@miaow1988 Would you please release the pretrained models of shufflenet v2 x1.0 and x2.0?

Hi, Can you please show me your solver file for training in caffe, thanks so much!

Hi,
In data augmentation, do you use random crop or random resized crop?
Thanks,
Yifan

How about your accuracy of 1.0x ?

I don't understand why

你好，我用shufflenet_v2_0.25的caffe模型跑imagenet val集分类精度只有0.37左右，
调用代码是这样的，

ransformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
transformer.set_transpose('data', (2,0,1)) # move image channels to outermost dimension
transformer.set_raw_scale('data', 1) # rescale from [0, 1] to [0, 255]
net.blobs['data'].reshape(1,3,224,224)

for file in filenames:
pic = os.path.join('data/ilsvrc12/val',file)
input = caffe.io.load_image(pic)
transformed_image = transformer.preprocess('data', input)
net.blobs['data'].data[...] = transformed_image

the first batch normalization layer is used to replace the mean subtract, but it should be used with affine=False.

this might be the issue that you can't reproduce the paper result.

hi,
I use your code with pytorch to train on imagenet. At the begin,the loss is very small,it is around 0.02.the speed of loss decrease is around 0.0001. do it ok?
thank you very much!

Traceback (most recent call last):
File "shufflenet_v2.py", line 224, in
net.convert_to_caffe(args.save_caffe)
File "shufflenet_v2.py", line 160, in convert_to_caffe
print(caffe_net.to_proto())
File "/home/orbosoham/freshCaffe/caffe/python/caffe/net_spec.py", line 193, in to_proto
top._to_proto(layers, names, autonames)
File "/home/orbosoham/freshCaffe/caffe/python/caffe/net_spec.py", line 97, in _to_proto
return self.fn._to_proto(layers, names, autonames)
File "/home/orbosoham/freshCaffe/caffe/python/caffe/net_spec.py", line 162, in _to_proto
assign_proto(layer, k, v)
File "/home/orbosoham/freshCaffe/caffe/python/caffe/net_spec.py", line 64, in assign_proto
is_repeated_field = hasattr(getattr(proto, name), 'extend')
AttributeError: group

Thanks for sharing those models, would u mind provide the X2.0 version of ShufflenetV2, thank you

I can not train the shufflenet to the accuracy in the paper, and there is a large margin between mine and that in the paper.
Could you please share your solver.prototxt?

Traceback (most recent call last):
File "/usr/local/lib/python3.6/dist-packages/caffe/net_spec.py", line 160, in _to_proto
_param_names[self.type_name] + '_param'), k, v)
KeyError: 'ShuffleChannel'

Have you modified the caffe source code to add a new layer ShuffleChannel?

If I don't load pre-trained pytorch model, the result of the converted model has diff within 0.001, but when I load the pre-trained model, the diff is very big. I don't know why, since the pre-trained model has been successfully loaded. BTW, the pretrained model is trained through multi-gpu.

When I convert the pytorch model to the caffe model, I found that 0.25x model can be converted. But more than 0.25x model can not be converted. I found that the problem is that when I executed the conversion code, gpu0 memory is filled up quickly, leading to errors.

F1031 20:37:10.787436 16330 cudnn_conv_layer.cpp:53] Check failed: status == CUDNN_STATUS_SUCCESS (4 vs. 0) CUDNN_STATUS_INTERNAL_ERROR
*** Check failure stack trace: ***
Aborted (core dumped)

I found that when the code is executed, there will be out of memory:
In def convert_to_caffe(self, name): in slim.py
caffe_net = caffe.Net(name + '.prototxt', caffe.TEST)

There is a conv layer at the last instead of a full connected layer. There isn't a dropout layer, I think it may lead to overfitting during training.
Actually I used the 1.0 model to train my dataset, and overfitting happens. (train acc 87 val acc 82)

as title

Hi, I downloaded the pretrained model in the release.
I tried to read the model (shufflenet_v2_x0.5.pth), the code

def test():
path = os.getcwd() + '/shufflenet_v2_x0.5.pth'
model = torch.load(path)
torch.save(model.state_dict(), os.getcwd() + '/params.pth')
model_object.load_state_dict(torch.load(os.getcwd() + '/params.pth'))

Error information here:

Traceback (most recent call last):
File "test.py", line 13, in
test()
File "test.py", line 9, in test
torch.save(model.state_dict(), os.getcwd() + '/params.pth')
AttributeError: 'collections.OrderedDict' object has no attribute 'state_dict'

the version of pytorch is 0.4.1. ( I am a newbee to pytorch and I am trying to convert the pytorch model to mxnet)

Could you kindly please provide the solution?

Thanks

hello,have you trained shufflenet_v2_x1.0 or shufflenet_v2_x1.5 in imagenet.

I use your code train a test model, just change the convolution layer to a linear layer, and a linear layer to the slim.py for conversion. When trying to convert the torch model to the caffe model, the mean diff is large.

I want to convert pytorch to caffe and support the linear layer. What other work needs to be done?

my codes:

# slim.py
if isinstance(m, nn.Linear):
    if m.bias is None:
        param = [dict(lr_mult=1, decay_mult=1)]
    else:
        param = [dict(lr_mult=1, decay_mult=1), dict(lr_mult=1, decay_mult=0)]
    inner_product_param = dict(
        num_output = m.out_features,
        bias_term = (m.bias is not None),
        weight_filler = dict(type='msra'),
    )
    layer = L.InnerProduct(
        layer,
        param=param,
        inner_product_param = inner_product_param,
    )
    caffe_net.tops[m.g_name] = layer
    return layer

Hi @miaow1988 ,

Nice work!
Would you please check the network config bellow for 32*32 dataset? Since only about 72% acc(1.5 version) would be obtained after 400 epoch using the same training paras as this repo.

    self.network_config = [
        g_name('data/bn', nn.BatchNorm2d(3)),
        slim.conv_bn_relu('stage1/conv', 3, in_channels, 1, 1, 1),   #3,2,1->1,1,1
        #g_name('stage1/pool', nn.MaxPool2d(3, 2, 0, ceil_mode=True)),  #removed
        (width_config[0], 2, 1, 4, 'b'),
        (width_config[1], 2, 1, 8, 'b'), # x16
        (width_config[2], 2, 1, 4, 'b'), # x32
        slim.conv_bn_relu('conv5', width_config[2], width_config[3], 1),
        g_name('pool', nn.AvgPool2d(4, 1)), #7->4
        g_name('fc', nn.Conv2d(width_config[3], self.num_classes, 1)),
    ]

Thanks very much for your help!

Why these is not depth wise convolution layer?

shufflenet_v2_x0.50, Top-1 Acc = 58.93%. This accuracy is 1.37% lower compared with the result in the official paper.

but I saw the paper, the x0.50 model only have 39.7% Top-1 Acc in Table 8.

In case of directly using the prototxt, please change the depthwise convolution layers type from Convolution into DepthwiseConvolution. After changing the 1.5x prototxt, the training speed up from ~14seconds/iter to 1s/iter.

Hi,
It is very nice of you sharing the code. Could you please also share the pretrained models?