26 KiB
26 KiB
LLM Prompts for Torchvision Finetuning Project
Prompt 1: Project Foundation Setup
Based on the project specification (`project-spec.md`), set up the initial project structure and tooling.
1. Create the following directory structure within the current directory:
```
├── configs/
├── data/
├── models/
├── utils/
├── tests/
├── scripts/
├── .gitignore # (Created in step 2)
├── pyproject.toml # (Created in step 3)
├── pre-commit-config.yaml # (Created in step 4)
├── README.md # (Created in step 7)
├── train.py # (Created in step 6)
└── test.py # (Created in step 6)
```
2. Initialize a git repository in the `torchvision-tutorial` directory.
3. Create a `.gitignore` file suitable for a Python project, ignoring directories like `data/`, `outputs/`, `logs/`, virtual environment folders (`.venv`), cache files (`__pycache__/`, `.pytest_cache/`, `.ruff_cache/`), and model checkpoints (`*.pth`).
4. Use `uv init` (or manually create) `pyproject.toml`. Specify Python 3.10. Add the following dependencies using `uv add`: `torch>=2.0`, `torchvision>=0.16`, `ruff`, `numpy`, `Pillow`, `pytest`.
5. Create `pre-commit-config.yaml`. Configure `ruff` for formatting (`ruff format`) and linting (`ruff check --select I --fix` for import sorting, and `ruff check --fix` for general linting).
6. Create empty `__init__.py` files in `configs/`, `models/`, `utils/`, and `tests/`.
7. Create empty placeholder files: `train.py`, `test.py`, `configs/base_config.py`, `utils/data_utils.py`, `models/detection.py`, `tests/conftest.py`.
8. Create a basic `README.md` with the project title "Torchvision Vibecoding Project" and a brief description based on `project-spec.md`.
9. Install pre-commit hooks (`pre-commit install`).
Prompt 2: Data Acquisition Script
Create a shell script `scripts/download_data.sh` that performs the following:
1. Checks if the target directory `data/PennFudanPed` already exists. If it does, print a message and exit.
2. Creates the `data/` directory if it doesn't exist.
3. Uses `wget` to download the Penn-Fudan dataset zip file (`PennFudanPed.zip`) from the specified URL (`https://www.cis.upenn.edu/~jshi/ped_html/PennFudanPed.zip`) into the `data/` directory.
4. Uses `unzip` to extract the contents of `PennFudanPed.zip` into the `data/` directory.
5. Removes the downloaded `PennFudanPed.zip` file after successful extraction.
6. Prints informative messages during download and extraction.
7. Make the script executable (`chmod +x`).
Ensure the `.gitignore` file correctly ignores the `data/` directory.
Prompt 3: Configuration System
Implement the configuration system using Python dictionaries:
1. In `configs/base_config.py`, define a Python dictionary named `base_config` containing placeholders or default values for common parameters:
* `data_root`: Path to the dataset root (e.g., `'data/PennFudanPed'`)
* `output_dir`: Base directory for outputs (e.g., `'outputs'`)
* `device`: Compute device (e.g., `'cuda'`)
* `num_classes`: Number of classes including background (e.g., `2` for Penn-Fudan)
* `batch_size`: Training batch size (e.g., `2`)
* `num_epochs`: Number of training epochs (e.g., `10`)
* `lr`: Learning rate (e.g., `0.005`)
* `momentum`: Optimizer momentum (e.g., `0.9`)
* `weight_decay`: Optimizer weight decay (e.g., `0.0005`)
* `lr_step_size`: Learning rate scheduler step size (e.g., `3`)
* `lr_gamma`: Learning rate scheduler gamma (e.g., `0.1`)
* `seed`: Random seed (e.g., `42`)
* `log_freq`: Logging frequency during training (e.g., `10`)
* `checkpoint_freq`: Checkpoint saving frequency (e.g., `1`)
2. In `configs/pennfudan_maskrcnn_config.py`, create a dictionary named `config`.
* Import the `base_config` from `configs.base_config`.
* Create the `config` dictionary, potentially starting with a copy of `base_config` (`config = base_config.copy()`).
