allow for image pretraining on video tokenizer

dreamer4/__init__.py

@@ -1,6 +1,7 @@
 from dreamer4.dreamer4 import (
     VideoTokenizer,
-    DynamicsWorldModel
+    DynamicsWorldModel,
+    AxialSpaceTimeTransformer
 )
 
 

dreamer4/dreamer4.py

@@ -1,4 +1,5 @@
 from __future__ import annotations
+from typing import Callable
 
 import math
 from math import ceil, log2
@@ -70,13 +71,19 @@ except ImportError:
 
 LinearNoBias = partial(Linear, bias = False)
 
+VideoTokenizerIntermediates = namedtuple('VideoTokenizerIntermediates', ('losses', 'recon'))
+
 TokenizerLosses = namedtuple('TokenizerLosses', ('recon', 'lpips', 'time_decorr', 'space_decorr'))
 
-WorldModelLosses = namedtuple('WorldModelLosses', ('flow', 'rewards', 'discrete_actions', 'continuous_actions'))
+WorldModelLosses = namedtuple('WorldModelLosses', ('flow', 'rewards', 'discrete_actions', 'continuous_actions', 'state_pred'))
 
 AttentionIntermediates = namedtuple('AttentionIntermediates', ('next_kv_cache', 'normed_inputs'))
 
-TransformerIntermediates = namedtuple('TransformerIntermediates', ('next_kv_cache', 'normed_time_inputs', 'normed_space_inputs'))
+TransformerIntermediates = namedtuple('TransformerIntermediates', ('next_kv_cache', 'normed_time_inputs', 'normed_space_inputs', 'next_rnn_hiddens'))
+
+Predictions = namedtuple('Predictions', ('flow', 'proprioception', 'state'))
+
+Embeds = namedtuple('Embeds', ['agent', 'state_pred'])
 
 MaybeTensor = Tensor | None
 
@@ -95,7 +102,7 @@ class Experience:
     lens: MaybeTensor = None
     is_truncated: MaybeTensor = None
     agent_index: int = 0
-    is_from_world_model: bool = True
+    is_from_world_model: bool | Tensor = True
 
     def cpu(self):
         return self.to(torch.device('cpu'))
@@ -123,6 +130,9 @@ def combine_experiences(
         if not exists(exp.is_truncated):
             exp.is_truncated = full((batch,), True, device = device)
 
+        if isinstance(exp.is_from_world_model, bool):
+            exp.is_from_world_model = tensor(exp.is_from_world_model)
+
     # convert to dictionary
 
     exps_dict = [asdict(exp) for exp in exps]
@@ -144,11 +154,15 @@ def combine_experiences(
 
     for field_values in all_field_values:
 
-        if is_tensor(first(field_values)):
+        first_value = first(field_values)
+
+        if is_tensor(first_value):
 
             field_values = pad_tensors_at_dim_to_max_len(field_values, dims = (1, 2))
 
-            new_field_value = cat(field_values)
+            cat_or_stack = cat if first_value.ndim > 0 else stack
+
+            new_field_value = cat_or_stack(field_values)
         else:
             new_field_value = first(list(set(field_values)))
 
@@ -189,6 +203,13 @@ def sample_prob(prob):
 def is_power_two(num):
     return log2(num).is_integer()
 
+def maybe(fn):
+    def inner(t, *args, **kwargs):
+        if not exists(t) or not exists(fn):
+            return None
+        return fn(t)
+    return inner
+
 # tensor helpers
 
 def is_empty(t):
@@ -1284,7 +1305,8 @@ class Attention(Module):
         pre_rmsnorm = True,
         gate_values = True,
         rmsnorm_query = False, # a paper claims that it is better to just norm only the keys https://openreview.net/forum?id=HkztQWZfl2
-        rmsnorm_key = True
+        rmsnorm_key = True,
+        value_residual = True
     ):
         super().__init__()
         self.norm = RMSNorm(dim) if pre_rmsnorm else Identity()
@@ -1323,6 +1345,14 @@ class Attention(Module):
         self.q_heads_rmsnorm = MultiHeadRMSNorm(dim_head, heads = query_heads) if rmsnorm_query else nn.Identity()
         self.k_heads_rmsnorm = MultiHeadRMSNorm(dim_head, heads = heads) if rmsnorm_key else nn.Identity()
 
+        # value residual
+
+        self.to_learned_value_residual_mix = nn.Sequential(
+            nn.Linear(dim, heads),
+            Rearrange('b n h -> b h n 1'),
+            nn.Sigmoid()
+        ) if value_residual else None
+
     def muon_parameters(self):
         # omit the queries and keys for now given what we learned from kimi 2 paper
 
@@ -1337,6 +1367,7 @@ class Attention(Module):
         kv_cache = None,
         return_intermediates = False,
         rotary_pos_emb = None,
+        residual_values = None,  # (b n h d)
         attend_fn: Callable | None = None
     ):
         tokens, inverse_packed_batch = pack_one(tokens, '* n d')
@@ -1349,6 +1380,17 @@ class Attention(Module):
 
         q, k, v = map(self.split_heads, (q, k, v))
 
