File size: 11,177 Bytes
dd9b4af | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 | import sys
sys.path.insert(0, "/mnt/steamdrive/openvla-micro")
sys.path.insert(0, "/home/the_one1/OmniVLA")
import math
from typing import Dict, List, Optional, Tuple, Union
import numpy as np
import timm
import torch
import torch.nn as nn
import torch.nn.functional as F
from PIL import Image
from timm.models.vision_transformer import VisionTransformer
from torchvision.transforms import Compose, Resize, Normalize
from transformers.modeling_outputs import CausalLMOutputWithPast
from modeling_openvla_micro import (
DinoSigLIPEncoder,
CombinedProjector,
ShimMLP,
unpack_tuple,
monkey_patch_featurizer,
_build_timm_transform,
)
class VisionBackboneWrapper:
"""Wrapper exposing API the training loop expects, using real encoder shapes."""
def __init__(self, encoder: DinoSigLIPEncoder):
self._encoder = encoder
self._num_images_in_input = 2
# Infer patch count from encoder output
device = next(encoder.parameters()).device
with torch.inference_mode():
dummy = torch.zeros(1, 3, 224, 224, device=device, dtype=next(encoder.parameters()).dtype)
dino_out = encoder.dino_featurizer(dummy)
if isinstance(dino_out, (list, tuple)):
dino_out = dino_out[0]
siglip_out = encoder.siglip_featurizer(dummy)
if isinstance(siglip_out, (list, tuple)):
siglip_out = siglip_out[0]
self._patches_per_img = dino_out.shape[1] + siglip_out.shape[1]
print(f"[VisionBackboneWrapper] patches_per_img = {self._patches_per_img}")
def get_num_patches(self) -> int:
return self._patches_per_img
def get_num_images_in_input(self) -> int:
return self._num_images_in_input
def set_num_images_in_input(self, n: int) -> None:
self._num_images_in_input = n
# SigLIP stats
SIGLIP_MEAN = torch.tensor([0.5, 0.5, 0.5]).reshape(1, 3, 1, 1)
SIGLIP_STD = torch.tensor([0.5, 0.5, 0.5]).reshape(1, 3, 1, 1)
# ImageNet stats (used by OpenVLA processor and DINOv2)
IMAGENET_MEAN = torch.tensor([0.485, 0.456, 0.406]).reshape(1, 3, 1, 1)
IMAGENET_STD = torch.tensor([0.229, 0.224, 0.225]).reshape(1, 3, 1, 1)
def _convert_pixel_values_for_siglip(pixel_values: torch.Tensor) -> torch.Tensor:
"""Undo ImageNet norm, apply SigLIP norm."""
device = pixel_values.device
m = IMAGENET_MEAN.to(device)
s = IMAGENET_STD.to(device)
sm = SIGLIP_MEAN.to(device)
ss = SIGLIP_STD.to(device)
return (pixel_values * s + m - sm) / ss
class OpenVLAMicroWrapper(nn.Module):
"""
Wraps openvla-micro (DinoSigLIPEncoder + CombinedProjector + Qwen2.5)
+ MLP hidden shim (896→2048→4096) into OmniVLA's forward interface.
Forward signature matches ``PrismaticForConditionalGeneration_MMNv1``
so that the existing training loop and ``run_forward_pass`` work as-is.
"""
def __init__(
self,
vision_encoder: DinoSigLIPEncoder,
projector: CombinedProjector,
llm: nn.Module,
hidden_shim: nn.Module,
tokenizer,
pad_token_id: int,
action_token_begin_idx: int = 151679,
):
super().__init__()
self.vision_encoder = vision_encoder
self.projector = projector
self.llm = llm
self.hidden_shim = hidden_shim
self.tokenizer = tokenizer
self.pad_token_id = pad_token_id
self.action_token_begin_idx = action_token_begin_idx
self.vision_backbone = VisionBackboneWrapper(vision_encoder)
self.llm_dim = 4096 # after hidden shim
self.vocab_size = llm.config.vocab_size
def get_input_embeddings(self) -> nn.Module:
if hasattr(self.llm, 'get_input_embeddings'):
return self.llm.get_input_embeddings()
if hasattr(self.llm, 'base_model') and hasattr(self.llm.base_model, 'model'):
return self.llm.base_model.model.embed_tokens
return self.llm.model.embed_tokens
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.Tensor] = None,
pixel_values: Optional[torch.FloatTensor] = None,
labels: Optional[torch.LongTensor] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
past_key_values: Optional[List[torch.FloatTensor]] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
# MoE/novelty projection (unused but accepted for compat)
proprio=None,
proprio_projector=None,
noisy_actions=None,
noisy_action_projector=None,
diffusion_timestep_embeddings=None,
use_film: bool = False,
# MMNv1-specific
attention_mask_label=None,
pixel_values_goal=None,
img_hist=None,
map_images=None,
obs_img=None,
modality_id=None,
goal_pose=None,
**kwargs,
) -> Union[Tuple, CausalLMOutputWithPast]:
_ = (attention_mask_label, pixel_values_goal, img_hist, map_images, obs_img, modality_id, goal_pose,
noisy_actions, noisy_action_projector, diffusion_timestep_embeddings, use_film)
