admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-07-06 17:41:47 +00:00
Code Issues Packages Projects Releases Wiki Activity

use patch tokens

Browse Source
This commit is contained in:
Pepijn
2025-08-31 20:52:00 +02:00
parent eff5b90542
commit d8c875e069
2 changed files with 73 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/lerobot/policies/rlearn/eval_script.py
+12 -1
View File
@@ -155,6 +155,7 @@ def predict_rewards_sliding(model, frames, language, max_seq_len=16, batch_size=
windows = []
frame_positions = [] # Track which temporal position each frame should use
left_pad_counts = [] # Number of left-pad (OOB) frames per window
for i in range(T):
start = max(0, i - L + 1)
@@ -162,8 +163,10 @@ def predict_rewards_sliding(model, frames, language, max_seq_len=16, batch_size=
if window.shape[0] < L:
pad_needed = L - window.shape[0]
pad = window[:1].expand(pad_needed, -1, -1, -1) # repeat first frame
pad = window[:1].expand(pad_needed, -1, -1, -1) # repeat first frame (clamp to frame 0)
window = torch.cat([pad, window], dim=0)
else:
pad_needed = 0
# IMPROVED FIX: Cycle through MLPs to get varied predictions throughout the episode
# This ensures we use all 16 frame-specific MLPs and get varied outputs
@@ -172,6 +175,7 @@ def predict_rewards_sliding(model, frames, language, max_seq_len=16, batch_size=
windows.append(window)
frame_positions.append(frame_pos)
left_pad_counts.append(pad_needed)
preds = np.zeros(T, dtype=float)
@@ -185,6 +189,13 @@ def predict_rewards_sliding(model, frames, language, max_seq_len=16, batch_size=
# Model returns (B, L) predictions for each temporal position
values = model.predict_rewards(batch) # torch.Tensor (B, L)
# Apply eval-time padding rule: predictions for left-padded (OOB) frames are zero
if values.dim() == 2 and len(left_pad_counts) >= (e - s):
for b_idx in range(e - s):
pad_n = left_pad_counts[s + b_idx]
if pad_n > 0:
values[b_idx, :pad_n] = 0.0
# Debug output removed - issue was identified and fixed
if values.dim() == 2:
src/lerobot/policies/rlearn/modeling_rlearn.py
+61 -42
View File
@@ -108,6 +108,9 @@ class RLearNPolicy(PreTrainedPolicy):
# Stronger temporal positional encoding
self.temporal_pos_embedding = nn.Parameter(torch.randn(config.max_seq_len, config.dim_model) * 0.1)
# Spatial (patch) positional encoding for patch tokens
self.max_patch_tokens = getattr(config, 'max_patch_tokens', 256)
self.spatial_pos_embedding = nn.Parameter(torch.randn(self.max_patch_tokens, config.dim_model) * 0.1)
# Single MLP processes all frames
self.frame_mlp = nn.Linear(config.dim_model, config.dim_model)
@@ -226,8 +229,8 @@ class RLearNPolicy(PreTrainedPolicy):
device = next(self.parameters()).device
frames = frames.to(device)
# Process video frames
video_embeds = self._encode_video_frames(frames).to(device) # (B, T, D_vision)
# Process video frames -> patch tokens per frame
video_patch_embeds = self._encode_video_frames(frames).to(device) # (B, T, P, D_vision)
# Language embeddings + mask
lang_embeds, mask = self._encode_language_tokens(commands, device)
@@ -237,10 +240,17 @@ class RLearNPolicy(PreTrainedPolicy):
# Project embeddings
lang_tokens = self.to_lang_tokens(lang_embeds)
video_tokens = self.to_video_tokens(video_embeds)
# Add temporal positional encoding (window-relative only)
T_video = video_tokens.shape[1]
video_tokens = video_tokens + self.temporal_pos_embedding[:T_video]
video_tokens = self.to_video_tokens(video_patch_embeds) # (B, T, P, D)
# Add temporal + spatial positional encoding (window-relative time + patch index)
Bv, T_video, P_video, Dm = video_tokens.shape
if P_video > self.spatial_pos_embedding.shape[0]:
raise ValueError(f"Number of patch tokens {P_video} exceeds max_patch_tokens {self.spatial_pos_embedding.shape[0]}")
t_pos = self.temporal_pos_embedding[:T_video] # (T, D)
p_pos = self.spatial_pos_embedding[:P_video] # (P, D)
