smichalowski/google_inception_v3_for_caffe

Google Inception (deepdream) v3 for Caffe

google_inception_v3_for_caffe's Issues

trained caffemodel on imagenet

Thank you so much for sharing. If you have completed training on imagenet, would you be able to upload the trained model? Thanks again!

inception_v3 for BVLC Caffe (without batch norm)

Hello,

Do you have the model (train_val.prototxt) for inception_v3 for "BVLC" Caffe without batch normalization?

Can you please attach that as well?

Thanks!

Where is ScaleBias layers?

Original BN paper uses Scale-bias y=kx+b after BatchNorm. Are you omitting them intentionally?

18 categories of imagenet

Hello, could you provide the 18 categories's sysnet and data preparation scripts you used for training the 11704.caffemodel, thanks.

digits model wrong?

Hey, I'm trying to use your model from the digits folder in DIGITS 5.0. Seems like there is a mistake somewhere because I'm using it verbatim but getting the following error:

Creating Layer mixed_3_tower_conv_2_relu
mixed_3_tower_conv_2_relu <- mixed_3_tower_conv_2_conv2d_bn
mixed_3_tower_conv_2_relu -> mixed_3_tower_conv_2_conv2d_relu
Setting up mixed_3_tower_conv_2_relu
Top shape: 15 96 14 14 (282240)
Memory required for data: 1148628540
Creating layer max_pool_mixed_3_pool
Creating Layer max_pool_mixed_3_pool
max_pool_mixed_3_pool <- ch_concat_mixed_2_chconcat_ch_concat_mixed_2_chconcat_0_split_2
max_pool_mixed_3_pool -> max_pool_mixed_3_pool
Setting up max_pool_mixed_3_pool
Top shape: 15 288 15 15 (972000)
Memory required for data: 1152516540
Creating layer ch_concat_mixed_3_chconcat
Creating Layer ch_concat_mixed_3_chconcat
ch_concat_mixed_3_chconcat <- max_pool_mixed_3_pool
ch_concat_mixed_3_chconcat <- mixed_3_conv_conv2d_relu
ch_concat_mixed_3_chconcat <- mixed_3_tower_conv_2_conv2d_relu
ch_concat_mixed_3_chconcat -> ch_concat_mixed_3_chconcat
== bottom[i]->shape(j) (15 vs. 14) All inputs must have the same shape, except at concat_axis.

Any ideas?

Any progress on google_inception_v3 re-training?

Hi, what is the final accuracy on you re-trained google_inception_v3 model?
Thanks!

questions about the test accuracy

Finally see someone training inception-v3 on Caffe :)
Refer to the google's inception-v3 paper (Figure 2), on epoch 53, the test top1 accuracy is above 0.7 (2*10^6 iteration ). Do you have any idea about this gap?

There are two data layer

Hi,

I have two questions.
1.
I want to know the reason why there are two data layer in your train.prototxt.

I tried to train your model in NVIDIA/caffe and faced this error.

I0827 18:18:25.026118 18597 net.cpp:94] Creating Layer ch_concat_mixed_3_chconcat
I0827 18:18:25.026121 18597 net.cpp:435] ch_concat_mixed_3_chconcat <- max_pool_mixed_3_pool
I0827 18:18:25.026126 18597 net.cpp:435] ch_concat_mixed_3_chconcat <- mixed_3_conv_conv2d_relu
I0827 18:18:25.026132 18597 net.cpp:435] ch_concat_mixed_3_chconcat <- mixed_3_tower_conv_2_conv2d_relu
I0827 18:18:25.026137 18597 net.cpp:409] ch_concat_mixed_3_chconcat -> ch_concat_mixed_3_chconcat
F0827 18:18:25.026154 18597 concat_layer.cpp:42] Check failed: top_shape[j] == bottom[i]->shape(j) (30 vs. 29) All inputs must have the same shape, except at concat_axis.
*** Check failure stack trace: ***
@ 0x7fb84eb81daa (unknown)
@ 0x7fb84eb81ce4 (unknown)
@ 0x7fb84eb816e6 (unknown)
@ 0x7fb84eb84687 (unknown)
@ 0x7fb84f2c78e7 caffe::ConcatLayer<>::Reshape()
@ 0x7fb84f1b234c caffe::Net<>::Init()
@ 0x7fb84f1b325a caffe::Net<>::Net()
@ 0x7fb84f2e877a caffe::Solver<>::InitTrainNet()
@ 0x7fb84f2e96fc caffe::Solver<>::Init()
@ 0x7fb84f2e9a43 caffe::Solver<>::Solver()
@ 0x7fb84f2dd659 caffe::Creator_RMSPropSolver<>()
@ 0x40f30c caffe::SolverRegistry<>::CreateSolver()
@ 0x4080dd train()
@ 0x405dcc main
@ 0x7fb84d6def45 (unknown)
@ 0x40659d (unknown)
@ (nil) (unknown)
Aborted (core dumped)

