MXNet-SSD 再识--Mobilenet_v2
1。加载预训练模型
除了MXNet-SSD 初识中介绍的参数外，在此还需要再添加另一个参数：
--pretrained2。修改/home/whut/yyCode/ssd/symbol/symbol_factory.py
elif network=='mobilenet_v2':
image_shape = '3,224,224'
network = 'mobilenet_v2'
from_layers = ['relu6_1_expand', 'relu6_4', '', '', '', '']
num_filters = [-1, -1, 512, 256, 256, 128]
strides = [-1, -1, 2, 2, 2, 2]
pads = [-1, -1, 1, 1, 1, 1]
sizes = [[.1, .141], [.2, .272], [.37, .447], [.54, .619], [.71, .79], [.88, .961]]
ratios = [[1, 2, .5], [1, 2, .5, 3, 1. / 3], [1, 2, .5, 3, 1. / 3], [1, 2, .5, 3, 1. / 3], \
[1, 2, .5], [1, 2, .5]]
normalizations = -1
steps = []
return locals()添加mobilenet_v2的相关说明。
3。开始训练
尝试一：batch-size=128，lr=0.005，--lr-steps='120,240'，--end-epoch=360
发现：过早的调整了学习率。
尝试二：batch-size=128，lr=0.005，--lr-steps='150,280'，--end-epoch=360
发现：可以尝试不改变学习率，让其自然生长。
尝试三：batch-size=128，lr=0.005，--lr-steps=''，--end-epoch=360
发现可以在200和250设置衰减，验证是否能够增加1%。
--finetune #从哪一个epoch开始，此时会冻住蛮多层，会从当时的学习率开始
--begin-epoch #从哪一个epoch开始训练，与从头训练类似
--resume #恢复训练此时，我的目的是为了对所有参数进行调整，所以选取resume
4。验证
python evaluate.py --network mobilenet_v2 --epoch 0毫无疑问会报错
bug-1
whut@whut-Z370-AORUS-Gaming-5:~/yyCode/ssd$ python evaluate.py --network mobilenet_v2 --epoch 0
['0']
[gpu(0)]
[20:59:21] src/io/iter_image_det_recordio.cc:281: ImageDetRecordIOParser: /home/whut/yyCode/ssd/data/val.rec, use 11 threads for decoding..
[20:59:21] src/io/iter_image_det_recordio.cc:334: ImageDetRecordIOParser: /home/whut/yyCode/ssd/data/val.rec, label padding width: 8
Traceback (most recent call last):
File "evaluate.py", line 115, in
voc07_metric=args.use_voc07_metric)
File "/home/whut/yyCode/ssd/evaluate/evaluate_net.py", line 116, in evaluate_net
data = [mx.random.uniform(-1.0, 1.0, shape=shape, ctx=ctx) for _, shape in mod.data_shapes]
File "/home/whut/anaconda2/lib/python2.7/site-packages/mxnet/ndarray/random.py", line 102, in uniform
[low, high], shape, dtype, ctx, out, kwargs)
File "/home/whut/anaconda2/lib/python2.7/site-packages/mxnet/ndarray/random.py", line 48, in _random_helper
return random(*params, shape=shape, dtype=dtype, ctx=ctx, out=out, **kwargs)
File "", line 58, in _random_uniform
File "/home/whut/anaconda2/lib/python2.7/site-packages/mxnet/_ctypes/ndarray.py", line 92, in _imperative_invoke
ctypes.byref(out_stypes)))
File "/home/whut/anaconda2/lib/python2.7/site-packages/mxnet/base.py", line 251, in check_call
raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: [20:59:22] include/mxnet/./base.h:388: Invalid context string [gpu(0)]
仿照train.py修改两个地方：
1）evaluate.py
#87行
ctx = [mx.gpu(int(i)) for i in args.gpu_id.split(',') if i.strip()]2）evaluate_net.py
#116行
data = [mx.random.uniform(-1.0, 1.0, shape=shape, ctx=ctx[0]) for _, shape in mod.data_shapes]bug-2
whut@whut-Z370-AORUS-Gaming-5:~/yyCode/ssd$ python evaluate.py --network mobilenet_v2 --epoch 0
['0']
[gpu(0)]
[21:17:09] src/io/iter_image_det_recordio.cc:281: ImageDetRecordIOParser: /home/whut/yyCode/ssd/data/val.rec, use 11 threads for decoding..
[21:17:09] src/io/iter_image_det_recordio.cc:334: ImageDetRecordIOParser: /home/whut/yyCode/ssd/data/val.rec, label padding width: 8
Traceback (most recent call last):
File "evaluate.py", line 115, in
voc07_metric=args.use_voc07_metric)
File "/home/whut/yyCode/ssd/evaluate/evaluate_net.py", line 92, in evaluate_net
load_net, args, auxs = mx.model.load_checkpoint(model_prefix, epoch)
File "/home/whut/anaconda2/lib/python2.7/site-packages/mxnet/model.py", line 439, in load_checkpoint
save_dict = nd.load('%s-%04d.params' % (prefix, epoch))
File "/home/whut/anaconda2/lib/python2.7/site-packages/mxnet/ndarray/utils.py", line 175, in load
ctypes.byref(names)))
File "/home/whut/anaconda2/lib/python2.7/site-packages/mxnet/base.py", line 251, in check_call
raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: [21:17:10] src/io/local_filesys.cc:199: Check failed: allow_null LocalFileSystem::Open "/home/whut/yyCode/ssd/model/ssd_mobilenet_v2_300-0000.params": No such file or directory
修改 model目录下的相应文件
whut@whut-Z370-AORUS-Gaming-5:~/yyCode/ssd$ python evaluate.py --network mobilenet_v2 --epoch 0
['0']
[gpu(0)]
[21:20:34] src/io/iter_image_det_recordio.cc:281: ImageDetRecordIOParser: /home/whut/yyCode/ssd/data/val.rec, use 11 threads for decoding..
[21:20:35] src/io/iter_image_det_recordio.cc:334: ImageDetRecordIOParser: /home/whut/yyCode/ssd/data/val.rec, label padding width: 8
[21:20:36] src/operator/nn/./cudnn/./cudnn_algoreg-inl.h:97: Running performance tests to find the best convolution algorithm, this can take a while... (setting env variable MXNET_CUDNN_AUTOTUNE_DEFAULT to 0 to disable)
[21:20:42] src/operator/nn/./cudnn/./cudnn_algoreg-inl.h:97: Running performance tests to find the best convolution algorithm, this can take a while... (setting env variable MXNET_CUDNN_AUTOTUNE_DEFAULT to 0 to disable)
INFO:root:Finished inference with 160 images
INFO:root:Finished with 332.948322 images per second
1: 0.93783993784
2: 0.979020979021
3: 1.0
4: 0.763636363636
5: 0.929292929293
6: 0.818181818182
7: 0.75
8: 0.984848484848
9: 0.927272727273
10: 1.0
11: 0.733044733045
12: 0.675324675325
13: 0.69696969697
14: 1.0
15: 1.0
16: 1.0
17: 0.909090909091
18: 1.0
mAP: 0.894695736362
5。demo
python demo.py --network mobilenet_v2相关主要参数：
--network
--images
--data-shape
--class-names实验结果：
