共计 8160 个字符，预计需要花费 21 分钟才能阅读完成。
COMP0090 Coursework 2 (originally the“group work”) –
Multi-task learning (LSA)
1 Introduction
Multi-task learning (MTL) is a field of machine learning which focuses on learning multiple tasks either
simultaneously or sequentially [1, 2, 3]. The aim of MTL is to leverage information from multiple tasks
in order to improve performance for either all the tasks being learnt, or for a single separate‘target task’.
In the case that target task(s) exist, auxiliary task(s) may be learnt prior to or simultaneously with the
target task(s). Under this definition,‘transfer learning’may be considered as a special case of MTL with
one auxiliary task and one target task, where the learning is sequential rather than simultaneous.
2 The project
In this project you will implement and evaluate a MTL framework for a target task of animal segmen-
tation. Learning auxiliary task(s) simultaneously with or prior to learning this target task may improve
segmentation performance. The choice and number of auxiliary tasks, as well as the decision of whether
learning is simultaneous or sequential, can all impact the target task performance in MTL. You will be
implementing a MTL framework and evaluating this with respect to a non-MTL baseline as well as to a
MTL framework that omits certain components (ablation study).
The goal of this project is to assess your ability to reason about design choices in deep learning ex-
periments, to implement deep learning frameworks, and to evaluate the developed deep learning models,
akin to conducting a research study. Your findings are to be compiled in a report which will follow a
similar structure to a research paper.
N.B. This assessment component is originally a group project. This late summer assessment (LSA)
has been adjusted and can also be completed by individuals, without the peer assessment.
2.1 The data set
The Oxford-IIIT Pet Data-set will be used in this project where a pre-processed version has been made
available in an open-source repository. The pre-processed data contains 3686 samples with three kinds
of task labels available for each sample: 1) binary classification of dog vs cat (‘binary.h5’); 2) animal
bounding boxes (‘bboxes.h5’); 3) animal segmentation masks (‘masks.h5’). In addition to these tasks,
other tasks that utilise self-supervision such as self-reconstruction of the images using auto-encoding,
predicting random rotations for images, predicting random flips for images etc. may be used, if suitable.
2.2 The“minimum required project”
The so-called minimum required project (MRP) (80%) for this group work consists of the following
points, which are expected to be found in your submission. Some of these points are discussed in more
detail to get you started.
- Deciding and justifying which MTL formulation to use
- Implementing a data loading pipeline
- Implementing a MTL framework
- Implementing a baseline network and an ablated version of MTL
- Experimenting for network comparison
1 - Describing implemented methods and conducted experiments
- Summarising obtained results and drawing conclusions
2.2.1 Deciding and justifying which MTL formulation to use
In this project we only focus on a MTL formulation with a single target task with potentially multiple
(≥1) auxiliary tasks. It should be noted that formulations with multiple target tasks or all tasks being
the target task are also valid MTL formulations. As introduced above, MTL may be formulated in one
of two ways: 1) simultaneous learning: auxiliary task(s) are learnt together with the target task; 2)
sequential/ transfer learning: auxiliary task(s) are learnt prior to the target task. In this project you
must decide which formulation you want to use as well as present a brief survey of both kinds of MTL
formulations (by citing previous work). The choice must be justified in the report.
Hint: some keywords that you can search for, are: 1)‘multi-task leanring’; 2)‘transfer learning’; 3)
‘parameter sharing’; 4)‘deep learning fine-tuning’; 5)‘shared weights’.
2.2.2 Implementing a data loading pipeline
For this working point you must implement one data generator/ loader to sequentially load batches of
data from h5 files, for the purpose of your project. For the loading pipeline in MTL, the output must be
compatible with the deep learning framework you use for this project (either TensorFlow or PyTorch).
