Semantic Segmentation of Images with

Convolutional Networks

Lecture Notes on Deep Learning

Avi Kak and Charles Bouman

Purdue University

Monday 12th October, 2020 15:10

Purdue University 1
 Preamble

At its simplest, the purpose of semantic segmentation is to assign correct labels

to the different objects in a scene, while localizing them at the same time.

At a more sophisticated level, a system that carries out semantic segmentation

should also output a symbolic expression that reflects an understanding of the
scene that is based on the objects found in the image and their spatial
relationships with one another. This more ambitious definition of semantic
segmentation is still largely unfulfilled in computer vision, but continues to remain
as a goal to aim for.

This lecture on semantic segmentation is based on the first definition — being

able to segment out different objects based on their identity and being able to
localize them at the same time.

Semantic segmentation obviously involves classification, but it is NOT

classification at the level of the entire image.

Purdue University 2
 Preamble (contd.)

Since the overarching goal is to segment out of an image the objects that belong
to different categories, the classification that is needed for semantic
segmentation is at the pixel level.

Pixel-level classification needed for semantic segmentation means that

now you cannot use the more widely known convolutional networks that
are meant for whole-image classification. These better-known networks push
the image data through progressively coarser abstractions until at the final output
the number of nodes equals the number of classes to which an image can belong.

For semantic segmentation, a network must still become aware of the abstractions
at a level higher than that of a pixel — otherwise how else would the network be
able to aggregate the pixels together for any object — but, at the same time, the
network must be able to map those abstractions back to the pixel level.

Purdue University 3
 Preamble (contd.)
This calls for using some sort of an encoder-decoder architecture for a neural
network. The encoder’s job would be to create high-level abstractions in an image
and the decoder’s job to map those back to the image.

That what’s mentioned above could be implemented in a fully

convolutional network was first demonstrated by Ronneberger, Fischer and
Brox in the paper ”U-Net: Convolutional Networks for Biomedical Image
Segmentation” that you can download from this link:

https://arxiv.org/abs/1505.04597

The two key ideas that the Unet network is based on are: (1) The concept of a
“Transpose Convolution”; and (2) notion of skip connections (which is something
that should be very familiar to you by this time). The transpose convolutions are
used to up-sample the abstractions till we get back to the pixel-level resolution.
And the skip connections are used as pathways between the encoder part and the
decoder part in order to harmonize them.
Purdue University 4
 Preamble (contd.)
The main goal of this lecture is present my implementation of the Unet
under the name mUnet, with the letter ’m’ standing for “multi’ for mUnet’s
ability to segment out multiple objects of different types simultaneously
from a given input image. The original Unet, on the other hand, outputs a
single output mask for the segmentation. By contrast, mUnet produces
multiple masks, one for each type of object.

Additionally, mUnet uses skip connections not only between the encoder part and
the decoder part, as in the original Unet, but also within the encoder and within
the decoder.

You will find the code for mUnet in Version 1.0.9 of DLStudio that you can
download from:

https://pypi.org/project/DLStudio/

Make sure your install of DLStudio is 1.0.9 or higher.

Purdue University 5
 Preamble (contd.)

Regarding the organization of this lecture, since basic to understanding the notion
of “transpose convolution” is the representation of a 2D convolution as a
matrix-vector product, the section that follows is devoted to that.

Subsequently, I’ll introduce you to my mUnet network for semantic segmentation

and explain in what sense this network goes beyond the now classic Unet.

That will be followed by a presentation of the PurdueShapes5MultiObject

dataset for experimenting with semantic segmentation.

Finally, I will show some semantic segmentation results on the

PurdueShapes5MultiObject dataset as obtained with the mUnet network.

Purdue University 6
 Outline
1 A Brief Survey of Networks for Semantic Segmentation 8
2 Transpose Convolutions 18
3 Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions 34
4 Using the stride Parameter of nn.ConvTranspose2d for
Upsampling 43
5 Including Padding in the nn.ConvTranspose2d Constructor 47
6 The Building Blocks for mUnet 51
7 The mUnet Network for Semantic Segmentation 56
8 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation 61
9 Training the mUnet 70
10 Testing mUnet on Unseen Data 72
Purdue University 7
 A Brief Survey of Networks for Semantic Segmentation

Outline

1 A Brief Survey of Networks for Semantic Segmentation 8

2 Transpose Convolutions 18
3 Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions 34
4 Using the stride Parameter of nn.ConvTranspose2d for
Upsampling 43
5 Including Padding in the nn.ConvTranspose2d Constructor 47
6 The Building Blocks for mUnet 51
7 The mUnet Network for Semantic Segmentation 56
8 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation 61
9 Training the mUnet 70
10 Testing mUnet on Unseen Data 72
Purdue University 8
 A Brief Survey of Networks for Semantic Segmentation

Generation 1 Networks for Semantic Segmentation

These fully convolutional networks for semantic segmentation involve
mostly the same steps as a CNN for image classification, the main
difference being that whereas the latter uses a linear layer at the end
of the network whose output yields the class label, the former ends in
a reduced resolution representation of the input image.
That such an approach for semantic segmentation could be made to
work with a fully convolutional network was first demonstrated in
2015 by Shelhamer et al. in
https://arxiv.org/pdf/1605.06211.pdf

Purdue University 9
 A Brief Survey of Networks for Semantic Segmentation

Generation 2 Networks
The second generation networks are based on the encoder-decoder
principle, with the job of the encoder being to create a low-res rep
that captures the semantic components in the input image and the
job of the decoder being to map those back to the pixel level. The
accuracy of the detail in the back-to-pixel mapping in the docoder is
aided by the skip links from the corresponding levels in the encoder.
Here is the original paper on this encoder-decoder approach:
https://arxiv.org/abs/1505.04597

Purdue University 10
 A Brief Survey of Networks for Semantic Segmentation

Generation 3 Networks
The third generation networks for semantic segmentation are based
on the idea of atrous convolution. [See slide 14 for what is meant by
atrous convolution.] The goal here is to detect semantically
meaningful blobs in an input at different scales without increasing the
number of learnable parameters.
Here is the paper that got everybody in the DL community excited
about the possibilities of atrous convolutions:
https://arxiv.org/abs/1606.00915

Purdue University 11
 A Brief Survey of Networks for Semantic Segmentation

Generation 4 Networks
These use atrous convolutions in an encode-decoder framework.
Here is the paper that showed how the atrous-convo based learning
could be carried in an encoder-decoder framework:
https://arxiv.org/abs/1706.05587

Purdue University 12
 A Brief Survey of Networks for Semantic Segmentation

A Possible Generation 5 Network

Here is my prediction for a network of the future based on atrous
convolutions. I base this prediction on the notion that two successive
applications of a rate M atrous convolution do not yield the same
information a single application of an atrous convolution at rate 2M.
Therefore, there is learnable information in creating multiple pyramids
based on successive applications of the same rate atrous convolution
— but with a very elevated computational burden.

Purdue University 13
 A Brief Survey of Networks for Semantic Segmentation

What is Atrous Convolution?

Atrous convolution is also known as the dilated convolution.

In order to understand atrous convolution, let’s first take another look

at regular convolution of a 16 × 16 input with a 3 × 3 kernel.

Here is the 16 × 16 input:

[[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]
[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]
[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]
[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]
[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]
[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]
[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]
[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]
[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]
[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]
[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]
[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]
[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]
[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]
[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]
[4. 4. 4. 4. 4. 4. 4. 4. 1. 1. 1. 1. 1. 1. 1. 1.]]

And here is the 3 × 3 kernel:

[[-1. 0. 1.]
[-1. 0. 1.]
[-1. 0. 1.]]

Purdue University 14
 A Brief Survey of Networks for Semantic Segmentation

What is Atrous Convolution (contd.)

Shown below is the result of convolving the 16 × 16 input with the

3 × 3 kernel. We obtain the result by sweeping the kernel over the
input.

Since no padding is being used, the output is smaller than the input,
the size of the output being 14 × 14.

As you can see, the feature extracted is the vertical edge in the
middle of the image.
[[0. 0. 0. 0. 0. 0. 9. 9. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 9. 9. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 9. 9. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 9. 9. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 9. 9. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 9. 9. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 9. 9. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 9. 9. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 9. 9. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 9. 9. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 9. 9. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 9. 9. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 9. 9. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 9. 9. 0. 0. 0. 0. 0. 0.]]

Purdue University 15
 A Brief Survey of Networks for Semantic Segmentation

What is Atrous Convolution (contd.)

Let’s consider Rate 2 atrous convolution. At this rate for the dilation
of the kernel, we alternate the original kernel entries with rows and
columns of zeros.

Shown below is the new kernel:

[[-1. 0. 0. 0. 1. 0.]
[ 0. 0. 0. 0. 0. 0.]
[-1. 0. 0. 0. 1. 0.]
[ 0. 0. 0. 0. 0. 0.]
[-1. 0. 0. 0. 1. 0.]
[ 0. 0. 0. 0. 0. 0.]]