pos = t_pos[:, None, :] + p_pos[None, :, :] # (T, P, D)
video_tokens = video_tokens + pos # broadcast over batch
# Flatten patch dimension for attention
video_tokens = rearrange(video_tokens, 'b t p d -> b (t p) d')
# Pack all tokens for attention
tokens, lang_video_packed_shape = pack((lang_tokens, register_tokens, video_tokens), 'b * d')
@@ -252,10 +262,11 @@ class RLearNPolicy(PreTrainedPolicy):
attended = self.decoder(tokens, mask=mask)
# Unpack and get video token features
_, _, attended_video_tokens = unpack(attended, lang_video_packed_shape, 'b * d')
# Process all frames with single MLP
frame_tokens = self.frame_mlp(attended_video_tokens) # (B, T, D)
_, _, attended_video_tokens = unpack(attended, lang_video_packed_shape, 'b * d') # (B, T*P, D)
# Restore (B, T, P, D) and pool patches per frame
attended_video_tokens = rearrange(attended_video_tokens, 'b (t p) d -> b t p d', t=T_video, p=P_video)
frame_tokens = attended_video_tokens.mean(dim=2) # (B, T, D)
frame_tokens = self.frame_mlp(frame_tokens)
# MLP predictor
video_frame_embeds = self.mlp_predictor(frame_tokens)
@@ -283,13 +294,13 @@ class RLearNPolicy(PreTrainedPolicy):
return batch
def _encode_video_frames(self, frames: Tensor) -> Tensor:
"""Encode video frames through DinoV3 to get per-frame embeddings.
"""Encode video frames through DinoV3 to get per-frame PATCH embeddings.
Args:
frames: (B, T, C, H, W)
Returns:
(B, T, D_vision)
(B, T, P, D_vision) where P is number of patch tokens per frame (excludes CLS)
"""
B, T, C, H, W = frames.shape
flat = rearrange(frames, 'b t c h w -> (b t) c h w')
@@ -315,40 +326,40 @@ class RLearNPolicy(PreTrainedPolicy):
# Process in batch through DINOv3 model
vision_outputs = self.vision_model(**inputs)
# Prefer mean-pooled patch tokens over pooler/CLS to ensure input-dependent variation
# Prefer patch tokens from last_hidden_state (exclude CLS at index 0)
if hasattr(vision_outputs, 'last_hidden_state') and vision_outputs.last_hidden_state is not None:
tokens = vision_outputs.last_hidden_state # (BT, N_tokens, D)
if tokens.dim() == 3 and tokens.shape[1] > 1:
# Exclude CLS/reg token at index 0, average over patch tokens
vision_features_flat = tokens[:, 1:, :].mean(dim=1)
patch_tokens_flat = tokens[:, 1:, :] # (BT, P, D)
else:
# Fallback to first token if only one token is present
vision_features_flat = tokens[:, 0]
# Only one token available → treat as single patch
patch_tokens_flat = tokens[:, :1, :]
elif hasattr(vision_outputs, 'pooler_output') and vision_outputs.pooler_output is not None:
vision_features_flat = vision_outputs.pooler_output # (BT, D)
# No per-patch tokens available, synthesize single patch from pooler
patch_tokens_flat = vision_outputs.pooler_output[:, None, :] # (BT, 1, D)
else:
raise RuntimeError("DINOv3 outputs do not contain last_hidden_state or pooler_output")
# Robustly reshape to (B, T, D): detect correct flatten order by maximizing temporal variance
# Robustly reshape to (B, T, P, D): detect correct flatten order by maximizing temporal variance (on patch-mean)
try:
cand1 = rearrange(vision_features_flat, '(b t) d -> b t d', b=B, t=T)
cand2 = rearrange(vision_features_flat, '(t b) d -> b t d', t=T, b=B)
# Compute mean temporal difference per sample
def mean_time_diff(x):
cand1 = rearrange(patch_tokens_flat, '(b t) p d -> b t p d', b=B, t=T)
cand2 = rearrange(patch_tokens_flat, '(t b) p d -> b t p d', t=T, b=B)
def mean_time_diff_4d(x):
if T <= 1:
return torch.tensor(0.0, device=x.device)
diffs = (x[:, 1:, :] - x[:, :-1, :]).pow(2).sum(dim=-1).sqrt()
x_mean = x.mean(dim=2) # (B, T, D)
diffs = (x_mean[:, 1:, :] - x_mean[:, :-1, :]).pow(2).sum(dim=-1).sqrt()
return diffs.mean()
diff1 = mean_time_diff(cand1)
diff2 = mean_time_diff(cand2)
vision_features = cand1 if diff1 >= diff2 else cand2
diff1 = mean_time_diff_4d(cand1)
