Trigger word detection from Zero

Course Q&A Deep Learning Specialization Sequence Models
1

Dear Community,

I tried to build a set of 4,000 examples from the sounds given in the course. And from 10-seconds Youtube audio snippets that I manually tagged, a development set. I did not get the result that was expected. I think the model cannot generalize based on the training set.

What can I do wrong?
Is there perhaps no detail that has been omitted from the curriculum?

Here you can see the history of the training:
Loss and F1 Score

IMG_20211105_000024
IMG_20211105_000024987×1383 94.3 KB

AUC and Accuracy
IMG_20211105_000047
IMG_20211105_000047986×1413 97.1 KB

Here you can see the output of the prediction and my result:
Prediction

IMG_20211105_214505
IMG_20211105_2145051080×687 118 KB

My result
IMG_20211105_214617
IMG_20211105_214617990×632 40.6 KB

Thank You for your help! :slightly_smiling_face:

2

@staff Could you help me, please?

3

I do not understand the data you presented.
You did not label the horizontal axis.
Why do some of the plots range from 0 to 100, and others range from 0 to 1400?

4

Because the 100 is epoch, the 1371 (1400) is datapoint from the audio. These are autorange ticks.

5

Where did your “output of the prediction” plot come from?

6

From the original pretrained model that is the sequence model part of deep learning specialisation.

Below that is the result of my model that I taught from zero.

7

How large are your training and test sets?
I’m not sure what you mean by “a development set”.

8

In the course the traning set is 4000 and the development/validation set is 26 large. In my case the training set size is 4000 and the validation set is 50.

By the way the the training set with which you teach the model. The development set with which you test the goodness of different models or hyperparameters, etc. The test set is used for the last test before the model is moved to the final live environment. This was very well explained in the specialization.

9

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?

10

Months have passed. I visited DeepLearning.AI at a number of contacts and they did not respond anywhere about their course. This is a bit of a characteristic of the company’s current mentality. They don’t pay attention to their students.