Task1: Training masked autoencoder on PASCAL VOC 2007 dataset Dataset preparation: Download the PASCAL VOC 2007 dataset from here. (http://host.robots.ox.ac.uk/pascal/VOC/voc2007/) Preprocess the dataset by resizing the images to a fixed size and normalizing the pixel values. Dataset split: (Run Task.1.3(a) with both the following splits and choose the best one for Task.1.3(b), Task.1.3(c) and next experiments.) Use 80-10-10, train-val-test split Use 70-10-20, train-val-test split Architecture: An autoencoder with three hidden layers with the following bottleneck dimension (on denoising task, add gaussian noise to the input image, and the output should be denoised image) 256 (run using both the splits in 2, and choose the best one for the further set of experiments) 128 64 32 16 Choose the best bottleneck dimension, and re-run the autoencoder using masking strategy: (mask the following % of pixels in the image, i.e., set the pixel value to (0,0,0)) 20% 40% 60% 80%. Plot reconstruction error for every autoencoder model. Evaluation: Report MSE (mean square error), MAE (mean absolute error) for all models. Comment on which metric is more useful in judging the quality . Visualize and compare the original images, masked images, and reconstructed images. Use any other metric of your choice (apart from MSE, MAE) to judge the image quality. Use the best split, and best masking strategy (with best encoding dimension) for Task 2. Task2: Fine-tuning a pre-trained autoencoder on STL-10 dataset Dataset: Download the STL-10 dataset from here (https://cs.stanford.edu/~acoates/stl10/)and pre-process as required. Use the encoder of the above-pretrained autoencoder (Refer to task 1.9) as a feature extractor. Build a downstream task classifier (a MLP) with (100% training samples) and choose the best one among a, and b for Task 2.4. # hidden_layer = 3 # hidden_layer = 5 Use the best from Task 2.3.a and 2.3.b, fine-tune the classifier on the STL-10 dataset with the following % of training samples 1% 10% 20% 40%. 60 % Prepare confusion matrix and AUC-ROC curve to evaluate the model performance (for Task 2.3.(a-b) and Task 2.4.(a-e)). Implement a different architecture than the architecture described in Task.2.3.(a-b), that improves the result. Task3: Report hyperparameters clearly for every model No. of epochs Learning rate Optimizer Loss function. Dimensions Report and justify your choice of best (for split, for bottleneck dimension, masking strategy etc.) Observations about results, any additional analysis over results.

Task1: Training masked autoencoder on PASCAL VOC 2007 dataset Dataset preparation: Download the PASCAL VOC 2007 dataset from here. (http://host.robots.ox.ac.uk/pascal/VOC/voc2007/) Preprocess the dataset by resizing the images to a fixed size and normalizing the pixel values. Dataset split: (Run Task.1.3(a) with both the following splits and choose the best one for Task.1.3(b), Task.1.3(c) and next experiments.) Use 80-10-10, train-val-test split Use 70-10-20, train-val-test split Architecture: An autoencoder with three hidden layers with the following bottleneck dimension (on denoising task, add gaussian noise to the input image, and the output should be denoised image) 256 (run using both the splits in 2, and choose the best one for the further set of experiments) 128 64 32 16 Choose the best bottleneck dimension, and re-run the autoencoder using masking strategy: (mask the following % of pixels in the image, i.e., set the pixel value to (0,0,0)) 20% 40% 60% 80%. Plot reconstruction error for every autoencoder model. Evaluation: Report MSE (mean square error), MAE (mean absolute error) for all models. Comment on which metric is more useful in judging the quality . Visualize and compare the original images, masked images, and reconstructed images. Use any other metric of your choice (apart from MSE, MAE) to judge the image quality. Use the best split, and best masking strategy (with best encoding dimension) for Task 2. Task2: Fine-tuning a pre-trained autoencoder on STL-10 dataset Dataset: Download the STL-10 dataset from here (https://cs.stanford.edu/~acoates/stl10/)and pre-process as required. Use the encoder of the above-pretrained autoencoder (Refer to task 1.9) as a feature extractor. Build a downstream task classifier (a MLP) with (100% training samples) and choose the best one among a, and b for Task 2.4. # hidden_layer = 3 # hidden_layer = 5 Use the best from Task 2.3.a and 2.3.b, fine-tune the classifier on the STL-10 dataset with the following % of training samples 1% 10% 20% 40%. 60 % Prepare confusion matrix and AUC-ROC curve to evaluate the model performance (for Task 2.3.(a-b) and Task 2.4.(a-e)). Implement a different architecture than the architecture described in Task.2.3.(a-b), that improves the result. Task3: Report hyperparameters clearly for every model No. of epochs Learning rate Optimizer Loss function. Dimensions Report and justify your choice of best (for split, for bottleneck dimension, masking strategy etc.) Observations about results, any additional analysis over results.

