C3 W3 Assignment - Training: low accuracy

Course Q&A Natural Language Processing NLP with Sequence Models
1

Although all tests are passed, my submission fails the “Training” portion of the grading.

My accuracy is only about .63 after 100 epochs, as opposed to .90 as suggested in the project. I’m using Adam(.01) as instructed. My model is learning, but not as quickly as the expected output.

Any thoughts about what I might be doing wrong?

Here is my output:
Step 1: Total number of trainable weights: 1780117
Step 1: Ran 1 train steps in 2.50 secs
Step 1: train CrossEntropyLoss | 4.16958475
Step 1: eval CrossEntropyLoss | 4.12566984
Step 1: eval Accuracy | 0.31596923

Step 100: Ran 99 train steps in 49.61 secs
Step 100: train CrossEntropyLoss | 3.53727651
Step 100: eval CrossEntropyLoss | 3.35681584
Step 100: eval Accuracy | 0.62917300

2

Hi @John8

The most common mistakes are made in data_generator() function but if your are absolutely sure it is correct, you might made some mistakes defining NER model with wrong layers? There is also a chance that there are small mistakes intrain_model() function. It is hard to advise by just seeing that model fails to train properly.

3

Thank you for the suggestions. It was an error in my NER model, wrong axis for LogSoftmax.