+        # handle maybe value residual
+
+        if exists(residual_values):
+            residual_values = rearrange(residual_values, '... n h d -> (...) h n d')
+
+            assert exists(self.to_learned_value_residual_mix)
+
+            learned_mix = self.to_learned_value_residual_mix(tokens)
+
+            v = v.lerp(residual_values, learned_mix)
+
         # qk rmsnorm
 
         q = self.q_heads_rmsnorm(q)
@@ -1425,6 +1467,29 @@ class SwiGLUFeedforward(Module):
 
         return self.proj_out(x)
 
+# rnn
+
+class GRULayer(Module):
+    def __init__(
+        self,
+        dim,
+        dim_out
+    ):
+        super().__init__()
+        self.norm = nn.RMSNorm(dim)
+        self.gru = nn.GRU(dim, dim_out, batch_first = True)
+
+    def forward(
+        self,
+        x,
+        prev_hiddens = None
+    ):
+        x = self.norm(x)
+
+        x, hiddens = self.gru(x, prev_hiddens)
+
+        return x, hiddens
+
 # axial space time transformer
 
 class AxialSpaceTimeTransformer(Module):
@@ -1432,6 +1497,7 @@ class AxialSpaceTimeTransformer(Module):
         self,
         dim,
         depth,
+        attn_heads = 8,
         attn_dim_head = 64,
         attn_softclamp_value = 50.,
         time_block_every = 4,
@@ -1440,7 +1506,9 @@ class AxialSpaceTimeTransformer(Module):
         num_residual_streams = 1,
         num_special_spatial_tokens = 1,
         special_attend_only_itself = False,  # this is set to True for the video tokenizer decoder (latents can only attend to itself while spatial modalities attend to the latents and everything)
-        final_norm = True
+        final_norm = True,
+        value_residual = True,               # https://arxiv.org/abs/2410.17897 - but with learned mixing from OSS
+        rnn_time = True
     ):
         super().__init__()
         assert depth >= time_block_every, f'depth must be at least {time_block_every}'
@@ -1461,6 +1529,24 @@ class AxialSpaceTimeTransformer(Module):
 
         self.time_rotary = Rotary1D(attn_dim_head)
 
+        # project initial for value residuals
+
+        self.value_residual = value_residual
+
+        if value_residual:
+            dim_inner = attn_dim_head * attn_heads
+
+            self.to_value_residual = nn.Sequential(
+                nn.RMSNorm(dim),
+                nn.Linear(dim, dim_inner, bias = False),
+                Rearrange('... (h d) -> ... h d', h = attn_heads)
+            )
+
+        # a gru layer across time
+
+        self.rnn_time = rnn_time
+        rnn_layers = []
+
         # transformer
 
         layers = []
@@ -1472,17 +1558,21 @@ class AxialSpaceTimeTransformer(Module):
             is_time_block = divisible_by(layer_index, time_block_every)
             is_time.append(is_time_block)
 
-            rearrange_to_attend = Rearrange('b t s d -> b s t d') if is_time_block else Identity()
+            rearrange_to_attend = Rearrange('b t s ... -> b s t ...') if is_time_block else Identity()
-            rearrange_from_attend = Rearrange('b s t d -> b t s d') if is_time_block else Identity()
+            rearrange_from_attend = Rearrange('b s t ... -> b t s ...') if is_time_block else Identity()
 
             layers.append(ModuleList([
                 rearrange_to_attend,
                 rearrange_from_attend,
-                hyper_conn(branch = Attention(dim = dim, dim_head = attn_dim_head, **attn_kwargs)),
+                hyper_conn(branch = Attention(dim = dim, heads = attn_heads, dim_head = attn_dim_head, value_residual = value_residual, **attn_kwargs)),
                 hyper_conn(branch = SwiGLUFeedforward(dim = dim, **ff_kwargs))
             ]))
 
+            rnn_layers.append(hyper_conn(branch = GRULayer(dim, dim)) if is_time_block and rnn_time else None)
+
         self.layers = ModuleList(layers)
+        self.rnn_layers = ModuleList(rnn_layers)
+
         self.is_time = is_time
 
         # final norm
@@ -1505,7 +1595,7 @@ class AxialSpaceTimeTransformer(Module):
     def forward(
         self,
         tokens, # (b t s d)
-        kv_cache: Tensor | None = None,  # (y 2 b h t d)
+        cache: TransformerIntermediates | None = None,
         return_intermediates = False
 
