Training your first Model

Question 1

What’s an epoch?

Answer 1

Epoch: One complete pass through the dataset

Question 2

What’s the rule of thumb when it comes to deep learning and recognizing categories?

Answer 2

If the human eye can recognize categories from the images, then a deep learning model should be able to do so too

Question 3

What’s overfitting and why should we avoid it?

Answer 3

It occurs when ML model learns not only the underlying patterns in the training data but also the noise and random fluctuations. As a result, the model performs exceptionally well on the training data but poorly on new, unseen data.

It is crucial to avoid overfitting to ensure that models generalize well and provide reliable predictions in real-world applications.

Question 4

How can Overfitting be avoided?

Answer 4

• More Training Data
• Simplifying the Model
• Regularization:
  
  • L1 and L2 Regularization: Penalty for large coefficients in the model’s objective function.
    - Dropout:
• Cross-Validation
• Early Stopping
• Pruning
• Data Augmentation

Question 5

Advantages of transfer learning

Answer 5

• Make use of pretrained models
• enables deep learning with small amounts of data

Question 6

What’s the key to avoid overfitting?

Answer 6

Resampling and test split

Question 7

How to avoid unintentional overfitting?

Answer 7

Split your data carefully into: train / validation / test sets

Question 8

Which tasks does the following code snippet accomplish?

learn = vision_learner(dls, resnet34, metrics=error_rate)

Answer 8

Define the metrics to track (here: error rate)

Define a vision learner with a resnet architecture

Question 9

Which statements are true about a DataLoaders object?

• It downloads the data from an URL
• It contains the training as well as validation data
• It contains the parameters of the final learned model for future access
• It specifies which data to use in training a deep learning model

Answer 9

• It contains the training as well as validation data
• It specifies which data to use in training a deep learning model

Question 10

Which statements are true about ImageNet?

• It is a classification task with >21k classes
• It is a neural network architecture with >14M parameters
• It contains >14M artificially generated images
• It can be used for pre-training image-based neural networks

Answer 10

• It is a classification task with >21k classes
• It can be used for pre-training image-based neural networks

Question 11

Can you solve non-image tasks using image classifiers? How well can you expect it to work?

Answer 11

Yes. If the data can be transformed to an image reasonably well, you can then apply standard image classification techniques. It delivers state-of-the-art results in some cases

Question 12

How can you counter overfitting in deep learning?

Answer 12

Validations sets can reduce the chance to overfit, though they need to be carefully designed for each use case. Otherwise you still may be overfitting unintentionally.

Question 13

Which library is mentioned as the most popular for low-level deep learning functionality?

A) TensorFlow
B) Keras
C) PyTorch
D) Scikit-learn

Answer 13

Answer: C) PyTorch

Question 14

What is the primary advantage of using transfer learning with pre-trained models?

A) It reduces the amount of labeled data needed for training.
B) It increases the complexity of the model architecture.
C) It requires extensive training time from scratch.
D) It eliminates the need for a GPU.

Answer 14

Answer: A) It reduces the amount of labeled data needed for training

Question 15

Given a dataset with 10,000 images, if you use a validation split of 20%, how many images will be used for training and how many for validation?

Answer 15

Training images: = 10,000×0.8=8,000
Validation images: 10,000×0.2=2,000

Question 16

If a model is trained for 5 epochs with a batch size of 32 and the dataset contains 1,024 samples, how many gradient updates (iterations) will be performed?

Answer 16

Total iterations per epoch:
1,024/32 =32
Total iterations for 5 epochs:
32×5 = 160

Question 17

Write the fastai code to create a learner with a ResNet34 architecture and train it for 1 epoch.

Answer 17

learn = vision_learner(dls, resnet34, metrics=error_rate)
learn.fine_tune(1)

Question 18

Explain the concept of transfer learning and its benefits in deep learning.

Answer 18

Transfer learning involves using a pre-trained model on a new, often smaller, dataset. The pre-trained model has already learned features from a large dataset (e.g., ImageNet), and these features can be useful for the new task. The benefits include reduced training time, lower need for large datasets, and often improved performance since the model starts with a good baseline of learned features .

Question 19

Describe the role and importance of GPUs in deep learning model training.

Answer 19

GPUs (Graphical Processing Units) are crucial in deep learning because they are optimized for the kind of matrix and tensor operations that are common in deep learning algorithms. They can process multiple parallel operations much faster than CPUs, significantly speeding up the training of large models .

Question 20

If a trained model shows a high error rate on the validation set but a low error rate on the training set, what issue is likely occurring and how can it be addressed?

Answer 20

Answer:
The model is likely overfitting, meaning it performs well on the training data but poorly on unseen validation data. To address overfitting, techniques such as regularization, dropout, data augmentation, or using a simpler model can be employed .

Question 21

Interpret the significance of the learning rate in the 1cycle training policy.

Answer 21

Answer:
The learning rate in the 1cycle training policy starts low, increases to a maximum, and then decreases again. This approach helps to navigate the loss landscape more effectively, avoiding sharp minima and promoting faster convergence. The high learning rate helps escape local minima, while the final low learning rate ensures the model settles into a smooth and well-generalized region of the loss landscape .