I don't know why this error occurs.

Thank you

Cannot run google inception v3 on any caffe

Hi,
Wen I try to make a run with that google inception v3 I get the following error:
ERROR: == bottom[i]->shape(j) (13 vs. 12) All inputs must have the same shape, except at concat_axis.
The latest caffe log lines are:

Setting up mixed_3_tower_conv_2_relu
Top shape: 24 96 12 12 (331776)
Memory required for data: 1402748352
Creating layer max_pool_mixed_3_pool
Creating Layer max_pool_mixed_3_pool
max_pool_mixed_3_pool <- ch_concat_mixed_2_chconcat_ch_concat_mixed_2_chconcat_0_split_2
max_pool_mixed_3_pool -> max_pool_mixed_3_pool
Setting up max_pool_mixed_3_pool
Top shape: 24 288 13 13 (1168128)
Memory required for data: 1407420864
Creating layer ch_concat_mixed_3_chconcat
Creating Layer ch_concat_mixed_3_chconcat
ch_concat_mixed_3_chconcat <- max_pool_mixed_3_pool
ch_concat_mixed_3_chconcat <- mixed_3_conv_conv2d_relu
ch_concat_mixed_3_chconcat <- mixed_3_tower_conv_2_conv2d_relu
ch_concat_mixed_3_chconcat -> ch_concat_mixed_3_chconcat
== bottom[i]->shape(j) (13 vs. 12) All inputs must have the same shape, except at concat_axis.

I've tried it on both nvcaffe 0.14 and nvcaffe 0.15
Do you have any idea?

deploy.prototxt

Can you send me the file of deploy.prototxt? When I try to write it , it always wrong. I need your help.Thanks.

CUDNN_STATUS_BAD_PARAM when trying to perform forward pass

Hi,
I'm interested in testing the speed and memory-consumption of your network, and so I downloaded the files linked to in the 'Training on TINY SET' section. I downloaded branch 0.15.5 of NVIDIA/caffe, and compiled it with CuDNN v.4 enabled. So far, so good. But, when I try the benchmark (caffe time) on deploy.prototxt, using the GPU, I get the following error:

Check failed: status == CUDNN_STATUS_SUCCESS (3 vs. 0) CUDNN_STATUS_BAD_PARAM
*** Check failure stack trace: ***
@ 0x7fc1a5c9e778 (unknown)
@ 0x7fc1a5c9e6b2 (unknown)
@ 0x7fc1a5c9e0b4 (unknown)
@ 0x7fc1a5ca1055 (unknown)
@ 0x7fc1ab5d39bd caffe::CuDNNConvolutionLayer<>::Forward_gpu()
@ 0x4099c2 time()
@ 0x405ca3 main
@ 0x7fc19e56bb45 (unknown)
@ 0x40644c (unknown)
@ (nil) (unknown)

Running it on the CPU works. Do you have any idea what could be causing this?
Thanks!

Different structure with paper?

Hi, I'm looking at the train_val.prototxt.
After conv_4_4_conv2d_relu blob, it should be another convolution 3x3/1 which is from paper.

So in paper image-conv-conv-conv(padded)-pool-conv-conv-conv
but in here image-conv-conv-conv(padded)-pool-conv-conv-pool

anyone can help me out?

scale_filler" is not recognized field name in BatchNorm

I keep getting the above error on running this prototxt through caffe. Have you seen this before?