walterma / gluon-faster-rcnn Goto Github PK

I want to call the model process prediction of training through C++, but I do not know how to do it, I hope to get your help, thank you!

@WalterMa
I found it very useful.
But i am not getting class scores very less like 0.5-0.8...
How can we increase the class scores close to 0.90 ...
Also Have you tried to increase the testing accuracy close to 90%.

I have only 2 classes.. I want something close to 90%+..

Can you suggest me some tuning steps to increase it.

Can i use mobilenet as basenet instead of vgg..
If yes, Where exactly should i randomize the weights..
is it something here
and here , Shoul di change something?

Is there any other place, where i need to change it.
And, what should be done to config.py file to not to include param file..this .

Also, How did you write the names of vgg16 layers. fixed_params. config.py line 113
What are the names of the layers of mobilenetv2_0.25 model?

please try to help provide me some hints regarding this...
Thank You..

Hi,
Is it possible to run the code on windows.?
Is it similar to the original faster rcnn in terms of accuracy ?
Did you do inference with this repo??

Hi ,I'm training gluon-cv's faster rcnn(vgg16 and VOC). But,the 7 epoch mAP is 0.52 ,I think it is lower than I expected.
Could you share the training log of your faster rcnn(vgg16)? So I can find some mistake about my training.

The previous link is invalid, could you reshare it? Thanks

Hello,
I have trained the model with the given configs. But i am getting only 80% accuracy on 2 object class data.

Is there any possibility to increase it further?? What parameters should i change to increase the accuracy??

I am using mobilenet 0.25 version as base network..

Hi,
I have done training on my own dataset and i got 70% accuracy after 4 epochs..
I want to visualize the output.. so i tried with demo script.. i gave 1 input image and tried with the trained model. i changed the class names in demo script.
but i got this error.. Could you please let me know whats the problem.. Thank You

Traceback (most recent call last):
File "demo_faster_rcnn.py", line 65, in
cls, scores, bboxes = net(data.as_in_context(ctx), im_info.as_in_context(ctx))
File "/home/ubuntu/anaconda3/lib/python3.6/site-packages/mxnet/gluon/block.py", line 413, in call
return self.forward(*args)
File "/home/ubuntu/anaconda3/lib/python3.6/site-packages/mxnet/gluon/block.py", line 629, in forward
return self.hybrid_forward(ndarray, x, *args, **params)
File "/home/ubuntu/gluon-faster-rcnn/rcnn/rcnn.py", line 69, in hybrid_forward
rois = self.proposal(rpn_cls_prob, rpn_bbox_pred, im_info)
File "/home/ubuntu/anaconda3/lib/python3.6/site-packages/mxnet/gluon/block.py", line 413, in call
return self.forward(*args)
File "/home/ubuntu/anaconda3/lib/python3.6/site-packages/mxnet/gluon/block.py", line 629, in forward
return self.hybrid_forward(ndarray, x, *args, **params)
File "/home/ubuntu/gluon-faster-rcnn/rcnn/proposal.py", line 32, in hybrid_forward
threshold=self.rpn_nms_threshold, rpn_min_size=self.rpn_min_size)
File "", line 82, in MultiProposal
File "/home/ubuntu/anaconda3/lib/python3.6/site-packages/mxnet/_ctypes/ndarray.py", line 92, in _imperative_invoke
ctypes.byref(out_stypes)))
File "/home/ubuntu/anaconda3/lib/python3.6/site-packages/mxnet/base.py", line 149, in check_call
raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: Cannot find argument 'cls_prob', Possible Arguments:

rpn_pre_nms_top_n : int, optional, default='6000'
Number of top scoring boxes to keep after applying NMS to RPN proposals
rpn_post_nms_top_n : int, optional, default='300'
Overlap threshold used for non-maximumsuppresion(suppress boxes with IoU >= this threshold
threshold : float, optional, default=0.7
NMS value, below which to suppress.
rpn_min_size : int, optional, default='16'
Minimum height or width in proposal
scales : tuple of , optional, default=[4,8,16,32]
Used to generate anchor windows by enumerating scales
ratios : tuple of , optional, default=[0.5,1,2]
Used to generate anchor windows by enumerating ratios
feature_stride : int, optional, default='16'
The size of the receptive field each unit in the convolution layer of the rpn,for example the product of all stride's prior to this layer.
output_score : boolean, optional, default=0
Add score to outputs
iou_loss : boolean, optional, default=0
Usage of IoU Loss
, in operator _contrib_MultiProposal(name="", feature_stride="16", ratios="(0.5, 1, 2)", rpn_min_size="16", scales="(8, 16, 32)", rpn_post_nms_top_n="300", rpn_pre_nms_top_n="6000", threshold="0.7", cls_prob="
[[[[9.2525011e-01 9.8686647e-01 9.9559492e-01 ... 9.6093690e-01
9.3473071e-01 8.3388972e-01]
[9.8144472e-01 9.9909139e-01 9.9984789e-01 ... 9.9409735e-01
9.8589975e-01 9.3193233e-01]
[9.8883343e-01 9.9964535e-01 9.9995410e-01 ... 9.9763453e-01
9.9354243e-01 9.5571983e-01]
...
[9.8543328e-01 9.9948043e-01 9.9991584e-01 ... 9.9969471e-01
9.9917930e-01 9.8851913e-01]
[9.7469234e-01 9.9870670e-01 9.9970120e-01 ... 9.9901140e-01
9.9767345e-01 9.7834754e-01]
[9.0814865e-01 9.8365211e-01 9.9276966e-01 ... 9.8466349e-01
9.7399849e-01 9.0880662e-01]]

