Mendukung platfrom, mobile app, web.

Draft
Word 250
Backgroun/ introduction 1 paragraph, thesis stratement, grammar, vocabulary
Body of pragraph
Topic sentences
Supporting ideas
Concluding sentences (optional)
Conclusion
Coherence
Cohesive

List of References

1.
Question 1
Which of the following is true about training your model using data parallelism technique? Check
all that are true.

1 point

The full data set is split up and subsets of the data are stored across multiple machines

Weights from different machines are aggregated and updated into a single model.

All of the data is on 1 master machine, and copies of the data are then distributed to machines
having different model architectures based on their capacity of processing the data.

The same model architectures are used on different machines, and each machine processes the
entire data set.

2.
Question 2
In TensorFlow version 2, tf.distribute.Strategy class supports _______. Check all that apply.

1 point

Graph Mode

Eager Mode
3.
Question 3
Which of the following are true of both MirroredStrategy and TPU Strategy? Check all that are
true.

1 point

Uses multiple machines

Variables are synchronized (mirrored) across each replica of the model

Uses a single machine

The same model is replicated on each core.

4.
Question 4
To modify training code to work with Mirrored Strategy, which of the following should we do?
Choose all that apply.

1 point

Put the code that creates, compiles and fits the model inside the scope of “with
strategy.scope()”.

Put code that creates the model object inside the scope of “with strategy.scope()”.

Adjust the batch size to equal the batch size per replica times the number of replicas

Increase the batch size as long as the number is 2^n (e.g. 64, 128, 256 etc).

5.
Question 5
To modify training code to work with distributed data, which of the following should we do?
Choose all that apply.

1 point
Use strategy.reduce to aggregate the losses across the replicas.

Replace the code that updates the model weights (calculating loss, calculating gradients, and
applying the gradients) so that each training step handles all replicas at once.

Use strategy.experimental_distribute_dataset to convert training and test sets into

distributed datasets.

Use strategy.run to run the code that updates the model weights (calculating loss, calculating
the gradients, and applying the gradients).

6.
Question 6
To use the TPU strategy, there are some steps that you’ll take before running the training code.
Please think about which line of code implements each step and choose the set of code that
performs these steps in this order.

1 Get the TPU address

2 Find the TPU cluster

3 Connect to the TPU cluster

4 Initialize the TPU cluster

5 Create your TPU strategy

In a Multi-Class classification scenario, your model can identify all the different items and
people that are present in a given input image.

1 point

False

True

2.
Question 2
Which of the following statements correctly describes the difference between object detection
and object localization?

1 point

They both are the same.

Object localization is where you get a bounding box around the main subject of the image, while
in object detection you get a bounding box around all of the objects within an image.

Object detection refers to detecting the object within an image, while object localization gives
us the bounding box around that object.

Object detectionis where you get a bounding box around the main subject of the image, while in
object localization you get a bounding box around all of the objects within an image.

3.
Question 3
What is the method that locates an object(s) by labelling the pixels, where each similar object(s)
is assigned to the same class? Type your response here (two words, all lower case).

1 point
Enter answer here
4.
Question 4
In the context of Transfer Learning, the initial training task where the model learns reusable
patterns is called a downstream task.

1 point

True

False

5.
Question 5
Check all the scenarios in which Transfer Learning could be beneficial.

1 point

When the task you want to perform is a sub-task of an already trained, larger, model.
When you don’t have enough data for the task you want to perform, which resembles another
same or similar, already trained task.

To reduce computation and processing cost

To ensure better performance

6.
Question 6
What is the name of the built-in TensorFlow layer-type which you can use to increase the
dimensions of a 2D image ?

1 point

SampleUp2D

SampleIncrease

UpSampling

UpSampling2D

7.
Question 7
You have an image of dimensions 48 x 48, and you want to upscale it to 240 x 240 using the
built-in TensorFlow layer-type which is used to perform such a task (mentioned in Question 6).
What will you pass in as size=____?

1 point
Enter answer here
8.
Question 8
Consider the following code:

What does “include_top=False” mean ?

1 point
It discards the first layer of ResNet50 when initializing my_layer using it.

It sets the top most layers as untrainable of ResNet50 when initializing my_layer using it.

It randomly sets up the weights, instead of using that of ImageNet, for the top most dense
layers of ResNet50 when initializing my_layer using it.

It discards the top most layers of ResNet50 when initializing my_layer using ResNet50.

9.
Question 9
What is the name of the technique used in the output dense layer that is used to predict
Bounding Boxes ? (Hint: It is a one word answer)

1 point
Enter answer here
10.
Question 10
Check all the statements that are true regarding Intersection Over Union (IoU), with regards
to Bounding Boxes.

1 point

The closer the value of IoU is to 0 the poorer is the prediction of the bounding box.

The closer the value of IoU is to 0 the better is the prediction of the bounding box.

IoU is the area of intersection of the two boxes (true and predicted) divided by the total union
area of the two boxes.

The values of IoU range from 0 to all possible positive values.