Code/utils.py

import torch
import torch.nn as nn
from data import _STD_ENERGY, _STD_FORCE_SCALE
from torch_scatter import scatter

from tqdm import tqdm

class ForceRMSELoss(nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, pred, target, batch):
        return scatter((pred - target).pow(2).sum(dim=-1), batch, reduce="mean", dim=0, dim_size=batch.max().item() + 1).sqrt().mean()

def train(model, device, train_loader, optimizer, criterion_energy, criterion_force, energy_weight=1.0, force_weight=1.0, clip_gradients=False, grad_clip_norm=1.0):
    model.train()

    total_energy_loss = 0.
    total_force_loss = 0.

    progress_bar = tqdm(train_loader, desc='Training', bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}')

    for batch in progress_bar:
        optimizer.zero_grad()
        data = batch.to(device, non_blocking=True)

        energies, forces, mask = model(data)

        energy_loss = criterion_energy(energies, data.y)
        
        force_loss = criterion_force(forces, data.y_force[mask], data.batch[mask])

        loss = energy_weight * energy_loss + force_weight * force_loss
        
        total_energy_loss += energy_loss.item()
        total_force_loss += force_loss.item()
        
        loss.backward()
        
        if clip_gradients:
            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=grad_clip_norm)
        
        optimizer.step()

        progress_bar.set_description(
                f"Training - Energy Loss: {energy_loss * _STD_ENERGY:.5f}, "
                f"Force Loss: {force_loss * _STD_FORCE_SCALE:.5f}")
        
    average_energy_loss = total_energy_loss / len(train_loader)
    average_force_loss = total_force_loss / len(train_loader)
    return average_energy_loss, average_force_loss

def evaluate(model, device, loader, criterion_energy, criterion_force):
    model.eval()

    total_energy_loss = 0.
    total_force_loss = 0.

    progress_bar = tqdm(loader, desc='Evaluating', bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}')

    for batch in progress_bar:
        data = batch.to(device, non_blocking=True)

        energies, forces, mask = model(data)

        energy_loss = criterion_energy(energies, data.y)
        force_loss = criterion_force(forces, data.y_force[mask], data.batch[mask])

        total_energy_loss += energy_loss.item()
        total_force_loss += force_loss.item()

        progress_bar.set_description(
                f"Evaluation - Energy Loss: {energy_loss * _STD_ENERGY:.5f}, Force Loss: {force_loss * _STD_FORCE_SCALE:.5f}")
        
    average_energy_loss = total_energy_loss / len(loader)
    average_force_loss = total_force_loss / len(loader)
    return average_energy_loss, average_force_loss