Running into error while calling predict

[nishithbsk]

Hi,

I'm getting this error: Find name bn_pool3_moving_inv_var that is not in the auxiliary states

Would you know how to resolve it?

Thanks!

[ashishgupta-skg]

@nishithbsk
I am also getting the same error. Were you able to solve it?

[zedomel]

You should rename it to bn_pool3_moving_var. Here a piece of code to do that:
`model = mx.model.FeedForward.load('deep3d', 50, mx.gpu(0))

model.aux_params['bn_pool3_moving_var'] = model.aux_params['bn_pool3_moving_inv_var']
model.aux_params['bn_pool2_moving_var'] = model.aux_params['bn_pool2_moving_inv_var']
model.aux_params['bn_pool4_moving_var'] = model.aux_params['bn_pool4_moving_inv_var']
model.aux_params['bn_pool1_moving_var'] = model.aux_params['bn_pool1_moving_inv_var']
del model.aux_params['bn_pool1_moving_inv_var']
del model.aux_params['bn_pool2_moving_inv_var']
del model.aux_params['bn_pool3_moving_inv_var']
del model.aux_params['bn_pool4_moving_inv_var']`

[dongcin]

@zedomel @nishithbsk I also met the problem .When I ran the Y = model.predict(test_iter) I got the warning the kernel will restart .My GPU memory is 2G .I want to know how to solve it .Can you help me ?Thank you

[AjayNandoriya]

I had the similar problem. I used above solution but then results were not consistent with author's result.
So, I used MxNet v0.7.0 with CuDNN V4.0 and cuda 8(cuda version not important)...then without any modification, it worked and produced same results as author's.

[zedomel]

AjayNandoriya, the solution above should be adopted when using older version of CuDNN. Some GPU's are not compatible with new versions of CuDNN, so the only way to get it works is to change parameters names to match older versions I guess. But as you said the results can differ from the author's.

[AjayNandoriya]

@zedomel , your solution works with CuDNN v5 and removes runtime errors but it produces different results. For CuDNN v4.0, no change in model-json or script is required.

[zedomel]

@AjayNandoriya thanks. The solution is for newer versions of CuDNN, but may be there are some other parameters to change to make it works and produce results as author's. At end, I have used the same solution as you MxNet 0.7.0 and CuDNN v4.0.

[philtomson]

Does anyone have a fork of this repo where they've made the changes to the json (and any other files) to get things working with newer versions of mxnet?

[philtomson]

I found the one fork of this repo that purportedly has updates for CUDA8 and CnDNN 5.x: https://github.com/motypas/deep3d

Unfortunately motypas did not enable issues for his/her fork of deep3d.

I definitely get further along when I run with that fork, but now I'm seeing:

Y = model.predict(test_iter)
[14:23:45] src/operator/tensor/./matrix_op-inl.h:162: Using target_shape will be deprecated.
[14:23:45] /home/phil/build/mxnet/dmlc-core/include/dmlc/./logging.h:308: [14:23:45] src/operator/tensor/./../elemwise_op_common.h:122: Check failed: assign(&dattr, (*vec)[i]) Incompatible attr in node elementwisesum0 at 4-th input: expected [1,33,360,864], got [1,33,80,192]

which makes me wonder if the model is equivalent to the original deep3d?

[liutianling]

@philtomson Can you tell me the meaning of parameter "--source" in convert_movie.py and how can I get it ? Before running the convert_movie.py, I just use parse.py and data.py generating the database. But I don;t know how to get the --source( default = "test_idx" ) .
Thanks!