SSD-Tensorflow 512x512 训练配置

搞了几天终于把这个给搞得差不多了,遇到的错误这里也记录一下:

一、配置【配置什么的300和512其实差不多,这里只举一个例子来分析一下】

 之前的文件修改什么的和300x300的一样:https://www.cnblogs.com/GrPhoenix/p/10018072.html

从自己训练的ssd_300_vgg模型开始训练ssd_512_vgg的模型

      因ssd_300_vgg中没有block12,又因为block7,block8,block9,block10,block11,中的参数张量两个网络模型中不匹配,因此ssd_512_vgg中这几个模块的参数不从ssd_300_vgg模型中继承,因此使用checkpoint_exclude_scopes命令指出。

         因为所有的参数均需要训练,因此不使用命令--trainable_scopes

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
1  #/bin/bash
  2  DATASET_DIR = / home / data / xxx / imagedata / xing_tf / train_tf /
  3  TRAIN_DIR = / home / data / xxx / model / xing300512_model /
  4  CHECKPOINT_PATH = / home / data / xxx / model / xing300_model / model.ckpt - 60000    #加载的ssd_300_vgg模型
  5  python3 . / train_ssd_network.py \
  6         - - train_dir = ${TRAIN_DIR} \
  7         - - dataset_dir = ${DATASET_DIR} \
  8         - - dataset_name = pascalvoc_2007 \
  9         - - dataset_split_name = train \
10         - - model_name = ssd_512_vgg \
11         - - checkpoint_path = ${CHECKPOINT_PATH} \
12         - - checkpoint_model_scope = ssd_300_vgg \
13         - - checkpoint_exclude_scopes = ssd_512_vgg / block7,ssd_512_vgg / block7_box,ssd_512_vgg / block8,ssd_512_vgg / block8_box,ssd_512_vgg / block9,ssd_512_vgg / block9_box,ssd_512_vgg / block10,ssd_512_vgg / block10_box,ssd_512_vgg / block11,ssd_512_vgg / b    lock11_box,ssd_512_vgg / block12,ssd_512_vgg / block12_box \
14         #--trainable_scopes=ssd_300_vgg/conv6,ssd_300_vgg/conv7,ssd_300_vgg/block8,ssd_300_vgg/block9,ssd_300_vgg/block10,ssd_300_vgg/block11,ssd_300_vgg/block4_box,ssd_300_vgg/block7_box,ssd_300_vgg/block8_box,ssd_300_vgg/block9_box,ssd_3    00_vgg/block10_box,ssd_300_vgg/block11_box \
15         - - save_summaries_secs = 28800  \
16         - - save_interval_secs = 28800  \
17         - - weight_decay = 0.0005  \
18         - - optimizer = adam \
19         - - learning_rate_decay_factor = 0.94  \
20         - - batch_size = 16  \
21         - - num_classes = 4  \
22         - gpu_memory_fraction = 0.8  \

另外由300转512后还需修改:

1. 首先修改ssd_vgg_512.py的训练类别

2.修改train_ssd_network.py的model_name

   修改为ssd_512_vgg

3. 修改nets/np_methods.py

    修改:将300改为512, 将类别改为自己数据的类别(+背景)

4. 修改preprocessing/ssd_vgg_preprocessing.py

    修改:将300改为512

5. 修改ssd_notbook.ipynb

   a  将文件中数字“300”改为“512”

其他修改可以参考:http://blog.csdn.net/liuyan20062010/article/details/78905517

二、我遇到的错误:

InvalidArgumentError (see above for traceback): Restoring from checkpoint failed. This is most likely due to a mismatch between the current graph and the graph from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original error:

Assign requires shapes of both tensors to match. lhs shape= [84] rhs shape= [8]
	 [[{{node save/Assign_20}} = Assign[T=DT_FLOAT, _class=["loc:@ssd_512_vgg/block12_box/conv_cls/biases"], use_locking=true, validate_shape=true, _device="/job:localhost/replica:0/task:0/device:GPU:0"](ssd_512_vgg/block12_box/conv_cls/biases, save/RestoreV2/_41)]]
	 [[{{node save/RestoreV2/_104}} = _Send[T=DT_FLOAT, client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device_incarnation=1, tensor_name="edge_110_save/RestoreV2", _device="/job:localhost/replica:0/task:0/device:CPU:0"](save/RestoreV2:52)]]

  这类的问题本质上来说还是自己的配置不对,这个问题我查了很久,最后发现实在是太simpleT-T。

  我的问题的话:在从300转到512的时候忘记改ssd_vgg_512.py的类别导致test的时候文件配置和训练的tensor  shape不匹配TT...

转载于:https://www.cnblogs.com/GrPhoenix/p/10039833.html

你可能感兴趣的:(SSD-Tensorflow 512x512 训练配置)