zhaoweicai/hwgq

Caffe implementation of accurate low-precision neural networks

hwgq's Issues

Exception: Only binarized weights supported when use deploy_bw.prototxt in FINN

Hi,
I would to generate an hardware design through FINN, passing it a prototxt of any BNN.
I don't understand if the folder https://github.com/zhaoweicai/hwgq/tree/master/examples/imagenet already contains prototxts suitable for FINN (as well as caffemodels in https://github.com/zhaoweicai/hwgq#models).
As the guide of FINN says, I have tried to run:
python FINN/bin/finn --device=pynqz1 --prototxt=FINN/inputs/deploy_bw.prototxt --mode=estimate
in which deploy_bw.prototxt is the prototxt contained in https://github.com/zhaoweicai/hwgq/tree/master/examples/imagenet/alex-hwgq-3ne-clip-poly-320k
but i get the exception:
....
File "/home/user/FINN/FINN/backend/fpga/backend_fpga.py", line 64, in passConvertToFPGALayers ret += [layers_fpga.FPGABipolarConvThresholdLayer(L)] File "/home/user/FINN/FINN/backend/fpga/layers_fpga.py", line 337, in __init__ raise Exception("Only binarized weights supported") Exception: Only binarized weights supported

Can you clarify me this doubt?
Thanks,
Sara

Share the VGG-small model on CIFAR-10

Hi,

Thanks for sharing your quantization flow. Is it possible to make the model and training script for the VGG-small model on CIFAR-10 available? Appreciate your help!

The weights of the trained model seem to not be binarized

Hi,
I trained the model VGG-Net on CIFAR10.
And then I tried to print the weights of the caffe-model. But there seem to be problems with the weights.
Hope you could give me some tips on these.
Many thanks.

The 'conv2_1' looks like,

[[-2.10925471e-02 6.74787164e-03 1.83824124e-03]
[ 1.17361760e-02 4.95712832e-02 1.15355672e-02]
[ 1.40904170e-03 1.91040952e-02 2.68753860e-02]]

[[ 7.47812912e-04 9.25247837e-03 -1.92467440e-02]
[ 6.26096362e-03 6.52389135e-03 -2.91604549e-02]
[ 1.54679930e-02 1.69047303e-02 -1.29050175e-02]]

[[-1.59913283e-02 -4.31000069e-03 -9.07981582e-03]
[-2.14854758e-02 4.27068589e-04 -3.39677893e-02]
[ 1.60318725e-02 -2.15732064e-02 -2.49724630e-02]]

[[ 2.94211088e-03 3.10821529e-03 -1.46567877e-02]
[-2.81586102e-03 1.56722255e-02 -4.10768725e-02]
[ 1.45177245e-02 8.73563997e-03 -2.80519202e-02]]

[[-3.02609615e-02 -2.83134021e-02 -3.68605666e-02]
[-3.66647467e-02 -1.59114692e-02 -2.45084912e-02]
[-1.47473682e-02 -3.27019729e-02 -1.81703269e-02]]

[[-2.29146797e-02 -4.16266685e-03 -1.24716444e-03]
[ 2.29197666e-02 1.97226536e-02 -1.27944229e-02]
[ 9.49834287e-03 2.72172503e-02 -2.50766743e-02]]

[[ 1.04283448e-02 2.07192469e-02 -1.10709341e-02]
[-7.62524176e-03 -3.04542063e-03 -9.67555027e-03]
[-4.35589701e-02 -3.34056690e-02 -6.37046574e-03]]

[[-2.52210582e-03 3.50272022e-02 -1.26641279e-03]
[ 1.27567565e-02 4.36151065e-02 2.20290497e-02]
[ 2.17617508e-02 3.94680649e-02 -8.48370232e-03]]

[[-2.48368196e-02 2.40874123e-02 3.73369120e-02]
[ 2.25116219e-02 3.33309509e-02 5.79682887e-02]
[-1.95888728e-02 3.29406597e-02 3.80768776e-02]]

[[-1.85104611e-03 -1.77885883e-03 -1.04705431e-02]
[ 2.15757657e-02 5.22872880e-02 2.86879446e-02]
[ 4.90898751e-02 6.27720580e-02 4.92752111e-03]]

[[-2.06267685e-02 7.74261449e-03 -2.70165708e-02]
[ 2.36926116e-02 4.17519510e-02 -1.96355823e-02]
[ 4.65140045e-02 3.26207504e-02 -1.86813949e-03]]

[[-2.06958782e-02 -1.99438701e-03 -5.76990540e-04]
[ 1.37184141e-02 1.84343923e-02 3.58887762e-02]
[ 1.97797827e-02 1.47607196e-02 9.19601228e-03]]

[[-3.14316433e-03 1.37923444e-02 2.48924959e-02]
[ 4.98341862e-04 1.73550507e-03 2.57142968e-02]
[ 1.15122795e-02 -3.53835919e-03 2.31737401e-02]]