     ): # (b t s d) | (y 2 b h t d)
@@ -1514,6 +1604,14 @@ class AxialSpaceTimeTransformer(Module):
 
         assert tokens.ndim == 4
 
+        # destruct intermediates to cache for attention and rnn respectively
+
+        kv_cache = rnn_prev_hiddens = None
+
+        if exists(cache):
+            kv_cache = cache.next_kv_cache
+            rnn_prev_hiddens = cache.next_rnn_hiddens
+
         # attend functions for space and time
 
         has_kv_cache = exists(kv_cache) 
@@ -1528,7 +1626,7 @@ class AxialSpaceTimeTransformer(Module):
         # prepare cache
 
         time_attn_kv_caches = []
-
+        rnn_hiddens = []
 
         if has_kv_cache:
             past_tokens, tokens = tokens[:, :-1], tokens[:, -1:]
@@ -1543,10 +1641,21 @@ class AxialSpaceTimeTransformer(Module):
 
         iter_kv_cache = iter(kv_cache)
 
+        rnn_prev_hiddens = default(rnn_prev_hiddens, (None,))
+
+        iter_rnn_prev_hiddens = iter(rnn_prev_hiddens)
+
         # rotary
 
         rotary_pos_emb = self.time_rotary(rotary_seq_len, offset = rotary_pos_offset)
 
+        # value residual
+
+        residual_values = None
+
+        if self.value_residual:
+            residual_values = self.to_value_residual(tokens)
+
         # normed attention inputs
 
         normed_time_attn_inputs = []
@@ -1556,10 +1665,22 @@ class AxialSpaceTimeTransformer(Module):
 
         tokens = self.expand_streams(tokens)
 
-        for (pre_attn_rearrange, post_attn_rearrange, attn, ff), layer_is_time in zip(self.layers, self.is_time):
+        for (pre_attn_rearrange, post_attn_rearrange, attn, ff), maybe_rnn, layer_is_time in zip(self.layers, self.rnn_layers, self.is_time):
 
             tokens = pre_attn_rearrange(tokens)
 
+            # maybe rnn for time
+
+            if layer_is_time and exists(maybe_rnn):
+
+                tokens, inverse_pack_batch = pack_one(tokens, '* t d')
+
+                tokens, layer_rnn_hiddens = maybe_rnn(tokens, next(iter_rnn_prev_hiddens, None)) # todo, handle rnn cache
+
+                tokens = inverse_pack_batch(tokens)
+
+                rnn_hiddens.append(layer_rnn_hiddens)
+
             # when is a axial time attention block, should be causal
 
             attend_fn = time_attend if layer_is_time else space_attend
@@ -1570,6 +1691,10 @@ class AxialSpaceTimeTransformer(Module):
 
             maybe_kv_cache = next(iter_kv_cache, None) if layer_is_time else None
 
+            # residual values
+
+            layer_residual_values = maybe(pre_attn_rearrange)(residual_values)
+
             # attention layer
 
             tokens, attn_intermediates = attn(
@@ -1577,6 +1702,7 @@ class AxialSpaceTimeTransformer(Module):
                 rotary_pos_emb = layer_rotary_pos_emb,
                 attend_fn = attend_fn,
                 kv_cache = maybe_kv_cache,
+                residual_values = layer_residual_values,
                 return_intermediates = True
             )
 
@@ -1611,7 +1737,8 @@ class AxialSpaceTimeTransformer(Module):
         intermediates = TransformerIntermediates(
             stack(time_attn_kv_caches),
             safe_stack(normed_time_attn_inputs),
-            safe_stack(normed_space_attn_inputs)
+            safe_stack(normed_space_attn_inputs),
+            safe_stack(rnn_hiddens)
         )
 
         return out, intermediates
@@ -1643,11 +1770,6 @@ class VideoTokenizer(Module):
         encoder_add_decor_aux_loss = False,
         decor_auxx_loss_weight = 0.1,
         decorr_sample_frac = 0.25,
-        nd_rotary_kwargs: dict = dict(
-            rope_min_freq = 1.,
-            rope_max_freq = 10000.,
-            rope_p_zero_freqs = 0.
-        ),
         num_residual_streams = 1,
     ):
         super().__init__()
@@ -1820,11 +1942,23 @@ class VideoTokenizer(Module):
 
     def forward(
         self,
-        video, # (b c t h w) 
+        video_or_image, # (b c t h w) | (b c h w)
         return_latents = False,
         mask_patches = None,
-        return_all_losses = False
+        return_intermediates = False,
     ):
+
+        # handle image pretraining
+
+        is_image = video_or_image.ndim == 4
+
+        if is_image:
+            video = rearrange(video_or_image, 'b c h w -> b c 1 h w')
+        else:
+            video = video_or_image
+
+        # shapes
+
         batch, _, time, height, width = video.shape
         patch_size, device = self.patch_size, video.device
 
@@ -1864,7 +1998,7 @@ class VideoTokenizer(Module):
 
         # encoder attention
 
-        tokens, (_, time_attn_normed_inputs, space_attn_normed_inputs) = self.encoder_transformer(tokens, return_intermediates = True)
+        tokens, (_, time_attn_normed_inputs, space_attn_normed_inputs, _) = self.encoder_transformer(tokens, return_intermediates = True)
 
         # latent bottleneck
 
@@ -1904,12 +2038,21 @@ class VideoTokenizer(Module):
             space_decorr_loss * self.decorr_aux_loss_weight
         )
 
