel,
          save_dirname=None,
          model_filename=None,
          params_filename=None,
          is_local=True):

    random.seed(100)
    fluid.default_startup_program().random_seed = 100

    if use_cuda and not fluid.core.is_compiled_with_cuda():
        return
    img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32')
    label = fluid.layers.data(name='label', shape=[1], dtype='int64')

    if nn_type == 'mlp':
        net_conf = mlp
    else:
        net_conf = conv_net

    if parallel:
        print("Not supported")
        return
    else:
        prediction, avg_loss, acc = net_conf(img, label)

    test_program = fluid.default_main_program().clone(for_test=True)

    optimizer = fluid.optimizer.Adam(learning_rate=0.001)
    optimize_ops, params_grads = optimizer.minimize(avg_loss)

    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()

    exe = fluid.Executor(place)

    train_reader = paddle.batch(
        paddle.reader.shuffle(
            paddle.dataset.mnist.train(), buf_size=500),
        batch_size=BATCH_SIZE)
    test_reader = paddle.batch(
        paddle.dataset.mnist.test(), batch_size=BATCH_SIZE)
    feeder = fluid.DataFeeder(feed_list=[img, label], place=place)

    def train_loop(main_program):
        exe.run(fluid.default_startup_program())

        PASS_NUM = 100
        for pass_id in range(PASS_NUM):
            for batch_id, data in enumerate(train_reader()):
                # train a mini-batch, fetch nothing
                exe.run(main_program, feed=feeder.feed(data))
                if (batch_id + 1) % 10 == 0:
                    acc_set = []
                    avg_loss_set = []
                    for test_data in test_reader():
                        acc_np, avg_loss_np = exe.run(
                            program=test_program,
                            feed=feeder.feed(test_data),
                            fetch_list=[acc, avg_loss])
                        acc_set.append(float(acc_np))
                        avg_loss_set.append(float(avg_loss_np))
                    # get test acc and loss
                    acc_val = numpy.array(acc_set).mean()
                    avg_loss_val = numpy.array(avg_loss_set).mean()
                    if float(acc_val
                             ) > 0.8:  # Smaller value to increase CI speed
                        if save_dirname is not None:
                            fluid.io.save_inference_model(
                                save_dirname, ["img"], [prediction],
                                exe,
                                model_filename=model_filename,
                                params_filename=params_filename)
                        return
                    else:
                        print((pass_id, batch_id + 1,
                                   float(avg_loss_val), float(acc_val)))
                        if math.isnan(float(avg_loss_val)):
                            sys.exit("got NaN loss, training failed.")
        raise AssertionError("Loss of recognize digits is too large")

    if is_local:
        train_loop(fluid.default_main_program())
    else:
	print("Not supported")


def main(use_cuda, parallel, nn_type, combine):
    save_dirname = None
    model_filename = None
    params_filename = None
    if not use_cuda and not parallel:
        save_dirname = "recognize_digits_" + nn_type + ".inference.model"
        if combine == True:
            model_filename = "__model_combined__"
            params_filename = "__params_combined__"

    train(
        nn_type=nn_type,
        use_cuda=use_cuda,
        parallel=parallel,
        save_dirname=save_dirname,
        model_filename=model_filename,
        params_filename=params_filename)

if __name__ == '__main__':
    # CPU version
    main(False, False, 'mlp', False)

    # GPU
    # main(True, False, 'mlp', False)