The deploy.prototxt I used is:
`
name: "CIFAR10"
input: "data"
input_shape {
dim: 1
dim: 3
dim: 32
dim: 32
}
layer {
name: "conv1_1"
type: "Convolution"
bottom: "data"
top: "conv1_1"
convolution_param {
num_output: 128
pad: 1
kernel_size: 3
stride: 1
bias_term: false
}
}
layer {
name: "bn1_2"
type: "BatchNorm"
bottom: "conv1_1"
top: "conv1_1"

batch_norm_param {
use_global_stats: true
}
}
layer {
name: "qt1_2"
type: "Quant"
bottom: "conv1_1"
top: "qt1_2"
quant_param {
forward_func: "hwgq"
backward_func: "relu"
centers: 0.538 centers: 1.076 centers: 1.614
clip_thr: 1.614
}
}
layer {
name: "conv1_2"
type: "Convolution"
bottom: "qt1_2"
top: "conv1_2"
convolution_param {
num_output: 128
pad: 1
kernel_size: 3
stride: 1
bias_term: false
}
}
layer {
name: "pool1"
type: "Pooling"
bottom: "conv1_2"
top: "pool1"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "bn2_1"
type: "BatchNorm"
bottom: "pool1"
top: "pool1"
batch_norm_param {
use_global_stats: true
}
}
layer {
name: "qt2_1"
type: "Quant"
bottom: "pool1"
top: "qt2_1"
quant_param {
forward_func: "hwgq"
backward_func: "relu"
centers: 0.538 centers: 1.076 centers: 1.614
clip_thr: 1.614
}
}
layer {
name: "conv2_1"
type: "Convolution"
bottom: "qt2_1"
top: "conv2_1"
convolution_param {
num_output: 256
pad: 1
kernel_size: 3
stride: 1
bias_term: false
}
}
layer {
name: "bn2_2"
type: "BatchNorm"
bottom: "conv2_1"
top: "conv2_1"
batch_norm_param {
use_global_stats: true
}
}
layer {
name: "qt2_2"
type: "Quant"
bottom: "conv2_1"
top: "qt2_2"
quant_param {
forward_func: "hwgq"
backward_func: "relu"
centers: 0.538 centers: 1.076 centers: 1.614
clip_thr: 1.614
}
}
layer {
name: "conv2_2"
type: "Convolution"
bottom: "qt2_2"
top: "conv2_2"
convolution_param {
num_output: 256
pad: 1
kernel_size: 3
stride: 1
bias_term: false
}
}
layer {
name: "pool2"
type: "Pooling"
bottom: "conv2_2"
top: "pool2"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "bn3_1"
type: "BatchNorm"
bottom: "pool2"
top: "pool2"
batch_norm_param {
use_global_stats: true
}
}
layer {
name: "qt3_1"
type: "Quant"
bottom: "pool2"
top: "qt3_1"
quant_param {
forward_func: "hwgq"
backward_func: "relu"
centers: 0.538 centers: 1.076 centers: 1.614
clip_thr: 1.614
}
}
layer {
name: "conv3_1"
type: "Convolution"
bottom: "qt3_1"
top: "conv3_1"
convolution_param {
num_output: 512
pad: 1
kernel_size: 3
stride: 1
bias_term: false
}
}
layer {
name: "bn3_2"
type: "BatchNorm"
bottom: "conv3_1"
top: "conv3_1"
batch_norm_param {
use_global_stats: true
}
}
layer {
name: "qt3_2"
type: "Quant"
bottom: "conv3_1"
top: "qt3_2"
quant_param {
forward_func: "hwgq"
backward_func: "relu"
centers: 0.538 centers: 1.076 centers: 1.614
clip_thr: 1.614
}
}
layer {
name: "conv3_2"
type: "Convolution"
bottom: "qt3_2"
top: "conv3_2"
convolution_param {
num_output: 512
pad: 1
kernel_size: 3
stride: 1
bias_term: false
}
}
layer {
name: "bn3"
type: "BatchNorm"
bottom: "conv3_2"
top: "conv3_2"
batch_norm_param {
use_global_stats: true
}
}

layer {
name: "relu3"
type: "ReLU"
bottom: "conv3_2"
top: "conv3_2"
}
layer {
name: "pool3"
type: "Pooling"
bottom: "conv3_2"
top: "pool3"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}

layer {
name: "fc4"
type: "InnerProduct"
bottom: "pool3"
top: "fc4"
inner_product_param {
num_output: 10
}
}
layer {
name: "prob"
type: "Softmax"
bottom: "fc4"
top: "prob"
}
`

HWGQ: binarized weights in your AlexNet_HWGQ_BW is not low precision(1 bit fixed point number)

Hi，

I read your paper and found something I don't understand.

1, How to set the bit number of the weight w in the code, and activation and grad bit number.

2, How to convert a trained weight file like AlexNet_HWGQ to a 1-bits or n-bit fixed-point number

3,I extracted the weights in the AlexNet_HWGQ_BW file. It does not seem to be the 1-bit fixed-point numbers mentioned in the paper. That is, w is not a low-precision weight, but a binary real value.

I am looking forward to your reply！ Thank you.

How to save the binarized model?

According to my understanding, the weights of the trained model were not binarized directly.
So, do I need to binarize the trained model with parameter 'scale' and then re-save the model?