|
| 1 | +# Copyright The PyTorch Lightning team. |
| 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 | +from abc import ABC |
| 15 | +from argparse import ArgumentParser |
| 16 | +from random import shuffle |
| 17 | +from warnings import warn |
| 18 | + |
| 19 | +import numpy as np |
| 20 | +import torch |
| 21 | +from torch.nn import functional as F |
| 22 | +from torch.utils.data import random_split |
| 23 | + |
| 24 | +import pytorch_lightning as pl |
| 25 | + |
| 26 | +try: |
| 27 | + from torchvision.datasets.mnist import MNIST |
| 28 | + from torchvision import transforms |
| 29 | +except Exception: |
| 30 | + from tests.base.datasets import MNIST |
| 31 | + |
| 32 | +try: |
| 33 | + import nvidia.dali.ops as ops |
| 34 | + import nvidia.dali.types as types |
| 35 | + from nvidia.dali.pipeline import Pipeline |
| 36 | + from nvidia.dali.plugin.pytorch import DALIClassificationIterator |
| 37 | +except (ImportError, ModuleNotFoundError): |
| 38 | + warn('NVIDIA DALI is not available') |
| 39 | + ops, types, Pipeline, DALIClassificationIterator = ..., ..., ABC, ABC |
| 40 | + |
| 41 | + |
| 42 | +class ExternalMNISTInputIterator(object): |
| 43 | + """ |
| 44 | + This iterator class wraps torchvision's MNIST dataset and returns the images and labels in batches |
| 45 | + """ |
| 46 | + |
| 47 | + def __init__(self, mnist_ds, batch_size): |
| 48 | + self.batch_size = batch_size |
| 49 | + self.mnist_ds = mnist_ds |
| 50 | + self.indices = list(range(len(self.mnist_ds))) |
| 51 | + shuffle(self.indices) |
| 52 | + |
| 53 | + def __iter__(self): |
| 54 | + self.i = 0 |
| 55 | + self.n = len(self.mnist_ds) |
| 56 | + return self |
| 57 | + |
| 58 | + def __next__(self): |
| 59 | + batch = [] |
| 60 | + labels = [] |
| 61 | + for _ in range(self.batch_size): |
| 62 | + index = self.indices[self.i] |
| 63 | + img, label = self.mnist_ds[index] |
| 64 | + batch.append(img.numpy()) |
| 65 | + labels.append(np.array([label], dtype=np.uint8)) |
| 66 | + self.i = (self.i + 1) % self.n |
| 67 | + return (batch, labels) |
| 68 | + |
| 69 | + |
| 70 | +class ExternalSourcePipeline(Pipeline): |
| 71 | + """ |
| 72 | + This DALI pipeline class just contains the MNIST iterator |
| 73 | + """ |
| 74 | + |
| 75 | + def __init__(self, batch_size, eii, num_threads, device_id): |
| 76 | + super(ExternalSourcePipeline, self).__init__(batch_size, num_threads, device_id, seed=12) |
| 77 | + self.source = ops.ExternalSource(source=eii, num_outputs=2) |
| 78 | + self.build() |
| 79 | + |
| 80 | + def define_graph(self): |
| 81 | + images, labels = self.source() |
| 82 | + return images, labels |
| 83 | + |
| 84 | + |
| 85 | +class DALIClassificationLoader(DALIClassificationIterator): |
| 86 | + """ |
| 87 | + This class extends DALI's original DALIClassificationIterator with the __len__() function so that we can call len() on it |
| 88 | + """ |
| 89 | + |
| 90 | + def __init__( |
| 91 | + self, |
| 92 | + pipelines, |
| 93 | + size=-1, |
| 94 | + reader_name=None, |
| 95 | + auto_reset=False, |
| 96 | + fill_last_batch=True, |
| 97 | + dynamic_shape=False, |
| 98 | + last_batch_padded=False, |
| 99 | + ): |
| 100 | + super().__init__(pipelines, size, reader_name, auto_reset, fill_last_batch, dynamic_shape, last_batch_padded) |
| 101 | + |
| 102 | + def __len__(self): |
| 103 | + batch_count = self._size // (self._num_gpus * self.batch_size) |
| 104 | + last_batch = 1 if self._fill_last_batch else 0 |
| 105 | + return batch_count + last_batch |
| 106 | + |
| 107 | + |
| 108 | +class LitClassifier(pl.LightningModule): |
| 109 | + def __init__(self, hidden_dim=128, learning_rate=1e-3): |
| 110 | + super().__init__() |
| 111 | + self.save_hyperparameters() |
| 112 | + |
| 113 | + self.l1 = torch.nn.Linear(28 * 28, self.hparams.hidden_dim) |
| 114 | + self.l2 = torch.nn.Linear(self.hparams.hidden_dim, 10) |
