dynamics.py

from typing import Dict, Any

import jax
import jax.numpy as jnp
import flax.linen as nn

from utils.nn import STTransformer


class DynamicsMaskGIT(nn.Module):
    """MaskGIT dynamics model"""

    model_dim: int
    num_latents: int
    num_blocks: int
    num_heads: int
    dropout: float
    mask_limit: float

    def setup(self):
        self.dynamics = STTransformer(
            self.model_dim,
            self.num_latents,
            self.num_blocks,
            self.num_heads,
            self.dropout,
        )
        self.patch_embed = nn.Embed(self.num_latents, self.model_dim)
        self.mask_token = self.param(
            "mask_token",
            nn.initializers.lecun_uniform(),
            (1, 1, 1, self.model_dim),
        )
        self.action_up = nn.Dense(self.model_dim)

    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:
        # --- Mask videos ---
        vid_embed = self.patch_embed(batch["video_tokens"])
        if training:
            rng1, rng2 = jax.random.split(batch["mask_rng"])
            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)
            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])
            mask = mask.at[:, 0].set(False)
            vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)
        else:
            mask = None

        # --- Predict transition ---
        act_embed = self.action_up(batch["latent_actions"])
        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))
        logits = self.dynamics(vid_embed)
        return dict(token_logits=logits, mask=mask)