The dilated kernel leads to the following output. Since we are still not
using any padding, the enlarged kernel results in a smaller output. Its
size is now 11 × 11.
[[0. 0. 0. 9. 9. 9. 9. 0. 0. 0. 0.]
[0. 0. 0. 9. 9. 9. 9. 0. 0. 0. 0.]
[0. 0. 0. 9. 9. 9. 9. 0. 0. 0. 0.]
[0. 0. 0. 9. 9. 9. 9. 0. 0. 0. 0.]
[0. 0. 0. 9. 9. 9. 9. 0. 0. 0. 0.]
[0. 0. 0. 9. 9. 9. 9. 0. 0. 0. 0.]
[0. 0. 0. 9. 9. 9. 9. 0. 0. 0. 0.]
[0. 0. 0. 9. 9. 9. 9. 0. 0. 0. 0.]
[0. 0. 0. 9. 9. 9. 9. 0. 0. 0. 0.]
[0. 0. 0. 9. 9. 9. 9. 0. 0. 0. 0.]
[0. 0. 0. 9. 9. 9. 9. 0. 0. 0. 0.]]

Purdue University 16
 A Brief Survey of Networks for Semantic Segmentation

What is Atrous Convolution (contd.)

The basic idea of an atrous convolution is to detect an image feature

at a specific scale. So if you carry out such convolutions at different
dilation rates, you will have a multi-scale representation of an image
for a given feature type, the type of the feature being determined by
the base form of the kernel.

My example on the last couple of slides does not do full justice to the
idea described above since the nature of grayscale variation in the
input pattern I used should be detectable at all scales (albeit with
different ’signatures’).

Nevertheless, you should be able extend that example in your own

head and imagine variants of my input pattern that would lead to
detection of the image feature in question at different scales.

Purdue University 17
 Transpose Convolutions

Outline

1 A Brief Survey of Networks for Semantic Segmentation 8

2 Transpose Convolutions 18
3 Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions 34
4 Using the stride Parameter of nn.ConvTranspose2d for
Upsampling 43
5 Including Padding in the nn.ConvTranspose2d Constructor 47
6 The Building Blocks for mUnet 51
7 The mUnet Network for Semantic Segmentation 56
8 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation 61
9 Training the mUnet 70
10 Testing mUnet on Unseen Data 72
Purdue University 18
 Transpose Convolutions

Revisiting the Convolution in a Convo Layer

As mentioned in the Preamble, semantic segmentation requires

pixel-level classification in an image, but the logic of this classification
must be driven by higher-level pixel groupings that are detected in an
image.

The Preamble also mentioned that this requires using an

encoder-decoder framework, with the encoder devoted to the
discovery of higher-level data abstractions and the decoder devoted to
the mapping of these abstractions back to the pixel level with the
help of transpose convolutions.

To understand what is meant by a transpose convolution, we must

first express a regular 2D convolution as a matrix-vector product. We
start with that on the next slide.

Purdue University 19
 Transpose Convolutions

Implementing a Convolution as a Matrix-Vector

Product
Consider convolving a 3 × 3 input pattern with a 2 × 2 kernel to
produce a 2 × 2 output:
1 2 3
a b 1xa+2xb+4xc+5xd 2xa+3xb+5xc+6xd
convolved with 4 5 6 =
c d 4xa+5xb+7xc+8xd 5xa+6xb+8xc+9xd
7 8 9

In our own mind’s eye we like to imagine 2D convolution as sweeping

the kernel over the input pattern, multiplying the two element-wise at
each position of the kernel, and summing the result.

However, for efficient computations, especially when you want to use

a GPU (GPUs are designed to carry out thousands of matrix-vector
products in parallel), it is best to implement a convolution as a
vector-matrix product.
Purdue University 20
 Transpose Convolutions

Convolution as a Matrix-Vector Product (contd.)

We can create a matrix-vector product implementation of the
convolution shown on the previous slide by either vectorizing the
kernel or the input. If we vectorize the kernel, the vector-matrix
product will look like:
_ _
| 1 2 4 5 |
| |
[a b c d] | 2 3 5 6 |
| |
| 4 5 7 8 |
| |
| 5 6 8 9 |
- -

This would produce a 4-element output vector that can be reshaped

into the desired 2 × 2 result.

As to how the elements of the inputs are arranged in each column of

the matrix shown above would depend on both the size of the kernel
and the size of the input.
Purdue University 21
 Transpose Convolutions

Convolution as a Matrix-Vector Product (contd.)

This slide shows the other approach to representing a convolution by

a matrix-vector product:
_ _ _ _
| a b 0 c d 0 0 0 0 | | 1 |
| | | |
| 0 a b 0 c d 0 0 0 | | 2 |
| | | |
| 0 0 0 a b 0 c d 0 | | 3 |
| | | |
| 0 0 0 0 a b 0 c d | | 4 |
- - | |
| 5 |
| |
| 6 |
| |
| 7 |
| |
| 8 |
| |
| 9 |
- -

This would again return an array of 4 elements that would need to be

reshaped into a 2 × 2 output.

I believe this approach to the representation of a 2D convolution by a

matrix-vector product is more popular — probably because it is easier
to write down the matrix with this approach.

Purdue University 22
 Transpose Convolutions

Convolution as a Matrix-Vector Product (contd.)

In the second approach shown on the previous slide, you lay down all
the rows of the kernel in the top row of the matrix and then you shift
it by one position to the right with the caveat that the shift needs to
be more than one position when, in the 2D visualization, the kernel
has reached the rightmost positions in each horizontal sweep over the
input.

Let C represent the matrix obtained in our matrix-vector product

representation of a 2D convolution. If x represents the input pattern
and y the output, we can write:

y = Cx

This relationship between the input/output for one convo layer is

displayed at left in the figure on the next slide.

Purdue University 23
 Transpose Convolutions

Convolution as a Matrix-Vector Product (contd.)

Figure: Forward data flow through one convo layer in a CNN. The depiction at left is for the case when the stride equals 1
and the one at right for the case when the stride equals 2.

Purdue University 24
 Transpose Convolutions

Using the Matrix CT to Backpropagate the Loss

What’s interesting is that the mathematics of the backpropagation of
the loss dictates that if the data flows forward through a convo layer
as y = Cx, then the loss backpropagates through the same layer as

Lossbackprop = CT Loss

as shown at left in the figure on the next slide.

In our discussion so far on representing a convolution by a

matrix-vector product, we implicitly assumed that the stride was
equal to one along both axes of the input. With this assumption,
except for the boundary effects, the output of the convo layer will be
of the same size as that of the input.

Let’s now consider what happens to the matrix C in our matrix-vector

product representation of a convolution when the stride equals 2
along both axes of the input
Purdue University 25
 Transpose Convolutions

Using the Matrix CT to Backpropagate the Loss

(contd.)

Figure: Backpropagating the loss through one convo layer in a CNN. The depiction at left is for the case when the stride
equals 1 and the one at right for the case when the stride equals 2.

Purdue University 26
 Transpose Convolutions

Convolution as a Matrix-Vector Product for a Larger Input

To best visualize what happens to the matrix C when the stride is

greater then 1, we need a slightly larger input. So let’s consider a
2 × 2 kernel and a 4 × 4 input as shown below for a stride that equals
1 as before along both axes of the input:
1 2 3 4
1a+2b+5c+6d 2a+3b+6c+7d 3a+4b+7c+8d
a b 5 6 7 8
convolved with = 5a+6b+9c+10d 6a+7b+10c+11d 7a+8b+11c+12d
c d [stride=(1,1)] 9 10 11 12
9a+10b+13c+14d 10a+11b+15c+16d 11a+12b+15c+16d
13 14 15 16

We can represent this 2D convolution with the matrix-vector product:

_ _ _ _
| a b 0 0 c d 0 0 0 0 0 0 0 0 0 0| | 1 |
| | | |
| 0 a b 0 0 c d 0 0 0 0 0 0 0 0 0| | 2 |
| | | |
| 0 0 a b 0 0 c d 0 0 0 0 0 0 0 0| | 3 |
| | | |
| 0 0 0 0 a b 0 0 c d 0 0 0 0 0 0| | 4 |
| | | |
| 0 0 0 0 0 a b 0 0 c d 0 0 0 0 0| | 5 |
| | | |
| 0 0 0 0 0 0 a b 0 0 c d 0 0 0 0| | 6 |
| | | |
| 0 0 0 0 0 0 0 0 a b 0 0 c d 0 0| | 7 |
| | | |
| 0 0 0 0 0 0 0 0 0 a b 0 0 c d 0| | 8 |
| | | |
| 0 0 0 0 0 0 0 0 0 0 a b 0 0 c d| | 9 |
- - | |
| 10|
| |
| 11|
| |
| 12|
| |
| 13|
| |
| 14|
| |
| 15|
| |
| 16|
- -

Purdue University 27
 Transpose Convolutions

Transitioning from Stride = (1, 1) to Stride = (2, 2)

In the matrix shown on the previous slide, as before, each row of the
matrix is a shifted version of the row above. While the shift is one
element for the consecutive sweep positions of the kernel vis-a-vis the
input image, the shift becomes 2 positions when the kernel has to
move from one row of the input to the next.

The matrix-vector product can again be presented by the notation

y = Cx where C is now the 9 × 16 matrix shown on the previous slide.

Let’s now consider the same convolution but with a stride of

(2, 2).