diff2 = mean_time_diff_4d(cand2)
patch_features = cand1 if diff1 >= diff2 else cand2
if self.training and torch.rand(1).item() < 0.05:
print(f"DINO reshape choice: {'(b t)->b t' if diff1 >= diff2 else '(t b)->b t'} | diff1={diff1.item():.6f}, diff2={diff2.item():.6f}")
except Exception:
# Fallback to default
vision_features = rearrange(vision_features_flat, '(b t) d -> b t d', b=B, t=T)
patch_features = rearrange(patch_tokens_flat, '(b t) p d -> b t p d', b=B, t=T)
# DEBUG: Analyze vision feature variability
# DEBUG: Analyze vision feature variability (use per-frame pooled features for readability)
if self.training and torch.rand(1).item() < 0.1: # 10% of training steps for more frequent debugging
with torch.no_grad():
print(f"\n🔍 DINOv3 VISION FEATURE DEBUG (B={B}, T={T}):")
@@ -395,7 +406,8 @@ class RLearNPolicy(PreTrainedPolicy):
else:
print(f" ✓ Batch samples have different first frames. Diff: {batch_first_frame_diff:.6f}")
# Check feature statistics
# Check feature statistics (pooled over patches)
vision_features = patch_features.mean(dim=2) # (B, T, D)
feature_mean = vision_features.mean().item()
feature_std = vision_features.std().item()
print(f"Feature stats: mean={feature_mean:.4f}, std={feature_std:.4f}")
@@ -440,7 +452,7 @@ class RLearNPolicy(PreTrainedPolicy):
print("=" * 50)
return vision_features
return patch_features
def _mask_from_lens(self, lens: Tensor) -> Tensor:
"""Create mask from sequence lengths."""
@@ -497,9 +509,9 @@ class RLearNPolicy(PreTrainedPolicy):
elif not isinstance(commands, list):
commands = [str(commands)] * B
# Process video frames through SigLIP2
# Process video frames through vision encoder (returns patch tokens)
vision_start = time.perf_counter()
video_embeds = self._encode_video_frames(frames).to(device) # (B, T_eff, D_vision)
video_patch_embeds = self._encode_video_frames(frames).to(device) # (B, T_eff, P, D_vision)
vision_time = time.perf_counter() - vision_start
# Language embeddings + mask
@@ -513,12 +525,18 @@ class RLearNPolicy(PreTrainedPolicy):
# Project embeddings
lang_tokens = self.to_lang_tokens(lang_embeds)
video_tokens = self.to_video_tokens(video_embeds)
video_tokens = self.to_video_tokens(video_patch_embeds) # (B, T, P, D)
# Add temporal positional encoding (window-relative only)
T_video = video_tokens.shape[1]
video_tokens = video_tokens + self.temporal_pos_embedding[:T_video]
# Add temporal + spatial positional encoding (window-relative only)
Bv, T_video, P_video, Dm = video_tokens.shape
if P_video > self.spatial_pos_embedding.shape[0]:
raise ValueError(f"Number of patch tokens {P_video} exceeds max_patch_tokens {self.spatial_pos_embedding.shape[0]}")
t_pos = self.temporal_pos_embedding[:T_video] # (T, D)
p_pos = self.spatial_pos_embedding[:P_video] # (P, D)
pos = t_pos[:, None, :] + p_pos[None, :, :] # (T, P, D)
video_tokens = video_tokens + pos
# Flatten patches into sequence tokens
video_tokens = rearrange(video_tokens, 'b t p d -> b (t p) d')
# Pack all tokens for attention [lang | register | video]
tokens, lang_video_packed_shape = pack((lang_tokens, register_tokens, video_tokens), 'b * d')
@@ -531,10 +549,11 @@ class RLearNPolicy(PreTrainedPolicy):
attended = self.decoder(tokens, mask=mask)
# Unpack and get video token features
_, _, attended_video_tokens = unpack(attended, lang_video_packed_shape, 'b * d')
# Process all frames with single MLP
frame_tokens = self.frame_mlp(attended_video_tokens) # (B, T, D)
_, _, attended_video_tokens = unpack(attended, lang_video_packed_shape, 'b * d') # (B, T*P, D)
# Restore (B, T, P, D) and pool patches per frame
attended_video_tokens = rearrange(attended_video_tokens, 'b (t p) d -> b t p d', t=T_video, p=P_video)
frame_tokens = attended_video_tokens.mean(dim=2) # (B, T, D)
frame_tokens = self.frame_mlp(frame_tokens)
# MLP predictor
video_frame_embeds = self.mlp_predictor(frame_tokens)