-        if not return_all_losses:
+        if not return_intermediates:
             return total_loss
 
-        losses = (recon_loss, lpips_loss, decorr_loss)
+        losses = TokenizerLosses(recon_loss, lpips_loss, time_decorr_loss, space_decorr_loss)
 
-        return total_loss, TokenizerLosses(*losses)
+        out = losses
+
+        # handle returning of reconstructed, and image pretraining
+
+        if is_image:
+            recon_video = rearrange(recon_video, 'b c 1 h w -> b c h w')
+
+        out = (losses, recon_video)
+
+        return total_loss, out
 
 # dynamics model, axial space-time transformer
 
@@ -1931,22 +2074,24 @@ class DynamicsWorldModel(Module):
         depth = 4,
         pred_orig_latent = True,   # directly predicting the original x0 data yield better results, rather than velocity (x-space vs v-space)
         time_block_every = 4,      # every 4th block is time
-        attn_kwargs: dict = dict(
+        attn_kwargs: dict = dict(),
-            heads = 8,
-        ),
         transformer_kwargs: dict = dict(),
+        attn_heads = 8,
         attn_dim_head = 64,
         attn_softclamp_value = 50.,
         ff_kwargs: dict = dict(),
+        use_time_rnn = True,
         loss_weight_fn: Callable = ramp_weight,
-        num_future_predictions = 8,                 # they do multi-token prediction of 8 steps forward
         prob_no_shortcut_train = None,              # probability of no shortcut training, defaults to 1 / num_step_sizes
         add_reward_embed_to_agent_token = False,
         add_reward_embed_dropout = 0.1,
+        add_state_pred_head = False,
+        state_pred_loss_weight = 0.1,
+        state_entropy_bonus_weight = 0.05,
         num_discrete_actions: int | tuple[int, ...] = 0,
         num_continuous_actions = 0,
         continuous_norm_stats = None,
-        multi_token_pred_len = 8,
+        multi_token_pred_len = 8,                   # they do multi-token prediction of 8 steps forward
         value_head_mlp_depth = 3,
         policy_head_mlp_depth = 3,
         latent_flow_loss_weight = 1.,
@@ -2066,6 +2211,25 @@ class DynamicsWorldModel(Module):
         self.pred_orig_latent = pred_orig_latent # x-space or v-space
         self.loss_weight_fn = loss_weight_fn
 
+        # state prediction, for state entropy bonus
+
+        self.add_state_pred_head = add_state_pred_head
+        self.state_pred_loss_weight = state_pred_loss_weight
+
+        self.should_pred_state = add_state_pred_head and state_pred_loss_weight > 0.
+
+        if self.should_pred_state:
+            self.state_pred_token = nn.Parameter(torch.randn(dim) * 1e-2)
+
+            self.to_state_pred = Sequential(
+                RMSNorm(dim),
+                nn.Linear(dim, num_latent_tokens * dim_latent * 2),
+                Rearrange('... (n d two) -> ... n d two', n = num_latent_tokens, two = 2)
+            )
+
+        self.state_entropy_bonus_weight = state_entropy_bonus_weight
+        self.add_state_entropy_bonus = self.should_pred_state and state_entropy_bonus_weight > 0.
+
         # reinforcement related
 
         # they sum all the actions into a single token
@@ -2147,6 +2311,7 @@ class DynamicsWorldModel(Module):
         self.transformer = AxialSpaceTimeTransformer(
             dim = dim,
             depth = depth,
+            attn_heads = attn_heads,
             attn_dim_head = attn_dim_head,
             attn_softclamp_value = attn_softclamp_value,
             attn_kwargs = attn_kwargs,
@@ -2155,6 +2320,7 @@ class DynamicsWorldModel(Module):
             num_special_spatial_tokens = num_agents,
             time_block_every = time_block_every,
             final_norm = False,
+            rnn_time = use_time_rnn,
             **transformer_kwargs
         )
 
@@ -2298,10 +2464,10 @@ class DynamicsWorldModel(Module):
         env,
         seed = None,
         agent_index = 0,
-        step_size = 4,
+        num_steps = 4,
         max_timesteps = 16,
         env_is_vectorized = False,
-        use_time_kv_cache = True,
+        use_time_cache = True,
         store_agent_embed = True,
         store_old_action_unembeds = True,
     ):
@@ -2338,9 +2504,14 @@ class DynamicsWorldModel(Module):
 
         episode_lens = full((batch,), 0, device = device)
 
+        # derive step size
+
+        assert divisible_by(self.max_steps, num_steps)
+        step_size = self.max_steps // num_steps
+
         # maybe time kv cache
 
-        time_kv_cache = None
+        time_cache = None
 
         step_index = 0
 
@@ -2349,14 +2520,14 @@ class DynamicsWorldModel(Module):
 
             latents = self.video_tokenizer(video, return_latents = True)
 