In order to understand how to modify C when stride = 2, it is best to

start with the C on the previous slide for the unit stride case and to
mentally partition this matrix vertically into sections, with each
section corresponding to one row of the output.
Purdue University 28
 Transpose Convolutions

Convolution as a Matrix-Vector Product for Stride = (2, 2)

For the stride = 1 case, the first three rows of the matrix C on Slide
27 correspond to the first row of the output, the second three rows to
the second row of the output. And the last three rows to the third
row of the output.

For the case of stride = 2, we zero out alternate rows in each

above-mentioned vertical section of the matrix shown on Slide 27
since those would correspond to the horizontal jumps as dictated by
the stride. This explains the all-zero rows for the 2nd and the 8th
rows of the matrix shown on the next slide for the case of stride = 2.

For the vertical jumps dictated by the striding action on the sweep
motions of the kernel, we must also zero out entire vertical partitions
of the matrix that was shown earlier for the stride = 1 case. This is
the reason for why the 4th, the 5th and the 6th rows are all zeroed
out in matrix shown on the next slide for the case of stride = 2.
Purdue University 29
 Transpose Convolutions

Convolution as a Matrix-Vector Product for Stride = (2, 2) (contd.)

Here is the matrix-vector product for a stride of 2 along both axes:
_ _ _ _
| a b 0 0 c d 0 0 0 0 0 0 0 0 0 0| | 1 |
| | | |
| 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0| | 2 |
| | | |
| 0 0 a b 0 0 c d 0 0 0 0 0 0 0 0| | 3 |
| | | |
| 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0| | 4 |
| | | |
| 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0| | 5 |
| | | |
| 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0| | 6 |
| | | |
| 0 0 0 0 0 0 0 0 a b 0 0 c d 0 0| | 7 |
| | | |
| 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0| | 8 |
| | | |
| 0 0 0 0 0 0 0 0 0 0 a b 0 0 c d| | 9 |
- - | |
| 10|
| |
| 11|
| |
| 12|
| |
| 13|
| |
| 14|
| |
| 15|
| |
| 16|
- -

If C again represents the 9 × 16 matrix that is shown above, the data

flow in the forward direction and the loss backpropagation can again
be expressed as before:
y = Cx
Lossbackprop = CT Loss
provided you account for the intentionally zeroed out rows in the
matrix C. The figures on the right in Slides 24 and 26 are for the case
when the stride equals 2.
Purdue University 30
 Transpose Convolutions

Reformatting the Output Vector for Stride = (2, 2)

As it is, the output vector y will have 9 elements in it (the same as

was the case when the stride was equal to 1). However, since we
intentionally deleted several rows of the matrix C, we would need to
drop those elements from y before we accept the output of the
matrix-vector product.
It is this deletion of the elements of y that makes the output of the
convo layer to have a size that, except for the boundary effects, would
be half the size of the input.
Note the bottom line here: When the stride is 2, the matrix C maps
the input data x to the output y that is roughly half the size of x.
And, going in the opposite direction, the transpose matrix CT maps
the loss in the layer where y resides to the layer where x resides. In
other words, CT maps the loss from a layer of some size to a
layer that is twice as large.
It isUniversity
Purdue this property of CT that we will focus on in the next slide. 31
 Transpose Convolutions

Focusing on the Upsampling Ability of CT

Based on the discussion so far, we can assign two different roles to

the matrices C and CT :
We can use C in an encoder since (when the strides are greater than 1)
it creates coarser abstractions from the data and that’s what an
encoder is supposed to do.
And we can use CT in the decoder since (when the strides are greater
than 1) it creates finer abstractions from coarser abstractions.

In other words, we can use matrices like C to create feature and

object level abstractions from the pixels and use matrices like CT to
map those abstractions back to the pixel level.

Given the importance of CT in encoder-decoder learning frameworks,

it has name of its own: It is known as the “Transpose Convolution”.

Purdue University 32
 Transpose Convolutions

A Useful Reference on “Convolution Arithmetic”

For deeper insights into the relationship between the convolutional
layers and the transposed convolutional layers, see the following 2018
article by Dumoulin and Visin:
https://arxiv.org/abs/1603.07285

This reference is a great resource for understanding how the output

shape depends on the input shape, the kernel shape, the zero padding
used, and the strides for both convolutional layers and the transposed
convolutional layers.

Purdue University 33
Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions

Outline

1 A Brief Survey of Networks for Semantic Segmentation 8

2 Transpose Convolutions 18
3 Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions 34
4 Using the stride Parameter of nn.ConvTranspose2d for
Upsampling 43
5 Including Padding in the nn.ConvTranspose2d Constructor 47
6 The Building Blocks for mUnet 51
7 The mUnet Network for Semantic Segmentation 56
8 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation 61
9 Training the mUnet 70
10 Testing mUnet on Unseen Data 72
Purdue University 34
Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions

PyTorch’s torch.nn.ConvTranspose2d Class

In PyTorch, the class that helps you carry out transpose convolutions
for regular images is torch.nn.ConvTranspose2d. This link will
take you directly to its API:

https://pytorch.org/docs/stable/nn.html#convtranspose2d

If this is your first exposure to transpose convolutions, you are likely

to be confused by the relationship between the kernel size, the value
you choose for padding, and the output size in the API of the
torch.nn.ConvTranspose2d. The goal of this section is to clarify
this relationship.

Starting with Slide 37, I’ll illustrate the above mentioned relationship
with an example that uses transpose convolution to dilate 4-channel
1 × 1 noise image into a 1-channel 4 × 4 image by using a 4 × 4
kernel. The code for this example is on the next slide.
Purdue University 35
Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions

Dilating a One Pixel Image into a 4 × 4 Image with ConvTranspose2d

batch_size = 1 ## (A)
input = torch.randn((batch_size, 4)).view(-1, 4, 1, 1) ## A 4-channel image of size 1x1 ## (B)
print(input) ## (C)
# tensor([[[[ 1.5410]],
#
# [[-0.2934]],
#
# [[-2.1788]],
#
# [[ 0.5684]]]])
print(input.shape) # (1, 4, 1, 1) ## (D)
#
# args: ch_in, ch_out, ker_size, stride, padding
trans_conop = nn.ConvTranspose2d(4, 2, 4, 1, 0) ## (E)
print(trans_conop.weight) ## (F)
# tensor([[[[-0.0157, 0.0468, -0.0534, -0.0347],
# [-0.1689, -0.1171, -0.0729, 0.0065],
# [ 0.0699, 0.1061, -0.1198, -0.0770],
# [ 0.0642, 0.1468, -0.0364, 0.1323]],
#
# [[-0.0285, 0.0187, 0.1601, -0.1640],
# [-0.1113, -0.0448, -0.0689, 0.1527], As shown in the output in line
# [-0.1146, -0.0814, -0.1235, -0.1656], (G), the shape of the kernel
# [-0.1032, 0.1520, 0.0789, 0.0857]]], weights is (4,2,4,4). We want
# the 4 input channels to go into
# 2 channels in the output and this
# [[[ 0.0093, -0.0906, 0.0299, -0.1651], must be accomplished with 4x4
# [-0.1277, -0.0911, 0.1115, 0.1036], kernel. The first two 4x4
# [-0.0784, -0.0064, 0.1131, 0.1757], matrices shown are for taking the
# [ 0.0702, 0.0239, 0.1185, -0.1041]], first channel of the input to
# each of the two channels of the
# [[ 0.0329, -0.1371, -0.1225, -0.0913], output. The same applies to the
# [ 0.0800, 0.0711, -0.1047, 0.0534], second and the third 4x4 matrics:
# [ 0.0970, -0.0223, 0.0067, 0.0410], the first of these will be applied
# [ 0.1097, 0.1697, -0.1362, -0.0648]]], to the second input channel and
# the result sent to the first
# output channel; and the second
# [[[ 0.0695, 0.1465, 0.1538, 0.1560], applied to the same second input
# [ 0.0352, -0.1537, 0.0163, -0.1106], channel but its output sent to the
# [-0.1647, 0.1571, 0.1344, -0.1763], second output channel. And so on.
# [ 0.0331, -0.0298, -0.0291, -0.0809]],
#
# [[ 0.0680, -0.1047, 0.0648, 0.0894],
# [ 0.1265, 0.0661, -0.1750, -0.1147]
# [ 0.0883, 0.0370, -0.1379, -0.1018],
# [ 0.1663, 0.1191, -0.0771, -0.0445]]],
#
#
# [[[-0.1684, -0.0032, -0.1331, -0.1364],
# [-0.0097, 0.0265, -0.0724, 0.1049],
# [-0.1076, 0.1604, 0.1211, -0.1491],
# [-0.0440, 0.0080, 0.0258, 0.0419]],
#
# [[ 0.0694, 0.0106, -0.0863, 0.0836],
# [-0.1696, -0.1048, -0.0443, -0.0861],
# [-0.0618, -0.1449, -0.0376, 0.0378],
# [-0.1152, -0.0091, 0.1265, -0.0182]]]], requires_grad=True)
print(trans_conop.weight.shape) # (4, 2, 4, 4) ## (G)
x = trans_conop(input) ## (H)
print(x) ## (I)
# tensor([[[[-0.2691, -0.2173, -0.4970, -0.4176],
# [-0.3000, 0.2013, -0.2167, 0.2852],
# [ 0.4334, -0.0808, -0.4369, 0.1342],
# [-0.0138, 0.2935, -0.0079, 0.4394]], Shown here is the 2-channel
# 4x4 output image
# [[-0.1775, 0.2879, 0.0771, -0.3884],
# [-0.5823, -0.3086, 0.2653, 0.4053],
# [-0.4478, -0.2971, 0.0715, -0.0391],
# [-0.6342, -0.0956, 0.3861, 0.2224]]]],
# grad_fn=<ConvTranspose2DBackward>)
print(x.shape) # (1, 2, 4, 4) ## (J)

