Deeplab tensorflow实现：培训自定义数据。如何选择参数？

我尝试使用我的图像和分割模板数据集来训练deeplabv3+。然而，官方的github项目（）只提供培训VOC、城市景观和ADE20K数据集。因此，尝试按照VOC数据集映射来构建我的数据集。我的目标是从视频中分割桌子。我从一个视频中得到了80个连续帧。图像形状为1920*1080。我使用“Labelme”为这80幅图像创建了表的掩码。视频帧中的表格类似。我将它们分开，60个用于培训，20个用于验证。结构类似于VOC数据集映射：

JPEG图像：对于RGB彩色图像

分段：对于包含train文件名、val文件名和trainval文件名的txt文件

SegmentationClass：用于带有彩色贴图的遮罩图像

之后，我修改了“download_and_convert_voc2012.sh”以删除颜色映射并创建tfrecord文件。然后，修改“data_gergenerator.py”以更改VOC的数据信息。换句话说，我更改了原始代码：

_PASCAL_VOC_SEG_INFORMATION = DatasetDescriptor(
    splits_to_sizes={
        'train': 1464, 
        'train_aug': 10582, 
        'trainval': 2913, 
        'val': 1449, 
    },
    num_classes=21,
    ignore_label=255,
)

到

下一步是使用下面的commond进行培训：

python3 deeplab/train.py \
    --logtostderr \
    --training_number_of_steps=10000 \
    --learning_rate_decay_step=500 \
    --train_split="train" \
    --base_learning_rate=0.0001 \
    --adan_learning_rate=0.0001 \
    --model_variant="xception_65" \
    --atrous_rates=6 \
    --atrous_rates=12 \
    --atrous_rates=18 \
    --output_stride=16 \
    --decoder_output_stride=4 \
    --train_crop_size="513,513" \
    --train_batch_size=1 \
    --dataset="pascal_voc_seg" \
    --tf_initial_checkpoint="/home/.../models/check_point/xception/model.ckpt.index" \
    --train_logdir="/home/.../models/train_log" \
    --dataset_dir="/home/.../models/research/deeplab/datasets/pascal_voc_seg/tfrecord"

初始检查点是“Exception65_coco_voc_trainval”，tfrecord文件的数据集是从我的80幅图像数据生成的。然而，在10000个步骤中，训练过程上下损失约4.8~5.5。我尝试降低学习率，但损失没有减少。我试图将“train_crop_size”改为“19201080”，但我的电脑坏了。我是否应该将图像大小调整到“500*500”左右？然而，我并没有改变我的形象。经过10000步的训练。让我展示最后100个步骤：

I0201 21:24:03.829240 140124578015040 learning.py:507] global step 9900: loss = 5.0399 (3.194 sec/step)
INFO:tensorflow:global step 9910: loss = 4.8355 (3.152 sec/step)
I0201 21:24:35.473860 140124578015040 learning.py:507] global step 9910: loss = 4.8355 (3.152 sec/step)
INFO:tensorflow:global step 9920: loss = 4.8337 (3.179 sec/step)
I0201 21:25:07.198329 140124578015040 learning.py:507] global step 9920: loss = 4.8337 (3.179 sec/step)
INFO:tensorflow:global step 9930: loss = 4.8411 (3.156 sec/step)
I0201 21:25:38.925366 140124578015040 learning.py:507] global step 9930: loss = 4.8411 (3.156 sec/step)
INFO:tensorflow:global step 9940: loss = 4.8540 (3.188 sec/step)
I0201 21:26:10.489563 140124578015040 learning.py:507] global step 9940: loss = 4.8540 (3.188 sec/step)
INFO:tensorflow:global step 9950: loss = 4.8426 (3.221 sec/step)
I0201 21:26:42.325965 140124578015040 learning.py:507] global step 9950: loss = 4.8426 (3.221 sec/step)
INFO:tensorflow:global step 9960: loss = 4.8415 (3.130 sec/step)
I0201 21:27:14.000005 140124578015040 learning.py:507] global step 9960: loss = 4.8415 (3.130 sec/step)
INFO:tensorflow:global step 9970: loss = 4.9121 (3.163 sec/step)
I0201 21:27:45.751608 140124578015040 learning.py:507] global step 9970: loss = 4.9121 (3.163 sec/step)
INFO:tensorflow:global step 9980: loss = 4.8351 (3.145 sec/step)
I0201 21:28:17.505195 140124578015040 learning.py:507] global step 9980: loss = 4.8351 (3.145 sec/step)
INFO:tensorflow:global step 9990: loss = 4.8401 (3.153 sec/step)
I0201 21:28:49.131623 140124578015040 learning.py:507] global step 9990: loss = 4.8401 (3.153 sec/step)
INFO:tensorflow:Recording summary at step 9990.
I0201 21:28:50.591169 140119415715584 supervisor.py:1050] Recording summary at step 9990.
INFO:tensorflow:global step 10000: loss = 4.8512 (3.149 sec/step)
I0201 21:29:21.474581 140124578015040 learning.py:507] global step 10000: loss = 4.8512 (3.149 sec/step)
INFO:tensorflow:Stopping Training.
I0201 21:29:21.475129 140124578015040 learning.py:777] Stopping Training.
INFO:tensorflow:Finished training! Saving model to disk.
I0201 21:29:21.475337 140124578015040 learning.py:785] Finished training! Saving model to disk.
/home/.../.local/lib/python3.6/site-packages/tensorflow/python/summary/writer/writer.py:386: UserWarning: Attempting to use a closed FileWriter. The operation will be a noop unless the FileWriter is explicitly reopened.
  warnings.warn("Attempting to use a closed FileWriter. "