-            _, (agent_embed, next_time_kv_cache) = self.forward(
+            _, (embeds, next_time_cache) = self.forward(
                 latents = latents,
                 signal_levels = self.max_steps - 1,
                 step_sizes = step_size,
                 rewards = rewards,
                 discrete_actions = discrete_actions,
                 continuous_actions = continuous_actions,
-                time_kv_cache = time_kv_cache,
+                time_cache = time_cache,
                 latent_is_noised = True,
                 return_pred_only = True,
                 return_intermediates = True
@@ -2364,11 +2535,13 @@ class DynamicsWorldModel(Module):
 
             # time kv cache
 
-            if use_time_kv_cache:
+            if use_time_cache:
-                time_kv_cache = next_time_kv_cache
+                time_cache = next_time_cache
 
             # get one agent
 
+            agent_embed = embeds.agent
+
             one_agent_embed = agent_embed[..., -1:, agent_index, :]
 
             # values
@@ -2423,6 +2596,17 @@ class DynamicsWorldModel(Module):
             elif len(env_step_out) == 5:
                 next_frame, reward, terminated, truncated, info = env_step_out
 
+            # maybe add state entropy bonus
+
+            if self.add_state_entropy_bonus:
+                state_pred_token = embeds.state_pred
+
+                state_pred = self.to_state_pred(state_pred_token)
+
+                state_pred_log_variance = state_pred[..., 1].sum()
+
+                reward = reward + state_pred_log_variance * self.state_entropy_bonus_weight
+
             # update episode lens
 
             episode_lens = torch.where(is_terminated, episode_lens, episode_lens + 1)
@@ -2601,7 +2785,7 @@ class DynamicsWorldModel(Module):
         ):
 
             with world_model_forward_context():
-                _, (agent_embeds, _) = self.forward(
+                _, (embeds, _) = self.forward(
                     latents = latents,
                     signal_levels = self.max_steps - 1,
                     step_sizes = step_size,
@@ -2613,7 +2797,7 @@ class DynamicsWorldModel(Module):
                     return_intermediates = True
                 )
 
-            agent_embeds = agent_embeds[..., agent_index, :]
+            agent_embeds = embeds.agent[..., agent_index, :]
 
         # maybe detach agent embed
 
@@ -2757,13 +2941,13 @@ class DynamicsWorldModel(Module):
         image_width = None,
         return_decoded_video = None,
         context_signal_noise = 0.1,       # they do a noising of the past, this was from an old diffusion world modeling paper from EPFL iirc
-        time_kv_cache: Tensor | None = None,
+        time_cache: Tensor | None = None,
-        use_time_kv_cache = True,
+        use_time_cache = True,
         return_rewards_per_frame = False,
         return_agent_actions = False,
         return_log_probs_and_values = False,
         return_for_policy_optimization = False,
-        return_time_kv_cache = False,
+        return_time_cache = False,
         store_agent_embed = True,
         store_old_action_unembeds = True
 
@@ -2852,7 +3036,7 @@ class DynamicsWorldModel(Module):
             # (2) decoding anything off agent embedding (rewards, actions, etc)
 
             take_extra_step = (
-                use_time_kv_cache or
+                use_time_cache or
                 return_rewards_per_frame or
                 store_agent_embed or
                 return_agent_actions
@@ -2893,7 +3077,7 @@ class DynamicsWorldModel(Module):
 
                 signal_levels_with_context = F.pad(signal_levels, (curr_time_steps, 0), value = self.max_steps - 1)
 
-                pred, (agent_embed, next_time_kv_cache) = self.forward(
+                pred, (embeds, next_time_cache) = self.forward(
                     latents = noised_latent_with_context,
                     signal_levels = signal_levels_with_context,
                     step_sizes = step_size,
@@ -2903,15 +3087,15 @@ class DynamicsWorldModel(Module):
                     discrete_actions = decoded_discrete_actions,
                     continuous_actions = decoded_continuous_actions,
                     proprio = noised_proprio_with_context,
-                    time_kv_cache = time_kv_cache,
+                    time_cache = time_cache,
                     latent_is_noised = True,
                     latent_has_view_dim = True,
                     return_pred_only = True,
                     return_intermediates = True,
                 )
 
-                if use_time_kv_cache and is_last_step:
+                if use_time_cache and is_last_step:
-                    time_kv_cache = next_time_kv_cache
+                    time_cache = next_time_cache
 
                 # early break if taking an extra step for agent embedding off cleaned latents for decoding
 
@@ -2920,8 +3104,10 @@ class DynamicsWorldModel(Module):
 
                 # maybe proprio
 
-                if has_proprio:
+                # maybe proprio
-                    pred, pred_proprio = pred
+
+                pred_proprio = pred.proprioception
+                pred = pred.flow
 