Purdue University 36
Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions

Explaining the Example

In line (B) of the code shown on the previous slide, the call
input = torch.randn((batch_size, 4)).view(-1, 4, 1, 1)

generates from a 4-dimensional noise vector a 1 × 1 image with 4

channels. The call torch.randn(batch size, 4) generates a 1 × 4 array of noise
samples that are drawn from a zero-mean and unit-variance Gaussian
distribution. As shown in line (A) on the previous slide, the variable
batch size is set to 1 for this demonstration. Subsequently, in
accordance with the explanation in the last section of my Week 6
slides, the invocation view(-1, 4, 1, 1) reshapes the 1 × 4 noise array into
a tensor of shape (1, 4, 1, 1), which can be interpreted as a 4-channel
one-pixel image. The first value in the shape (1, 4, 1, 1) is for the
batch-size of 1. The output of the code in line (B) is shown in the
commented-out block just after line (C).
The statement in line (E) and what comes after that are explained on
the next slide.
Purdue University 37
Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions

Explaining the Example (contd.)

In line (E) of the example code in Slide 36, we call the constructor of
torch.nn.ConvTranspose2d class as follows:

trans_conop = nn.ConvTranspose2d(4, 2, 4, 1, 0)

where the arguments supplied carry the following meanings in the

left-to-right order:
input channels = 4
output channels = 2
kernel size = (4,4)
stride = (1,1)
padding = (0,0)

For the last three arguments, if the two values in a tuple are identical,
we can supply the argument as a scalar.

The constructor call shown above means that a (4, 4) kernel will be
used for the transpose convolution with our 4-channel 1 × 1 input
image.
Purdue University 38
Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions

Explaining the Example (contd.)

As the PyTorch documentation page for torch.nn.ConvTranspose2d
says, before actually carrying out the transpose convolution operation,
the input image is padded with N zeros on “both” sides where N is
given by
N = dilation * (kernel_size - 1) - padding

where dilation is a constructor parameter for the ConvTranspose2d

class whose default value is 1.

Shown below are the values for N, which is the number of padding
zeros for “both” sides of the input image:
kernel_size = (4,4)
dilation = 1
_______________________________________
padding | N
---------------------------------------
0 | 3
1 | 2
2 | 1
---------------------------------------

Purdue University 39
Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions

Explaining the Example (contd.)

As to what is meant by “both sides” with regard to the padding of
the input image with N zeros, shown below are the padded versions
of our one-pixel input image for different values for the padding
constructor parameter:
padding = 0:
N = 3
padded version of the 1x1 input_image:
0 0 0 0 0 0 0
0 0 0 0 0 0 0
0 0 0 0 0 0 0
0 0 0 1 0 0 0
0 0 0 0 0 0 0
0 0 0 0 0 0 0
0 0 0 0 0 0 0
padded input_image size: 7x7
output_image size: 4x4
-----------------------------------------------------------
padding = 1
N = 2
padded version of the 1x1 input_image:
0 0 0 0 0
0 0 0 0 0
0 0 1 0 0
0 0 0 0 0
0 0 0 0 0
padded input_image size: 5x5
output_image size: 2x2

-----------------------------------------------------------
padding = 2
N = 1
padded version of the 1x1 input_image:
0 0 0
0 1 0
0 0 0
padded input_image size: 3x3
output_image size: ERROR

Purdue University 40
Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions

Calculating the Size of the Output Image

Based on what is shown on the previous slide, the reader should now
be able to make sense of the output shown in the commented out
portion after line (I) on Slide 36. That shows a 2-channel 4 × 4
output image obtained with the transpose convolution operator
invoked with the statement in line (H) on the same slide.

The reason why the output image is of size 4 × 4 when padding=0 is

because the padded input-image as shown on the previous slide for
this case is of size 7 × 7. So a 4 × 4 kernel can be fit at four different
locations width-wise and height-wise as the kernel sweeps over the
padded input image. And that gives us a 4 × 4 output.

The discussion so far has been on explaining the code as shown on

Slide 36. As the reader knows, we used padding=0 in that example. In the
next slide, let’s see what happens when we use padding=1 and padding=2.

Purdue University 41
Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions

Calculating the Size of the Output Image (Contd.)

If we want to use padding=1, the constructor call for the torch.nn.ConvTranspose2d

class would become as shown below (assuming that we do not change
the other parameters):
trans_conop = nn.ConvTranspose2d(4, 2, 4, 1, 1)

In this case, as shown in the middle in Slide 40, the padded-input

image will be of size 5 × 5. That would allow for only two different
shifts for the 4 × 4 kernel for the production of the output. Therefore,
in this case the size of the output image would of size 2 × 2.

On the other hand, if we had set the padding parameter to 2, the

padded input image would like what is shown at the bottom on the
previous slide. This is too small for a 4 × 4 kernel to fit in. So in this
case, we get an error message from PyTorch.

Purdue University 42
 Using the stride Parameter of nn.ConvTranspose2d for
Upsampling

Outline

1 A Brief Survey of Networks for Semantic Segmentation 8

2 Transpose Convolutions 18
3 Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions 34
4 Using the stride Parameter of nn.ConvTranspose2d for
Upsampling 43
5 Including Padding in the nn.ConvTranspose2d Constructor 47
6 The Building Blocks for mUnet 51
7 The mUnet Network for Semantic Segmentation 56
8 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation 61
9 Training the mUnet 70
10 Testing mUnet on Unseen Data 72
Purdue University 43
 Using the stride Parameter of nn.ConvTranspose2d for
Upsampling

The stride Parameter for Upsampling

For most applications of the nn.ConvTranspose2d class, you are
likely to use the stride parameter for bulk of the upsampling effect.
Whereas the kernel size has mostly an additive effect on the size of
the output, the stride parameter has a multiplicative effect.
Assume that the stride along both image axes is the same and set to
s. Assume the same for the kernel and let its size be denoted k.
If Hin and Win denote the height and the width of the image at the
input to a transpose convolution operator and Hout and Wout the
same at the output, the relationship between the sizes at the input
and the output is given by
Hout = (Hin − 1) ∗ s + k
Wout = (Win − 1) ∗ s + k

where I have also assumed that you are using the default choices for
padding (= 0), for output padding (= 0), and for dilation (= 1).
Purdue University 44
 Using the stride Parameter of nn.ConvTranspose2d for
Upsampling

An Upsampling Example with Stride

The next slide shows an example of upsampling achieved with a stride
of size 2 × 2 and a kernel of size 3 × 3 on an input image of size 4 × 4.

In line (A), we set the batch size to 1. With its first argument set to
1, the call in line (B) generates a random-number vector of size 32
that is then shaped into a 2-channel 4 × 4 image. We want to feed
this image into a nn.ConvTranspose2d operator and check the
output image for its size.

The nn.ConvTranspose2d operator in line (E) has its stride set to

2 × 2 and its kernel set to be of size 3 × 3. The formulas shown on
the previous slide say that this should result in the output image
produced by the operator to be of size 9 × 9. Note that the operator
in line (E) also sets the number of output channels to 4. This is
borne out by the output displayed for the call to the operator in line
(F) and for the shape of the output displayed in line (G).
Purdue University 45
 Using the stride Parameter of nn.ConvTranspose2d for
Upsampling

Example of Upsampling with Stride

import torch
import torch.nn as nn
torch.manual_seed(0)
batch_size = 1 ## (A)
input = torch.randn((batch_size, 32)).view(-1, 2, 4, 4) ## (B)
print("\n\n\nThe 32-dimensional noise vector shaped as an 2-channel 4x4 image:\n\n")
print(input) ## (C)
# tensor([[[[ 0.7298, -1.8453, -0.0250, 1.3694],
# [ 2.6570, 0.9851, -0.2596, 0.1183],
# [-1.1428, 0.0376, 2.6963, 1.2358],
# [ 0.5428, 0.5255, 0.1922, -0.7722]],
#
# [[ 1.6268, 0.1723, -1.6115, -0.4794],
# [-0.1434, -0.3173, 0.9671, -0.9911],
# [ 0.5436, 0.0788, 0.8629, -0.0195],
# [ 0.9910, -0.7777, 0.3140, 0.2133]]]])
#
print(input.shape) ## (1, 2, 4, 4) ## (D)

# ch_in ch_out kernel stride padding

trans_conop = nn.ConvTranspose2d( 2, 4, 3, 2, 0 ) ## (E)
x = trans_conop(input) ## (F)

print(x) ## (G)