* Update specific values as needed for this experiment (e.g., ensure `data_root`, `num_classes` are correct for Penn-Fudan). Add a `config_name` key, e.g., `'pennfudan_maskrcnn_v1'`. This name will be used for naming output folders.
Prompt 4: Core Data Loading (Torch Dataset)
Implement the core dataset loading logic in `utils/data_utils.py`:
1. Import necessary libraries: `os`, `torch`, `PIL.Image`, `numpy`, `torch.utils.data`.
2. Create a class `PennFudanDataset(torch.utils.data.Dataset)`:
* In `__init__(self, root, transforms)`:
* Store `root` and `transforms`.
* Load all image file paths from `root/PNGImages` and sort them.
* Load all mask file paths from `root/PedMasks` and sort them. Ensure alignment between images and masks.
* In `__getitem__(self, idx)`:
* Load the image using `PIL.Image.open` from the path at index `idx`.
* Load the corresponding mask using `PIL.Image.open`.
* Convert the PIL mask image to a numpy array.
* Identify unique object instances in the mask (0 is background). Each unique non-zero value corresponds to a distinct pedestrian instance.
* Generate binary masks for each instance.
* From the binary masks, calculate bounding boxes (`[xmin, ymin, xmax, ymax]`) for each instance. Exclude instances with zero area.
* Create a `target` dictionary containing:
* `boxes`: A `torch.FloatTensor` of shape `(N, 4)` where N is the number of instances.
* `labels`: A `torch.Int64Tensor` of shape `(N,)` where all labels are `1` (for pedestrian).
* `masks`: A `torch.UInt8Tensor` of shape `(N, H, W)`.
* `image_id`: A `torch.Int64Tensor` containing `idx`.
* `area`: A `torch.FloatTensor` containing the area of each bounding box.
* `iscrowd`: A `torch.UInt8Tensor` of shape `(N,)` where all values are `0`.
* Apply the `transforms` to the image and target if `transforms` is not None.
* Return the transformed image and target.
* In `__len__(self)`:
* Return the total number of images.
Prompt 5: Data Utilities (Transforms and Collate)
Add utility functions to `utils/data_utils.py`:
1. Import `torchvision.transforms.v2 as T`.
2. Create a function `get_transform(train)`:
* Initialize a list of transforms.
* Always include `T.ToImage()` and `T.ToDtype(torch.float32, scale=True)`.
* If `train` is `True`, add data augmentation transforms like `T.RandomHorizontalFlip(p=0.5)`. (Keep it simple for now, just horizontal flip).
* Return `T.Compose(transforms)`.
3. Create a function `collate_fn(batch)`:
* This function takes a list of tuples (image, target) and batches them correctly.
* It should return a tuple of `(list(images), list(targets))`. Use `tuple(zip(*batch))` for this.
Prompt 6: Model Definition
Implement the model loading function in `models/detection.py`:
1. Import `torch`, `torchvision`, `torchvision.models.detection`, `torchvision.models.detection.faster_rcnn`, `torchvision.models.detection.mask_rcnn`.
2. Create a function `get_maskrcnn_model(num_classes, pretrained=True, pretrained_backbone=True)`:
* Load a pre-trained Mask R-CNN model. Use `torchvision.models.detection.maskrcnn_resnet50_fpn_v2` with `weights=MaskRCNN_ResNet50_FPN_V2_Weights.DEFAULT` if `pretrained` is True, otherwise `weights=None`. Set `weights_backbone=ResNet50_Weights.DEFAULT` if `pretrained_backbone` is true and `pretrained` is false, otherwise `weights_backbone=None`.
* Get the number of input features for the classifier (`in_features_box`).
* Replace the bounding box predictor head (`model.roi_heads.box_predictor`) with a new `FastRCNNPredictor` instance having `in_features_box` and `num_classes`.
* Get the number of input features for the mask classifier (`in_features_mask`).
* Replace the mask predictor head (`model.roi_heads.mask_predictor`) with a new `MaskRCNNPredictor` instance having `in_features_mask`, 256 hidden layers (default), and `num_classes`.
* Return the modified model.