                 # unpack pred
 
@@ -2958,6 +3144,8 @@ class DynamicsWorldModel(Module):
             # take care of the rewards by predicting on the agent token embedding on the last denoising step
 
             if return_rewards_per_frame:
+                agent_embed = embeds.agent
+
                 one_agent_embed = agent_embed[:, -1:, agent_index]
 
                 reward_logits = self.to_reward_pred.forward_one(one_agent_embed, id = 0)
@@ -2968,6 +3156,8 @@ class DynamicsWorldModel(Module):
             # maybe store agent embed
 
             if store_agent_embed:
+                agent_embed = embeds.agent
+
                 one_agent_embed = agent_embed[:, -1:, agent_index]
                 acc_agent_embed = safe_cat((acc_agent_embed, one_agent_embed), dim = 1)
 
@@ -3060,10 +3250,10 @@ class DynamicsWorldModel(Module):
         if not has_at_least_one(return_rewards_per_frame, return_agent_actions, has_proprio):
             out = video if return_decoded_video else latents
 
-            if not return_time_kv_cache:
+            if not return_time_cache:
                 return out
 
-            return out, time_kv_cache
+            return out, time_cache
 
         # returning agent actions, rewards, and log probs + values for policy optimization
 
@@ -3093,10 +3283,10 @@ class DynamicsWorldModel(Module):
 
             gen.values = decoded_values
 
-        if not return_time_kv_cache:
+        if not return_time_cache:
             return gen
 
-        return gen, time_kv_cache
+        return gen, time_cache
 
     def forward(
         self,
@@ -3115,7 +3305,7 @@ class DynamicsWorldModel(Module):
         discrete_action_types = None,    # (na)
         continuous_action_types = None,  # (na)
         proprio = None,                  # (b t dp)
-        time_kv_cache = None,
+        time_cache = None,
         return_pred_only = False,
         latent_is_noised = False,
         return_all_losses = False,
@@ -3265,9 +3455,13 @@ class DynamicsWorldModel(Module):
 
         agent_tokens = repeat(agent_tokens, 'b ... d -> b t ... d', t = time)
 
+        # empty token
+
+        empty_token = agent_tokens[:, :, 0:0]
+
         # maybe reward tokens
 
-        reward_tokens = agent_tokens[:, :, 0:0]
+        reward_tokens = empty_token
 
         if exists(rewards):
             two_hot_encoding = self.reward_encoder(rewards)
@@ -3307,6 +3501,13 @@ class DynamicsWorldModel(Module):
             else:
                 noised_proprio = proprio
 
+        # maybe state prediction token
+
+        if self.should_pred_state:
+            state_pred_token = repeat(self.state_pred_token, 'd -> b t 1 d', b = batch, t = time)
+        else:
+            state_pred_token = empty_token
+
         # maybe create the action tokens
 
         if exists(discrete_actions) or exists(continuous_actions):
@@ -3331,11 +3532,11 @@ class DynamicsWorldModel(Module):
             action_tokens = torch.zeros_like(agent_tokens[:, :, 0:1])
 
         else:
-            action_tokens = agent_tokens[:, :, 0:0] # else empty off agent tokens
+            action_tokens = empty_token # else empty off agent tokens
 
         # main function, needs to be defined as such for shortcut training - additional calls for consistency loss
 
-        def get_prediction(noised_latents, noised_proprio, signal_levels, step_sizes_log2, action_tokens, reward_tokens, agent_tokens, return_agent_tokens = False, return_time_kv_cache = False):
+        def get_prediction(noised_latents, noised_proprio, signal_levels, step_sizes_log2, state_pred_token, action_tokens, reward_tokens, agent_tokens, return_agent_tokens = False, return_time_cache = False):
 
             # latents to spatial tokens
 
@@ -3384,15 +3585,15 @@ class DynamicsWorldModel(Module):
 
             # pack to tokens for attending
 
-            tokens, packed_tokens_shape = pack([flow_token, space_tokens, proprio_token, registers, action_tokens, reward_tokens, agent_tokens], 'b t * d')
+            tokens, packed_tokens_shape = pack([flow_token, space_tokens, proprio_token, state_pred_token, registers, action_tokens, reward_tokens, agent_tokens], 'b t * d')
 
             # attention
 
-            tokens, (next_time_kv_cache, *_) = self.transformer(tokens, kv_cache = time_kv_cache, return_intermediates = True)
+            tokens, intermediates = self.transformer(tokens, cache = time_cache, return_intermediates = True)
 
             # unpack
 
-            flow_token, space_tokens, proprio_token, register_tokens, action_tokens, reward_tokens, agent_tokens = unpack(tokens, packed_tokens_shape, 'b t * d')
+            flow_token, space_tokens, proprio_token, state_pred_token, register_tokens, action_tokens, reward_tokens, agent_tokens = unpack(tokens, packed_tokens_shape, 'b t * d')
 
             # pooling
 
@@ -3404,42 +3605,54 @@ class DynamicsWorldModel(Module):
 
             if self.has_proprio:
                 pred_proprio = self.to_proprio_pred(proprio_token)
+            else:
+                pred_proprio = None
 
-                pred = (pred, pred_proprio)
+            # maybe state pred
+
+            if self.should_pred_state:
+                pred_state = self.to_state_pred(state_pred_token)
+            else:
+                pred_state = None
 
