Description
Introduction
Last quarter, we explored the traditional machine learning algorithms, practiced with each type of algorithm by building models for three numerical datasets using Python’s Scikit-Learn library, and presented the results using a web-based application.
This quarter, we will continue our machine learning journey and shift our focus to deep learning i.e., deep neural networks and their application. In this first homework, we will get some hands-on experience with the Convolutional Neural Network (CNN), which are commonly used in computer vision tasks, and its implementation using TensorFlow. In this assignment, you will be designing, training and evaluating CNN models for a wafer classification task using free GPU/TPU resources available on Google Cloud.
Google CoLab
Google Colab is a Python development environment that runs in the browser using Google Cloud. It is similar to Jupyter Notebook but with the latest tensorflow package pre-installed. You can learn about CoLab’s basic features here. We will be mainly touching three features of Colab in this homework:
• Online python editing environment based on Jupyter.
• Virtual machine with private file system that’s associated with your Google account. • Free GPU and TPU resources.
Wafer Map Failure Pattern Analysis
In semiconductor manufacturing process, batches of functional circuits are fabricated on a slice of silicon material which we call a wafer. After going through complicated test stages, the wafer is cut (diced) into pieces, each containing one copy of the circuit. Each of these pieces is called a die. In companies like Intel and Qualcomm, one of the important ways to get revenue is to ensure a high yield of their manufactured wafers. Yield is the fraction of dies on the yielding wafers that are not discarded during the manufacturing process. The yield per wafer is quantified as the number of working (good/passing) dies divided by the total number of dies fabricated on that wafer.
Wafer maps provide visual details that locate the bad/failing dies on the wafer. By analyzing the failure pattern on a wafer, experienced engineers can often identify issues in manufacturing and/or test process that caused the failure and take actions accordingly in order to increase quality of production and ultimately to increase revenue.
Problem Formulation
For this homework, we will build and train a convolutional neural network model for classifying the failure types of given wafers using Tensorflow. You can follow the coding template and descriptions in the provided Colab notebook to implement this assignment.
1
Column Count
Wafer Map
Description
A list of wafer maps, each is a 64-by-64 Numpy array with data type “float32”.
Value -1.0 denotes good/passing,
1.0 denotes bad/failing,
0.0 denotes no die.
A list of labels, each is a string denotes the type of failure pattern from 6 categories:
“pass”, “edge”, “deform”, “total_loss”, “crack”, “nodule”
3300
Failure Type 3300
Table 1: Part One Data Set: Wafers Generated by Min Jian Yang (W21 157B TA)
Column
Lot Name Wafer Number TSMC Index Die Count
Count Description
904 A list of strings, each denoting the manufacturing lot name of a wafer. 904 A list of integers, each identifying a wafer in its lot.
904 A list of integers, each indexing a wafer for TSMC internal usage. 904 A list of integers, each denoting the number of dies on a wafer.
Wafer Map 904 Failure Type 904
Same format as in Table.1
but with 2 different shapes: 30-by-34, and 27-by-25 Same format as in Table.1
but from 4 categories: “Center”, “Edge-Loc”, “Loc”, “Scratch”
1
Table 2: Part Two Data Set: Real-world Wafers from TSMC [1]
Part one: Classification with a simple data set
We start with a simple data set generated by TAs. The data is stored as a Pandas DataFrame (a table). TAs provided the code for downloading the data set from a shared URLs to your Google Colab file system.
1.1 Data Set
This simple data set contains 3300 wafer maps and their labels, as shown in Table.1. There are 6 types of known labels: pass, total_loss, edge, nodule, crack and deform. The first 300 wafers in the data set are assigned with known labels and we will use these as our training and validation data set. The remaining 3000 wafers are labeled as ’unknown’ and we will use our trained CNN model to predict their labels. The accuracy of this prediction is our test accuracy to report on the unknown data. We expect to see a fairly high accuracy ( 95%) obtained on this first data set. You need to submit your prediction results as a csv file to GradeScope for grading.
