README.md

---
title: Ice Cream Dataset + Counter
emoji: 🍦
colorFrom: yellow
colorTo: red
sdk: gradio
sdk_version: 6.5.1
app_file: app.py
pinned: false
---

# 🍦 Ice Cream Dataset + Counter — Gradio Space

This is the **Gradio + ZeroGPU-compatible** version of the uploaded Ice Cream Counter project. It does **not** use Docker, FastAPI, Uvicorn, or a custom HTML frontend.

## What it does

1. Upload training freezer photos.
2. Define your product classes.
3. Annotate each ice cream with bounding boxes.
4. Train an RT-DETR object detector.
5. Upload one new freezer image.
6. Get the total count, per-product counts, confidence scores, and an annotated result image.

The model is still:

`PekingU/rtdetr_r50vd`

No YOLO is used.

## Create the Hugging Face Space

Create a new Space and choose:

- **SDK:** Gradio
- **Hardware:** ZeroGPU or a dedicated GPU is recommended for training

The app uses `@spaces.GPU` for training and counting, so it also boots correctly when the Space hardware is **ZeroGPU**.

Then upload these files/folders:

```text
app.py
requirements.txt
README.md
training/
```

You do **not** need:

```text
Dockerfile
FastAPI
Uvicorn
static/index.html
app/main.py
```

## Persistent dataset and model

The app stores:

```text
images/
dataset.json
model/
generated_dataset/
```

When `/data` is available and writable, the app automatically uses:

```text
/data/icecream_counter/
```

You can also explicitly set:

```text
DATA_DIR=/data/icecream_counter
```

in the Space variables.

### Important Hugging Face storage point

A normal Space filesystem is not permanent storage across every rebuild/restart. If you need the dataset and trained model to survive Space restarts/rebuilds, attach **persistent storage** to the Space or move the data/model to an external persistent service.

The Gradio conversion itself does not change this storage rule.

## Annotation workflow

In the **Annotate** tab the actual uploaded training image is displayed directly using Gradio's Image component. This avoids browser canvas/JavaScript issues that can make the preview appear black.

1. Select a training image.
2. The real image appears in the preview.
3. Read the image dimensions shown below it.
4. Select the product class.
5. Enter the bounding box in original-image pixel coordinates:
   - X (left)
   - Y (top)
   - Width
   - Height
6. Click **Save Box**.
7. Saved boxes are drawn in red on the real image.
8. Repeat for every ice cream.
9. Use **Delete Box** or **Clear All Boxes** when needed.

This version prioritizes a reliable visible image over the previous JavaScript canvas approach.

## Training

The app creates an 80/20 COCO train/validation split from the annotated images and starts the existing RT-DETR training script.

Default settings:

```text
Epochs: 30
Batch size: 2
Learning rate: 1e-5
```

For an initial test on a small dataset, use fewer epochs such as 2–5. Once everything works, increase the epochs.

A GPU Space is strongly recommended.


## Training fix

This release includes a CUDA-device fix for RT-DETR's contrastive-denoising training path. On some
Transformers/PyTorch combinations, the denoising class-index tensor can remain on CPU while the
RT-DETR class embedding is on CUDA, producing:

`RuntimeError: Expected all tensors to be on the same device ... cpu ... cuda:0`

The training script now moves nested target tensors explicitly and patches the RT-DETR denoising
helper so its target tensors follow the class-embedding device. The `num_labels=10` vs. checkpoint
`80` message is expected when fine-tuning the COCO-pretrained checkpoint for 10 custom classes;
`ignore_mismatched_sizes=True` intentionally reinitializes the classification heads.

## Counting

After training, open the **Count** tab and upload one image.

The result contains:

```json
{
  "total": 31,
  "counts": {
    "cone": 6,
    "correto": 7,
    "cornetto": 12,
    "magnum": 6
  }
}
```

The result image also shows the detected bounding boxes and confidence values.

## Default classes

```text
cornetto
magnum
correto
cone
cup
sandwich
stick
other
```

You can change them from the Dataset tab.

## Local test

```bash
pip install -r requirements.txt
python app.py
```

Then open:

```text
http://localhost:7860
```

## Recommended Space setup

For the first deployment:

1. Create the Space as **Gradio**.
2. Upload `app.py`, `requirements.txt`, `README.md`, and `training/`.
3. Wait for dependencies to install.
4. Open the Dataset tab.
5. Upload 2+ training images.
6. Annotate them.
7. Start with 2–5 epochs to verify training.
8. After the model finishes, test the Count tab.
9. For serious training, attach a GPU and persistent storage.



### RT-DETR training compatibility
The trainer disables RT-DETR contrastive denoising by default because some Transformers releases can create CPU class-index tensors while the embedding is on CUDA. The detector's normal supervised detection loss remains enabled. A device-safe denoising patch is also included for future re-enablement.


## Fixed22 training patch

This version fixes the RT-DETR CPU/CUDA denoising crash by wrapping the denoising class embedding as a real `torch.nn.Module` and moving its index tensor to the embedding weight device before `nn.Embedding` is called. A plain Python function wrapper is intentionally not used because Transformers expects the embedding to remain a module.