Method forward

gridencoder/grid.py:207–224

(self, inputs, embeddings, bound=1)

        return f"GridEncoder: input_dim={self.input_dim} num_levels={self.num_levels} level_dim={self.level_dim} resolution={self.base_resolution} -> {int(round(self.base_resolution * self.per_level_scale ** (self.num_levels - 1)))} per_level_scale={self.per_level_scale:.4f} params={tuple(self.embeddings.shape)} gridtype={self.gridtype} align_corners={self.align_corners}"
    
    def forward(self, inputs, embeddings, bound=1):
        # inputs: [..., input_dim], normalized real world positions in [-bound, bound]
        # return: [..., num_levels * level_dim]
        input_embeddings = torch.cat([embeddings, self.embeddings], dim=0)

        inputs = (inputs + bound) / (2 * bound) # map to [0, 1]
        
        #print('inputs', inputs.shape, inputs.dtype, inputs.min().item(), inputs.max().item())

        prefix_shape = list(inputs.shape[:-1])
        inputs = inputs.view(-1, self.input_dim)

        outputs = grid_encode(inputs, input_embeddings, self.offsets, self.per_level_scale, self.base_resolution, inputs.requires_grad, self.gridtype_id, self.align_corners)
        outputs = outputs.view(prefix_shape + [self.output_dim])

        #print('outputs', outputs.shape, outputs.dtype, outputs.min().item(), outputs.max().item())

        return outputs