1# Copyright 2020 Huawei Technologies Co., Ltd 2# 3# Licensed under the Apache License, Version 2.0 (the "License"); 4# you may not use this file except in compliance with the License. 5# You may obtain a copy of the License at 6# 7# http://www.apache.org/licenses/LICENSE-2.0 8# 9# Unless required by applicable law or agreed to in writing, software 10# distributed under the License is distributed on an "AS IS" BASIS, 11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12# See the License for the specific language governing permissions and 13# limitations under the License. 14# ============================================================================ 15import os 16import random 17import argparse 18import numpy as np 19from resnet import resnet50 20 21import mindspore.common.dtype as mstype 22import mindspore.context as context 23import mindspore.dataset as de 24import mindspore.dataset.transforms.c_transforms as C 25import mindspore.dataset.vision.c_transforms as vision 26import mindspore.nn as nn 27from mindspore import Tensor 28from mindspore.communication.management import init 29from mindspore.nn.optim.momentum import Momentum 30from mindspore.ops import functional as F 31from mindspore.ops import operations as P 32from mindspore.train.callback import ModelCheckpoint, CheckpointConfig, LossMonitor 33from mindspore.train.model import Model 34from mindspore.context import ParallelMode 35 36random.seed(1) 37np.random.seed(1) 38de.config.set_seed(1) 39 40parser = argparse.ArgumentParser(description='Image classification') 41parser.add_argument('--run_distribute', type=bool, default=False, help='Run distribute') 42parser.add_argument('--device_num', type=int, default=1, help='Device num.') 43parser.add_argument('--do_train', type=bool, default=True, help='Do train or not.') 44parser.add_argument('--do_eval', type=bool, default=False, help='Do eval or not.') 45parser.add_argument('--epoch_size', type=int, default=1, help='Epoch size.') 46parser.add_argument('--batch_size', type=int, default=4, help='Batch size.') 47parser.add_argument('--num_classes', type=int, default=10, help='Num classes.') 48parser.add_argument('--checkpoint_path', type=str, default=None, help='Checkpoint file path') 49parser.add_argument('--dataset_path', type=str, default="/var/log/npu/datasets/cifar", help='Dataset path') 50args_opt = parser.parse_args() 51 52device_id = int(os.getenv('DEVICE_ID')) 53 54data_home = args_opt.dataset_path 55 56context.set_context(mode=context.GRAPH_MODE, device_target="Ascend") 57context.set_context(device_id=device_id) 58 59 60def create_dataset(repeat_num=1, training=True): 61 data_dir = data_home + "/cifar-10-batches-bin" 62 if not training: 63 data_dir = data_home + "/cifar-10-verify-bin" 64 ds = de.Cifar10Dataset(data_dir) 65 66 if args_opt.run_distribute: 67 rank_id = int(os.getenv('RANK_ID')) 68 rank_size = int(os.getenv('RANK_SIZE')) 69 ds = de.Cifar10Dataset(data_dir, num_shards=rank_size, shard_id=rank_id) 70 71 resize_height = 224 72 resize_width = 224 73 rescale = 1.0 / 255.0 74 shift = 0.0 75 76 # define map operations 77 random_crop_op = vision.RandomCrop((32, 32), (4, 4, 4, 4)) # padding_mode default CONSTANT 78 random_horizontal_op = vision.RandomHorizontalFlip() 79 resize_op = vision.Resize((resize_height, resize_width)) # interpolation default BILINEAR 80 rescale_op = vision.Rescale(rescale, shift) 81 normalize_op = vision.Normalize((0.4465, 0.4822, 0.4914), (0.2010, 0.1994, 0.2023)) 82 changeswap_op = vision.HWC2CHW() 83 type_cast_op = C.TypeCast(mstype.int32) 84 85 c_trans = [] 86 if training: 87 c_trans = [random_crop_op, random_horizontal_op] 88 c_trans += [resize_op, rescale_op, normalize_op, 89 changeswap_op] 90 91 # apply map operations on images 92 ds = ds.map(operations=type_cast_op, input_columns="label") 93 ds = ds.map(operations=c_trans, input_columns="image") 94 95 # apply repeat operations 96 ds = ds.repeat(repeat_num) 97 98 # apply shuffle operations 99 ds = ds.shuffle(buffer_size=10) 100 101 # apply batch operations 102 ds = ds.batch(batch_size=args_opt.batch_size, drop_remainder=True) 103 104 return ds 105 106 107class CrossEntropyLoss(nn.Cell): 108 def __init__(self): 109 super(CrossEntropyLoss, self).__init__() 110 self.cross_entropy = P.SoftmaxCrossEntropyWithLogits() 111 self.mean = P.ReduceMean() 112 self.one_hot = P.OneHot() 113 self.one = Tensor(1.0, mstype.float32) 114 self.zero = Tensor(0.0, mstype.float32) 115 116 def construct(self, logits, label): 117 label = self.one_hot(label, F.shape(logits)[1], self.one, self.zero) 118 loss_func = self.cross_entropy(logits, label)[0] 119 loss_func = self.mean(loss_func, (-1,)) 120 return loss_func 121 122 123if __name__ == '__main__': 124 if not args_opt.do_eval and args_opt.run_distribute: 125 context.set_auto_parallel_context(device_num=args_opt.device_num, parallel_mode=ParallelMode.DATA_PARALLEL) 126 context.set_auto_parallel_context(all_reduce_fusion_split_indices=[140]) 127 init() 128 129 context.set_context(mode=context.GRAPH_MODE) 130 epoch_size = args_opt.epoch_size 131 net = resnet50(args_opt.batch_size, args_opt.num_classes) 132 loss = CrossEntropyLoss() 133 opt = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), 0.01, 0.9) 134 135 model = Model(net, loss_fn=loss, optimizer=opt, metrics={'acc'}) 136 137 if args_opt.do_train: 138 dataset = create_dataset(epoch_size) 139 batch_num = dataset.get_dataset_size() 140 config_ck = CheckpointConfig(save_checkpoint_steps=batch_num * 5, keep_checkpoint_max=10) 141 ckpoint_cb = ModelCheckpoint(prefix="train_resnet_cifar10", directory="./", config=config_ck) 142 loss_cb = LossMonitor() 143 model.train(epoch_size, dataset, callbacks=[ckpoint_cb, loss_cb]) 144 145 if args_opt.do_eval: 146 eval_dataset = create_dataset(1, training=False) 147 res = model.eval(eval_dataset) 148 print("result: ", res) 149 checker = os.path.exists("./normal_memreuse.ir") 150 assert checker, True 151