I am trying to evaluate different Integrated Gradients methods on my RoBERTa based model, and I came to a paper introducing “Sequential Integrated Gradients” with this github repo:
text

I understand to get the IGs we need to pass input_embeddings not ids to the model so it’s able to compute gradients but in the given code the author is producing the embeddings like:

getattr(model, model_name).embeddings.word_embeddings(input_ids)

and later generate positional and type embeddings (also attention mask) manually and finally summing them and applying the LayerNorm and dropout layers to pass to the model using this wrapper:

class ForwardModel(nn.Module):
    def __init__(self, model, model_name):
        super().__init__()
        self.model = model
        self.model_name = model_name

    def forward(
        self,
        input_embed,
        attention_mask=None,
        position_embed=None,
        type_embed=None,
        return_all_logits=False,
    ):
        embeds = input_embed + position_embed
        if type_embed is not None:
            embeds += type_embed

        # Get predictions
        embeds = getattr(self.model, self.model_name).embeddings.dropout(
            getattr(self.model, self.model_name).embeddings.LayerNorm(embeds)
        )
        pred = self.model(
            inputs_embeds=embeds,
            attention_mask=attention_mask,
        )[0]

        # Return all logits or just maximum class
        if return_all_logits:
            return pred
        else:
            return pred.max(1).values

however when I pass the model just “embeds” itself using argument “inputs_embeds” I get the same output as passing “input_ids” using “input_ids” argument, it means the model handles the position and type embeddings itself. So adding them manually to input_embeddings brings a different output. It seems they did it maybe because the model bypasses those embeddings if the input_embeddings instead of ids are passed. I appreciate if someone clarifies.