File size: 2,107 Bytes
fc19214 7b3fbf5 dc5e11f fc19214 7b3fbf5 fc19214 7b3fbf5 fc19214 7b3fbf5 fc19214 7b3fbf5 fc19214 dc5e11f 7b3fbf5 fc19214 7b3fbf5 fc19214 7b3fbf5 fc19214 7b3fbf5 fc19214 d730a0a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 |
---
library_name: transformers
license: creativeml-openrail-m
base_model:
- facebook/detr-resnet-50-panoptic
datasets:
- FriedParrot/a-large-scale-fish-dataset
language:
- en
---
# Model Card for Fish Segmentation (Fine-Tuned DETR)
This is a **fine-tuned DETR model (`facebook/detr-resnet-50-panoptic`)** adapted for **fish detection and segmentation**.
The model performs **multi-task prediction** including:
* **Classification** (fish species recognition)
* **Bounding Box prediction**
* **Segmentation masks**
It has **42.9M parameters** and is trained on the **[A Large Scale Fish Dataset](https://www.kaggle.com/datasets/crowww/a-large-scale-fish-dataset)** from Kaggle.
The copy of this dataset on hugging face is available [here](https://huggingface.co/datasets/FriedParrot/a-large-scale-fish-dataset)
## Model Sources
* **Base model**: [facebook/detr-resnet-50-panoptic](https://huggingface.co/facebook/detr-resnet-50-panoptic)
* **Fine-tuned model**: [FriedParrot/fish-segmentation-simple](https://huggingface.co/FriedParrot/fish-segmentation-simple)
* **Training dataset**: [A Large Scale Fish Dataset](https://www.kaggle.com/datasets/crowww/a-large-scale-fish-dataset)
* **Source code & tutorials**: [GitHub Repository](https://github.com/FRIEDparrot/fish-segmentation)
> [!note]
> This model is fully compatible with `AutoModelForObjectDetection`, `AutoProcessor`, and Hugging Face Trainer.
> Unlike the first model (`fish-segmentation-model`), this one does **not** require custom config classes.
## Training Details
* **Hardware**: NVIDIA RTX 4090 (48GB VRAM)
* **CUDA**: 12.8
* **Framework**: PyTorch + Hugging Face Transformers
* **Batch size**: use 8 as train batch sizes
* **Training strategy**: Direct fine-tuning of DETR with minimal modifications
## Results & Example Predictions
Since its a fine-tuned model, the accuracy is really high, and also classification accuracy can reach about 100%.
The predicted bounding box and masks are also very accurate :
|