The Role of Model Architecture and Scale in Predicting Molecular Properties: Insights from Fine-Tuning RoBERTa, BART, and LLaMA

This repository is the official implementation of [Modeling of Time-varying Wireless Communication Channel with Fading and Shadowing](Not yet uploaded).

Requirements

To install requirements:

pip install -r requirements_cuda118.txt

📋 The experiments were done under CUDA 11.8

Dataset

1. Data : data.py
2. Data Hyper-paremeters : data_hp.py

📋 You don't have to run those files

Training

1. Train Nakagami 1 from Random Init : python run_n.py
2. Train Nakagami 1 with N_G_var(10x-16) from Random Init : python run_ne2.py
3. Train Log-Normal 1 from Random Init : python run_ln.py Then you have to open and work through './DataAnalysis/0_Nakagami1_Eval.ipynb' and './DataAnalysis/1_LogNormal1_Eval.ipynb' to select the Global Best (and Median) Nakagami 1 (Log-Normal 1).

The git repository already have Nakagami_XXX.h5 and LogNormal_XXX.h5, but those are from the research paper.

You need to replaced those with yours if you want to try with your owns. 4. Train Nakagami 2 from Random Init : python run_n2.py 5. Train Nakagami 2 from Global Best (and Median) Nakagami 1 : python run_n1_n2.py. Please check the file before run. 6. Train Log-Normal 2 from Random Init : python run_ln2.py 7. Train Log-Normal 2 from Global Best (and Median) Nakagami 2 : python run_ln1_ln2.py. Please check the file before run.

📋 You can control the DMDN model's and its training hyper-parameters with model.py and model_hp.py

Evaluation

Move to DataAnalysis

1. Detail information and the visualizations of the data : Data Visualization.ipynb
2. Evaluate and select the Global Best (and Median) Nakagami 1 : 0_Nakagami1_Eval.ipynb
3. Evaluate and select the Global Best (and Median) Log-Normal 1 : 1_LogNormal1_Eval.ipynb
4. Evaluate Nakagami 2 cases : 2_Nakagami1_Nakagami2_Eval.ipynb
5. Evaluate Log-Normal 2 cases : 3_LogNormal1_LogNormal2_Eval.ipynb

Contributing

📋 MIT

Authors' Note

Please use this code only for social goods and positive impact.