Skip to content

Latest commit

 

History

History
81 lines (58 loc) · 3.95 KB

README.md

File metadata and controls

81 lines (58 loc) · 3.95 KB

CV-Starter: Image Classification Using CNN

git clone https://github.com/AIFep-FDU/CV-Starter.git

Environment Setup:

  1. Python Environment: Ensure you have Python installed (preferably Python 3.x).
  2. Dependencies: Install required dependencies using pip:
pip install torch torchvision pytorchsummary numpy matplotlib tqdm
  1. Colab: no need to pip

Code Flow:

  1. Model Definition:
  • The code defines a Convolutional Neural Network (CNN) model for image classification.
  • ResNet-18 architecture is used as an example, but you can analyze and discuss other CNN architectures as well.
  1. Training and Testing:
  • The code includes functions for training and testing the CNN model.
  • It utilizes datasets provided by PyTorch and DataLoader for efficient data handling.
  1. Data Prepare Augmentation:
  • Create a dataset class to read VOC2007 annotations and get its sub-images for classification.
  • You need to design data augmentation techniques to improve accuracy. This may include rotations, flips, or other transformations.
  1. Optimizer and Learning Rate Scheduler:
  • Stochastic Gradient Descent (SGD) optimizer with momentum is used.
  • Learning rate scheduling is implemented using ReduceLROnPlateau scheduler.
  1. Experimental Requirements:
    • Complete the model training and validation code.
    • Crop the image directly and design data augmentation (including normalization) techniques to enhance accuracy.
    • Analyze and discuss the design of other optimizers and learning rate schedules.
    • Analyze and discuss different CNN or other model (based on timm) architectures.

Filling in the Blanks:

  1. Custom Dataset Class:

    • Complete the __getitem__ method to crop the image based on the bounding box and apply transformations.

  2. Model Architecture:

    • Adjust the ResNet model architecture by modifying the BasicBlock and ResNet classes.
    • Consider using other network designs or directly using models from libraries like timm for comparison.

  3. Optimizer and Scheduler Configuration:

    • Modify the optimizer and learning rate scheduler settings in the main function to find the optimal parameters for training.
    • Fill in the code for updating the learning rate scheduler after each epoch.

  4. Normalization Values:

    • Adjust the normalization values in the train_transforms and test_transforms to better suit the VOC2007 dataset.
    • Obtain these values through statistical analysis or use predefined values from similar datasets.

  5. Loss Function and Backpropagation:

    • In the model_training function, fill in the code for calculating the loss using a suitable loss function and performing backpropagation.

  6. Model Output and Loss Calculation in Testing:

    • In the model_testing function, fill in the code for obtaining the model's output and calculating the loss during the testing phase.

  7. Data Augmentation:

    • Design image augmentation techniques inside the train_transforms to improve model performance.
    • Consider including transformations like random cropping, flipping, rotation, and color jitter.

Running the Script:

  • Execute the script main.py from the command line: python main.py.
  • Adjust parameters and configurations as needed for experimentation.
  • For colab, just click the "run" button.

Submission:

After training your model, you can execute the python infer.py script to generate predictions for the 'test_1k' dataset. Simply zip the output and submit it directly to the CodaLab Competition.

Submit the code along with an experimental report (one page, No involution) to the e-learning platform.

Note:

  • Modify and adjust the arguments, configurations, and experimental setup as necessary.
  • Provide detailed explanations in the experimental report, covering code implementation, experimental setup, results, analysis, and discussions based on the specified requirements.