Prompt 7: Basic Training Script Structure (train.py)
Set up the basic structure for `train.py`:
1. Import `torch`, `argparse`, `importlib`, `os`, `random`, `numpy`.
2. Import necessary components from `utils.data_utils` (`PennFudanDataset`, `get_transform`, `collate_fn`) and `models.detection` (`get_maskrcnn_model`).
3. Define a `main()` function.
4. Inside `main()`:
* Use `argparse` to create a parser that accepts one required argument: `--config`, the path to the configuration Python file (e.g., `configs/pennfudan_maskrcnn_config.py`).
* Parse the arguments.
* Load the configuration dictionary dynamically from the specified file path using `importlib`. For example:
```python
spec = importlib.util.spec_from_file_location("config_module", args.config)
config_module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(config_module)
config = config_module.config
```
* Set random seeds for reproducibility using `random.seed`, `np.random.seed`, `torch.manual_seed` based on `config['seed']`. If using CUDA, also set `torch.cuda.manual_seed_all`.
* Determine the device (`torch.device(config['device'] if torch.cuda.is_available() else 'cpu')`). Print the device being used.
* Create the output directory structure based on `config['output_dir']` and `config['config_name']`. E.g., `output_path = os.path.join(config['output_dir'], config['config_name'])`. Create this directory and subdirectories like `checkpoints` if they don't exist (`os.makedirs(..., exist_ok=True)`).
* Instantiate the `PennFudanDataset` for training (`dataset_train`) using `config['data_root']` and `get_transform(train=True)`.
* Instantiate the `DataLoader` for training (`data_loader_train`) using `dataset_train`, `config['batch_size']`, `shuffle=True`, `num_workers=4` (or appropriate number), and `collate_fn=collate_fn`.
* Instantiate the model using `get_maskrcnn_model(num_classes=config['num_classes'])`.
* Move the model to the determined `device`.
5. Add the standard Python entry point guard (`if __name__ == "__main__":`) to call `main()`.
Prompt 8: Minimal Training Step (train.py)
Extend `train.py` within the `main()` function to perform a single training step:
1. After moving the model to the device, instantiate the optimizer. Use `torch.optim.SGD` with parameters from the config (`lr`, `momentum`, `weight_decay`). Pass `model.parameters()` to the optimizer.
2. Set the model to training mode: `model.train()`.
3. Fetch *one* batch from `data_loader_train`. Use `next(iter(data_loader_train))`.
4. Move images and targets to the `device`. Remember images is a list of tensors and targets is a list of dicts. Iterate through them.
```python
images = list(image.to(device) for image in images)
targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
```
5. Perform the forward pass: `loss_dict = model(images, targets)`. Note that in training mode, Mask R-CNN returns a dictionary of losses.
6. Calculate the total loss: `losses = sum(loss for loss in loss_dict.values())`.
7. Perform the backward pass:
* Zero gradients: `optimizer.zero_grad()`.
* Backpropagate: `losses.backward()`.
* Update weights: `optimizer.step()`.
8. Print the `loss_dict` and the total `losses` tensor for this single step.
9. **(Important)** For now, after this single step, you can add `print("Single training step completed.")` and `return` or `sys.exit()` within `main()` to prevent further execution until the full loop is implemented.
Prompt 9: Logging Integration
Integrate file and console logging:
1. Create a new file `utils/log_utils.py`.
2. Import `logging` and `os`.
3. Define a function `setup_logging(log_dir, config_name)`:
* Create the `log_dir` if it doesn't exist.
* Define the log file path (e.g., `os.path.join(log_dir, f"{config_name}_train.log")`).
* Configure the root logger using `logging.basicConfig`:
* Set `level=logging.INFO`.
* Set `format='%(asctime)s [%(levelname)s] %(message)s'`.
* Set `datefmt='%Y-%m-%d %H:%M:%S'`.
* Provide handlers:
* A `logging.FileHandler` writing to the log file path.
* A `logging.StreamHandler` writing to `sys.stdout`.
4. In `train.py`:
* Import `logging` and `setup_logging` from `utils.log_utils`.
* Immediately after creating the `output_path`, call `setup_logging(output_path, config['config_name'])`.