[[9.0745032e-01 9.8171026e-01 9.9309546e-01 ... 9.4973421e-01
9.1728598e-01 8.1418854e-01]
[9.7337264e-01 9.9846858e-01 9.9970394e-01 ... 9.9094427e-01
9.7995251e-01 9.1768110e-01]
[9.8243284e-01 9.9936765e-01 9.9990177e-01 ... 9.9625152e-01
9.9037081e-01 9.4557309e-01]
...
[9.7682333e-01 9.9898654e-01 9.9980742e-01 ... 9.9938107e-01
9.9859077e-01 9.8415011e-01]
[9.6041822e-01 9.9727988e-01 9.9929476e-01 ... 9.9794215e-01
9.9601054e-01 9.7078675e-01]
[8.7193352e-01 9.7200722e-01 9.8639816e-01 ... 9.7515827e-01
9.6249181e-01 8.8967586e-01]]

[[5.2806801e-01 5.3886396e-01 5.5010569e-01 ... 5.2669793e-01
5.2231640e-01 5.0962281e-01]
[5.3768706e-01 5.5899465e-01 5.7622200e-01 ... 5.5065542e-01
5.4214233e-01 5.2825642e-01]
[5.4670048e-01 5.8282024e-01 6.0394657e-01 ... 5.6064773e-01
5.5870861e-01 5.4004127e-01]
...
[5.3445053e-01 5.7814318e-01 6.0343522e-01 ... 5.9942263e-01
5.9564185e-01 5.6060779e-01]
[5.3276056e-01 5.7079929e-01 5.9411222e-01 ... 5.9032643e-01
5.8910215e-01 5.5843079e-01]
[5.2759832e-01 5.5251533e-01 5.7285386e-01 ... 5.6627262e-01
5.6415069e-01 5.4235542e-01]]

...

[[1.6489255e-01 5.4860741e-02 2.7824294e-02 ... 1.1167015e-01
1.5454119e-01 2.7348977e-01]
[7.4070774e-02 1.0540956e-02 3.4242510e-03 ... 3.8886167e-02
6.8945184e-02 1.8131968e-01]
[6.2780201e-02 6.9735665e-03 1.9518270e-03 ... 2.3429820e-02
4.5364555e-02 1.4465846e-01]
...
[8.8465296e-02 1.4774417e-02 5.4020169e-03 ... 1.1072693e-02
1.9699827e-02 8.5111000e-02]
[1.4203803e-01 3.7630506e-02 2.2168955e-02 ... 3.7781410e-02
5.5475168e-02 1.4689194e-01]
[2.8729475e-01 1.7887905e-01 1.5341425e-01 ... 1.8521468e-01
2.1700267e-01 3.0219343e-01]]

[[7.6535888e-02 1.3174757e-02 4.5662634e-03 ... 3.9024629e-02
6.6420421e-02 1.6296616e-01]
[1.8801216e-02 8.8067626e-04 1.5799509e-04 ... 5.9979130e-03
1.4321312e-02 6.7583486e-02]
[1.2135372e-02 3.7127602e-04 5.5430377e-05 ... 2.5837927e-03
6.8379878e-03 4.4826828e-02]
...
[1.6011752e-02 5.7889975e-04 1.1127151e-04 ... 3.9832355e-04
9.8138128e-04 1.2958074e-02]
[2.6029671e-02 1.4883390e-03 3.9916934e-04 ... 1.2645581e-03
2.7250603e-03 2.4580965e-02]
[1.0069502e-01 1.9385004e-02 9.5264316e-03 ... 1.9384181e-02
3.1448375e-02 1.0509791e-01]]

[[4.3997696e-01 3.9228746e-01 3.6535779e-01 ... 4.2972672e-01
4.3877992e-01 4.6890491e-01]
[4.0322891e-01 3.3196816e-01 3.0024055e-01 ... 3.9013031e-01
4.0616569e-01 4.5436901e-01]
[4.0472379e-01 3.2188171e-01 2.8730047e-01 ... 3.8169590e-01
3.9514536e-01 4.5027012e-01]
...
[4.1938949e-01 3.4382537e-01 3.1303972e-01 ... 3.3924583e-01
3.4913608e-01 4.1282722e-01]
[4.3390730e-01 3.7113073e-01 3.4658235e-01 ... 3.7191394e-01
3.8552991e-01 4.3190357e-01]
[4.6732298e-01 4.3559861e-01 4.2490643e-01 ... 4.4331262e-01
4.5451128e-01 4.7383672e-01]]]]
<NDArray 1x18x37x37 @gpu(0)>")

I tried to train the network in its default configuration, which works like expected.

If I set the detault.hybridize parameter in utils/config.py to False, the network does'nt learn anything anymore:

INFO:root:[Epoch 4][Batch 8269], Speed: 5.551862 samples/sec, RPNLogLoss=3.406618, RPNSmoothL1Loss=0.463918, RCNNLogLoss=0.758454, RCNNSmoothL1Loss=0.295339
INFO:root:[Epoch 4] Training cost: 2994.834264, RPNLogLoss=3.406611, RPNSmoothL1Loss=0.463785, RCNNLogLoss=0.758340, RCNNSmoothL1Loss=0.295269
INFO:root:[Epoch 4] Validation:
aeroplane=0.000000
...
sofa=0.000000
train=0.000000
tvmonitor=0.000000
mAP=0.000009

Does someone have an idea, why this is, or wether it is reproducable on other systems ?

I am using mxnet 1.3.0 and cudnn 7.1.2 with cuda 9.0

Best Regards
Stephan