A console/terminal pipeline for training and running an Xception image classifier with PyTorch, supporting Windows 11 and Ubuntu Linux 24.04. The engine scripts live under src/examples/imagenet/ and support several architectures through --arch (xception by default). The example dataset plate contains steel-plate surface defect images in 6 classes (CR, IN, PA, PS, RS, SC) and ships with a trained Xception checkpoint, so inference works out of the box.

 

[Folder Structure]

  • CLAUDE.md / AGENT.md: program specification (identical content).
  • CHANGELOG.md: change log.
  • copyright.txt: copyright notice.
  • 1610.02357.pdf: Xception reference paper.
  • data/{dataset}/: labels_map.txt (class list, regenerated by training) plus train/, val/ and test/ in torchvision ImageFolder layout (one subfolder per class).
  • model/{dataset}/: model_best.pth.tar (best checkpoint), checkpoint.pth.tar (latest), logs/ (TensorBoard).
  • tool/tools.py: shared helpers.
  • src/: Python entry points, engine (examples/imagenet/), launcher scripts and this document.
  • venv/: Python virtual environment, created by step 01.

[Prerequisites]

  • Operating system: Windows 11 or Ubuntu Linux 24.04.
  • Python 3 and pip.
  • An NVIDIA GPU is recommended for training; inference also runs on CPU.

[Instructions]

Step 1: Create the virtual environment (CPU)

Delete the old venv/, create a fresh one and install the packages listed in src/requirements.txt. The script aborts loudly if pip fails.

 

Windows: src\01_venv_create.bat

Linux: bash src/01_venv_create.sh

Step 2: Create the virtual environment (GPU)

Same as step 1, but swaps torch for the CUDA 12.8 (cu128) build, for machines with an NVIDIA GPU (driver CUDA version 12.8 or later, check with nvidia-smi).

 

Windows: src\01_venv_create_gpu.bat

Linux: bash src/01_venv_create_gpu.sh

Step 3: Activate the virtual environment

Activate the venv. On Linux, source the script.

 

Windows: src\02_venv_activate.bat

Linux: source src/02_venv_activate.sh

Step 4: Train the model

Train the image classifier. Defaults: dataset plate, architecture xception, 500 epochs, batch size 2, image size 512, learning rate 0.001. Training resumes from model_best.pth.tar when that file is present.

 

Windows: src\03_train.bat

Linux: bash src/03_train.sh

Step 5: Watch the training curves

Open the training logs in model/{dataset}/logs with TensorBoard. It accepts two positional arguments, [dataset] and [port], defaulting to plate and 6066.

 

Windows: src\04_tensorboard.bat

Linux: bash src/04_tensorboard.sh

Step 6: Classify a single image

Classify one image, by default data/plate/test/IN/IN-1.png.

 

Windows: src\05_inference.bat

Linux: bash src/05_inference.sh

Step 7: Classify a whole folder

Classify every image in a folder (data/{dataset}/test by default) and report Acc@1 and Acc@5.

 

Windows: src\06_inference_folder.bat

Linux: bash src/06_inference_folder.sh

Step 8: Folder inference with underkill and overkill statistics

Same as step 7, but additionally reports underkill, overkill and right-rate statistics.

 

Windows: src\07_inference_folder_topk.bat

Linux: bash src/07_inference_folder_topk.sh

Step 9: Remove the virtual environment

Delete the venv/ folder.

 

Windows: src\99_venv_clear.bat

Linux: bash src/99_venv_clear.sh

[Python Entry Points]

  • src/train.py: --dataset plate --arch xception --batch-size 2 --image-size 512 --epochs 500 --learning-rate 0.001 --workers 4 [--num-classes N] [--resume PATH] [--pretrained PATH]. The class count is derived from the class folders under data/{dataset}/train. Training resumes from model/{dataset}/model_best.pth.tar when it exists; pass an empty --resume to train from scratch. Checkpoints land in model/{dataset}/ and TensorBoard logs in model/{dataset}/logs/ (cleared per run).
  • src/inference.py: --dataset plate --arch xception --source FILE|FOLDER --batch-size 1 --image-size 512 [--num-classes N] [--weights PATH] [--topk-right] [--show]. A file source classifies one image; a folder source reports Acc@1 and Acc@5 over all images; --topk-right switches folder mode to the underkill/overkill/right-rate variant; --show displays each image and requires a desktop session.

[GPU]

The default PyPI torch wheel is CPU-only on Windows. For an NVIDIA GPU, install the build matching your driver's CUDA version (check with nvidia-smi), for example CUDA 12.8:

pip uninstall -y torch torchvision

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128

The engine auto-detects CUDA and falls back to CPU when it is unavailable.

 

[References]

 

This SDK is built in AppForAI - AI Dev Tools.

 

Purchase license separately:  USD 400, permanent authorization, single APP authorization, single machine authorization, one-year activation, one-year download, one-year update, one-year email technical support.