# tensor([[[[-4.4152e-02, 1.2544e-01, -1.2656e-01, 9.6014e-02, -2.2811e-01, -2.7850e-01, 7.4894e-02, -2.1415e-01, 1.2354e-01],

# [ 2.4731e-02, -1.6544e-01, -1.3804e-01, 2.2369e-01, -2.8944e-01, -4.7890e-02, 6.9339e-02, -2.5903e-01, 1.0654e-01],
# [ 3.6436e-01, -3.7065e-01, 6.3544e-01, -3.4491e-02, -2.9987e-01, 1.6315e-01, 4.4707e-03, -2.8466e-01, 8.7493e-02],
# [ 1.0370e-01, -4.5312e-01, 2.8402e-01, -2.0143e-01, 6.5421e-02, -1.5800e-02, -1.2667e-01, -7.0222e-02, -8.5209e-03],
# [-3.0782e-01, -9.1773e-02, 1.3661e-01, -1.0635e-01, 4.5713e-01, -1.3314e-01, 5.2770e-01, -1.1149e-01, 1.4809e-01],
# [-1.1116e-01, 1.1756e-01, -1.8814e-01, -5.9404e-02, 1.2867e-01, -4.6027e-01, 2.7642e-01, -2.3950e-01, 7.7300e-02],
# [-1.1877e-04, 1.1310e-01, 5.9378e-02, -2.0451e-01, -4.0972e-02, -2.5601e-01, 2.0241e-01, -6.2370e-02, 7.7144e-03],
# [ 2.7385e-05, -1.3653e-01, -1.3757e-02, -1.3173e-01, 4.5617e-02, -8.2951e-02, -8.5837e-02, 6.2223e-02, -1.4210e-01],
# [-4.3369e-02, -1.3176e-01, -8.8378e-03, -6.3154e-02, 2.0169e-02, -7.9904e-02, -9.0935e-03, -4.3222e-03, -1.5681e-01]],
#
# [[-6.3428e-02, -1.9877e-01, 8.4883e-02, -5.5194e-02, 1.8710e-02, 2.0991e-01, -5.2714e-02, 8.7261e-02, -5.6352e-03],
# [ 3.0607e-01, -1.8582e-01, -1.5720e-01, -1.0889e-01, -3.6943e-01, 2.1707e-01, 1.1601e-01, 1.2883e-01, 1.1567e-01],
# [-4.9457e-02, 1.3517e-01, -3.1747e-01, 1.4358e-01, -1.5253e-01, -2.2648e-01, -1.1735e-01, 4.7839e-02, 3.2453e-02],
# [ 2.9721e-01, 1.4327e-01, 2.8931e-01, 8.9015e-02, 1.8221e-01, -1.4263e-01, -1.5140e-01, 1.3982e-01, 2.1605e-02],
# [ 2.9850e-01, -2.0616e-01, -9.8867e-02, -6.7920e-02, -2.4457e-01, 4.6328e-03, -4.7882e-01, -3.9504e-02, -4.9977e-02],
# [-6.2585e-02, -1.2643e-01, -8.2678e-02, -8.6435e-03, 4.3937e-01, 9.0001e-03, 3.5033e-01, 6.0396e-02, 9.9465e-02],
# [-2.9405e-02, -4.0695e-02, -1.3474e-01, 1.1429e-01, 3.3068e-01, -9.3352e-02, 5.3747e-02, -9.2056e-02, -3.4831e-02],
# [ 1.9825e-01, -1.0853e-01, -1.0243e-02, 1.2991e-01, 1.1680e-01, -3.3269e-02, -5.1290e-02, -6.4512e-02, -6.4124e-02],
# [ 1.9561e-01, 3.8178e-02, -1.1138e-01, -6.7620e-02, 4.1267e-02, 1.1471e-02, -1.5264e-02, 4.4058e-02, 2.8338e-02]],
#
# [[-2.4940e-01, 2.8603e-01, 2.1782e-01, -6.6619e-02, 5.7356e-02, -4.9506e-02, 1.5497e-01, 1.6988e-01, 3.2393e-01],
# [ 9.2061e-02, -9.0965e-02, -1.6751e-01, 2.9311e-01, 1.4122e-01, 1.7886e-01, 3.0424e-01, -4.8374e-02, 1.0782e-01],
# [-3.4120e-01, 4.7060e-01, 5.4504e-01, 3.4156e-01, 3.3260e-01, -7.7028e-02, -1.2933e-01, -1.7688e-01, 7.6152e-02],
# [ 3.9559e-01, -2.1855e-01, 2.3644e-01, -1.2267e-02, 2.5593e-02, 6.2246e-02, 1.2039e-01, 1.2731e-01, 1.1114e-01],
# [-1.4819e-01, -2.3600e-01, -5.2384e-01, -3.8295e-02, -4.8354e-01, 5.5731e-01, 2.0288e-01, 4.8694e-02, 2.1953e-01],
# [-7.0234e-02, 1.8975e-01, 6.4450e-02, 7.6979e-02, 3.5079e-01, -2.7943e-01, 2.2640e-01, -5.8386e-02, 9.5471e-02],
# [ 3.3609e-03, 3.9622e-01, 1.9816e-01, 7.6436e-02, -1.4130e-01, 3.2905e-04, -2.3993e-01, -8.2371e-02, -1.6212e-01],
# [ 1.0082e-01, -3.3432e-02, 1.6210e-01, 6.7511e-02, 1.1597e-01, 4.5647e-02, -1.8666e-02, 1.6465e-01, 7.4799e-02],
# [ 2.3834e-02, 1.7224e-01, -9.4561e-03, -6.9326e-02, -3.4250e-03, 1.1134e-01, 1.6478e-01, 1.8597e-01, 1.6523e-01]],
#
# [[ 1.2324e-01, -6.7804e-02, -4.1464e-01, 3.9657e-02, -4.1462e-01, -3.2299e-01, 1.5135e-02, -3.4607e-01, -1.1564e-01],
# [-3.4457e-02, -2.7444e-01, -3.6810e-01, -3.7973e-01, -6.1502e-02, 2.5991e-02, -7.0435e-02, 4.2427e-02, -2.9984e-01],
# [-3.4998e-02, -4.2079e-01, 3.6114e-01, -1.0331e-01, -2.7160e-01, -1.2306e-01, -7.8796e-01, -4.3289e-01, 4.2614e-02],
# [-1.5327e-01, 1.4117e-01, -5.0262e-01, -1.7961e-02, -1.9241e-01, -3.0302e-01, -2.7004e-01, -3.1703e-02, -1.0608e-01],
# [-4.5616e-01, -2.5729e-01, -3.6639e-02, -2.6968e-01, 2.6512e-01, -2.6265e-01, -1.6591e-01, -3.5190e-01, -2.7374e-01],
# [-1.2856e-01, -3.4800e-01, -4.9452e-02, -1.6643e-01, -8.8080e-02, 2.6840e-02, -5.5462e-01, -2.6119e-02, -3.1550e-01],
# [ 1.7508e-01, 2.6454e-02, -4.1953e-01, -2.9197e-01, -2.7528e-01, -3.6171e-01, 7.2971e-02, -1.8341e-01, -3.9131e-02],
# [-8.3295e-02, -2.1962e-01, -3.5403e-01, -1.3091e-02, -1.4763e-01, -1.7753e-01, -1.9705e-01, -2.6927e-01, -7.8970e-02],
# [-1.7471e-01, -1.5918e-01, -7.2195e-02, -2.5484e-01, -2.0372e-01, -1.6250e-01, -6.8606e-04, -7.4715e-02, -2.2458e-01]]]],
# grad_fn=<ConvTranspose2DBackward>)
#
print(x.shape) ## (1, 4, 9, 9) ## (H)

Purdue University 46
 Including Padding in the nn.ConvTranspose2d Constructor

Outline

1 A Brief Survey of Networks for Semantic Segmentation 8

2 Transpose Convolutions 18
3 Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions 34
4 Using the stride Parameter of nn.ConvTranspose2d for
Upsampling 43
5 Including Padding in the nn.ConvTranspose2d Constructor 47
6 The Building Blocks for mUnet 51
7 The mUnet Network for Semantic Segmentation 56
8 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation 61
9 Training the mUnet 70
10 Testing mUnet on Unseen Data 72
Purdue University 47
Including Padding in the nn.ConvTranspose2d Constructor

Using the padding Parameter

As mentioned at the beginning of the previous section, the stride
parameter has a multiplicative effect on the size of the output image
in relation to that of the input image. To fine-tune the output-size
enlargement achieved with stride, you can use the padding
parameter when constructing an instance of the
nn.ConvTranspose2d operator.
Recall that whereas the stride parameter in the nn.Conv2d
constructor causes an input image to be downsampled while the
image is subject to a convolutional operation, the effect of the same
parameter in the nn.ConvTranspose2d constructor is exactly the
opposite, in the sense that it causes the input image to be upsampled.
The same thing is true for the padding parameter. For the
nn.Conv2d case, it causes the input image to be enlarged by the
extent of padding on both sides. However, for the
nn.ConvTranspose2d case, it shrinks the input image
correspondingly.
Purdue University 48
Including Padding in the nn.ConvTranspose2d Constructor

Using the padding Parameter (Contd.)