培训后，我使用下面的commond导出模型：

python3 deeplab/export_model.py \
  --logtostderr \
  --checkpoint_path="/home/.../models/train_log/model.ckpt-10000" \
  --export_path="/home/.../models/mode_export/table_seg_graph.pb" \
  --model_variant="xception_65" \
  --atrous_rates=6 \
  --atrous_rates=12 \
  --atrous_rates=18 \
  --output_stride=16 \
  --decoder_output_stride=4 \
  --num_classes=2 \
  --crop_size=513 \
  --crop_size=513 \
  --inference_scales=1.0

def run_visualization(url):
  try:
    original_im = Image.open('/home/.../models/research/deeplab/datasets/pascal_voc_seg/Tabledata/JPEGImages/v00019.jpg')
  except IOErr:
    ...

我得到了权重模型“table_seg_graph.pb”。在此之后，我尝试修改“deeplab_demo.ipynb”以使用此模型文件。我删除了关于解压tar文件的代码，并在下面添加代码以加载pb文件：

class DeepLabModel(object):
  def __init__(self, tarball_path):
    ...
    with tf.gfile.GFile(tarball_path, "rb") as f:
       graph_def = tf.GraphDef()
       graph_def.ParseFromString(f.read())
    ...
  ...

我还将模型下载部分更改为模型路径：

download_path = '/home.../models/mode_export/table_seg_graph.pb'
MODEL = DeepLabModel(download_path)

然后，我加载表格图像以测试模型：

python3 deeplab/export_model.py \
  --logtostderr \
  --checkpoint_path="/home/.../models/train_log/model.ckpt-10000" \
  --export_path="/home/.../models/mode_export/table_seg_graph.pb" \
  --model_variant="xception_65" \
  --atrous_rates=6 \
  --atrous_rates=12 \
  --atrous_rates=18 \
  --output_stride=16 \
  --decoder_output_stride=4 \
  --num_classes=2 \
  --crop_size=513 \
  --crop_size=513 \
  --inference_scales=1.0

def run_visualization(url):
  try:
    original_im = Image.open('/home/.../models/research/deeplab/datasets/pascal_voc_seg/Tabledata/JPEGImages/v00019.jpg')
  except IOErr:
    ...

然而，结果是什么都没有。它没有检测到任何东西遮罩分区全部为黑色。 由于视频具有复制权，我无法向您显示图像结果

我想我应该把我的1920*1080调整到513*513。否则，我不知道。谁能告诉我应该选择什么样的参数

我应该选择什么“萎缩率”

我应该给出什么“学习率”

---

我应该给出什么样的“裁剪尺寸”？

我也有同样的问题，没有调整尺寸的推断1920*1080有什么解决方案吗？