allow for adding an RNN before time attention, but need to handle caching still

dreamer4/dreamer4.py

@@ -1470,7 +1470,8 @@ class AxialSpaceTimeTransformer(Module):
         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,
-        value_residual = True                # https://arxiv.org/abs/2410.17897 - but with learned mixing from OSS
+        value_residual = True,               # https://arxiv.org/abs/2410.17897 - but with learned mixing from OSS
+        rnn_time = False
     ):
         super().__init__()
         assert depth >= time_block_every, f'depth must be at least {time_block_every}'
@@ -1504,6 +1505,11 @@ class AxialSpaceTimeTransformer(Module):
                 Rearrange('... (h d) -> ... h d', h = attn_heads)
             )
 
+        # a gru layer across time
+
+        self.rnn_time = rnn_time
+        rnn_layers = []
+
         # transformer
 
         layers = []
@@ -1525,7 +1531,14 @@ class AxialSpaceTimeTransformer(Module):
                 hyper_conn(branch = SwiGLUFeedforward(dim = dim, **ff_kwargs))
             ]))
 
+            rnn_layers.append(ModuleList([
+                nn.RMSNorm(dim),
+                nn.GRU(dim, dim, batch_first = True)
+            ]) 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
@@ -1605,10 +1618,23 @@ 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_modules, 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_modules):
+                rnn_prenorm, rnn = maybe_rnn_modules
+
+                rnn_input, inverse_pack_time = pack_one(tokens, '* t d')
+
+                rnn_out, rnn_hiddens = rnn(rnn_prenorm(rnn_input)) # todo, handle rnn cache
+
+                rnn_out = inverse_pack_time(rnn_out)
+
+                tokens = rnn_out + tokens
+
             # when is a axial time attention block, should be causal
 
             attend_fn = time_attend if layer_is_time else space_attend

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "dreamer4"
-version = "0.1.15"
+version = "0.1.16"
 description = "Dreamer 4"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }