From 73f0c9f680429840c79cbb2e0a856fa62cf38444 Mon Sep 17 00:00:00 2001 From: Dun Liang Date: Thu, 21 Mar 2019 15:53:15 +0800 Subject: [PATCH] fix some hyper parameters --- fluid/PaddleCV/deeplabv3+/README.md | 9 +++++---- fluid/PaddleCV/deeplabv3+/reader.py | 31 +++++++++++++++++++++++++---- fluid/PaddleCV/deeplabv3+/train.py | 6 +++--- 3 files changed, 35 insertions(+), 11 deletions(-) diff --git a/fluid/PaddleCV/deeplabv3+/README.md b/fluid/PaddleCV/deeplabv3+/README.md index b9990a20..9f86b1da 100644 --- a/fluid/PaddleCV/deeplabv3+/README.md +++ b/fluid/PaddleCV/deeplabv3+/README.md @@ -72,20 +72,21 @@ python train.py --help 以上命令用于测试训练过程是否正常，仅仅迭代了50次并且使用了1的batch size，如果需要复现 原论文的实验，请使用以下设置： ``` +CUDA_VISIBLE_DEVICES=0 \ python ./train.py \ - --batch_size=8 \ + --batch_size=4 \ --parallel=True \ --norm_type=gn \ --train_crop_size=769 \ - --total_step=90000 \ + --total_step=500000 \ --base_lr=0.001 \ --init_weights_path=deeplabv3plus_gn_init \ --save_weights_path=output \ --dataset_path=$DATASET_PATH ``` -如果您的显存不足，可以尝试减小`batch_size`，同时等比例放大`total_step`, 保证相乘的值不变，这得益于Group Norm的特性，改变 `batch_size` 并不会显著影响结果，而且能够节约更多显存, 比如您可以设置`--batch_size=4 --total_step=180000`。 +如果您的显存不足，可以尝试减小`batch_size`，同时等比例放大`total_step`, 缩小`base_lr`, 保证相乘的值不变，这得益于Group Norm的特性，改变 `batch_size` 并不会显著影响结果，而且能够节约更多显存, 比如您可以设置`--batch_size=2 --total_step=1000000 --base_lr=0.0005`。 -如果您希望使用多卡进行训练，可以同比增加`batch_size`，减小`total_step`, 比如原来单卡训练是`--batch_size=4 --total_step=180000`，使用4卡训练则是`--batch_size=16 --total_step=45000` +如果您希望使用多卡进行训练，可以同比增加`batch_size`，减小`total_step`, 增加`base_lr`, 比如原来单卡训练是`--batch_size=4 --total_step=500000`，使用4卡训练则是`--batch_size=16 --total_step=125000 --base_lr=0.004` ### 测试 执行以下命令在`Cityscape`测试数据集上进行测试： diff --git a/fluid/PaddleCV/deeplabv3+/reader.py b/fluid/PaddleCV/deeplabv3+/reader.py index d420f0a2..b424318b 100644 --- a/fluid/PaddleCV/deeplabv3+/reader.py +++ b/fluid/PaddleCV/deeplabv3+/reader.py @@ -9,7 +9,7 @@ import six default_config = { "shuffle": True, "min_resize": 0.5, - "max_resize": 2, + "max_resize": 4, "crop_size": 769, } @@ -73,6 +73,16 @@ class CityscapeDataset: if self.index >= len(self.label_files): self.reset() + def color_augmentation(self, img): + r = np.random.rand() + img = img / 255.0 + if r < 0.5: + img = np.power(img, 1 - r) + else: + img = np.power(img, 1 + (r - 0.5) * 2) + img = img * 255.0 + return img + def get_img(self): shape = self.config["crop_size"] while True: @@ -90,9 +100,22 @@ class CityscapeDataset: break if shape == -1: return img, label, ln - random_scale = np.random.rand(1) * (self.config['max_resize'] - - self.config['min_resize'] - ) + self.config['min_resize'] + + if np.random.rand() > 0.5: + range_l = 1 + range_r = self.config['max_resize'] + else: + range_l = self.config['min_resize'] + range_r = 1 + + if np.random.rand() > 0.5: + assert len(img.shape) == 3 and len( + label.shape) == 3, "{} {}".format(img.shape, label.shape) + img = img[:, :, ::-1] + label = label[:, :, ::-1] + img = self.color_augmentation(img) + + random_scale = np.random.rand(1) * (range_r - range_l) + range_l crop_size = int(shape / random_scale) bb = crop_size // 2 diff --git a/fluid/PaddleCV/deeplabv3+/train.py b/fluid/PaddleCV/deeplabv3+/train.py index 799a1900..473ce6ff 100755 --- a/fluid/PaddleCV/deeplabv3+/train.py +++ b/fluid/PaddleCV/deeplabv3+/train.py @@ -21,10 +21,10 @@ parser = argparse.ArgumentParser() add_arg = lambda *args: utility.add_arguments(*args, argparser=parser) # yapf: disable -add_arg('batch_size', int, 2, "The number of images in each batch during training.") +add_arg('batch_size', int, 4, "The number of images in each batch during training.") add_arg('train_crop_size', int, 769, "Image crop size during training.") -add_arg('base_lr', float, 0.0001, "The base learning rate for model training.") -add_arg('total_step', int, 90000, "Number of the training step.") +add_arg('base_lr', float, 0.001, "The base learning rate for model training.") +add_arg('total_step', int, 500000, "Number of the training step.") add_arg('init_weights_path', str, None, "Path of the initial weights in paddlepaddle format.") add_arg('save_weights_path', str, None, "Path of the saved weights during training.") add_arg('dataset_path', str, None, "Cityscape dataset path.") -- GitLab