Course 5 - Week 3 - Trigger Word Detection : Training from Scratch

Course Q&A Deep Learning Specialization Sequence Models
1

Hello,

With Trigger-word-detection assignment, I accomplished good detection result with
Pre-trained model provided in the notebook. As shown below.

Test-Pre-Trained-Model
Test-Pre-Trained-Model806×666 86.1 KB

Then, I started additional target of training from scratch.

But test result after training shows much worse result than expected.
Test-Trained-From-Scratch

Can I get better result by just increasing even larger ‘nsamples’ like 4000,
and more epochs like 10,000 ?

Or there should be other things to do for ‘Training from Scratch’ showing better result?

Is there anyone successfully trained a new model with that notebook?

Thanks,
Peter

2

Today I increased ‘nsamples’ to 1,600. And epochs to 4,000.
But Trained results does not show improvement.

(Training was done on local Host by copying the assignment folder.
And it took about 6 hours on Linux host with RTX3060 GPU)

3

Increased the learning_rate to 0.0001, then got better result.

Train_210527
Train_210527813×203 32.4 KB

4

I have a same problem. I used Youtube data for training. But I used all the parameters as they were in the course. I tried to change the parameters but it didn’t get any better. I used data audmentation, a larger set of training, but nothing got better. Can my data or spectrogtam be the cause of the error?

Please help me.