Channel Pruning for SSD

[jiaxiang-wu]

In this PR, we have updated the implementation of ChannelPrunedRmtLearner so that it is now capable of compression SSD models. Below is the detailed results:

Model: SSD-VGG-16
Uncompressed model's mAP: 77.64%

Pruned Layers	Prune Ratio	FLOPs	mAP
Backbone-only	0.2	67.34%	77.43%
Backbone-only	0.3	53.58%	77.05%
Backbone-only	0.4	41.63%	75.47%
Backbone-only	0.5	31.56%	73.29%
All layers	0.2	66.50%	76.98%
All layers	0.3	52.32%	76.42%
All layers	0.4	39.96%	74.89%
All layers	0.5	29.47%	72.48%

Usage:

# backbone-only
$ ./scripts/run_docker.sh nets/vgg_at_pascalvoc_run.py -n=4 \
    --learner=chn-pruned-rmt \
    --cpr_prune_ratio=0.2 \
    --batch_size=8

# all layers
$ ./scripts/run_docker.sh nets/vgg_at_pascalvoc_run.py -n=4 \
    --learner=chn-pruned-rmt \
    --cpr_prune_ratio=0.2 \
    --batch_size=8 \
    --cpr_skip_op_names=multibox_head

P.S.: A minor update to utils/get_idle_gpus.py. It now returns top-k GPUs with least memory occupation.

[CoderGenJ]

运行环境：
（1）运行指令：./scripts/run_local.sh nets/vgg_at_pascalvoc_run.py -n=4
--learner=chn-pruned-rmt
--cpr_prune_ratio=0.2
--batch_size=8
--cpr_skip_op_names=multibox_head
（2）运行环境：将ssd vgg16.ckpt（官方提供百度云盘）放入了./models文件中，运行之后出现
NotFoundError (see above for traceback): Restoring from checkpoint failed. This is most likely due to a Variable name or other graph key that is missing from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original error:

Key model/ssd300/additional_layers/conv10/conv10_1/bias not found in checkpoint
[[node model/save/RestoreV2 (defined at /home/Eric/work_space/model_pruning/PocketFlow/learners/channel_pruning_rmt/learner.py:358) = RestoreV2[dtypes=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, ..., DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT], _device="/job:localhost/replica:0/task:0/device:CPU:0"](_arg_model/save/Const_0_0, model/save/RestoreV2/tensor_names, model/save/RestoreV2/shape_and_slices)]]

但是在graph.pbtxt中找到了相关节点
请问如何解决？