zsef123 / efficientnets-pytorch Goto Github PK

I have converted Tensorflow weights to Pytorch weights perfectly using your code in EfficientNet-b0, but I got some wrong from EfficientNet-b1 to EfficientNet-b3.
I would like to ask if you have completed the code, and can you help fix these bugs?
Traceback (most recent call last):
File "convert.py", line 117, in
convert_MBConv(mbconv, key_to_param(tf_mbconv))
File "convert.py", line 68, in convert_MBConv
convert_se(m.se, tf_params[10:14])
File "convert.py", line 55, in convert_se
convert_conv(m.se[1], tf_params[0], tf_params[1])
File "convert.py", line 19, in convert_conv
m.weight.data = torch.from_numpy(np.transpose(weight, (3, 2, 0, 1)))
File "/home/xux/anaconda3/envs/tf1.9_py3.6/lib/python3.6/site-packages/numpy/core/fromnumeric.py", line 639, in transpose
return _wrapfunc(a, 'transpose', axes)
File "/home/xux/anaconda3/envs/tf1.9_py3.6/lib/python3.6/site-packages/numpy/core/fromnumeric.py", line 56, in _wrapfunc
return getattr(obj, method)(*args, **kwds)
ValueError: axes don't match array

I download eval_ckpt_main.py and some other files from Official TF Repo, I don't know if they are necessary. Can you give me some help?

How can I use the code for training on multi GPUs?

Hi

what is the meaning of depth_div used in new_ch?

https://github.com/zsef123/EfficientNets-PyTorch/blob/master/models/effnet.py#L81

HI @zsef123
Great work!
You said in another issue that you are gonna implement the stochastic depth. Just a quick note that original Efficient Net uses linear schedule which is the same as the Stochastic depth paper, where Layer 0 has drop rate 0 and the last layer has drop rate 0.2.

Thank you!

I just use your model code, and report the error like this

AttributeError: module 'torch.nn' has no attribute 'Identity'

Can you please help me, how can I plot those figure on papar:)

self.skip always equal False?

Hi,

Thanks for your code. Did you do some test about the code? Specifically, what accuracy did your get using the code?

Thanks a lot