FIT5217代写、Python程序语言代写

日期：2022-05-30 08:02

FIT5217 (S1 2022) - Assignment 2

Marks Worth 80 marks, and 20% of all marks for the unit

Due Date Friday, 27 May 2022, 11:55 PM

Extension An extension could be granted under some circumstances. A special consideration

application form must be submitted. Please refer to the university webpage on

special consideration.

Lateness For all assessment items handed in after the official due date, and without an agreed

extension, a 10% penalty applies to the student’s mark for each day after the due

date (including weekends) for up to 7 days. Assessment items handed in after 7

days without special consideration will not be considered.

Authorship This is an individual assessment. All work must be your own. All submissions will

be placed through Turnitin. This makes plagiarism remarkably easy to identify for

us.

Submission 3 files: a Jupyter notebook file, a pdf print of the Jupyter notebook, and a 4 page

pdf report need to be submitted. All answers in the report should be typed (no

scanned documents). The name of the files must be report_012345678.pdf,

code_012345678.ipynb, code_012345678.pdf where “012345678” is replaced by

your own student ID. The report should be A4 size with standard margins and 11

point.

Table 1: Instructions for FIT5217 (S1 2022) - Assignment 2

1

Introduction

Came back home after a long day, opened your fridge, you got a few ingredients in there, but have

no clue how to cook something with them. You reach out for your and call your mom, but she

does not answer your call. She had enough of you already! You reach out for your , search for

a cooking App, but surprisingly there is nothing in the App Store. 1 Well, you are a scientist, why

not build one yourself? You could train a deep neural model to do the job for you. This is simply a

Sequence-to-Sequence modelling task, mapping ingredients to a recipe.

Part 1 - Literature Review (Total 25 Marks)

No real research project could be done without a proper literature review. But due to time limit,

you will be only reviewing 4 papers, writing 2 paragraphs about each. Your summary should cover,

? Paper Summary: 1 paragraph summary [Maximum 100 words] of the work (could include

problem statement, novelty, method, key results). [2 Marks for each paper]

? Potential Ideas: 1 paragraph [Maximum 100 words] on the potential ideas you can borrow

from this paper for the recipe generation task. This could be anything from evaluation metrics,

to technical components to fix a potential undesired behaviour, an alternative modelling

framework for the task. Note, this is not a commitment to what you will do in Part 2, but

to assess your capability of identifying potential ideas for future extensions of your work. [3

Marks for each paper]

? The structure of your report for this part should have clear headings and be easy to navigate

and read. It should also follow the word limits stated above. [1.25 Marks for each paper]

NOTE: This should not exceed 1-1.5 page of your 4-page report. Font size is 11, and page is A4.

Paper: Chloé Kiddon, Luke Zettlemoyer, Yejin Choi. Globally Coherent Text Generation

with Neural Checklist Models. In Proceedings of the 2016 Conference on Empirical Methods

in Natural Language Processing.

Paper: Antoine Bosselut, Omer Levy, Ari Holtzman, Corin Ennis, Dieter Fox, Yejin Choi.

Simulating Action Dynamics with Neural Process Networks. In Proceedings of the 2018

International Conference on Learning Representations.

Paper: Abigail See, Peter J. Liu, Christopher D. Manning. Get To The Point: Summarization

with Pointer-Generator Networks. In Proceedings of the 2017 Annual Meeting of the Association

for Computational Linguistics.

Paper: Ximing Lu, Peter West, Rowan Zellers, Ronan Le Bras, Chandra Bhagavatula, Yejin

Choi. NeuroLogic Decoding: (Un)supervised Neural Text Generation with Predicate Logic

Constraints. In Proceedings of the 2021 Conference of the North American Chapter of

the Association for Computational Linguistics.

1Could someone actually check if there is an App for improvising food recipes? Share this in our Ed forum if you

find anything.

2

Part 2 - Neural Chef Assistant (Total 55 Marks)

The task is to train a deep neural model for recipe generation. The dataset we will use is from

the Kiddon et al paper you reviewed above. Download the dataset from Moodle. You should use

PyTorch for this part (need CE’s approval if you want to use a different deep learning library). The

test data should not be used for training. For each sample only use the ingredients and recipes, and

ignore the titles, servings, categories. A sample from the data is demonstrated in Table 2.

We want you to first build two baseline models:

? Baseline 1: Sequence-to-Sequence model without attention [4 Marks]

? Baseline 2: Sequence-to-Sequence model with attention [5 Marks]

And then extend Baseline 2 and build 2 other models [each worth 8 Marks] by choosing to do two

of the followings:

? Preprocessing data instead of using the original text (e.g., replacing all numbers with a unified

token NUM, etc)

? Using pretrained embeddings (e.g., word2vec, etc)

? Separating or sharing the embedding matrices for encoder and decoder

? Stacking up more layers in encoder or decoder

? Pretraining the encoder and decoder by training on Ingredients-to-Ingredients, or Recipes-toRecipes.

This is to train the encoder and decoder networks by reconstructing the same input

at the output (a.k.a. auto-encoding), and re-using the pretrained encoder and decoder for

Ingredients-to-Recipes.