1.2 Workflow
The following guiding tasks will help you implement the code and achieve the required accuracy. You are provided with a Colab notebook with incomplete code blocks for you to implement as well as question prompts for you to answer.
- (2 pts) Download the first data set and visualize the failure types by randomly selecting one wafer map from each failure type. Generate a single plot showing one wafer map image for each of the six types.
- (3 pts) Use at least three training and validation split ratio and show the validation prediction accuracy for each split ratio. Explain why the accuracy rises or drops in each configuration. Which ratio did you choose at the end?
- (2 pts) Provide the definitions for the following hyper-parameters.
1. (1 pt) Number of filters of a CNN layer 2. (1 pt) Kernel size of a CNN layer
2
- (2 pts) Why is the provided network using a Softmax activation function in the final dense layer? Can we use a different activation function instead?
- (272B only: 2 pts) We’ve learned Binary Cross Entropy loss function in class. Why is the provided netowrk using the Categorical Cross Entropy as loss function? How are the two related?
- (8 pts) Make these four modification to the model provided (one at a time) and explain how they affect the accuracy.
1. (2 pts) Add Dropout layers and explain its effect
2. (2 pts) Add more CNN layers and explain its effect
3. (2 pts) Add L2 regularization to CNN layers and explain its effect 4. (2 pts) Add L2 regularization to Dense layers and explain its effect - (6 pts) Use image augmentation to generate 100 more images for each failing class (total of five, excluding the ’pass’) and retrain your model with the added images. Did the perfor- mance of the retrained model improve or decrease? Image augmentation could be rotation, horizontal/vertical flip, etc.
- (2 pts) Show the confusion matrix for the validation predictions (include the axis values).
- (2 pts) Use the model with the best validation accuracy to predict labels for the unknown wafers. Submit your prediction csv file to GradeScope to see your accuracy. Are you able to achieve 95% accuracy on the unknown dataset?
*NOTE: you must build your model layer-by-layer. You are not allowed to call built-in models in Tensorflow (e.g. Xception, VGG16, ResNet…)
2 Part two: Classification with the TSMC data set
At this point, you should already have a good understanding of a CNN model. We will use the architecture you have configured from part one to train and evaluate on a real-world wafer data set from TSMC [1].
2.1 Data Set
The second data set contains 904 labeled wafer maps as shown in Table.2. There are 4 types of labels: Center, Edge, Edge-Loc, and Scratch.
2.2 Workflow
- (2 pts) Download the second data set and visualize the class types as you did in part one.
- (8 pts) Prepare data, use the same model architecture as in part one to train and evaluate this TSMC data set, and show the confusion matrix for the validation predictions. Are the accuracy results the same across all the categories?
- (4 pts) Were you able to get a good accuracy (> 95%) on this TSMC data set? If not, explain why the same model performed worse?
- (272B only: 10 pts) Study the paper where the TSMC data set was first proposed [2]. Drawing inspirations from this paper, provide possible solutions to improve the classification accuracy of this data set. (Even better if you come up with your original ideas.)
3
What to Turn In
Save your Google Colab notebook as .ipybn file and submit it to Gradescope: https://www. gradescope.com/courses/351582. Course entry code: 5V3YRX
The Colab notebook should contain the following parts for grading:
• Completed and functional code cells. (10 pts as specified in Colab)
• Answers to all the questions and diagrams/plots listed in the Problem Formulation. (41 pts for 157B, 53 pts for 272B, as specified above)
Note: Do not close the cell outputs after running the code cells, so that all logging and plots will be saved in the notebook for grading.
References
- [1] Ming-Ju Wu, Jyh-Shing R. Jang, and Jui-Long Chen. MIR-WM811K. http://mirlab.org/ dataSet/public/. 2015.
- [2] Ming-Ju Wu, Jyh-Shing R. Jang, and Jui-Long Chen. “Wafer Map Failure Pattern Recognition and Similarity Ranking for Large-Scale Data Sets”. In: IEEE Transactions on Semiconductor Manufacturing 28.1 (2015), pp. 1–12. doi: 10.1109/TSM.2014.2364237.
4