| 115 | + |
| 116 | + def forward(self, x): |
| 117 | + x = x.view(x.size(0), -1) |
| 118 | + x = torch.relu(self.l1(x)) |
| 119 | + x = torch.relu(self.l2(x)) |
| 120 | + return x |
| 121 | + |
| 122 | + def split_batch(self, batch): |
| 123 | + return batch[0]["data"], batch[0]["label"].squeeze().long() |
| 124 | + |
| 125 | + def training_step(self, batch, batch_idx): |
| 126 | + x, y = self.split_batch(batch) |
| 127 | + y_hat = self(x) |
| 128 | + loss = F.cross_entropy(y_hat, y) |
| 129 | + return loss |
| 130 | + |
| 131 | + def validation_step(self, batch, batch_idx): |
| 132 | + x, y = self.split_batch(batch) |
| 133 | + y_hat = self(x) |
| 134 | + loss = F.cross_entropy(y_hat, y) |
| 135 | + self.log('valid_loss', loss) |
| 136 | + |
| 137 | + def test_step(self, batch, batch_idx): |
| 138 | + x, y = self.split_batch(batch) |
| 139 | + y_hat = self(x) |
| 140 | + loss = F.cross_entropy(y_hat, y) |
| 141 | + self.log('test_loss', loss) |
| 142 | + |
| 143 | + def configure_optimizers(self): |
| 144 | + return torch.optim.Adam(self.parameters(), lr=self.hparams.learning_rate) |
| 145 | + |
| 146 | + @staticmethod |
| 147 | + def add_model_specific_args(parent_parser): |
| 148 | + parser = ArgumentParser(parents=[parent_parser], add_help=False) |
| 149 | + parser.add_argument('--hidden_dim', type=int, default=128) |
| 150 | + parser.add_argument('--learning_rate', type=float, default=0.0001) |
| 151 | + return parser |
| 152 | + |
| 153 | + |
| 154 | +def cli_main(): |
| 155 | + pl.seed_everything(1234) |
| 156 | + |
| 157 | + # ------------ |
| 158 | + # args |
| 159 | + # ------------ |
| 160 | + parser = ArgumentParser() |
| 161 | + parser.add_argument('--batch_size', default=32, type=int) |
| 162 | + parser = pl.Trainer.add_argparse_args(parser) |
| 163 | + parser = LitClassifier.add_model_specific_args(parser) |
| 164 | + args = parser.parse_args() |
| 165 | + |
| 166 | + # ------------ |
| 167 | + # data |
| 168 | + # ------------ |
| 169 | + dataset = MNIST('', train=True, download=True, transform=transforms.ToTensor()) |
| 170 | + mnist_test = MNIST('', train=False, download=True, transform=transforms.ToTensor()) |
| 171 | + mnist_train, mnist_val = random_split(dataset, [55000, 5000]) |
| 172 | + |
| 173 | + eii_train = ExternalMNISTInputIterator(mnist_train, args.batch_size) |
| 174 | + eii_val = ExternalMNISTInputIterator(mnist_val, args.batch_size) |
| 175 | + eii_test = ExternalMNISTInputIterator(mnist_test, args.batch_size) |
| 176 | + |
| 177 | + pipe_train = ExternalSourcePipeline(batch_size=args.batch_size, eii=eii_train, num_threads=2, device_id=0) |
| 178 | + train_loader = DALIClassificationLoader(pipe_train, size=len(mnist_train), auto_reset=True, fill_last_batch=False) |
| 179 | + |
| 180 | + pipe_val = ExternalSourcePipeline(batch_size=args.batch_size, eii=eii_val, num_threads=2, device_id=0) |
| 181 | + val_loader = DALIClassificationLoader(pipe_val, size=len(mnist_val), auto_reset=True, fill_last_batch=False) |
| 182 | + |
| 183 | + pipe_test = ExternalSourcePipeline(batch_size=args.batch_size, eii=eii_test, num_threads=2, device_id=0) |
| 184 | + test_loader = DALIClassificationLoader(pipe_test, size=len(mnist_test), auto_reset=True, fill_last_batch=False) |
| 185 | + |
| 186 | + # ------------ |
| 187 | + # model |
| 188 | + # ------------ |
| 189 | + model = LitClassifier(args.hidden_dim, args.learning_rate) |
| 190 | + |
| 191 | + # ------------ |
| 192 | + # training |
| 193 | + # ------------ |
| 194 | + trainer = pl.Trainer.from_argparse_args(args) |
| 195 | + trainer.fit(model, train_loader, val_loader) |
| 196 | + |
| 197 | + # ------------ |
| 198 | + # testing |
| 199 | + # ------------ |
| 200 | + trainer.test(test_dataloaders=test_loader) |
| 201 | + |
| 202 | + |
| 203 | +if __name__ == "__main__": |
| 204 | + cli_main() |
0 commit comments