README.md

metadata

title: TrainAI
emoji: 👁
colorFrom: pink
colorTo: yellow
sdk: gradio
sdk_version: 5.31.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: You can train any simple models

🧠 Universal CPU AI Trainer ⚙️

Welcome to the Universal CPU AI Trainer! This Hugging Face Space allows you to:

• Define AI Tasks: Choose from Tabular Classification, Tabular Regression, or basic Image Classification.
• Select Model Families: Experiment with classical Scikit-learn models or simpler PyTorch Neural Networks (MLPs, basic CNNs).
• Configure Datasets:
  • Generate synthetic datasets with configurable rows, features, and characteristics.
  • Let the "AI Assistant" (heuristic rules) suggest dataset parameters.
  • Upload your own datasets (CSV, JSON, Parquet).
• Design Neural Networks: For PyTorch MLPs, specify hidden layers and get suggestions for target parameter counts (10k - 1M).
• Train Models on CPU: All training happens on the free CPU tier. Be patient with larger models or datasets!
• Evaluate & Download: Get basic evaluation metrics and download your trained models (PKL, ONNX for Scikit-learn; PT for PyTorch).

⚠️ Important Considerations for CPU Training:

• Performance: This Space runs on a free CPU tier. Training complex models (especially Neural Networks with >100k parameters) or large datasets will be SLOW. An epoch can take minutes to hours.
• Memory Limits: The free tier has limited RAM (~15GB). Very large datasets or models might cause the Space to crash.
• Toy Examples: The "Basic Image Classification" task uses randomly generated pixel data, not real images. It's for demonstrating the CNN pipeline structure on CPU.
• Experimental: This is a tool for learning and experimentation, not for production-grade model training.

How to Use

1. Tab 1: Define Task & Model

   • Select your desired Task Type (e.g., Tabular Classification).
   • Choose a Model Family (Scikit-learn or PyTorch).
   • Select the Specific Model.
   • If using PyTorch NNs:
     • Select a Target Parameter Range (e.g., "Small (10k-50k)").
     • For MLPs, configure Hidden Layers or use the "Suggest MLP Layers" button (after defining a dataset in Tab 2 for better dimension estimates).

2. Tab 2: Configure Dataset

   • Choose to Generate a new dataset or Upload your own.
   • Generation: Specify rows, features, etc., or use the "AI suggest" checkbox.
   • Upload: Provide your CSV, JSON, or Parquet file.
   • Enter the Target Column Name from your dataset.
   • Click "Generate & Preview Dataset" or let the upload complete.

3. Tab 3: Train Model & Get Results

   • Adjust Training Hyperparameters (Epochs, Batch Size, Learning Rate - primarily for NNs).
   • Select the desired Model Output Format.
   • Click "🚀 Train Model".
   • Monitor the Training Log.
   • View Evaluation Metrics, Model Parameters, and (for PyTorch) a Loss Curve.
   • Download your trained model using the Download Trained Model button.

Model Output Formats

• Scikit-learn:
  • .pkl: Python pickle file containing the Scikit-learn pipeline (preprocessor + model).
  • .onnx: Open Neural Network Exchange format. The exported ONNX model includes the preprocessing steps and expects raw input matching the original training data structure.
• PyTorch:
  • .pt: PyTorch file. For MLPs trained on tabular data, this bundles the model's state_dict and the Scikit-learn preprocessor used. For CNNs, it's typically the state_dict.

Want More Power? Clone & Upgrade!

If training is too slow or you hit resource limits:

1. Go to this Space's main page.
2. Click the three dots (⋮) menu and select "Duplicate this Space."
3. On the creation page, choose upgraded Space Hardware (e.g., better CPU or a GPU - these are paid options).
4. Create your new, more powerful Space! (You'll likely need to re-upload/re-generate data).

Development & Contributions

This Space is built with Python, Gradio, Scikit-learn, and PyTorch.

• Main Application Logic: app.py
• Dependencies: requirements.txt

Feel free to explore the code, suggest improvements, or report issues!

License

This project is licensed under the Apache License 2.0. See the LICENSE file for details.