amirgholami / squeezenext Goto Github PK

Can we convert train_val.prototxt to deploy.prototxt for inference purpose ?

I try to make a graph file on Movidius NCS and this process is finished with next warning:

"Check failed: target_blobs.size() == source_layer.blobs_size() (3 vs. 5) Incompatible number of blobs for layer BatchNorm1".

The topology was 1.0-SqNxt-23v5. The prototxt file was prepared according to https://github.com/BVLC/caffe/wiki/Using-a-Trained-Network:-Deploy

I just wonder if you have the weight file for pretrained network on Imagnenet

Is this official implementation?

Hello

I failed to load the pre-trained models in Intel Caffe 1.1.0 (commit commit a3d5b022fe026e9092fc7abc7654b1162ab9940d).

I get the following error message:
I0613 13:24:30.011922 22693 net.cpp:261] Creating Layer Convolution7
I0613 13:24:30.011926 22693 net.cpp:1498] Convolution7 <- Convolution6
I0613 13:24:30.011931 22693 net.cpp:1498] Convolution7 <- Convolution2_p
I0613 13:24:30.011937 22693 net.cpp:1472] Convolution7 -> Eltwise1
F0613 13:24:30.011960 22693 blob.cpp:260] Check failed: count_ == other.count() (25088 vs. 1568)
*** Check failure stack trace: ***

I am hoping to convert your models to Keras/Tensorflow to try them out on a smaller scale problem using transfer learning.

Hi,

I downloaded your '2.0-SqNxt-23v5' caffemodel and tried converting it to CoreML model.
However, Core ML Tools said, "Caffe layer 'Scale1' is defined in the .portotxt file but is missing from the .caffemodel file".
Would you tell me your '2.0-SqNxt-23v5' model(_iter_150136.caffemodel) has 'Scale1' layer or not?

Hello!

Thank you all for the clear and complete paper you wrote! I managed to implement SqueezeNext in Tensorflow and am using it as a part of a CNN+RNN+Attention model to do OCR. Earlier, I designed the CNN part inspired by Tiny Darknet and made this structure:

type, filters, filter size, stride
conv 16 3x3 2x2
conv 32 3x3 2x2
conv 32 1x1 1x1
conv 64 3x3 1x1
conv 32 1x1 1x1
conv 128 3x3 2x1
conv 64 1x1 1x1
conv 256 3x3 1x1
conv 64 1x1 1x1
conv 512 3x3 2x1
conv 128 1x1 1x1
conv 1024 3x3 2x1
conv 64 1x1 2x1

Now, I implemented the 1.0-SqNxt-23v5 model (with some added strides to make sure the output is the same as the original model; from 128x64 input to 32x1 output). This results in about the same performance (in terms of accuracy and speed) on my target platform, an ARMv7 mobile chip (about the speed of a Raspberry Pi 2), and 1.5x slower on a GPU. Any intuitions if this result is to be expected? I would expect it to be either faster or give better results than my previous, simpler, model. Or should I just not expect your model to work as good on my very different dataset? (text lines)

Also, I noticed the training took a lot longer (about 3 times) to converge. Is that to be expected due to the massively increased network depth or is your network less optimal when running with batches on a GPU?

Difference between 1x_v1 and 1x_v1-G version of squeeze next, why you are getting better accuracy on G version. Do you have results for CIFAR-10 dataset?