Upload README.md

README.md

@@ -1,26 +1,60 @@
----
-tags:
-- ml-intern
----
+# 籽粒分类模型（大米品种分类） / Grain Seed Classification Model
 
-# chaosbee997/rice-seed-classifier
+This repository contains the training script and (eventually) the fine-tuned model for classifying **rice grain varieties** from seed images.
 
-<!-- ml-intern-provenance -->
-## Generated by ML Intern
+## Dataset
 
-This model repository was generated by [ML Intern](https://github.com/huggingface/ml-intern), an agent for machine learning research and development on the Hugging Face Hub.
+- **Source:** [`nateraw/rice-image-dataset`](https://huggingface.co/datasets/nateraw/rice-image-dataset)
+- **Size:** 75,000 RGB images (250×250)
+- **Classes (5):**
+  1. Arborio
+  2. Basmati
+  3. Ipsala
+  4. Jasmine
+  5. Karacadag
+- **License:** CC0-1.0
 
-- Try ML Intern: https://smolagents-ml-intern.hf.space
-- Source code: https://github.com/huggingface/ml-intern
+## Model
 
-## Usage
+- **Architecture:** ResNet-18 (`microsoft/resnet-18`) — ~11M parameters, lightweight and fast
+- **Task:** Multi-class image classification
 
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
+## How to train
 
-model_id = "chaosbee997/rice-seed-classifier"
-tokenizer = AutoTokenizer.from_pretrained(model_id)
-model = AutoModelForCausalLM.from_pretrained(model_id)
+Run the provided script on a GPU (e.g. a10g-large or t4-small via Hugging Face Jobs, or Google Colab):
+
+```bash
+pip install transformers datasets torch accelerate evaluate pillow trackio
+
+export HF_MODEL_REPO=chaosbee997/rice-seed-classifier
+export HF_TOKEN=your_huggingface_token
+
+python train.py
+```
+
+Or submit via Hugging Face Jobs (requires GPU credits):
+
+```bash
+huggingface-cli job run \
+  --script train.py \
+  --hardware a10g-large \
+  --timeout 4h \
+  --dependencies "transformers,datasets,torch,accelerate,evaluate,pillow,trackio"
 ```
 
-For non-causal architectures, replace `AutoModelForCausalLM` with the appropriate `AutoModel` class.
+## Expected results
+
+- Typical fine-tuning on this dataset with ResNet-18 yields **> 95% accuracy** within 3-5 epochs.
+
+## Extending to other crops
+
+The same script works for any `datasets.ImageFolder`-style dataset. To add peanut, corn, wheat, etc.:
+
+1. Collect or find an image dataset with folder-per-class structure.
+2. Upload it to Hugging Face Hub or point `load_dataset` to a local path.
+3. Update `MODEL_NAME` if you want a different backbone (e.g. `microsoft/resnet-34`, `google/mobilenet_v2_1.0_224`).
+4. Run `train.py`.
+
+## License
+
+Apache-2.0