Database System Concepts

7th Edition

ISBN:9780078022159

Author:Abraham Silberschatz Professor, Henry F. Korth, S. Sudarshan

Publisher:Abraham Silberschatz Professor, Henry F. Korth, S. Sudarshan

Chapter1: Introduction

Section: Chapter Questions

Problem 1PE

See similar textbooks

Related questions

Question

Task1: Training masked autoencoder on PASCAL VOC 2007 dataset

Dataset preparation:
1. Download the PASCAL VOC 2007 dataset from here. (http://host.robots.ox.ac.uk/pascal/VOC/voc2007/)
2. Preprocess the dataset by resizing the images to a fixed size and normalizing the pixel values.
Dataset split: (Run Task.1.3(a) with both the following splits and choose the best one for Task.1.3(b), Task.1.3(c) and next experiments.)
1. Use 80-10-10, train-val-test split
2. Use 70-10-20, train-val-test split
Architecture: An autoencoder with three hidden layers with the following bottleneck dimension (on denoising task, add gaussian noise to the input image, and the output should be denoised image)
1. 256 (run using both the splits in 2, and choose the best one for the further set of experiments)
2. 128
3. 64
4. 32
5. 16
Choose the best bottleneck dimension, and re-run the autoencoder using masking strategy: (mask the following % of pixels in the image, i.e., set the pixel value to (0,0,0))
1. 20%
2. 40%
3. 60%
4. 80%.
Plot reconstruction error for every autoencoder model.
Evaluation: Report MSE (mean square error), MAE (mean absolute error) for all models. Comment on which metric is more useful in judging the quality .
Visualize and compare the original images, masked images, and reconstructed images.
Use any other metric of your choice (apart from MSE, MAE) to judge the image quality.
Use the best split, and best masking strategy (with best encoding dimension) for Task 2.

Task2: Fine-tuning a pre-trained autoencoder on STL-10 dataset

Dataset: Download the STL-10 dataset from here (https://cs.stanford.edu/~acoates/stl10/)and pre-process as required.
Use the encoder of the above-pretrained autoencoder (Refer to task 1.9) as a feature extractor.
Build a downstream task classifier (a MLP) with (100% training samples) and choose the best one among a, and b for Task 2.4.
1. # hidden_layer = 3
2. # hidden_layer = 5
Use the best from Task 2.3.a and 2.3.b, fine-tune the classifier on the STL-10 dataset with the following % of training samples
1. 1%
2. 10%
3. 20%
4. 40%.
5. 60 %
Prepare confusion matrix and AUC-ROC curve to evaluate the model performance (for Task 2.3.(a-b) and Task 2.4.(a-e)).
Implement a different architecture than the architecture described in Task.2.3.(a-b), that improves the result.

Task3:

Report hyperparameters clearly for every model
1. No. of epochs
2. Learning rate
3. Optimizer
4. Loss function.
5. Dimensions
Report and justify your choice of best (for split, for bottleneck dimension, masking strategy etc.)
Observations about results, any additional analysis over results.

Expert Solution

Trending now

This is a popular solution!

Step by step

Solved in 3 steps

SEE SOLUTION Check out a sample Q&A here

Knowledge Booster

Learn more about

Need a deep-dive on the concept behind this application? Look no further. Learn more about this topic, computer-science and related others by exploring similar questions and additional content below.