? Adding the copy mechanism from Abigail et al, to encourage a better coverage of ingredients

in the recipes.

? Any other idea from the papers reviewed in Part 1.

Note: The Jupyter notebook you submit should contain all the code and outputs needed to answer

different sections of your report (see below). Make sure your code has clear markdowns that are

easy to navigate and locate which part of the notebook does what [5 Marks]. The discussion about

the results should go into the corresponding sections of the pdf report.

Sections to include in your report: The structure of your report should have clear headings and

be easy to navigate and read [5 Marks]. Avoid having several plots and tables if you could present

them compactly in a single plot or table. Use the space wisely and be creative. Font size is 11, and

page is A4. The report should cover the followings:

2.1 Model & Training Configurations. [2 Marks] For all your 4 models, use LSTM with

hidden_size 256 as encoder and decoder, teacher_forcing_ratio 1, Adam as the optimizer with

default hyperparameters, dropout rate 0.1, and set the maximum length to 150. You can choose

the dimensionality of the word embeddings, number of training iterations, batch size, etc. But stick

to what you choose in all your 4 models (do not change this across models). Report these model &

training configurations along with training time (i.e., minutes) for each model explicitly in a table.

3

You need to also report what decoding algorithm you are using. You also need to explicitly provide

the detail of the hardware used for training the model (if used Google colab, just write Google

Colab).

Title: chiles rellenos casserole

categories: vegetarian mexican main dish vegetables

servings: 10

ingredients: 2 cn whole green chili peppers 4 c milk 3 c sharp cheddar cheese

3/4 c all-purpose flour 4 green onions, sliced 1/4 ts salt 3 c shredded mozzarella

cheese 2 cn green chili salsa 6 eggs

recipe: split chili peppers lengthwise and remove seeds and pith . spread chilies

in a single layer in a greased 9x13-inch baking dish . sprinkle cheddar cheese ,

green onions , and 1-1/2 cups of the mozzarella cheese over chilies . in a bowl ,

beat eggs , milk , flour , and salt together until smooth . pour over chilies and

cheese . bake in a 325 degrees oven for 50 minutes or until a knife inserted in

custard comes out clean . meanwhile , mix salsa with the remaining 1-1/2 cups

mozzarella cheese . sprinkle over casserole and return to oven for 10 minutes or

until cheese melts . let stand for 5 minutes before serving .

END RECIPE

Table 2: A sample from data.

2.2 Data Statistics. [2 Marks] Have a table with the statistics from training, dev, test sets you

are using. This should report the statistics separately for ingredients, and recipes (i.e., have them

as two separate rows). Statistics you need to report should include (but not limited to): number

of samples, vocabulary size, min/max/average lengths of ingredients and recipes. Also explicitly

mention if you have used the entire training data, or due to computational resource limitations used

only a subset (if so, what percentage of the full training data you used?).

2.3 Data Preprocessing. [1 Marks] Mention if you have done any normalization or cleaning

(i.e., removing tags, etc), or you just fed the data as-is into your model.

2.4 Analysis. [5 Marks] Include Training and Dev set loss plots (x-axis being the iterations,

y-axis being the loss) for the 4 models. Discuss the plots: why one model performed better than

the other, do you observe any sign of overfitting, etc. Try to use color and line styles to creatively

and compactly present this in a single plot (i.e., you can use dashed and solid line styles to denote

training and dev, and 4 different colors to represent different models). The axes need to have labels,

and the plot needs a legend. Make sure axes ticks and labels are readable at 100% zoom.

2.5 Quantitative Evaluation. [5 Marks] Check Table 1 of Kiddon et al. They use 4 metrics to

quantitatively evaluate the generated recipes from their models on test set. Briefly describe these 4

metrics, report them for your 4 models. You can use NLTK to calculate BLEU and METEOR, but

need to implement the other two metrics (straightforward). Use a table similar to Kiddon et al to

compactly present this. Discuss the results, how model compares, is something working surprisingly

well? Have you outperformed any of the two baseline models? If not, what would you try in future?

4

2.6 Qualitative Evaluation. [5 Marks] Have a table (similar to Figure 4 of Kiddon et al) where

you report the generated output from each of your models for the following two ingredients lists:

? Ingredients Sample 1: 10 oz chopped broccoli, 2 tbsp butter, 2 tbsp flour, 1/2 tsp salt, 1/4

tsp black pepper, 1/4 tsp ground nutmeg, 1 cup milk, 1 1/2 cup shredded swiss cheese, 2 tsp

lemon juice, 2 cup cooked cubed turkey, 4 oz mushrooms, 1/4 cup grated Parmesan cheese, 1

can refrigerated biscuits

? Ingredients Sample 2: 2 lb cream cheese, 1 3/4 cups sugar, 5 eggs, 15 oreo cookies, 1/2 tsp

vanilla, 1/2 tsp almond extract

Discuss how models’ outputs compare. Do you observe anything wrong (i.e., repetition, extra

ingredients, etc)? Could you speculate why this happened, and what would you try differently

in future to address this? Do you see any correlation between the quantitative metrics and the

qualitative behavior of the models?

5