             # returning
 
+            predictions = Predictions(pred, pred_proprio, pred_state)
+
+            embeds = Embeds(agent_tokens, state_pred_token)
+
             if not return_agent_tokens:
-                return pred
+                return predictions
 
-            if not return_time_kv_cache:
+            if not return_time_cache:
-                return pred, agent_tokens
+                return predictions, embeds
 
-            return pred, (agent_tokens, next_time_kv_cache)
+            return predictions, (embeds, intermediates)
 
         # curry into get_prediction what does not change during first call as well as the shortcut ones
 
-        _get_prediction = partial(get_prediction, action_tokens = action_tokens, reward_tokens = reward_tokens, agent_tokens = agent_tokens)
+        _get_prediction = partial(get_prediction, state_pred_token = state_pred_token, action_tokens = action_tokens, reward_tokens = reward_tokens, agent_tokens = agent_tokens)
 
         # forward the network
 
-        pred, (encoded_agent_tokens, next_time_kv_cache) = _get_prediction(noised_latents, noised_proprio, signal_levels, step_sizes_log2, return_agent_tokens = True, return_time_kv_cache = True)
+        pred, (embeds, intermediates) = _get_prediction(noised_latents, noised_proprio, signal_levels, step_sizes_log2, return_agent_tokens = True, return_time_cache = True)
 
         if return_pred_only:
             if not return_intermediates:
                 return pred
 
-            return pred, (encoded_agent_tokens, next_time_kv_cache)
+            return pred, (embeds, intermediates)
 
         # pack the predictions to calculate flow for different modalities all at once
 
         if self.has_proprio:
-            pred, for_flow_loss_packed_shape = pack(pred, 'b t *')
+            packed_pred, for_flow_loss_packed_shape = pack((pred.flow, pred.proprioception), 'b t *')
 
             noised, _ = pack((noised_latents, noised_proprio), 'b t *')
             data, _ = pack((latents, proprio), 'b t *')
             noise, _ = pack((noise, proprio_noise), 'b t *')
         else:
+            packed_pred = pred.flow
             noised = noised_latents
             data = latents
 
@@ -3458,9 +3671,10 @@ class DynamicsWorldModel(Module):
                 pred = fn(noised, noised_proprio, *args, **kwargs)
 
                 if self.has_proprio:
-                    pred, _ = pack(pred, 'b t *')
+                    packed_flow, _ = pack((pred.flow, pred.proprioception), 'b t *')
+                    return packed_flow
 
-                return pred
+                return pred.flow
             return inner
 
         wrapped_get_prediction = maybe_pack_unpack(_get_prediction)
@@ -3527,12 +3741,12 @@ class DynamicsWorldModel(Module):
             # need to convert x-space to v-space
 
             if is_x_space:
-                pred = (pred - noised) / (1. - first_times)
+                packed_pred = (packed_pred - noised) / (1. - first_times)
                 maybe_shortcut_loss_weight = (1. - first_times) ** 2
 
         # mse loss
 
-        flow_losses = F.mse_loss(pred, pred_target, reduction = 'none')
+        flow_losses = F.mse_loss(packed_pred, pred_target, reduction = 'none')
 
         flow_losses = flow_losses * maybe_shortcut_loss_weight # handle the (1-t)^2 in eq(7)
 
@@ -3564,6 +3778,8 @@ class DynamicsWorldModel(Module):
 
         if exists(rewards):
 
+            encoded_agent_tokens = embeds.agent
+
             if rewards.ndim == 2: # (b t)
                 encoded_agent_tokens = reduce(encoded_agent_tokens, 'b t g d -> b t d', 'mean')
 
@@ -3584,6 +3800,18 @@ class DynamicsWorldModel(Module):
             else:
                 reward_loss = reduce(reward_losses, '... mtp -> mtp', 'mean') # they sum across the prediction steps (mtp dimension) - eq(9)
 
+        # maybe autoregressive state prediction loss
+
+        state_pred_loss = self.zero
+
+        if self.should_pred_state:
+            pred_latent, latent_to_pred = pred.state[:, :-1], latents[:, 1:]
+
+            pred_latent_mean, pred_latent_log_var = pred_latent.unbind(dim = -1)
+            pred_latent_var = pred_latent_log_var.exp()
+
+            state_pred_loss = F.gaussian_nll_loss(pred_latent_mean, latent_to_pred, var = pred_latent_var)
+
         # maybe autoregressive action loss
 
         discrete_action_loss = self.zero
@@ -3651,7 +3879,7 @@ class DynamicsWorldModel(Module):
 
         # handle loss normalization
 
-        losses = WorldModelLosses(flow_loss, reward_loss, discrete_action_loss, continuous_action_loss)
+        losses = WorldModelLosses(flow_loss, reward_loss, discrete_action_loss, continuous_action_loss, state_pred_loss)
 
         if exists(self.flow_loss_normalizer):
             flow_loss = self.flow_loss_normalizer(flow_loss, update_ema = update_loss_ema)
@@ -3671,7 +3899,8 @@ class DynamicsWorldModel(Module):
             flow_loss * self.latent_flow_loss_weight +
             (reward_loss * self.reward_loss_weight).sum() +
             (discrete_action_loss * self.discrete_action_loss_weight).sum() + 
-            (continuous_action_loss * self.continuous_action_loss_weight).sum()
+            (continuous_action_loss * self.continuous_action_loss_weight).sum() +
+            (state_pred_loss * self.state_pred_loss_weight)
         )
 
         if not return_all_losses:

