Update README.md

README.md

@@ -1,13 +1,130 @@
 ---
 license: cc-by-4.0
 ---
-
 # PubChemQCR
 
-### The paper, code, and more details will be released soon!
+<p align="center">
+  <img src="images/dataset_new_v2_png.png" width="450">
+</p>
 
 ### Description
-PubChemQCR dataset contains the DFT relaxation trajectory of ~3.5 million small molecules, which can facilitate the development of ML force field models.
+PubChemQCR dataset contains the DFT relaxation trajectory of ~3.5 million small molecules, which can facilitate the development of ML force field models. The dataset is split into two portions, a subset and a full set. Both sets share the same test set but have unique training and validation sets. The dataloader has a flag that needs to be set to true when the subset is desired.
+
+### Data Loading
+#### Flags
+- 'root' : Path to directory containing LMDB files
+- 'stage' : Which DFT Stage to load
+    - "1st" : Load DFT 1st Data
+    - "1st_smash" : Loads only the DFT 1st Data calculated with SMASH
+    - "2nd" : Load DFT 2nd Data
+    - "mixing" : Load DFT 1st Data & DFT 2nd Data
+    - "pm3" : Load PM3 Data 
+    - "hf" : Load HF Data
+    
+- 'total_traj' : If true the entire trajectory of a molecule is loaded
+- 'SubsetOnly' : If true then only the subset is loaded
+
+#### Dataset Loading
+```python
+from data import LMDBDataLoader, _STD_ENERGY, _STD_FORCE_SCALE
+
+root = '/path/to/lmdb/dir'
+batch_size = 128
+num_workers = 16
+stage = '1st'
+total_traj = True
+SubsetOnly = True
+
+loader = LMDBDataLoader(root=root, batch_size=batch_size, num_workers=num_workers, stage=stage, total_traj=total_traj, SubsetOnly=SubsetOnly)
+
+train_set = loader.train_loader()
+val_set = loader.val_loader()
+test_set = loader.test_loader()
+```
+### Training Procedure
+#### Important
+- Full dataset training requires a few model functionality to work. See example models for indepth usage.
+- Some molecules have atoms not connected if a cutoff is too small, these nodes need to be removed with 'torch_geometric.utils.remove_isolated_nodes'
+    - Example usage:
+        ```python 
+        edge_index, _, mask = remove_isolated_nodes(edge_index, num_nodes=data.num_nodes)
+        pos = data.pos[mask]
+        z = data.x[mask]
+        batch = data.batch[mask]
+        ```
+        - num_nodes flag is needed as without it the function may infer a smaller number of atoms which will cause an error
+- The original batch size needs to be saved for all scatter operations. In rare instances, the whole molecule is removed and passing the original batch size into the scatter function will ensure that molecule gets the value 0. Without it you will get errors.
+    - Example usage:
+        ```python
+            batch_size = data.batch.max().item() + 1
+        ```
+    
+        ```python
+            out = scatter(h, batch, dim=0, dim_size=batch_size, reduce='sum').squeeze()
+        ```
+- Some models require gradient norm clipping in order to prevent loss explosion for some samples. I found gradient clipping to 1.0 was sufficient, but potential clipping values were not thoroughly explored
+    - The log for some epochs may show high losses in the force training phase due to single conformer explosion. 
+    - These high losses are because a single conformer produces a very high loss. Gradient clipping prevents the model form overreacting to these outliers. This only occurs in the training force loss phase, validation should remain normal. 
+    
+    
+### Main Procedure
+```python
+import torch
+import torch.nn as nn
+from torch.optim import Adam
+from models.schnet import SchNet
+from utils import train, evaluate, ForceRMSELoss
+from data import LMDBDataLoader, _STD_ENERGY, _STD_FORCE_SCALE
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+root = '/path/to/lmdb/dir'
+batch_size = 128
+num_workers = 16
+stage = '1st'
+total_traj = True
+SubsetOnly=True
+
+loader = LMDBDataLoader(root=root, batch_size=batch_size, num_workers=num_workers, stage=stage, total_traj=total_traj, SubsetOnly=SubsetOnly)
+
+train_set = loader.train_loader()
+val_set = loader.val_loader()
+test_set = loader.test_loader()
+
+hidden_channels = 128
+num_gaussians = 128
+num_filters = 128
+
+batch_size = 128
+num_interactions = 4
+cutoff = 4.5
+
+model = SchNet(num_gaussians=num_gaussians, num_filters=num_filters, hidden_channels=hidden_channels, num_interactions=num_interactions, cutoff=cutoff)
+model = model.to(device)
+
+max_epochs = 100
+
+params = [param for _, param in model.named_parameters() if param.requires_grad]
+lr = 5e-4
+weight_decay = 0.0
+
+optimizer = Adam([{'params' : params},], lr=lr, weight_decay=weight_decay)
+criterion_energy = nn.L1Loss()
+
+criterion_force = ForceRMSELoss()
+
+for epoch in range(max_epochs):
+     
+    train_energy_loss, train_force_loss = train(model, device, train_set, optimizer, criterion_energy, criterion_force)
+
+    val_energy_loss, val_force_loss = evaluate(model, device, val_set, criterion_energy, criterion_force)
+
+    print(f"#IN#Epoch {epoch + 1}, Train Energy Loss: {train_energy_loss * _STD_ENERGY:.5f}, Val Energy Loss: {val_energy_loss * _STD_ENERGY:.5f}, Train Force Loss: {train_force_loss * _STD_FORCE_SCALE:.5f}, Val Force Loss: {val_force_loss * _STD_FORCE_SCALE:.5f}")
+
+test_energy_loss, test_force_loss = evaluate(model, device, test_set, criterion_energy, criterion_force)
+
+print(f'Test Energy Loss: {test_energy_loss * _STD_ENERGY:.5f}, Test Force Loss: {test_force_loss * _STD_FORCE_SCALE:.5f}')
+```
 
 ### License
 This dataset is a processed version prepared by the TAMU DIVE Lab, based on the raw DFT trajectory data originally created by Maho Nakata from RIKEN.