Tf-pose-estimation: Error when running train.py

[Armen001]

when I run the train.py by
python3 train.py --datapath ../../annotations/annotations --batchsize=64 --lr=0.001 --modelpath bla --imgpath ../../images/
after printing out the below info
[2018-02-28 13:19:19,630] [train] [INFO] Restore pretrained weights...
the following errors occur:
2018-02-28 13:19:24.316570: W tensorflow/core/framework/op_kernel.cc:1198] Data loss: not an sstable (bad magic number)
........
Traceback (most recent call last):
File "/home/armen/.local/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1350, in _do_call
return fn(*args)
File "/home/armen/.local/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1329, in _run_fn
status, run_metadata)
File "/home/armen/.local/lib/python3.6/site-packages/tensorflow/python/framework/errors_impl.py", line 473, in __exit__
c_api.TF_GetCode(self.status.status))
tensorflow.python.framework.errors_impl.DataLossError: not an sstable (bad magic number)
[[Node: save_1/RestoreV2_2 = RestoreV2[dtypes=[DT_FLOAT], _device="/job:localhost/replica:0/task:0/device:CPU:0"](_arg_save_1/Const_0_0, save_1/RestoreV2_2/tensor_names, save_1/RestoreV2_2/shape_and_slices)]]

And when we run the train.py with the "--model=cmu" flag:
python3 train.py --model=cmu --datapath ../../annotations/annotations --batchsize=64 --lr=0.001 --modelpath bla --imgpath ../../images/
we get this error
OSError: Failed to interpret file '../models/numpy/openpose_coco.npy' as a pickle
@ildoonet is it familiar to you?

[ouceduxzk]

check your openpose_coco.npy file, is it valid ?

[Armen001]

I have no idea how to check it, but I downloaded it by this scrip:
https://github.com/ildoonet/tf-pose-estimation/blob/master/models/numpy/download.sh
That's why I think it should be valid.

[ouceduxzk]

its a npy file, a big one, if you just download it in one second, then it probably means you are not downloading the correct file, please have some common knowledge about a weight file, and try to solve such simple problems by yourself, for example, check the url in download.sh , paste into your brower and download that way, does it make any difference? then if it is , it means your file is downloaded wrong. Otherwise, it is some other problem. That's how people solve easy problems before asking

[scyoyo]

I also meet with this problem and waste one day to solve that. The thing is that wget in downlaod.sh doesn't work on mediafire.com but it will create a fake file that looks like working. Hope the author could fix this bug.

[ouceduxzk]

I like the buggy code here, people needs to solve those small problems themselves, which are not that hard and they can also avoid that in their own code

[Armen001]

Sorry for my english. I am a student and just getting experience in the domain. Though I finally was able to train. Thanks a lot @ouceduxzk ouceduxzk, you were rigth. The openpose_coco.npy was not valid, it was an html file. The error wich I get from mobilenet was again from the same mistake.

[daisukelab]

Summarizing here so that no more people lost by the download.sh...

1. Open your browser.
2. Copy and paste following address to your browser, one by one.
3. You will see DOWNLOAD button on the page, just click and download...

http://www.mediafire.com/file/kibz0x9e7h11ueb/mobilenet_v1_0.75_224.ckpt.data-00000-of-00001
http://www.mediafire.com/file/t8909eaikvc6ea2/mobilenet_v1_0.75_224.ckpt.index
http://www.mediafire.com/file/6jjnbn1aged614x/mobilenet_v1_0.75_224.ckpt.meta

Or we can simply visit tensorflow/models/research/slim/nets/mobilenet_v1.

Anyway thank you @ildoonet for sharing such a great work!