* Replace `print` statements used for informational output (like device used, starting training) with `logging.info()`.
* Log the loaded configuration dictionary at the beginning of `main()`.
* Log the losses calculated in the single training step using `logging.info(f"Step Loss Dict: {loss_dict}")` and `logging.info(f"Step Total Loss: {losses.item()}")`.
Prompt 10: Full Training Loop (train.py)
Implement the full training loop in `train.py`, replacing the single-step logic:
1. Import `time`.
2. **(Remove the `return` or `sys.exit()` added in Prompt 8).**
3. After creating the optimizer, create a learning rate scheduler (optional but good practice). Use `torch.optim.lr_scheduler.StepLR` with parameters from the config (`lr_step_size`, `lr_gamma`).
4. Add outer loop for epochs: `for epoch in range(config['num_epochs']):`
* Log the start of the epoch: `logging.info(f"--- Epoch {epoch+1}/{config['num_epochs']} ---")`.
* Set model to train mode: `model.train()`.
* Initialize variables to track epoch loss or metrics if needed.
* Add inner loop for batches: `for i, (images, targets) in enumerate(data_loader_train):`
* Move data to device.
* Perform forward pass: `loss_dict = model(images, targets)`.
* Calculate total loss: `losses = sum(loss for loss in loss_dict.values())`.
* Perform backward pass (zero grad, backward, step).
* Log batch loss periodically (e.g., every `config['log_freq']` iterations):
```python
if (i + 1) % config['log_freq'] == 0:
loss_str = f"Epoch {epoch+1}, Iter {i+1}/{len(data_loader_train)}, Loss: {losses.item():.4f}"
# Optional: Add individual losses from loss_dict to the log string
logging.info(loss_str)
```
* After the inner loop (end of epoch), step the learning rate scheduler: `lr_scheduler.step()`.
5. Log the total training time after the epoch loop finishes.
Prompt 11: Checkpointing (train.py)
Add model checkpointing capabilities to `train.py`:
1. Define the checkpoints directory: `checkpoint_dir = os.path.join(output_path, 'checkpoints')`. Create it if it doesn't exist.
2. **(Optional but Recommended: Resume Training Logic)** Before the epoch loop:
* Check if any checkpoints exist in `checkpoint_dir`. Find the latest one (e.g., based on epoch number in the filename).
* If a checkpoint is found:
* Log that training is resuming from the checkpoint.
* Load the checkpoint using `torch.load()`.
* Load the `model.state_dict()`.
* Load the `optimizer.state_dict()`.
* Load the starting epoch number (epoch from checkpoint + 1).
* Load the `lr_scheduler.state_dict()`.
* Handle potential device mismatches if loading a checkpoint saved on a different device (use `map_location`).
* If no checkpoint is found, initialize `start_epoch = 0`.
* Modify the epoch loop to start from `start_epoch`: `for epoch in range(start_epoch, config['num_epochs']):`
3. Inside the epoch loop, after the training batches are processed (e.g., at the end of the epoch):
* Check if the current epoch number satisfies the checkpoint frequency (e.g., `(epoch + 1) % config['checkpoint_freq'] == 0` or if it's the last epoch).
* If it does, construct the checkpoint filename (e.g., `checkpoint_epoch_{epoch+1}.pth`).
* Create a dictionary containing the state:
```python
checkpoint = {
'epoch': epoch + 1,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'scheduler_state_dict': lr_scheduler.state_dict(),
'config': config # Optional: save config used for this checkpoint
}
```
* Save the checkpoint dictionary using `torch.save()` to the `checkpoint_dir`.
* Log that a checkpoint has been saved.
Prompt 12: Evaluation Integration (Setup and Basic Function)
Prepare for and implement a basic evaluation function:
1. Add `pycocotools` (or `pycocotools-windows`) to `pyproject.toml` using `uv add` if planning full mAP later, otherwise skip for now. Add `torchvision`'s `references/detection/` utilities if needed (might require separate download/copying, let's avoid this complexity initially if possible).
