Image_Classifier

In-class Practice: Gemstone Image Classification

Practical Platform and Environment Setup:

1. Alibaba Magic: You don't need to configure the environment separately
2. Others:

• Python Environment: Ensure you have Python installed (preferably Python 3.x).
• *Dependencies: Install required dependencies using pip:

pip install opencv-python numpy pillow matplotlib torch torchvision

Code Flow:

1. Data Preprocessing:

• Select the computing device (GPU or CPU) through torch.device ("CUDA" if torch.cuda.is_availability () else "CPU").
• Set training related parameters, such as input image size, number of categories , dataset path , label dictionary , etc.
• Code check and decompress the dataset archive_trainzip. If the data does not exist, decompress it.
• Generate train.txt and eval.txt, respectively storing the image paths and corresponding labels for the training set and validation set.

2. Model Initialization:

• Use torch.nn Module defines a Convolutional Neural Network (CNN) for image classification.
• Mainly composed of multiple nn Conv2d, nn.MaxPool2d, and nn Composed of Linear layers, it achieves feature extraction and classification.
• Using CrossEntropyLoss as the loss function and Adam as the optimizer.

3. Training and Evaluation:

• The code iterates through num_ epochs, performs forward propagation, loss calculation, gradient feedback on the training data, and updates the model parameters using an optimizer.
• After each epoch, calculate and print the loss and accuracy of the training sets.
• Save the best model weights using torch.
• Load the optimal model parameters, evaluate the model on the validation set, and output the accuracy. Use models to predict and visualize data on the test set.

Filling in the Blanks:

1. MyCNN Class:

• Fill in the initialization for the Convolutional Neural Network model.

2. Model training:

• Fill in the missing code in the model training module to enable the program to run smoothly.

Experimental Requirements:

1. Implementing Convolutional Neural Network (CNN) with Supplementary Code - 1.5 points

2. Supplement the model training code to ensure the program runs smoothly, avg_acc ≥ 0.4 - 1.5 points

Total Score: 3 points

Other exploration:

1. On the basis of the successful construction of convolutional neural networks, other networks can be attempted to be constructed

2. Using accuracy on the validation set as the evaluation metric, consider optimizing to improve accuracy

Submission:

Submit both the code and the running results in a single zip archive named "学号_姓名_课堂练习2.ipynb".

ddl: 2025/4/14 23:59