import torch
import torch.nn as nn
import math

pe_scaling = 0.1 # Hyperparameter, 0.1 value was used from training
    
class PositionalEncoding(nn.Module):
    def __init__(self, d_model, max_len):
        super().__init__()
        self.register_buffer("pe", self._generate_pe(max_len, d_model))

    def _generate_pe(self, max_len, d_model):
        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len, dtype=torch.float32).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = (pe + 1) / 2 
        pe = pe_scaling * pe
        return pe.unsqueeze(0)

    def forward(self, x):
        x = x + self.pe[:, :x.size(1), :].to(x.device)
        return x