Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

Question about accuracy --> 0

See original GitHub issue

Hi,

a question: I applied OpenNRE both with our own dataset (created via DBpedia) and KBP37. When calling train_demo.py, eg. with the KBP37 datasets, during Epoch 1, loss is around 1.7-2.0, and accuracy at 0.18. Then at the end of Epoch 1, loss goes to ‘inf’, and starting from Epoch 2 accuracy is 0.0, and it stays there.

The same thing happens with my other dataset (based on DBpedia).

I am new to DL, is there something I do wrong, or is the dataset too small, …?

Thank you very much for the feedback, Gerhard

bash $ python2.7 train_demo.py 
Dataset Name kbp37 

Pre-processed files exist. Loading them...
Finish loading
Total relation fact: 12628
Pre-processed files exist. Loading them...
Finish loading
Total relation fact: 3295
Start training...
2018-10-23 10:02:47.737580: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
Calculating weights_table...
train_demo.py:99: RuntimeWarning: divide by zero encountered in divide
  _weights_table = 1 / (_weights_table ** 0.05)
Finish calculating
###### Epoch 0 ######
epoch 0 step 78 time 0.79 | loss: inf, not NA accuracy: 0.178651, accuracy: 0.17848177804
Average iteration time: 0.860235
Testing...
Calculating weights_table...
Finish calculating
[TEST] step 20 | not NA accuracy: 0.000000, accuracy: 0.036012
[TEST] auc: 0.0468855111775
Finish testing
Best model, storing...
Finish storing
###### Epoch 1 ######
epoch 1 step 78 time 0.88 | loss: nan, not NA accuracy: 0.000000, accuracy: 0.000949
Average iteration time: 0.822871
Testing...
Calculating weights_table...
Finish calculating
[TEST] step 20 | not NA accuracy: 0.000000, accuracy: 0.036012
[TEST] auc: 0.0468855111775
Finish testing
###### Epoch 2 ######
epoch 2 step 78 time 0.87 | loss: nan, not NA accuracy: 0.000000, accuracy: 0.000949
Average iteration time: 0.844874
Testing...
Calculating weights_table...
Finish calculating
[TEST] step 20 | not NA accuracy: 0.000000, accuracy: 0.036012
[TEST] auc: 0.0468855111775
Finish testing
###### Epoch 3 ######
epoch 3 step 78 time 0.86 | loss: nan, not NA accuracy: 0.000000, accuracy: 0.000949
Average iteration time: 0.885802
Testing...
Calculating weights_table...
Finish calculating
[TEST] step 20 | not NA accuracy: 0.000000, accuracy: 0.036012
[TEST] auc: 0.0468855111775
Finish testing
###### Epoch 4 ######
epoch 4 step 78 time 0.83 | loss: nan, not NA accuracy: 0.000000, accuracy: 0.000949
Average iteration time: 0.877218
Testing...
Calculating weights_table...
Finish calculating
[TEST] step 20 | not NA accuracy: 0.000000, accuracy: 0.036012
[TEST] auc: 0.0468855111775
Finish testing

Issue Analytics

State:
Created 5 years ago
Comments:12

Top GitHub Comments

1reaction

gwohlgencommented, Jul 30, 2019

Hi @pvthuy ah, I just see the variable name is _weights_table.

For example in train_demo.py:

            #_weights_table = np.zeros((self.rel_tot), dtype=np.float32)
            _weights_table = np.ones((self.rel_tot), dtype=np.float32)

0reactions

pvthuycommented, Jul 30, 2019

Hi @gwohlgen , could you let me know which part of codes that you modified? I couldn’t find _weights_matrix in the code.

Thanks!

Top Results From Across the Web

What's the best way to represent accuracy of a (0,0) target

Please be sure to answer the question. Provide details and share your research! ... Asking for help, clarification, or responding to other answers ......

771 questions with answers in ACCURACY | Science topic

Review and cite ACCURACY protocol, troubleshooting and other methodology information | Contact experts in ACCURACY to get answers.

Why is the accuracy for my Keras model always 0 when ...

Regarding your second question, 'accuracy' indeed can be used only for classification (as it measures the percentage of the correct labels).

Solved Baseline Accuracy 0/3 points (graded) Now, uncomment

Question : Baseline Accuracy 0/3 points (graded) Now, uncomment the relevant lines in main.py and report the training and validation accuracies of each ......

Why Accuracy is zero - MATLAB Answers - MathWorks

I showed in https://www.mathworks.com/matlabcentral/answers/894727-why-accuracy-is-zero#comment_1699059 that you can get up to 19% with those exact same ...