what tool did you use to convert mxnet to caffe ?

[GarrickLin]

I guess you are using MXNet2Caffe and it hard coded batchnorm eps as 0.001 but eps=2e-05 in mxnet. I will test it.

[nttstar]

@GarrickLin Thank you. I didn't find out it. Please tell me the result once you finish it~

[GarrickLin]

After replacing eps from 0.001 to 2e-05 in prototxt file, LFW result improved at 99.50%. There is still remaining some difference.

[nttstar]

Can you share the converting scripts here?

[GarrickLin]

model.prototxt.txt

[nttstar]

@GarrickLin I mean the modified MXNet2Caffe program, can you share it? At least the input layers, PRelu layers must be adjusted to match our pretrained model.

[GarrickLin]

GarrickLin/MXNet2Caffe/blob/master/prototxt_basic.py

Or you can just replace all 0.001 to 2e-05 in model.prototxt.

[nttstar]

@GarrickLin Thank you very much. Then other caffe guys can use it.

[GarrickLin]

Can you achieve the same result as mxnet model after modifying eps ?

[nttstar]

@GarrickLin I do not have time to test using Caffe model... I think someone else may want to.

[GarrickLin]

GarrickLin/MXNet2Caffe/blob/master/mxnet2caffe.py

I have fix the prelu support, and now the results are exactly the same.

[nttstar]

@GarrickLin Very promising, thank you! Which model did you test? Resnet50?

[nttstar]

@GarrickLin Can I copy your code to this repo?

[GarrickLin]

Yes you can

[GarrickLin]

I have just tested with LResNet34E-IR

[anguoyang]

hi@GarrickLin @nttstar, thank you all for your great help. I have tried your tool to convert LResNet50, it generated only the input layer for caffe:

name: "mxnet-mdoel" layer { name: "data" type: "Input" top: "data" input_param { shape: { dim: 1 dim: 3 dim: 64 dim: 64 } } }

[nttstar]

@GarrickLin Can you help?

[ysc703]

@GarrickLin @nttstar I got the same file like anguoyang, and the message is as follow: 0, op:null , name:data -> data 1, op:_copy , name:id -> id use shared weight -> data {'bottom': ['data'], u'inputs': [[0, 0, 0]], u'name': u'id', u'op': u'copy', 'params': ['data'], 'share': True, 'top': u'id'} Warning! Unknown mxnet op:_copy_

Can you give more details about how to convert the insightface model from mxnet to caffe? Thank a lot!

[YoWhatever]

@ysc703 Same problem! @GarrickLin could you help?

[GarrickLin]

I am really busy nowadays. I'll take care of this as soon as I have time.

[myfighterforever]

@YoWhatever @ysc703 @anguoyang I have met the same problem. Did you fix it?

[GarrickLin]

LResNet50.prototxt.txt what you need to do is just skip those layers

[ysc703]

@GarrickLin Thanks for the the proto file and the recommendation. It worked after I added some codes in json2prototxt.py and prototxt_basic.py as fellow: In json2prototxt.py _if str(node_i['op']) == '_copy' or str(node_i['op']) == '_minus_scalar' or \ str(node_i['op']) == '_mul_scalar' or str(nodei['op']) == 'Dropout': continue In prototxt_basic.py _if info[attrstr]['act_type'] == 'prelu': ... txtfile.write(' type: "PReLU"\n')

Although the bottom names of conv0 layer and pre_fc1 layer need to be modified, it is really nice. Thanks a lot!

[aitaoquanyi]

@ysc703 can you share your modify code? thanks!

[ewrfcas]

maybe this: in "json2prototxt.py": if str(node_i['op']) == '_copy' or str(node_i['op']) == '_minus_scalar' or str(node_i['op']) == 'mul_scalar' or str(node_i['op']) == '_mul_scalar' or str(node_i['op']) == 'Dropout' or (str(node_i['op']) == 'null' and str(node_i['name']) != 'data'): continue

in "prototxt_basic.py": def LeakyReLU(txt_file, info): if info[attrstr]['act_type'] == 'prelu': txt_file.write('layer {\n') txt_file.write(' bottom: "%s"\n' % info['bottom'][0]) txt_file.write(' top: "%s"\n' % info['top']) txt_file.write(' name: "%s"\n' % info['top']) txt_file.write(' type: "PReLU"\n') txt_file.write('}\n') txt_file.write('\n') else: raise Exception("unsupported Activation")

I am not sure about whether "elu_param { alpha: 0.25 }" is required for PReLU. The bottom names of conv0 and pre_fc1 should be change to "data" and "bn1", because some layers are removed.

