Training script
Closed this issue · 3 comments
Hello, @XinArkh and Thanks for sharing your implementation. Actually, I am implementing this paper and I think the order of the concatenation is not correct w.r.t the paper (See fig5 in the paper)
Lines 146 to 147 in cd0a58f
I think should be replaced like this:
self.res5c_branch2a_feat = tf.concat([self.res5c_delta_x, self.res5c_delta_y, self.res5c_delta_z, self.bn5c_branch2a, self.res5c_bone_length],
.
I want also to ask you if you can share the training script?
Hi! Thank you for your enthusiasm!
I have checked it again and found it a bit of misleading in figure 5.
In this implementation, the order of the components to concat is the same as that in the original caffe code provided by the author:
layer {
name: "res5c_branch2a_feat"
type: "Concat"
bottom: "res5c_branch2a"
bottom: "res5c_delta_x"
bottom: "res5c_delta_y"
bottom: "res5c_delta_z"
bottom: "res5c_bone_length"
top: "res5c_branch2a_feat"
concat_param {
axis: 1
}
}
(line 1212 - line 1224)
Maybe it would be helpful if you contact the author to ask for the original model for further research.
As for the training script: I'm so sorry but I don't have it either. Looking forward to your contribution!
@XinArkh Thanks for your answer. I have contacted the authors many times and they did not answer me. I have some questions and requests:
The figure 5 also contains another problem regarding your jupyter notebook because:
The output size of the resNet50 part is w/16, h/16 as indicated in the jupyter notebook 368/16 =23. However in the figure 5 the size were indicated h/8, w/8 which is 46!!!!!?? The same for the last output where figure 5 indicate h/4 but in the code h/8=46 # h/4 = 92 !!! So, I don't know with all these errors how the authors got good accuracy and real time
- Did you get the 30 frames per second ?
- Could you share the code (I know maybe your not allowed to do that)?
- Could you share the caffe weights?
Thanks
That's a good question. I guess the author made a mistake calculating the scale number in figure 5, otherwise it would be contradictory to the source code.
I can get 10~12 (mostly >11) fps in my lab computer with a gtx1060.
I'm afraid I'm not permitted to give you the model directly. I remember I have experienced this too. The author replied to me several weeks after I requested. They just deal with plenty of request mails at one time. Be patient and good luck!