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 #/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...
AIEarth是一个由众多领域内专家博主共同打造的学术平台，旨在建设一个拥抱智慧未来的学术殿堂！【平台地址：https://devpress.csdn.net/aiearth】 很高兴认识你！加入我们共同进步！
