Update app.py

app.py

@@ -64,6 +64,7 @@ def load_model(model_choice):
     model = GraphDiT(
         model_config_path=model_config_path,
         data_info_path=data_info_path,
+        # model_dtype=torch.float16,
         model_dtype=torch.float32,
     )
     ### test
@@ -238,7 +239,6 @@ with gr.Blocks(title="Polymer Design with GraphDiT") as iface:
     # Main Description
     gr.Markdown("""
     ## Introduction
-
     Input the desired gas barrier properties for CH₄, CO₂, H₂, N₂, and O₂ to generate novel polymer structures. The results are visualized as molecular graphs and represented by SMILES strings if they are successfully generated. Note: Gas barrier values set to 0 will be treated as `NaN` (unconditionally). If the generation fails, please retry or increase the number of repetition attempts.
     """)
 
@@ -254,15 +254,10 @@ with gr.Blocks(title="Polymer Design with GraphDiT") as iface:
     with gr.Accordion("🔍 Model Description", open=False):
         gr.Markdown("""
         ### GraphDiT: Graph Diffusion Transformer
-
         GraphDiT is a graph diffusion model designed for targeted molecular generation. It employs a conditional diffusion process to iteratively refine molecular structures based on user-specified properties.
-
         We have collected a labeled polymer database for gas permeability from [Membrane Database](https://research.csiro.au/virtualscreening/membrane-database-polymer-gas-separation-membranes/). Additionally, we utilize unlabeled polymer structures from [PolyInfo](https://polymer.nims.go.jp/).
-
         The gas permeability ranges from 0 to over ten thousand, with only hundreds of labeled data points, making this task particularly challenging.
-
         We are actively working on improving the model. We welcome any feedback regarding model usage or suggestions for improvement.
-
         #### Currently, we have two variants of Graph DiT:
         - **Graph DiT (trained on labeled + unlabeled)**: This model uses both labeled and unlabeled data for training, potentially leading to more diverse/novel polymer generation.
         - **Graph DiT (trained on labeled)**: This model is trained exclusively on labeled data, which may result in higher validity but potentially less diverse/novel outputs.