Skip to content

Difference between model context in paper and implementation #369

Description

@acard5meg

Hello, thank you for open sourcing your work and weights along with the extensive documentation. I am working through real-world implementation and noticed in internvla_n1_agent_realworld.py the following code is used to generate image context around line 190 shown in the first code preview below. This function makes the history always begin at 0 and give num_history spaced indices up to episode_idx-1 resulting in history id lists such as [0, 4,9, 13, 18, 22, 27, 32]. In Figure 2 of your paper, it seems like the context should be the previous num_history images. Could you explain whether we want to give this kind of context or the history id should be defined like the second code option?

history_id = np.unique(np.linspace(0, self.episode_idx - 1, self.num_history, dtype=np.int32)).tolist()

Image
min_indx = max(0, self.episode_idx - self.num_history)

max_indx = self.episode_idx

history_id = [i for i in range(min_indx, max_indx)]

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions