Machine Learning for Beginners : Artificial Neural Network Program to Recognize Hand Shapes that Form Rock, Paper, or Scissors by DICODING🌈🪄
I have successfully completed the ANN project. The detailed explanation of this project is as follows:
Rock paper scissors (commonly scissors, paper, rock or stone in Australia and New Zealand)is an intransitive hand game, usually played between two people, in which each player simultaneously forms one of three shapes with an outstretched hand. These shapes are "rock" (a closed fist), "paper" (a flat hand), and "scissors" (a fist with the index finger and middle finger extended, forming a V). The earliest form of "rock paper scissors"-style game originated in China and was subsequently imported into Japan, where it reached its modern standardized form, before being spread throughout the world in the early 20th century.
The name "rock paper scissors" is simply a translation of the Japanese words for the three gestures involved in the game, though the Japanese name for the game is different. The Japanese name is jan-ken-pon.
A simultaneous, zero-sum game, it has three possible outcomes: a draw, a win or a loss. A player who decides to play rock will beat another player who has chosen scissors ("rock crushes scissors" or "breaks scissors" or sometimes "blunts scissors"), but will lose to one who has played paper ("paper covers rock"); a play of paper will lose to a play of scissors ("scissors cuts paper"). If both players choose the same shape, the game is tied and is usually replayed to break the tie.
Rock paper scissors is often used as a fair choosing method between two people, similar to coin flipping, drawing straws, or throwing dice in order to settle a dispute or make an unbiased group decision. Unlike truly random selection methods, however, rock paper scissors can be played with some degree of skill by recognizing and exploiting non-random behavior in opponents.
-
Library Import: we can started by importing essential libraries such as TensorFlow, ImageDataGenerator, and other libraries needed to build and train a machine learning model.
-
Dataset Import: The dataset containing images of hands forming Rock, Paper, and Scissors was downloaded from Dicoding Academy. It was then extracted using the zipfile function.
-
ImageDataGenerator: ImageDataGenerator was used to perform data augmentation. Data augmentation aims to increase artificial data with transformations like rotation, shifting, and shear. This helps the model to be more generalizable and less prone to overfitting on the training data.
The parameters used in ImageDataGenerator include:
- rescale: scales the pixel values of the image from [0-255] to [0-1].
- rotation_range: randomly rotates the image within a specific range (degrees).
- width_shift_range, height_shift_range: randomly shifts the image on the horizontal and vertical axes.
- shear_range: randomly bends the image.
- zoom_range: randomly zooms in or out (reduces) the image.
- horizontal_flip: randomly flips the image horizontally.
- validation_split: splits the dataset into a training set (60%) and a validation set (40%).
-
Splitting the Dataset: ImageDataGenerator was then used to create generators for the training and validation sets. The training set was used to train the model, while the validation set was used to evaluate the model's performance during training.
-
Sequential Model: We can built the model using a sequential approach, where layers were added sequentially. The model consists of:
- Convolutional Layer (Conv2D): This layer extracts spatial features from the image. In this project, we can used 3 Conv2D layers, each with 32, 64, and 128 filters. Each layer was followed by MaxPooling2D to reduce the data dimension and prevent overfitting.
- Flatten Layer: This layer converts the data from a 3-dimensional tensor to a 1-dimensional tensor before entering the hidden layer.
- Dense Layer (Dense): The Dense layer acts as a fully-connected layer. we can used 2 Dense layers with 512 and 256 neurons each to learn high-level representations of the features extracted by the convolutional layer.
- Dropout Layer (Dropout): The Dropout layer is used to prevent overfitting by randomly disabling neurons during training. we can used dropout with a rate of 0.5.
- Output Layer (Dense): The output layer uses the softmax activation function to predict the probability of the three classes (Rock, Paper, Scissors).
-
Compile Model: The model was compiled with the Adam optimizer, categorical_crossentropy loss function, and accuracy metrics. The Adam optimizer helps to accelerate the training process to find optimal weights. Categorical crossentropy is used as the loss function for multi-class classification problems. Accuracy is used as a metric to measure the model's performance.
-
Early Stopping: The EarlyStopping callback was used to automatically stop training when the accuracy on the validation set did not improve significantly for 5 consecutive epochs. This prevents overfitting and saves training time.
-
Model Training: The model was trained using data from the training set. We can used a batch size of 32 and 100 epochs. During training, the accuracy on the training and validation sets was monitored.
-
Save Model: After training was completed, the model was saved with the name "rps_model.keras".
-
Total Training Time: The time required to train the model for 100 epochs was calculated and displayed in minutes.
-
Plot Accuracy and Loss: Plots of accuracy and loss on the training and validation sets were created using the Plotly library. These plots help to visualize how the model's performance changes during training. Ideally, the accuracy on the training set should increase and the loss should decrease with the epoch. The accuracy on the validation set is expected to be not too different from the accuracy on the training set, which indicates that the model is not overfitting.
-
Model Testing: To test the model, we can upload images of hands forming Rock, Paper, or Scissors. The model will predict the class of the image.
- Total training time: 17.65 minutes
- Model Training accuracy is 95%