output_hidden_states = output_hidden_states or False
return_dict = return_dict if return_dict is not None else True
use_cache = use_cache and not self.training
B = input_ids.shape[0]
# ---------- 1. Input embeddings ----------
embed_fn = self.get_input_embeddings()
input_embeds = embed_fn(input_ids) # (B, seq_len, 896)
input_embeds = input_embeds.to(self.llm.dtype)
# ---------- 2. Action mask ----------
if labels is not None:
all_actions_mask = self._action_mask(labels)
all_actions_mask_3d = all_actions_mask.unsqueeze(-1)
input_embeds = input_embeds * ~all_actions_mask_3d
else:
all_actions_mask = None
# ---------- 3. Vision features ----------
if pixel_values is not None:
num_patches_per_img = self.vision_backbone.get_num_patches()
num_imgs = pixel_values.shape[1] // 3 # 6 channels → 2 images
patch_feats = []
for i in range(num_imgs):
img = pixel_values[:, i*3:(i+1)*3, :, :] # (B, 3, 224, 224)
feats = self._encode_image(img)
patch_feats.append(feats)
projected = torch.cat(patch_feats, dim=1) # (B, N*num_imgs, 896)
# Optionally append proprio
if proprio_projector is not None and proprio is not None:
proprio = proprio.reshape(B, -1)
proprio_feats = proprio_projector(proprio).unsqueeze(1)
proprio_feats = proprio_feats.to(self.llm.dtype)
projected = torch.cat([projected, proprio_feats], dim=1)
else:
projected = None
# Cast projected features to LLM dtype
if projected is not None:
projected = projected.to(self.llm.dtype)
# ---------- 4. Build multimodal embeddings ----------
if projected is not None:
multimodal_embeds = torch.cat(
[input_embeds[:, :1, :], projected, input_embeds[:, 1:, :]], dim=1
)
# Build attention mask
if attention_mask is not None:
vis_mask = torch.full(
(B, projected.shape[1]), True,
dtype=attention_mask.dtype, device=attention_mask.device,
)
multimodal_attn_mask = torch.cat(
[attention_mask[:, :1], vis_mask, attention_mask[:, 1:]], dim=1
)
else:
multimodal_attn_mask = None
else:
multimodal_embeds = input_embeds
multimodal_attn_mask = attention_mask
# ---------- 5. Run LLM ----------
llm_out = self.llm(
input_ids=None,
attention_mask=multimodal_attn_mask,
inputs_embeds=multimodal_embeds,
labels=None,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=True,
)
# ---------- 6. Apply hidden shim ----------
if output_hidden_states and llm_out.hidden_states is not None:
shimmed = tuple(
self.hidden_shim(hs) if i == len(llm_out.hidden_states) - 1 else hs
for i, hs in enumerate(llm_out.hidden_states)
)
hidden = shimmed
else:
hidden = llm_out.hidden_states
# ---------- 7. Return ----------
if not return_dict:
return llm_out.logits, hidden, llm_out.past_key_values
return CausalLMOutputWithPast(
loss=llm_out.loss,
logits=llm_out.logits,
past_key_values=llm_out.past_key_values,
hidden_states=hidden,
attentions=llm_out.attentions,
)
def _action_mask(self, labels: torch.Tensor) -> torch.Tensor:
"""Same logic as PrismaticForConditionalGeneration_MMNv1."""
BEGIN = self.action_token_begin_idx
current_mask = (labels >= BEGIN) & (labels < BEGIN + 256)
next_mask = torch.zeros_like(labels, dtype=torch.bool)
for b in range(labels.shape[0]):
action_positions = torch.where(current_mask[b])[0]
if len(action_positions) == 0:
continue
first, last = action_positions[0], action_positions[-1]
if last + 1 < labels.shape[1] and labels[b, last + 1] < BEGIN:
next_mask[b, last + 1] = True
return current_mask | next_mask
def _encode_image(self, img: torch.Tensor) -> torch.Tensor:
"""
Encode a single (B, 3, 224, 224) image through DinoSigLIPEncoder.
Pixel values are assumed to be normalized with ImageNet stats.
"""
B = img.shape[0]
device = img.device
dtype = img.dtype
dino_out = self.vision_encoder.dino_featurizer(img)
if isinstance(dino_out, (list, tuple)):
dino_out = dino_out[0]
# SigLIP needs different normalization
siglip_input = _convert_pixel_values_for_siglip(img)
siglip_out = self.vision_encoder.siglip_featurizer(siglip_input)
if isinstance(siglip_out, (list, tuple)):
siglip_out = siglip_out[0]
# Zero-pad to joint dim and concat along sequence
D_total = self.vision_encoder.total_embed_dim # 1152
dino_padded = torch.zeros(B, dino_out.shape[1], D_total, device=device, dtype=dtype)
dino_padded[:, :, :dino_out.shape[-1]] = dino_out
siglip_padded = torch.zeros(B, siglip_out.shape[1], D_total, device=device, dtype=dtype)
siglip_padded[:, :, :siglip_out.shape[-1]] = siglip_out
combined = torch.cat([dino_padded, siglip_padded], dim=1) # (B, 458, 1152)
projected = self.projector(combined) # (B, 458, 896)
return projected
|