2. Create `utils/eval_utils.py`. Import `torch`, `logging`, `time`.
3. Define a function `evaluate(model, data_loader, device)`:
* Set model to evaluation mode: `model.eval()`.
* Initialize placeholder storage for results (e.g., list for losses).
* Log the start of evaluation.
* Start a timer.
* Use `with torch.no_grad():` context manager.
* Loop through the `data_loader`:
* Move images and targets to the `device`.
* Perform forward pass: `outputs = model(images, targets)` (in eval mode, it might return predictions OR losses depending on model/targets presence, check Torchvision docs. Let's assume for now targets ARE provided to eval loader and it returns losses similar to train).
* If it returns losses: Calculate `loss = sum(l for l in outputs.values())`. Store `loss.item()`.
* If it returns predictions (list of dicts with 'boxes', 'labels', 'scores', 'masks'): Store these predictions. (This path is more complex for metrics).
* **(Simplification for now):** Calculate the average loss across all evaluation batches.
* Log the average evaluation loss and the time taken.
* Return the average loss (or a dictionary of metrics if implementing more later).
4. In `train.py`:
* Import the `evaluate` function.
* After instantiating `dataset_train` and `data_loader_train`, also create `dataset_val` and `data_loader_val`. Use `get_transform(train=False)` for the validation dataset. You might need to modify `PennFudanDataset` or create a split if the dataset doesn't have predefined splits (Penn-Fudan doesn't, so maybe use a subset for validation - e.g., first/last N samples, requires modifying `PennFudanDataset` or using `torch.utils.data.Subset`). Let's use a simple Subset approach for now:
```python
# In train.py, after creating dataset_train
indices = torch.randperm(len(dataset_train)).tolist()
val_split = int(0.1 * len(dataset_train)) # 10% validation
dataset_train = torch.utils.data.Subset(dataset_train, indices[:-val_split])
dataset_val = torch.utils.data.Subset(PennFudanDataset(config['data_root'], get_transform(train=False)), indices[-val_split:]) # Re-instance dataset for val transforms
data_loader_val = torch.utils.data.DataLoader(
dataset_val, batch_size=config['batch_size'], shuffle=False,
num_workers=4, collate_fn=collate_fn
)
# Adjust train loader if using Subset on original dataset_train instance
data_loader_train = torch.utils.data.DataLoader(
dataset_train, batch_size=config['batch_size'], shuffle=True, # Should shuffle subset indices
num_workers=4, collate_fn=collate_fn
)
```
* Inside the epoch loop, after the training batches and `lr_scheduler.step()`, call `val_metrics = evaluate(model, data_loader_val, device)`.
* Log the returned validation metrics.
* **(Best Checkpoint Logic - Add Later):** Keep track of the best validation metric seen so far. If the current epoch's metric is better, save a special 'best_model.pth' checkpoint, overwriting the previous best. Also, modify the periodic checkpoint saving to potentially include the validation metric in the filename.
Prompt 13: Testing Script (test.py)
Implement the `test.py` script for evaluating a trained model:
1. Copy the argument parsing, config loading, device setup, dataset/dataloader creation (use `get_transform(train=False)`), and model instantiation logic from `train.py` into a `main()` function in `test.py`.
2. Add a required argument `--checkpoint` to `argparse` to specify the path to the `.pth` checkpoint file to load.
3. Create a test dataset/dataloader (`dataset_test`, `data_loader_test`). Penn-Fudan doesn't have a standard test split, so reuse the validation split logic (or a dedicated test split if created earlier) for demonstration. Ensure `shuffle=False`.
4. Load the specified checkpoint using `torch.load(args.checkpoint, map_location=device)`.
5. Load the `model_state_dict` from the checkpoint into the model. Handle potential `module.` prefix if the model was saved using `DataParallel`.
6. Import and call the `evaluate` function from `utils.eval_utils` using the `model`, `data_loader_test`, and `device`.
7. Log or print the evaluation results returned by the `evaluate` function.
8. Ensure logging is set up similarly to `train.py` (maybe log to `test.log`).
Prompt 14: Unit Tests (tests/)
Create basic unit tests using `pytest`:
1. In `tests/conftest.py` (optional): Define fixtures if needed, e.g., a fixture to provide a temporary directory or a minimal config dictionary.