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "dreamer4"
-version = "0.1.12"
+version = "0.1.24"
 description = "Dreamer 4"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

tests/test_dreamer.py

@@ -15,7 +15,8 @@ def exists(v):
 @param('condition_on_actions', (False, True))
 @param('num_residual_streams', (1, 4))
 @param('add_reward_embed_to_agent_token', (False, True))
-@param('use_time_kv_cache', (False, True))
+@param('add_state_pred_head', (False, True))
+@param('use_time_cache', (False, True))
 @param('var_len', (False, True))
 def test_e2e(
     pred_orig_latent,
@@ -28,7 +29,8 @@ def test_e2e(
     condition_on_actions,
     num_residual_streams,
     add_reward_embed_to_agent_token,
-    use_time_kv_cache,
+    add_state_pred_head,
+    use_time_cache,
     var_len
 ):
     from dreamer4.dreamer4 import VideoTokenizer, DynamicsWorldModel
@@ -71,12 +73,13 @@ def test_e2e(
         pred_orig_latent = pred_orig_latent,
         num_discrete_actions = 4,
         attn_dim_head = 16,
+        attn_heads = heads,
         attn_kwargs = dict(
-            heads = heads,
             query_heads = query_heads,
         ),
         prob_no_shortcut_train = prob_no_shortcut_train,
         add_reward_embed_to_agent_token = add_reward_embed_to_agent_token,
+        add_state_pred_head = add_state_pred_head,
         num_residual_streams = num_residual_streams
     )
 
@@ -123,7 +126,7 @@ def test_e2e(
         image_width = 128,
         batch_size = 2,
         return_rewards_per_frame = True,
-        use_time_kv_cache = use_time_kv_cache
+        use_time_cache = use_time_cache
     )
 
     assert generations.video.shape == (2, 3, 10, 128, 128)
@@ -617,9 +620,9 @@ def test_cache_generate():
         num_residual_streams = 1
     )
 
-    generated, time_kv_cache = dynamics.generate(1, return_time_kv_cache = True)
+    generated, time_cache = dynamics.generate(1, return_time_cache = True)
-    generated, time_kv_cache = dynamics.generate(1, time_kv_cache = time_kv_cache, return_time_kv_cache = True)
+    generated, time_cache = dynamics.generate(1, time_cache = time_cache, return_time_cache = True)
-    generated, time_kv_cache = dynamics.generate(1, time_kv_cache = time_kv_cache, return_time_kv_cache = True)
+    generated, time_cache = dynamics.generate(1, time_cache = time_cache, return_time_cache = True)
 
 @param('vectorized', (False, True))
 @param('use_pmpo', (False, True))
@@ -643,7 +646,9 @@ def test_online_rl(
         dim_latent = 16,
         patch_size = 32,
         attn_dim_head = 16,
-        num_latent_tokens = 1
+        num_latent_tokens = 1,
+        image_height = 256,
+        image_width = 256,
     )
 
     world_model_and_policy = DynamicsWorldModel(
@@ -677,10 +682,12 @@ def test_online_rl(
 
     # manually
 
+    dream_experience = world_model_and_policy.generate(10, batch_size = 1, store_agent_embed = store_agent_embed, return_for_policy_optimization = True)
+
     one_experience = world_model_and_policy.interact_with_env(mock_env, max_timesteps = 8, env_is_vectorized = vectorized, store_agent_embed = store_agent_embed)
     another_experience = world_model_and_policy.interact_with_env(mock_env, max_timesteps = 16, env_is_vectorized = vectorized, store_agent_embed = store_agent_embed)
 
-    combined_experience = combine_experiences([one_experience, another_experience])
+    combined_experience = combine_experiences([dream_experience, one_experience, another_experience])
 
     # quick test moving the experience to different devices
 
@@ -803,3 +810,22 @@ def test_epo():
 
     fitness = torch.randn(16,)
     dynamics.evolve_(fitness)
+
+def test_images_to_video_tokenizer():
+    import torch
+    from dreamer4 import VideoTokenizer, DynamicsWorldModel, AxialSpaceTimeTransformer
+
+    tokenizer = VideoTokenizer(
+        dim = 512,
+        dim_latent = 32,
+        patch_size = 32,
+        image_height = 256,
+        image_width = 256,
+        encoder_add_decor_aux_loss = True
+    )
+
+    images = torch.randn(2, 3, 256, 256)
+    loss, (losses, recon_images) = tokenizer(images, return_intermediates = True)
+    loss.backward()
+
+    assert images.shape == recon_images.shape