As in the previous section, assume that the stride along both image
axes is the same and set to s and that the same is true of the kernel
size that we represent by k. Continuing with this assumption, let the
padding along both axes also be the same and let’s represent it by p.
If Hin and Win denote the height and the width of the image at the
input to a transpose convolution operator and Hout and Wout the
same at the output, the relationship between the sizes at the input
and the output is given by
Hout = (Hin − 1) ∗ s + k − 2 ∗ p
Wout = (Win − 1) ∗ s + k − 2 ∗ p
where I have also assumed that you are using the default choices for
output padding (= 0), and for dilation (= 1).
Shown on the next slide is an example similar to the one on Slide 46,
except that now we also include padding (= 1) in line (E), with causes
the size of the output image to change from 9 × 9 to 7 × 7 as shown in
Purdue
lineUniversity
(H) 49
Including Padding in the nn.ConvTranspose2d Constructor

Example of Upsampling with Stride and Padding

import torch
import torch.nn as nn
torch.manual_seed(0)
batch_size = 1 ## (A)
input = torch.randn((batch_size, 32)).view(-1, 2, 4, 4) ## (B)
print("\n\n\nThe 32-dimensional noise vector shaped as an 2-channel 4x4 image:\n\n")
print(input) ## (C)
# tensor([[[[ 0.5769, -0.1692, 1.1887, -0.1575],
# [-0.0455, 0.6485, 0.5239, 0.2180],
# [ 1.9029, 0.3904, 0.0331, -1.0234],
# [ 0.7335, 1.1177, 2.1494, -0.9088]],
#
# [[-0.1434, -0.1947, 1.4903, -0.7005],
# [ 0.1806, 1.3615, 2.0372, 0.6430],
# [ 1.6953, 2.0655, 0.2578, -0.5650],
# [ 0.9278, 0.4826, -0.8298, 1.2678]]]])

print(input.shape) ## (1, 2, 4, 4) ## (D)

# ch_in ch_out kernel stride padding
trans_conop = nn.ConvTranspose2d( 2, 4, 3, 2, 1 ) ## (E)

x = trans_conop(input) ## (F)
print(x) ## (G)

# tensor([[[[-1.4602e-01, 3.4987e-02, -4.9057e-02, -2.1430e-01, -1.1586e-01, 1.3547e-01, -8.5486e-02],

# [ 3.6219e-02, -2.1292e-02, 9.6213e-02, 3.3451e-02, 3.8808e-01, -4.7209e-01, -1.2999e-02],
# [-3.9848e-02, -1.3260e-01, -5.1958e-02, -9.8983e-03, 1.1389e-02, 1.4386e-02, -4.3486e-02],
# [ 2.1806e-01, -5.7146e-02, 3.0747e-01, -3.9247e-01, 8.6604e-02, -3.7835e-01, -1.2648e-01],
# [-1.9795e-01, 1.7612e-01, 3.1326e-02, 1.7133e-02, -4.5113e-02, 6.3181e-02, 4.0513e-02],
# [ 3.7797e-01, -4.1503e-01, 1.4547e-01, -4.7569e-01, -3.7961e-02, 5.1364e-02, -1.1437e-01],
# [-9.3304e-02, -7.3418e-02, -1.7575e-01, -1.4605e-01, -4.0518e-01, 3.2027e-01, 1.5138e-01]],
#
# [[ 1.5954e-01, 1.4756e-01, 1.1634e-01, 1.3148e-01, 2.3051e-01, 2.5622e-01, 1.0570e-01],
# [ 1.5098e-01, 1.8500e-01, 2.9588e-01, 3.0268e-01, 9.4988e-02, 1.7306e-01, 2.9879e-01],
# [ 1.3169e-01, 2.0943e-01, 1.9719e-01, 4.1675e-01, 2.0525e-01, 3.5726e-01, 1.5688e-01],
# [ 2.3013e-01, 2.8313e-01, 1.4138e-01, 1.3809e-02, -1.4228e-01, -2.8762e-01, 7.0353e-03],
# [ 2.7535e-01, 5.2515e-01, 1.9836e-01, 3.3347e-01, 1.3785e-01, 2.0058e-01, 5.9918e-02],
# [-9.6119e-02, -2.8195e-01, -1.5659e-01, -6.8087e-02, -8.4310e-02, -2.9354e-02, 4.2759e-01],
# [ 1.9229e-01, 1.7491e-01, 2.0401e-01, -7.2777e-02, 2.3285e-01, 3.9534e-01, 1.0732e-01]],
#
# [[-9.4420e-02, -3.9821e-02, -4.0897e-02, -3.6939e-02, 6.1995e-02, 2.1703e-01, -1.0549e-01],
# [-1.5969e-03, 2.5784e-01, 1.9578e-03, 3.7270e-01, 3.5862e-01, -2.0336e-01, -8.5515e-02],
# [-3.6023e-03, 5.1168e-02, 8.9228e-02, 3.1974e-01, 1.8428e-01, 3.5481e-01, 3.3411e-02],
# [-6.5154e-02, 4.2694e-01, 3.4081e-01, 2.6115e-02, 2.7636e-01, -2.5520e-01, 1.2862e-01],
# [ 3.0369e-02, 4.2223e-01, 1.9858e-01, 3.4078e-01, -1.9767e-04, 3.4468e-02, -1.8808e-02],
# [ 2.6914e-01, 3.8698e-01, 1.6409e-01, 6.9061e-02, -3.4989e-01, 3.1591e-01, 1.0464e-01],
# [ 2.7803e-02, 1.1538e-01, -5.9130e-02, -9.0690e-02, -3.0726e-01, 2.9721e-02, 2.0266e-01]],
#
# [[ 3.1302e-02, -1.7105e-02, 3.9255e-02, 1.2681e-03, 2.6324e-01, 8.2351e-02, -3.6774e-02],
# [ 1.3670e-01, -1.3665e-01, 2.7974e-01, 1.5993e-01, 2.7132e-01, 1.6712e-01, 2.2299e-01],
# [ 9.2896e-02, -8.1275e-03, 2.5529e-01, -8.3837e-02, 3.5914e-01, 1.0724e-01, 1.5666e-01],
# [ 3.2818e-01, 3.0762e-01, 2.5367e-01, 4.6363e-01, -8.1398e-02, 3.6939e-01, -9.5386e-02],
# [ 2.7898e-01, -2.4000e-01, 3.6618e-01, 1.5419e-01, 1.0282e-01, 7.0281e-02, 1.6960e-03],
# [ 1.4526e-01, 1.8005e-01, -3.1231e-02, 6.9455e-02, -1.4039e-02, 1.6355e-01, 2.1294e-01],
# [ 1.8852e-01, 6.4715e-02, 1.1374e-01, 1.7605e-01, -1.0454e-01, -4.8272e-01, 2.7392e-01]]]],
# grad_fn=<ConvTranspose2DBackward>)
#
print(x.shape) ## (1, 4, 7, 7) ## (H)

Purdue University 50
 The Building Blocks for mUnet

Outline

1 A Brief Survey of Networks for Semantic Segmentation 8

2 Transpose Convolutions 18
3 Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions 34
4 Using the stride Parameter of nn.ConvTranspose2d for
Upsampling 43
5 Including Padding in the nn.ConvTranspose2d Constructor 47
6 The Building Blocks for mUnet 51
7 The mUnet Network for Semantic Segmentation 56
8 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation 61
9 Training the mUnet 70
10 Testing mUnet on Unseen Data 72
Purdue University 51
 The Building Blocks for mUnet

The Two Building Blocks for the mUnet Network

The building blocks of mUnet are the two network classes SkipBlockDN
and SkipBlockUP, the former for the encoder in which the size of the
image becomes progressively smaller as the data abstraction level
becomes higher and the latter for the decoder whose job is to map
the abstractions produced by the encoder back to the pixel level.

We need two different types of building blocks on account of the

different needs of the encoder and the decoder. The network
SkipBlockDN is based on regular convolutional layers that when used
with strides will give us a progression of reduced resolution
representations for an image.

And the network SkipBlockUP for the decoder is based on transpose

convolutions that will gradually increase the resolution in the output
of the encoder until it is at the same level as the original image.

Purdue University 52
 The Building Blocks for mUnet

The Two Building Blocks for mUnet (contd.)

In the code that I’ll show for SkipBlockDN and SkipBlockUP, do not
be confused by the significance of the skip connections in these two
classes. These are not as essential as the skip connections you will see
in mUnet.

I’m using the skip connections in SkipBlockDN and SkipBlockUP to

make the learning in these building blocks a bit more efficient. As it
turns out, there is a well-known semantic segmentation network out
there called Tiramisu that also has skip connections in the encoder
and the decoder.

