the forward_propagation、predict function batched forward and the signal_predict function, same input but predict output different. #45
Comments
|
So according to this: It seems in keras the same cell-state is used for both predicting and training. Do you need single_predict to use the same cell-state as well? |
|
I think it should be same cell. how about keras predict use one imge? |
|
So according to this: Keras (by default) resets its state after each sequence of batches. |
|
Sorry, now the result seems worse, predict and simgle_predict function output: training... forward_propagation inputdata------------------------------------ |
|
I fell that, batched forward the forward_propagation and predict is same fuction, but forward_propagation fuction is best,so predict is seems unnecessary. But the single_predict is the most common function, because we all predict with one image or one record. So concentrate on the single_predict function. |
|
Hi, I just added: To the Neural_Network class. In commit 05a9daf I am still debugging why there is a slight difference between the two. |
|
break!!!
But this net sturct is ok! I don't know why? |
|
Checking now. Are you testing in debug mode or release mode? |
I only test release mode. Need to test debug model? |
|
I do not (that would take very long too!). Just checking so I could replicate it |
|
Did it crash on the first call to predict? (Are you using malloc to create the neuralnetwork?) |
|
Yes, i create the neuralnetwork. Now use the first network struct, every time break, in release mode. I am test debug model. |
|
My example: |
|
I was able to reproduce the error. |
OK. How about this compile error? |
|
I added a fix for the predict_bug. 2405b82 |
Yes, now the forward_propagation and the predict function is same. We should focus on the single_predict function. |
|
How about them? Four days have no updata. You are busying in your work? |
|
Oh so sorry! I thought I had responded to you. I have been working on this branch: |
|
Hi! I just added max_pooling: #49 Next steps: Is single-predict still giving different results? |
yes, it is still a problem. Perhaps, you data is samll, but in my data, the litte different will result in the big different output data,this is the three function input and output, the single_predict the output is so big: Neural Network architecture: training... forward_propagation MAPE loss: 0.0321961 forward_propagation output predict data------------------------------------ predict inputdata------------------------------------ predict output predict data------------------------------------ single_predict inputdata------------------------------------ |
|
I think you shoul be the next steps: |
|
I found the issue with single_predict, some of the data wasn't being copied correctly in recurrent layers. |
I test it, it seems NO:
|
|
Did you use Example: |
last test, i have not use the copy_training_data_to_single_predict, so it seems not work. other problem, what the copy_training_data_to_single_predict function's parameter mean, 0 is ok? |
|
I believe single_predict is working: Added minor changes as of b812d74
Test: Output: |
Yes, now the predic and single_predict function, the same input and same output, but compare to forward_propagation function, them is not. It should be this, or still something wrong with it?
|
|
In windows predict and forward_propagation use the same cell-state so they should be different if you run them one after another. Do you need/want predict to have a seperate cell-state from forward_propagation? |
oh. It means it should be this, predict and forward_propagation? I don't know whether predict should be seperate cell-state, but if i load the model, whether have the same result as not load? |
|
Load is only saving the weights but it also needs to save the inputs/outputs of each layer. (However it is currently giving different results so its something I have to fix) |
|
OK. Thank you very much. |
|
So the network is saving variables and weights. |
|
This issue has been resolved, |
No description provided.
The text was updated successfully, but these errors were encountered: