Course 3 Module 1 Programming Assignment

Course Q&A Professional Certificate:PyTorch for Deep Learning PyTorch: Advanced Architectures and Deployment
, ,
1

I pass all test. Why my model just have 0.56 Val acc after 20 epochs? Is this error?

image
image825Ă—393 11.7 KB

1 Like
2

Hi @Datyespro,

That result is actually quite normal for this specific assignment setup, so don’t worry, it does not seem to be an error in your code. (I just ran the assignment myself and got similar results)

The validation accuracy of ~56% is expected given the constraints here. As mentioned in Section 1.6 - Evaluating the Classifier, the primary goal of this assignment is to practice “building models with modular components” and understanding the construction process, rather than achieving high performance.

There are a few reasons why the accuracy isn’t higher:

  1. Data Limitations: You are using a very small, curated subset of the clothing-dataset-small. Deep learning models (especially CNNs) typically require massive datasets to generalize well.
  2. Simplified Pipeline: To keep the assignment runnable in a standard environment, you aren’t asked to use extensive data augmentation, regularization, or hyperparameter tuning required for production-level accuracy.
  3. Model Scale: The custom MobileNetLikeClassifier you built is a simplified version of larger, pre-trained networks like ResNet or the full MobileNetV2.

The real “win” here is that you successfully implemented the building blocks, stacked the backbone, and got the pipeline to train. Passing the unit tests confirms you have built these custom architectures, which is the main learning objective.

Best,
Mubsi

3 Likes
3

@Datyespro

if one focus on this results irrespective of assignment or general project one is working on, and you get. the results you showed (by the way even I got the same results in assignment)

one important thing to learn from an ai architectural perspective, one needs to more worry about val loss and train loss being 1.18, meaning the dataset one is working upon or has created ai model, is not sufficient, as the staff mentioned the goal for them was to more about execution of running ai model and it’s implementation.

but when you will applying implementation in real world! data, you need to look at the following

  1. if you have. enough data of you are getting a val loss of 1.18 and accuracy of 56, it is indicative of model being underfitting.
  2. as staff already mentioned in his response about less use data augmentation, the underfittting is most probably because we are working on a very small dataset, that would mean not able to detect more non-linear complexity of the dataset. Another reason also being creation of a much simpler model.

So you are very right in looking at every perspectives of ai model implementation and their respective outcomeand make sure when you create! any model these points to consider in learning data pattern relative to ai model creation..

Regards

DP

1 Like