2. Create `tests/test_config.py`:
* Write a test function `test_load_config()` that attempts to load the `configs/pennfudan_maskrcnn_config.py` file and asserts that the loaded object is a dictionary and contains expected keys (e.g., `data_root`, `num_classes`).
3. Create `tests/test_model.py`:
* Import `get_maskrcnn_model` from `models.detection`.
* Write a test function `test_model_creation()`:
* Call `get_maskrcnn_model(num_classes=2, pretrained=False)`.
* Assert that the returned object is an instance of `torchvision.models.detection.MaskRCNN`.
* Check the output features of `model.roi_heads.box_predictor.cls_score` and `model.roi_heads.mask_predictor.mask_fcn_logits` to ensure they match the requested `num_classes`.
4. Create `tests/test_data_utils.py`:
* Import `PennFudanDataset`, `get_transform` from `utils.data_utils`.
* **(Challenge):** Testing dataset loading often requires actual data or a mock structure. For now, write a test function `test_dataset_instantiation()` that:
* Instantiates `PennFudanDataset` pointing to the *actual* downloaded data path (this makes the test dependent on `download_data.sh` being run first). Use `get_transform(train=False)`.
* Asserts that `len(dataset)` returns a positive number (e.g., 170 for Penn-Fudan).
* Gets the first item using `dataset[0]`.
* Asserts that the returned item is a tuple of (Tensor, dict).
* Asserts that the target dictionary contains the required keys (`boxes`, `labels`, `masks`, etc.) and that they have plausible shapes/types (e.g., `target['boxes']` is a FloatTensor, `target['labels']` is an Int64Tensor).
Prompt 15: Pre-commit Integration for Tests
Update `pre-commit-config.yaml` to run `pytest`:
1. Add a new repo section for `pytest`:
```yaml
- repo: local
hooks:
- id: pytest
name: pytest
entry: pytest -v # Add flags as needed, e.g., -x to stop on first failure
language: system
types: [python]
pass_filenames: false # pytest discovers files itself
# Optional: Specify files/directories if needed
# files: ^tests/
```
2. Ensure `pytest` is installed in the environment where pre-commit runs (it should be via `uv`).
3. Run `pre-commit run --all-files` to test the new hook.
Prompt 16: Refinement and Documentation
Perform final refinements and update documentation:
1. **Error Handling:** Review `train.py` and `test.py`. Add `try...except` blocks around critical sections like data loading, model forward/backward passes, and checkpoint loading/saving. Log errors appropriately.
2. **Config Validation:** Add checks at the beginning of `train.py`/`test.py` to validate essential config values (e.g., check if paths exist, types are correct).
3. **Evaluation Metric:** If only average loss was implemented in `evaluate`, attempt to integrate a proper metric like mAP using `torchvision.ops.box_iou` and potentially adapting logic from Torchvision's evaluation scripts or `pycocotools`. Update the `evaluate` function return value and logging. Update the "best model" saving logic in `train.py` to use this metric.
4. **Data Augmentation:** Add more relevant data augmentations to `get_transform(train=True)` in `utils/data_utils.py` (e.g., color jitter, resizing/cropping strategies suitable for object detection). Ensure transforms handle bounding boxes/masks correctly (use `torchvision.transforms.v2` which generally does).
5. **README.md:** Significantly expand `README.md`:
* Include project goals.
* Detailed **Setup** instructions (clone repo, install `uv`, run `uv sync`, run `scripts/download_data.sh`, install pre-commit hooks).
* **Configuration:** Explain the config files.
* **Training:** How to run `train.py` with a config file. Mention output directories and checkpoints. Explain how to resume training.
* **Testing:** How to run `test.py` with a config and checkpoint file.
* **Project Structure:** Briefly describe the purpose of each directory.
* **Dependencies:** List main dependencies.
* **(Optional) Results:** Mention expected performance or show sample outputs.
6. **Code Quality:** Run `ruff format .` and `ruff check . --fix` one last time to ensure code style and quality. Run `pre-commit run --all-files` to ensure all hooks pass.