x-space and v-space prediction in dynamics model

dreamer4/dreamer4.py

@@ -509,6 +509,7 @@ class DynamicsModel(Module):
         num_spatial_tokens = 32,   # latents were projected into spatial tokens, and presumably pooled back for the final prediction (or one special one does the x-prediction)
         num_register_tokens = 8,   # they claim register tokens led to better temporal consistency
         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(
             dim_head = 64,
@@ -532,9 +533,14 @@ class DynamicsModel(Module):
         assert divisible_by(dim, 2)
         dim_half = dim // 2
 
+        self.num_signal_levels = num_signal_levels
+        self.num_step_sizes = num_step_sizes
+
         self.signal_levels_embed = nn.Embedding(num_signal_levels, dim_half)
         self.step_sizes_embed = nn.Embedding(num_step_sizes, dim_half)
 
+        self.pred_orig_latent = pred_orig_latent
+
         # they sum all the actions into a single token
 
         self.action_learned_embed = Parameter(torch.randn(dim) * 1e-2)
@@ -573,6 +579,32 @@ class DynamicsModel(Module):
         step_sizes = None           # (b t)
     ):
 
+        assert not (exists(signal_levels) ^ exists(step_sizes))
+
+        flow_matching = exists(signal_levels)
+
+        # flow matching if `signal_levels` passed in
+
+        if flow_matching:
+
+            noise = torch.randn_like(latents)
+
+            interp = rearrange(signal_levels.float() / self.num_signal_levels, 'b t -> b t 1')
+
+            orig_latents = latents
+
+            latents = noise.lerp(latents, interp)
+
+            # allow for original velocity pred
+            # x-space as in paper is in else clause
+
+            if not self.pred_orig_latent:
+                pred_target = flow = latents - noise
+            else:
+                pred_target = latents
+
+        # latents to spatial tokens
+
         space_tokens = self.latents_to_spatial_tokens(latents)
 
         # pack to tokens
@@ -584,14 +616,13 @@ class DynamicsModel(Module):
 
         # determine signal + step size embed for their diffusion forcing + shortcut
 
-        assert not (exists(signal_levels) ^ exists(step_sizes))
-
         if exists(signal_levels):
             signal_embed = self.signal_levels_embed(signal_levels)
             step_size_embed = self.step_sizes_embed(step_sizes)
 
             flow_token = cat((signal_embed, step_size_embed), dim = -1)
             flow_token = rearrange(flow_token, 'b t d -> b t d')
+
         else:
             flow_token = registers[..., 0:0, :]
 
@@ -619,7 +650,12 @@ class DynamicsModel(Module):
 
         pooled = reduce(space_tokens, 'b t s d -> b t d', 'mean')
 
-        return self.to_pred(pooled)
+        pred = self.to_pred(pooled)
+
+        if not flow_matching:
+            return pred
+
+        return F.mse_loss(pred, pred_target)
 
 # dreamer
 

tests/test_dreamer.py

@@ -1,7 +1,11 @@
 import pytest
+param = pytest.mark.parametrize
 import torch
 
-def test_e2e():
+@param('pred_orig_latent', (False, True))
+def test_e2e(
+    pred_orig_latent
+):
     from dreamer4.dreamer4 import VideoTokenizer, DynamicsModel
 
     tokenizer = VideoTokenizer(512, dim_latent = 32, patch_size = 32)
@@ -13,10 +17,10 @@ def test_e2e():
     latents = tokenizer(x, return_latents = True)
     assert latents.shape[-1] == 32
 
-    dynamics = DynamicsModel(512, dim_latent = 32, num_signal_levels = 500, num_step_sizes = 32)
+    dynamics = DynamicsModel(512, dim_latent = 32, num_signal_levels = 500, num_step_sizes = 32, pred_orig_latent = pred_orig_latent)
 
     signal_levels = torch.randint(0, 500, (2, 4))
     step_sizes = torch.randint(0, 32, (2, 4))
 
-    pred = dynamics(latents, signal_levels = signal_levels, step_sizes = step_sizes)
-    assert pred.shape == latents.shape
+    flow_loss = dynamics(latents, signal_levels = signal_levels, step_sizes = step_sizes)
+    assert flow_loss.numel() == 1