On the other hand, the skip connections that you will see in mUnet are
shortcuts between the corresponding levels of abstractions in the
encoder and the decoder. Those skip connections are critical to the
operation of the encoder-decoder combo if the goal is to see
pixel-level semantic segmentation results.
Purdue University 53
 The Building Blocks for mUnet

SkipBlockDN — The Definition

class SkipBlockDN(nn.Module):
def __init__(self, in_ch, out_ch, downsample=False, skip_connections=True):
super(DLStudio.SemanticSegmentation.SkipBlockDN, self).__init__()
self.downsample = downsample
self.skip_connections = skip_connections
self.in_ch = in_ch
self.out_ch = out_ch
self.convo1 = nn.Conv2d(in_ch, out_ch, 3, stride=1, padding=1)
self.convo2 = nn.Conv2d(in_ch, out_ch, 3, stride=1, padding=1)
self.bn1 = nn.BatchNorm2d(out_ch)
self.bn2 = nn.BatchNorm2d(out_ch)
if downsample:
self.downsampler = nn.Conv2d(in_ch, out_ch, 1, stride=2)
def forward(self, x):
identity = x
out = self.convo1(x)
out = self.bn1(out)
out = torch.nn.functional.relu(out)
if self.in_ch == self.out_ch:
out = self.convo2(out)
out = self.bn2(out)
out = torch.nn.functional.relu(out)
if self.downsample:
out = self.downsampler(out)
identity = self.downsampler(identity)
if self.skip_connections:
if self.in_ch == self.out_ch:
out += identity
else:
out[:,:self.in_ch,:,:] += identity
out[:,self.in_ch:,:,:] += identity
return out
Purdue University 54
 The Building Blocks for mUnet

SkipBlockUP — The Definition

class SkipBlockUP(nn.Module):
def __init__(self, in_ch, out_ch, upsample=False, skip_connections=True):
super(DLStudio.SemanticSegmentation.SkipBlockUP, self).__init__()
self.upsample = upsample
self.skip_connections = skip_connections
self.in_ch = in_ch
self.out_ch = out_ch
self.convoT1 = nn.ConvTranspose2d(in_ch, out_ch, 3, padding=1)
self.convoT2 = nn.ConvTranspose2d(in_ch, out_ch, 3, padding=1)
self.bn1 = nn.BatchNorm2d(out_ch)
self.bn2 = nn.BatchNorm2d(out_ch)
if upsample:
self.upsampler = nn.ConvTranspose2d(in_ch, out_ch, 1, stride=2, dilation=2,
output_padding=1, padding=0)
def forward(self, x):
identity = x
out = self.convoT1(x)
out = self.bn1(out)
out = torch.nn.functional.relu(out)
if self.in_ch == self.out_ch:
out = self.convoT2(out)
out = self.bn2(out)
out = torch.nn.functional.relu(out)
if self.upsample:
out = self.upsampler(out)
identity = self.upsampler(identity)
if self.skip_connections:
if self.in_ch == self.out_ch:
out += identity
else:
out += identity[:,self.out_ch:,:,:]
return out
Purdue University 55
 The mUnet Network for Semantic Segmentation

Outline

1 A Brief Survey of Networks for Semantic Segmentation 8

2 Transpose Convolutions 18
3 Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions 34
4 Using the stride Parameter of nn.ConvTranspose2d for
Upsampling 43
5 Including Padding in the nn.ConvTranspose2d Constructor 47
6 The Building Blocks for mUnet 51
7 The mUnet Network for Semantic Segmentation 56
8 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation 61
9 Training the mUnet 70
10 Testing mUnet on Unseen Data 72
Purdue University 56
 The mUnet Network for Semantic Segmentation

The Architecture of mUnet

This network is called mUnet because it is intended for segmenting out
multiple objects simultaneously form an image. [A weaker reason for
”Multi” in the name of the class is that it uses skip connections not
only across the two arms of the ”U”, but also also along the arms.]
Shown on the next slide is the implementation code for mUnet that is
in the inner class SemanticSegmentation of the DLStudio module.
As is the case with all such classes, the constructor of the class mUnet
defines the components elements that are subsequently used in the
definition of forward() to create a network.
Note how I save in lines labeled (A), (B), (C) a portion of the data
that is passing through the different levels of the encoder. Also note
how I subsequently mix in this saved data at the corresponding levels
in the decoder in lines (D), (E), and (F) of the code.

Purdue University 57
 The mUnet Network for Semantic Segmentation

mUnet – The Definition

class mUnet(nn.Module):
def __init__(self, skip_connections=True, depth=16):
super(DLStudio.SemanticSegmentation.mUnet, self).__init__()
self.depth = depth // 2
self.conv_in = nn.Conv2d(3, 64, 3, padding=1)
self.pool = nn.MaxPool2d(2, 2)
## For the DN arm of the U:
self.bn1DN = nn.BatchNorm2d(64)
self.bn2DN = nn.BatchNorm2d(128)
self.skip64DN_arr = nn.ModuleList()
for i in range(self.depth):
self.skip64DN_arr.append(DLStudio.SemanticSegmentation.SkipBlockDN(64, 64, skip_connections=skip_co
self.skip64dsDN = DLStudio.SemanticSegmentation.SkipBlockDN(64, 64, downsample=True, skip_connections=s
self.skip64to128DN = DLStudio.SemanticSegmentation.SkipBlockDN(64, 128, skip_connections=skip_connectio
self.skip128DN_arr = nn.ModuleList()
for i in range(self.depth):
self.skip128DN_arr.append(DLStudio.SemanticSegmentation.SkipBlockDN(128, 128, skip_connections=skip
self.skip128dsDN = DLStudio.SemanticSegmentation.SkipBlockDN(128,128, downsample=True, skip_connections

## For the UP arm of the U:

self.bn1UP = nn.BatchNorm2d(128)
self.bn2UP = nn.BatchNorm2d(64)
self.skip64UP_arr = nn.ModuleList()
for i in range(self.depth):
self.skip64UP_arr.append(DLStudio.SemanticSegmentation.SkipBlockUP(64, 64, skip_connections=skip_co
self.skip64usUP = DLStudio.SemanticSegmentation.SkipBlockUP(64, 64, upsample=True, skip_connections=ski
self.skip128to64UP = DLStudio.SemanticSegmentation.SkipBlockUP(128, 64, skip_connections=skip_connectio

(Continued on the next slide .....)

Purdue University 58
 The mUnet Network for Semantic Segmentation

mUnet – The Definition (contd.)

(...... continued from the previous slide)
self.skip128UP_arr = nn.ModuleList()
for i in range(self.depth):
self.skip128UP_arr.append(DLStudio.SemanticSegmentation.SkipBlockUP(128, 128, skip_connections=skip
self.skip128usUP = DLStudio.SemanticSegmentation.SkipBlockUP(128,128, upsample=True, skip_connections=s
self.conv_out = nn.ConvTranspose2d(64, 5, 3, stride=2,dilation=2,output_padding=1,padding=2)

def forward(self, x):

## Going down to the bottom of the U:
x = self.pool(torch.nn.functional.relu(self.conv_in(x)))
for i,skip64 in enumerate(self.skip64DN_arr[:self.depth//4]):
x = skip64(x)
num_channels_to_save1 = x.shape[1] // 2
save_for_upside_1 = x[:,:num_channels_to_save1,:,:].clone() ## (A)
x = self.skip64dsDN(x)
for i,skip64 in enumerate(self.skip64DN_arr[self.depth//4:]):
x = skip64(x)
x = self.bn1DN(x)
num_channels_to_save2 = x.shape[1] // 2
save_for_upside_2 = x[:,:num_channels_to_save2,:,:].clone() ## (B)
x = self.skip64to128DN(x)
for i,skip128 in enumerate(self.skip128DN_arr[:self.depth//4]):
x = skip128(x)
x = self.bn2DN(x)
num_channels_to_save3 = x.shape[1] // 2

(Continued on the next slide .....)

Purdue University 59
 The mUnet Network for Semantic Segmentation

mUnet – The Definition (contd.)

(...... continued from the previous slide)

save_for_upside_3 = x[:,:num_channels_to_save3,:,:].clone() ## (C)

for i,skip128 in enumerate(self.skip128DN_arr[self.depth//4:]):
x = skip128(x)
x = self.skip128dsDN(x)
## Coming up from the bottom of U on the other side:
x = self.skip128usUP(x)
for i,skip128 in enumerate(self.skip128UP_arr[:self.depth//4]):
x = skip128(x)
x[:,:num_channels_to_save3,:,:] = save_for_upside_3 ## (D)
x = self.bn1UP(x)
for i,skip128 in enumerate(self.skip128UP_arr[:self.depth//4]):
x = skip128(x)
x = self.skip128to64UP(x)
for i,skip64 in enumerate(self.skip64UP_arr[self.depth//4:]):
x = skip64(x)
x[:,:num_channels_to_save2,:,:] = save_for_upside_2 ## (E)
x = self.bn2UP(x)
x = self.skip64usUP(x)
for i,skip64 in enumerate(self.skip64UP_arr[:self.depth//4]):
x = skip64(x)
x[:,:num_channels_to_save1,:,:] = save_for_upside_1 ## (F)
x = self.conv_out(x)
return x

Purdue University 60
 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation

Outline

1 A Brief Survey of Networks for Semantic Segmentation 8

2 Transpose Convolutions 18
3 Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions 34
4 Using the stride Parameter of nn.ConvTranspose2d for
Upsampling 43
5 Including Padding in the nn.ConvTranspose2d Constructor 47
6 The Building Blocks for mUnet 51
7 The mUnet Network for Semantic Segmentation 56
8 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation 61
9 Training the mUnet 70
10 Testing mUnet on Unseen Data 72
Purdue University 61
 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation

The PurdueShapes5MultiObject Dataset

I have created an annotated dataset, PurdueShapes5MultiObject, for
the training and testing of networks for semantic segmentation —
especially if the goal is to segment out multiple objects
simultaneously. Each image in the dataset is of size 64 × 64.

