error of "Incorrect shape" in example/cpp/image-classification-predict.cc

[VincentChong123]

prediction using image-classification-predict.cc return error below for my input 224x224 color image, which complies with size setting in image-classification-predict.cc. Please kindly advise.

[src/c_api/c_predict_api.cc]MXPredCreatePartialOut(177):CHECK
[00:17:33] /N/software/mxnet/mxnet-master/dmlc-core/include/dmlc/logging.h:235: [00:17:33] src/operator/./concat-inl.h:152: Check failed: (dshape[j]) == (tmp[j]) Incorrect shape[2]: (1,320,15,15). (first input shape: (1,576,14,14)

Thanks!

[stardut]

I met the error too.You can change your data_shape's value to solve it. What is your data_shape's value?

[VincentChong123]

Hi @stardust, thanks for your support.
data_shape's equivalence, input_shape_data[4] = {1, 3, 224,224} as set below

   //image-classification-predict.cc
    // Image size and channels
    int width  = 224;
    int height = 224;
    int channels = 3;

    const mx_uint input_shape_indptr[2] = { 0, 4 };
    // ( trained_width, trained_height, channel, num)
    const mx_uint input_shape_data[4] = { 1,
                                        static_cast<mx_uint>(channels),
                                        static_cast<mx_uint>(width),
                                        static_cast<mx_uint>(height) };
    PredictorHandle out = 0;  // alias for void *

    //-- Create Predictor
    MXPredCreate((const char*)json_data.GetBuffer(),
                 (const char*)param_data.GetBuffer(),
                 static_cast<size_t>(param_data.GetLength()),
                 dev_type,
                 dev_id,
                 num_input_nodes,
                 input_keys,
                 input_shape_indptr,
                 input_shape_data,
                 &out);

[taoari]

I think this is caused by the pooling behavior with 3x3/2 structures. This affects the CaffeNet, GoogleNet, Inception-BN models, but you can try VGG16, or ResNet50. They work good.

I did not notice this issue, and opened another similar issue #2718 , I think the reasons are the same. Hope that people will quickly fix this soon.

Or you can revert to the version tagged 20160531, it works good.

[VincentChong123]

@taoari, thanks. Will try and report soon.
On Jul 16, 2016 1:02 PM, "taoari" [email protected] wrote:

I think this is caused by the pooling behavior with 3x3/2 structures. This
affects the CaffeNet, GoogleNet, Inception-BN models, but you can try
VGG16, or ResNet50. They work good.

I did not notice this issue, and opened another similar issue #2718
#2718 , I think the reasons are the
same. Hope that people will quickly fix this soon.

Or you can revert to the version tagged 20160531, it works good.

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
#2709 (comment), or mute
the thread
https://github.com/notifications/unsubscribe-auth/AQyyHLZfvA3rPntuGfTL9n5QnspLvDbfks5qWGXUgaJpZM4JNAcK
.

[VincentChong123]

Hi @taoari, same error for inception_bn at mxnet-20160531 + 20160713's sub-modules (dmlc-core, ps-lite, ms-shadow). Will check with VGG16, or ResNet50

[taoari]

@weishengchong In my case, it works. See #2718 for my description.

It is this PR #2585 that changes the behavior. When I revert to one commit before that PR, it also works well.

I am not using the C++ file image-classification-predict.cc, you can use the notebook predict-with-pretrained-model.ipynb for a quick check.

[VincentChong123]

@taoari

This affects the CaffeNet, GoogleNet, Inception-BN models, but you can try VGG16, or ResNet50. They work good.

I confirmed image-classification-predict.cc + inception_v3 works good using model below
https://github.com/dmlc/mxnet-model-gallery/blob/master/imagenet-1k-inception-v3.md

[taoari]

@weishengchong Maybe, but Inception_BN does not work. I did not try Inception-V3 yet.

3x3/2 pooling behavior is changed, it does not mean that models using 3x3/2 will not work.

For example, ResNet50 will behave like 112->55->28, while in He's paper it should work like 112->56->28. But ResNet50 still runs, and the final results might slightly different. I am not sure to what degree it will hurts the final accuracy.

[itijyou]

What about using the 3x3/2 pooling with pad=1? Pre-trained models may still degrade in accuracy though. There seems an offset of one pixel between this implementation and the previous.

[phunterlau]

This issue is closed due to lack of activity in the last 90 days. Feel free to reopen if this is still an active issue. Thanks!