Hint: some keywords that you can search for, are: 1)‘multi-task learning’; 2)‘multi-output models’;
3)‘python yield statement’; 4)‘python generator’; 5)‘tensorflow data generator’; 6)‘pytorch data loader’. - import h5py
- import os
- import numpy as np
4 number of samples to load from h5 file
- num_samples = 256
7 define the path for the h5 files (in this case we use the train set)
- h5_save_path = r’COMP0090_CW/train’
10 - img_path = r’images.h5’
- mask_path = r’masks.h5’
- bbox_path = r’bboxes.h5’
- bin_path = r’binary.h5’
15 - total_samples = 2210
17 generate random indices which are to be sampled from the dataset
- inds_to_sample = np.sort(np.random.choice(total_samples , size=num_samples , replace=False
))
20 define a function to load only samples at the specified indices
- def load_data_from_h5(path , inds_to_sample):
- with h5py.File(path ,’r’) as file:
- key = list(file.keys())[0]
- elems = filekey
- return elems
27 load data at the randomly generated indices
- masks = load_data_from_h5(os.path.join(h5_save_path , mask_path), inds_to_sample)
- imgs = load_data_from_h5(os.path.join(h5_save_path , img_path), inds_to_sample)
- bboxes = load_data_from_h5(os.path.join(h5_save_path , bbox_path), inds_to_sample)
- binarys = load_data_from_h5(os.path.join(h5_save_path , bin_path), inds_to_sample)
Listing 1: Simple loading
Note: if you fail to implement a suitable loading scheme, you can still proceed with the remainder
of this project by loading all or part of the data directly into memory using the code snippet above,
however, you will loose marks for this working point for doing so.
2.2.3 Implementing a MTL framework
Your MTL framework must follow one of the two schemes:
2 - Simultaneous learning: The auxiliary task(s) and target task are learnt simultaneously, where
your network‘shares parameters’in some layers whereas some layers are specific to the task and
non-shared (usually output layers). Your model may then be considered to have multiple output
branches, each learning a separate task. Once trained, you will only be required to evaluate your
model for the target task. You have freedom to implement any network that you deem to be
suitable, however, please note that you will not be marked for the complexity of your network, but
simply for the correctness of the parameter sharing and multi-output behaviour. - Sequential/ transfer learning: The auxiliary task is to be learnt prior to learning the target task.
Parameters are again shared between the two learning stages (e.g. you can learn the auxiliary
task, then freeze parameters in all layers except the last after which you can edit the output layer
according to your target task and train). After training, you will only need to evaluate your model
for the target task. You have freedom to implement any network that you deem to be suitable,
however, please note that you will not be marked for the complexity of your network, but simply
for the correctness of the parameter fixing and layer editing behaviour.
Other things to consider include, but are not limited to, network architecture, training strategies,
regularisers, hyperparamters for both types of tasks. Describe and justify these details in the report, if
considered relevant.
2.2.4 Implementing a baseline network and an ablated version of MTL
The baseline network must be a network capable of performing only your target task. Discuss what
should be considered a fair comparison, in aspects such as network architecture and/or training strategies,
compared with the MTL network.
For the ablation study, you must perform at least one ablation study by adjusting a hyperparameter
of the MTL network. Discuss why this is an interesting ablation study to do. Examples of hyperparam-
eters that can be ablated include 1) weighting(s) of auxiliary task(s); and 2) parameters controlling the
“quantity”or“speed”of the transfer.
2.2.5 Experimenting for network comparison
Design experiments which 1) compare the baseline to the MTL, and 2) compare the ablated version to
the MTL. Things to consider may include, but not limited to, planning data sets and model development,
metrics to use, how to quantitatively and statistically compare these metrics. Describe and justify these
options in the report, if relevant.
2.3 An“open-ended question”
To be awarded the remaining 20% in this project, you need to come up with a new research/study question
to answer, an open-ended question (OEQ). Novelty is encouraged in this part. The submission should
clearly identify and generate a hypothesis (i.e. the research/study question), design experiments that
produce results to answer this question and analyse the obtained experimental results for quantitative
conclusion. This part needs to be built on the MRP and relevant to MTL. Describe the question,
experiment and results clearly and cohesively with the MRP in the report.
Some example research/study questions are given as follows.
What other auxiliary tasks can be used
What exactly in the auxiliary task is used to help the target task
Is there any other things that can help the target task
What can be used to control the“transfer”between the target and auxiliary tasks
…… - Submission and Marking Criteria
Each submission should include 1) one report in a PDF file and 2) one huddle of code in a single zip file