The program that generates the PurdueShapes5MultiObject dataset

is an extension of the program that generated the PurdueShapes5
dataset you saw earlier in the context of object detection and
localization. Whereas each image in the PurdueShapes5 dataset
contained a single randomly scaled and randomly oriented object at
some random location, each image in the PurdueShapes5MultiObject
can contain a random number of up to five objects.

The images in the PurdueShapes5MultiObject dataset come with two

annotations, the masks for each of the objects and the bounding
boxes. Even when the objects are overlapping in the images, the
masks remain distinct in a mask array whose shape is (5, 64, 64).
Purdue University 62
 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation

Some Example Images from the

PurdueShapes5MultiObject Dataset
The next four slides show some example images from the
PurdueShapes5MultiObject dataset.

Note in particular that each image comes with a multi-valued mask

for the object it contains. The gray values assigned to the objects in
the mask are according to the following rule:
rectangle : 50 oval : 200
triangle : 100 star : 250
disk : 150

A most important aspect of mUnet learning is that even when the

objects are overlapping in the input RGB image, the 5-layer deep
mask array associated with the input image is “aware” of the full
extent of each object in the input image. The mask array in each
object-type-specific channel at the input is used as the target for the
prediction in the corresponding channel of the output for calculating
theUniversity
Purdue loss. 63
 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation

Example Images from the PurdueShapes5MultiObject Dataset (contd.)

Figure: Left: A dataset image; Middle: Its multi-valued mask; Right: BBoxes for the objects

Figure: Left: A dataset image; Middle: Its multi-valued mask; Right: BBoxes for the objects

Purdue University 64
 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation

Example Images from the PurdueShapes5MultiObject Dataset (contd.)

Figure: Left: A dataset image; Middle: Its multi-valued mask; Right: BBoxes for the objects

Figure: Left: A dataset image; Middle: Its multi-valued mask; Right: BBoxes for the objects

Purdue University 65
 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation

Example Images from the PurdueShapes5MultiObject Dataset (contd.)

Figure: Left: A dataset image; Middle: Its multi-valued mask; Right: BBoxes for the objects

Figure: Left: A dataset image; Middle: Its multi-valued mask; Right: BBoxes for the objects

Purdue University 66
 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation

Example Images from the PurdueShapes5MultiObject Dataset (contd.)

Figure: Left: A dataset image; Middle: Its multi-valued mask; Right: BBoxes for the objects

Figure: Left: A dataset image; Middle: Its multi-valued mask; Right: BBoxes for the objects

Purdue University 67
 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation

Downloading the PurdueShapes5MultiObject

Dataset
This dataset is available in the following files in the “data’
subdirectory of the “Examples” directory of the DLStudio distribution
(version 1.0.9). You will see the following archive files there:
PurdueShapes5MultiObject-10000-train.gz
PurdueShapes5MultiObject-1000-test.gz
PurdueShapes5MultiObject-20-train.gz
PurdueShapes5MultiObject-20-test.gz

You will find the two smaller datasets, with just 20 images each,
useful for debugging your code. You would want your own training
and the evaluation scripts to run without problems on these two
datasets before you let them loose on the larger datasets.

Purdue University 68
 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation

Data Format Used for the

PurdueShapes5MultiObject Dataset
Each 64 × 64 image in the dataset is stored using the following
format:
Image stored as the list:
[R, G, B, mask_array, mask_val_to_bbox_map]
where
R : is a 4096 element list of int values for the red component
of the color at all the pixels
B : the same as above but for the blue component of the color
G : the same as above but for the green component of the color
mask_array : is an array whose shape is (5,64,64). Each 64x64 "plane"
in the array is a mask corresponding to the first index
for that plane. The mask assignments are as follows:
mask_array[0] : rectangle [val stored: 50]
mask_array[1] : triangle [val stored: 100]
mask_array[2] : disk [val stored: 150]
mask_array[3] : oval [val stored: 200]
mask_array[4] : star [val stored: 250]
The values stored in the masks are different for the
different shapes as shown in the third column above.
mask_val_to_bbox_map : is a dictionary that tells us what bounding-box
rectangle to associate with each shape in the image. To
illustrate what this dictionary looks like, assume that an image
contains only one rectangle and only one disk, the dictionary
in this case will look like:
mask values to bbox mappings: {200: [],
250: [],
100: [],
50: [[56, 20, 63, 25]],
150: [[37, 41, 55, 59]]}
Should there happen to be two rectangles in the same image,
the dictionary would then be like:
mask values to bbox mappings: {200: [],
250: [],
100: [],
50: [[56, 20, 63, 25], [18, 16, 32, 36]],
150: [[37, 41, 55, 59]]}

Purdue University 69
 Training the mUnet

Outline

1 A Brief Survey of Networks for Semantic Segmentation 8

2 Transpose Convolutions 18
3 Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions 34
4 Using the stride Parameter of nn.ConvTranspose2d for
Upsampling 43
5 Including Padding in the nn.ConvTranspose2d Constructor 47
6 The Building Blocks for mUnet 51
7 The mUnet Network for Semantic Segmentation 56
8 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation 61
9 Training the mUnet 70
10 Testing mUnet on Unseen Data 72
Purdue University 70
 Training the mUnet

Training mUnet with MSE Loss

Shown below are the training loss values calculated in the script
semantic segmentation.py in the Examples directory of the distro. I
executed the script with the batch size set to 4, number of epochs to
6, the learning rate to 10−4 , and the momentum to 0.9.
[epoch:1,batch: 1000] MSE loss: 453.626
[epoch:1,batch: 2000] MSE loss: 445.210
[epoch:2,batch: 1000] MSE loss: 426.408
[epoch:2,batch: 2000] MSE loss: 414.780
[epoch:3,batch: 1000] MSE loss: 413.198
[epoch:3,batch: 2000] MSE loss: 398.390
[epoch:4,batch: 1000] MSE loss: 396.653
[epoch:4,batch: 2000] MSE loss: 389.081
[epoch:5,batch: 1000] MSE loss: 388.792
[epoch:5,batch: 2000] MSE loss: 387.379
[epoch:6,batch: 1000] MSE loss: 386.174
[epoch:6,batch: 2000] MSE loss: 381.055

What’s interesting is that if you print out the MSE loss for several
iterations when the training has just started, you will see loss numbers
much smaller than what are shown above. However, it is easy to
verify that those represent states of the network in which it has
overfitted to the training data. A classic example of overfitting is low
training error but large error on unseen data.
Purdue University 71
 Testing mUnet on Unseen Data

Outline

1 A Brief Survey of Networks for Semantic Segmentation 8

2 Transpose Convolutions 18
3 Understanding the Relationship Between Kernel Size, Padding,
and Output Size for Transpose Convolutions 34
4 Using the stride Parameter of nn.ConvTranspose2d for
Upsampling 43
5 Including Padding in the nn.ConvTranspose2d Constructor 47
6 The Building Blocks for mUnet 51
7 The mUnet Network for Semantic Segmentation 56
8 The PurdueShapes5MultiObject Dataset for Semantic
Segmentation 61
9 Training the mUnet 70
10 Testing mUnet on Unseen Data 72
Purdue University 72
 Testing mUnet on Unseen Data

Results on Unseen Test Data

As mentioned previously on Slide 36, the PurdueShapes5MultiObject

dataset comes with a test dataset that contain 1000 instances of
multi-object images produced by the same randomizing program that
generates the training dataset.

Shown in the next three slides are the some typical results on this
unseen test dataset. You can see these results for yourself by
executing the script semantic segmentation.py in the Examples
directory of the distribution.

The top row in each display is the combined result from all the five
output channels. The bounding boxes from the dataset are simply
overlaid into this composite output. The second row shows the input
to the network. The semantic segmentations are shown in the lower
five rows.

Purdue University 73
 Testing mUnet on Unseen Data

Results on Unseen Test Data (contd.)

(a) Lower five rows show semantic segmentation. Row 3: rect;

(b) Lower five rows show semantic segmentation. Row 3:
Row 4: triangle; Row 5: disk; Row 6: oval; Row 7: star
rect; Row 4: triangle; Row 5: disk; Row 6: oval; Row 7: star
Purdue University 74
 Testing mUnet on Unseen Data

Results on Unseen Test Data (contd.)

(a) Lower five rows show semantic segmentation. Row 3: rect;(b) Lower five rows show semantic segmentation. Row 3:
Row 4: triangle; Row 5: disk; Row 6: oval; Row 7: star rect; Row 4: triangle; Row 5: disk; Row 6: oval; Row 7: star

Purdue University 75
 Testing mUnet on Unseen Data

Results on Unseen Test Data (contd.)

Purdue University 76