Training Your First Model

Question 1

Main steps in training a basic image classifier?

Answer 1

Import libraries, load data, define labelling function, create DataLoaders, define model, training

Question 2

Code to create a model in a computer vision task?

Answer 2

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

Question 3

What is dls (DataLoaders) object?

Answer 3

Contains both training and validation data and specifies which data is used in training

Question 4

What is ImageNet?

Answer 4

> 14M photos, >21k classes. dataset used for pre-training image-based NN

Question 5

Can you solve non-image tasks using image classifiers?

Answer 5

Yes, if the data can be transformed to an image reasonably well

Question 6

from fastai.vision.all import *

Answer 6

Imports everything (*) from the fastai.vision library

Question 7

First line of code for all fastai notebooks

Answer 7

pip install fastbook

Question 8

path = untar_data(URLs.PETS)/’images’

Answer 8

• Downloads (compressed) dataset from the URL stored in URLs.PETS
  – Unpacks the dataset and returns the local path to the dataset
  – In particular, the ‘images’ subfolder contains the images of the dataset

Question 9

fdef is_cat(x): return x[0].isupper()

Answer 9

Defines labelling function, Here: If the file name starts with a capital letter (e.g., “Abyssinian_1.jpg”), it contains a cat

Question 10

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

Answer 10

– Creates a vision learner, i.e., we use images as inputs
– Specifies the data (dls), architecture (resnet34) and metric (error_rate)
– ResNets achieve state-of-the art results
– resnet: “Residual Neural Network”, specific model architecture
– 34 in resnet34 refers to the number of layers of the network (alternatives: 18, 50, 101 and 152)

Question 11

learn.fine_tune(1)

Answer 11

– This fits the model (more precisely: fine-tunes it)
– Per default (shown here):
– One epoch only on randomly initialized model head
– Number of epochs requested (here: 1) to fit the entire model, but updating weights of later layers faster than earlier layers

Question 12

One complete pass through the dataset

Answer 12

Epoch

Question 13

Advantages of Transfer Learning

Answer 13

Faster Convergence, Reduced Data Requirements, Better Performance, Reduced Computational Cost

Question 14

Which ML tasks can be transformed to image classification?

Answer 14

Time Series Classification, Text Classification, Genomic Sequence Analysis, Sensor Data Analysis, Graph Data Analysis, Tabular Data Classification

Question 15

Name five areas where deep learning is now the best tool in the world

Answer 15

Computer Vision, Natural Language Processing (NLP), - Speech Recognition, Autonomous Vehicles, Healthcare and Drug Discovery

Question 16

What was the name of the first device that was based on the principle of the artificial neuron?

Answer 16

Perceptron. It was developed by Frank Rosenblatt in 1958.

Question 17

How we avoid overfitting in deep learning?

Answer 17

Model is trained on 80% of available labeled data. Metrics are only computed on the validation data (20%). Resampling. Checking for hyperparameters

Question 18

What is test split?

Answer 18

Part of the data, only evaluated after choosing the hyperparameters

Question 19

What is transfer learning?

Answer 19

ML technique where model developed for one task is used as the starting point for a model on a different task

Question 20

What structure deep learning model have?

Answer 20

Layered. Later layers can find more complex features by combining simpler ones from earlier levels

Question 21

Types of layers

Answer 21

Input, hidden, output

Question 22

What is DataBlock?

Answer 22

The DataBlock class is the building block to preprocess the data. It is somewhat similar to task im mlr3.