C3W3 Assignment training time

[anoopebey]

Hi, Is it normal to have a quite high training time per epoch for the model?

[Deepti_Prasad]

Depends on your steps per epoch code. make sure you have recalled it correctly for both training and validation set correctly. check ungraded lab for hints to do the correction.

If not able to find, let me know.

Regards
DP

[anoopebey]

The steps per epoch was already defined in the colab and from the answer hints it seems to be the correct no. of steps (125).

Best Regards,
Anoop

[Deepti_Prasad]

Not for the cell you are sharing. Go back to previous cells where you need to define steps per epoch for both dataset as per batch size

[memoros77]

Hello, I having the same problem and in the last few days, there are several post with the same problem. Maybe @Deepti_Prasad you should review the notebook.

I passed all test, and increase the learning rate but I still have the same problem.

Greetings

[memoros77]

ok the problem is the batch size that isn’t part of the exercises. You should clarify that in the notebook.

[Deepti_Prasad]

If one reads all the instructions properly, can understand this part. Good that you got the catch and debugged yourself.

[memoros77]

@Deepti_Prasad Hello, I Passed the Assigment and download the h5 file

but when I try to summit assigment I got the error:

Could you help me? Please…

Greetings!

[Deepti_Prasad]

You might have passed the test, but the assignment is expecting you a desired score which you have not according to the grade. So retweak your assignment by going through each cell where changes can be made

[gent.spah]

Yes as Deepti says go back to your assignment and check all the cells, it might be the case that you iou score is right but some other cell code is not as expected. Also make sure you have the latest assignment with the default naming of the file!

[memoros77]

Hello, thank you very much for your answer:

There is a dependency problem when I install packages:

I got the correct conv block

I got correct the downsampling path:

I got correct the expected output of the FCN-Encoder:

The model behave as expected, decreasing the loss in training and validation datasets

I got a IOU score a least > than 0.6

I’ve downloaded the last homework colab.

Please, Could you Check if there is a problem with the assigment?

Greetings

[Deepti_Prasad]

Hello @memoros77,

Click on either mine or @gent.spah name and then message to send your downloaded notebook. Also kindly rename the assignment by your name as memoros copy while sending. Do not post the notebook or codes here. It is against community guidelines.

Regards
DP

[Deepti_Prasad]

Hello memories,

Can you go through ungraded labs once. Also sharing a similar post thread where issues could be resolved by some of the model adjust, you have few similar errors, kindly go through them make the correction. if issue still persist, send the updated notebook.

in your convolution, you are using unit of 3, kindly stick to unit, pool_size and stride of 2 for all the conv_block.

C3W3-Assignment-Image Segmentation of Handwritten Digits-Submission Error

Hello, as per discussion, you can save a copy of your work and re do your assignment. While copy and pasting some of the codes, you have edited the grader cells which can also cause grader failure due to metadata editing. Kindly refrain from copy and pasting codes. Every notebook gets updated with the metadata grader has assigned and when you copy and paste codes from outside, you will never pass the grades. Also this kind of practice is not allowed. Remember the graders will know when you…

Regards
DP

[anoopebey]

Hi, The problem seems to be arising from the GPU session not being activated in Colab. It seems that the colab session does not activate the GPU when running with TF 2.8. The same code runs with TF 2.15 and the time per epoch seems to be as expected. But, during assignment submission it runs into an issue as the autograder is not compatible with TF 2.15

[Deepti_Prasad]

Anoop,

Can you show me the screenshot of your submission grader output.

Your model could have issue as I can see your loss is not going towards 0.

Kindly have a look into your model algorithm once. Check your model compile what loss and optimizer you used. If your have followed the instructions mentioned in the assignment and as mentioned in post comment to memorose, that link also provide necessary guidance.

If still stuck then let me know

Regards
DP

[anoopebey]

Hi Deepti,

This is the colab output and the submission grader output I received.

[memoros77]

@anoopebey you are rigth I have de same problem, the colab doesn’t activate the GPU when running TF 2.8 that is required only works with TF 2.15.

That’s could be the problem with the autograder.

@Deepti_Prasad and @gent.spah

In response of the post thread:

“While copy and pasting some of the codes, you have edited the grader cells which can also cause grader failure due to metadata editing.”

Yes I have copy and paste code but from my local enviroment, because is more efficient and cheaper work there, we have to pay double, coursera and google colab to pass this assigntment.

“In your encoder cell, Zero padding’s layer is not followed as per the instructions given.”

The colab is given us the correct padding, we don’t have to do anything:

x = tf.keras.layers.ZeroPadding2D(((0, 0), (0, 96-input_width)))(img_input)

“In the decoder cell, the data_format is incorrect.” It is Ok

“Also while I was checking your codes you choice of activation in the decoder grader cell is incorrect.” It is Ok

I have taken a lots of courses from deeplearning and it’s first time that I pass all test and get the wrigth answer and the grader hasn’t given a correct evaluation neither a good retroalimentation of the error

The loss is decreasing and the accuracy is increasing and got the correct IOU this is frustrating

I’m not a full time student, I work 10 hours a day and I try to find time to finish the courses. I don’t have time to guess which of all the parameters match yours, much less if the self-evaluator doesn’t tell us where we have the error.

[anoopebey]

Hi @memoros77 , the workaround that worked for me was to run the code in a local machine with TF 2.10 and save the model. The model generated in this way was accepted by the auto-grader. No change in code was required.

[Deepti_Prasad]

Hello @memoros77,

corrections required.

1. in def conv_block
   your data_format code should be the recall function of image_ordering and not image_ordering.

2. in def FCN8
   In your Block codes as instructions given

3. We recommend keeping the pool size and stride parameters constant at 2.
   Please use unit also 2 and not 3 or all the conv_block

4. in the same def FCN8
   Upsample o above and crop any extra pixels introduced
   kindly remove the below codes
   use_bias = False, data_format=IMAGE_ORDERING

5. load the pool 4 prediction and do a 1x1 convolution to reshape it to the same shape of o above, remove data_format.

6. remove use_bias = False, data_format=IMAGE_ORDERING from code line
   upsample the resulting tensor of the operation you just did

7. in the below code line, again for data_Format use the recall function of image_ordering and not image_ordering
   load the pool 3 prediction and do a 1x1 convolution to reshape it to the same shape of o above

8. Remove use_bias = False, data_format=IMAGE_ORDERING from code line upsample up to the size of the original image.

9. for model compile statement, use Adam optimizer instead of SGD, momentum=0.9, nesterov=True

10. the epoch you used is too high, try using the same number of epoch as shown in the expected output which is 70.

Do these corrections and let me know once your issue is resolved.

Sorry for the delay in response as had other notebooks to review.

Regards
DP

[memoros77]

Deepti thank you very much for your feedback, I will make the changes and let you know if it works