[class8hawk]

@ewrfcas thanks

[GarrickLin]

MXNet2Caffe is updated now

[GarrickLin]

@ewrfcas yes you should manually modify the prototxt file

[nwdxbx]

@nttstar @GarrickLin when I use your MXNet2Caffe code to convert mxnet LMobileNetE to caffe . It appear: Check failed: status == CUDNN_STATUS_SUCCESS (4 vs. 0) CUDNN_STATUS_INTERNAL_ERROR. But when I convert mxnet model-r50-am-lfw to caffe,it's good. Did you fix it?

[GarrickLin]

@xubenxiang Sorry, I didn't try LMobileNetE.

[nwdxbx]

@GarrickLin when you free,can you try it. Thanks

[anguoyang]

@xubenxiang I recommend you to use mxnet instead, as the conversion tools are always with bugs.

[GarrickLin]

@xubenxiang where is the model file ?

[nwdxbx]

@GarrickLin I'm sorry. I don't know how can give you the model file. I planed use github,the github's file size limit of 100.00MB，but my param file is 115.8M. External network is restricted too.

[GarrickLin]

@xubenxiang LMobileNetE has depthwise convolution which is not well optimized in caffe. Make sure you really need this model in caffe, or you can get faster performance in mxnet.

[nwdxbx]

@GarrickLin Thank you very much.I'll give up LMobileNetE

[WeicongChen]

This line is wrong @GarrickLin https://github.com/GarrickLin/MXNet2Caffe/blob/d77bc61c99e3fbc03e79bcccfac31aa0827de2fe/prototxt_basic.py#L34 When no_bias is true, bias_term should be false.

[GarrickLin]

@juicecwc thank you, will you submit a PR ?

[WeicongChen]

ok.

[nwdxbx]

@GarrickLin Have you test the caffe model? I test the pretrained caffe-r34-amf model,but the accruacy is very low. Probe score matrix size: 3530 3530 distractor score matrix size: 3530 1000000 Done loading. Time to compute some stats! Finding top distractors! Done sorting distractor scores Making gallery! Done Making Gallery! Allocating ranks (1000080) Rank 1: 0.232038 I don't kown where is wrong?can you give me some advice?

[GarrickLin]

@xubenxiang I didn't run a LFW or megaface benchmark. But I have tested with the output of caffe and mxnet models, the results are exactly the same for the first 5~6 decimal places. You can try following tips:

1. try to test the output of both model with the same image input
2. make sure you did a preprocessing before sending an image tensor to caffe model, since mxnet doing the normlization during inference, this might be a little bit difference
3. make sure your input data layout is NCHW, which is the same as mxnet but different with traditional caffe model

[nwdxbx]

@GarrickLin I got feature code is here: def do_flip(data): for idx in xrange(data.shape[0]): data[idx,:,:] = np.fliplr(data[idx,:,:])

class Resnet_caffe(object): def init(self,deploy_net=deploy_net,caffe_model=caffe_model): caffe.set_device(0) caffe.set_mode_gpu() self.net = caffe.Net(deploy_net,caffe_model,caffe.TEST)

def get_feature(self,img):
    img = cv2.cvtColor(img,cv2.COLOR_BGR2RGB)
    img = cv2.resize(img,(112,112))
    #img = img/127.5 -1
    img = np.transpose(img,(2,0,1))

    embedding = None
    for flipid in [0,1]:
        _img = np.copy(img)
        if flipid==1:
            do_flip(_img)
        #self.net.blobs['data'].data[0,...] = _img  #np.expand_dims(_img,axis=0)
        self.net.blobs['data'].data[...] = np.expand_dims(_img,axis=0)
        out = self.net.forward()
        _embedding = self.net.blobs['fc1'].data[0].flatten()
        if embedding is None:
            embedding = _embedding
        else:
            embedding +=_embedding
    _norm = np.linalg.norm(embedding)
    embedding /= _norm

    return embedding

do you find some wrong?

[GarrickLin]

you comment this line

#img = img/127.5 -1

which should be uncomment, and fix a bug

img = (img-127.5) / 128.

[nwdxbx]

@GarrickLin 2. since mxnet doing the normlization during inference, this might be a little bit difference but I see in author's code ,I did't find this operator img = img/127.5 -128.This is you say mxnet doing the normlization during inference?

[GarrickLin]

mxnet doing this during inference

img = (img-127.5) / 128.

but caffe not

[nwdxbx]

@GarrickLin I see. Thank you very much.

[marcosly]

@GarrickLin Have you succeed to convert the mobilenetv1 from mxnet to caffe model? The result seems different. Thx

[jeehyun100]

Hello. I have tried verification.py using your caffe model. Accuracy is 50%. I think I made big mistake in my caffe prediction code. So Can I get source for caffe inferrence you made?

[KaleidoZhouYN]

@marcosly did you find the reason? I found it dosen't work on mobilefacenet model.

[marcosly]

@KaleidoZhouYN it seems that the attr setting doesn't fix well.Some modification need to be done in MXNet2Caffe manually.It could be a large work.So I suggest to train it in caffe framwork