hongshaorou/dreamer4
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Revert "address https://github.com/lucidrains/dreamer4/issues/1"

Browse Source 
This reverts commit e23a5294ec2f49d58d3ccb936c498eb86939fa96.
This commit is contained in:
lucidrains 2025-10-02 12:25:05 -07:00
parent e23a5294ec
commit 8d1cd311bb
2 changed files with 18 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

34
dreamer4/dreamer4.py
View File

@ -528,11 +528,11 @@ class DynamicsModel(Module):
dim_latent,
num_signal_levels = 500,
num_step_sizes = 32,
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
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_is_clean_latents = 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
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,
heads = 8,
@ -561,7 +561,7 @@ class DynamicsModel(Module):
self.signal_levels_embed = nn.Embedding(num_signal_levels, dim_half)
self.step_sizes_embed = nn.Embedding(num_step_sizes, dim_half)
self.pred_is_clean_latents = pred_is_clean_latents
self.pred_orig_latent = pred_orig_latent
# they sum all the actions into a single token
@ -611,15 +611,19 @@ class DynamicsModel(Module):
noise = torch.randn_like(latents)
times = signal_levels.float() / self.num_signal_levels
interp = rearrange(signal_levels.float() / self.num_signal_levels, 'b t -> b t 1')
times = rearrange(times, 'b t -> b t 1')
orig_latents = latents
flow = latents - noise
latents = noise.lerp(latents, interp)
latents = noise.lerp(latents, times)
# allow for original velocity pred
# x-space as in paper is in else clause
noised_latents = latents
if not self.pred_orig_latent:
pred_target = flow = latents - noise
else:
pred_target = latents
# latents to spatial tokens
@ -673,15 +677,7 @@ class DynamicsModel(Module):
if not flow_matching:
return pred
# x-space vs v-space
if self.pred_is_clean_latents:
denoised_latent = pred
pred_flow = (denoised_latent - noised_latents) / (1. - times)
else:
pred_flow = pred
return F.mse_loss(pred_flow, flow)
return F.mse_loss(pred, pred_target)
# dreamer

6
tests/test_dreamer.py
View File

@ -2,9 +2,9 @@ import pytest
param = pytest.mark.parametrize
import torch
@param('pred_is_clean_latents', (False, True))
@param('pred_orig_latent', (False, True))
def test_e2e(
pred_is_clean_latents
pred_orig_latent
):
from dreamer4.dreamer4 import VideoTokenizer, DynamicsModel
@ -17,7 +17,7 @@ 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, pred_is_clean_latents = pred_is_clean_latents)
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))