File size: 15,839 Bytes
4f07533 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 | # Copyright 2025 TeleAI Rhodes Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Processor for PRTS built on Qwen3-VL (hub / trust_remote_code; no prts package required)."""
from __future__ import annotations
import logging
from typing import Optional, Union
import numpy as np
import torch
from transformers.feature_extraction_utils import BatchFeature
from transformers.image_utils import ImageInput
from transformers.processing_utils import (
ImagesKwargs,
MultiModalData,
ProcessingKwargs,
ProcessorMixin,
Unpack,
VideosKwargs,
)
from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
from transformers.utils.logging import get_logger
from transformers.video_utils import VideoInput
ACTION_START_TOKEN = "<|action_start|>"
ACTION_PLACEHOLDER_TOKEN = "<|action_pad|>"
ACTION_END_TOKEN = "<|action_end|>"
CRL_GOAL_REPR_TOKEN = "<|goal_repr|>"
CRL_OBS_REPR_TOKEN = "<|obs_repr|>"
VISION_START_TOKEN = "<|vision_start|>" # beginning of vision input
IMAGE_PLACEHOLDER_TOKEN = "<|image_pad|>" # image placeholder
VIDEO_PLACEHOLDER_TOKEN = "<|video_pad|>" # video placeholder
logger = get_logger(__name__)
if not logger.handlers:
handler = logging.StreamHandler()
handler.setLevel(logging.INFO)
handler.setFormatter(logging.Formatter("%(levelname)s:%(name)s:%(message)s"))
logger.addHandler(handler)
class Qwen3VLVideosProcessorKwargs(VideosKwargs, total=False):
pass
class Qwen3VLImagesKwargs(ImagesKwargs):
min_pixels: Optional[int]
max_pixels: Optional[int]
patch_size: Optional[int]
temporal_patch_size: Optional[int]
merge_size: Optional[int]
class Qwen3VLProcessorKwargs(ProcessingKwargs, total=False):
images_kwargs: Qwen3VLImagesKwargs
videos_kwargs: Qwen3VLVideosProcessorKwargs
_defaults = {
"text_kwargs": {
"padding": False,
"return_token_type_ids": False,
"return_mm_token_type_ids": False,
},
"videos_kwargs": {"return_metadata": True},
}
class PRTS_Qwen3VLProcessor(ProcessorMixin):
r"""
Constructs a PRTS processor which wraps a Qwen3-VL image processor and a Qwen2 tokenizer into a single processor.
This processor is built independently (not inheriting from Qwen3VLProcessor) to avoid tight coupling,
while maintaining compatibility with Qwen3-VL's timestamp-based video processing approach.
[`PRTS_Qwen3VLProcessor`] offers all the functionalities needed for PRTS model with:
- Action token handling (discrete and continuous)
- State token handling for proprioceptive inputs
- Expert trigger tokens for flow matching action prediction
- Qwen3-VL compatible image/video processing with timestamp-based video handling
Args:
image_processor ([`Qwen2VLImageProcessor`], *optional*):
The image processor is a required input.
tokenizer ([`Qwen2TokenizerFast`], *optional*):
The tokenizer is a required input.
video_processor ([`Qwen3VLVideoProcessor`], *optional*):
The video processor is a required input.
chat_template (`str`, *optional*):
A Jinja template which will be used to convert lists of messages in a chat into a tokenizable string.
"""
attributes = ["image_processor", "tokenizer", "video_processor"]
image_processor_class = "AutoImageProcessor"
video_processor_class = "AutoVideoProcessor"
tokenizer_class = ("Qwen2Tokenizer", "Qwen2TokenizerFast")
def __init__(self, image_processor=None, tokenizer=None, video_processor=None,
chat_template=None, **kwargs):
# Initialize base ProcessorMixin
super().__init__(image_processor, tokenizer, video_processor, chat_template=chat_template)
# Get image/video tokens from tokenizer
self.image_token = "<|image_pad|>" if not hasattr(tokenizer, "image_token") else tokenizer.image_token
self.video_token = "<|video_pad|>" if not hasattr(tokenizer, "video_token") else tokenizer.video_token
self.image_token_id = (
tokenizer.image_token_id
if getattr(tokenizer, "image_token_id", None)
else tokenizer.convert_tokens_to_ids(self.image_token)
)
self.video_token_id = (
tokenizer.video_token_id
if getattr(tokenizer, "video_token_id", None)
else tokenizer.convert_tokens_to_ids(self.video_token)
)
# Qwen3-VL vision tokens
self.vision_start_token = (
"<|vision_start|>" if not hasattr(tokenizer, "vision_start_token") else tokenizer.vision_start_token
)
self.vision_end_token = (
"<|vision_end|>" if not hasattr(tokenizer, "vision_end_token") else tokenizer.vision_end_token
)
self.vision_start_token_id = (
tokenizer.vision_start_token_id
if getattr(tokenizer, "vision_start_token_id", None)
else tokenizer.convert_tokens_to_ids(self.vision_start_token)
)
self.vision_end_token_id = (
tokenizer.vision_end_token_id
if getattr(tokenizer, "vision_end_token_id", None)
else tokenizer.convert_tokens_to_ids(self.vision_end_token)
)
prts_special_tokens = [
ACTION_START_TOKEN,
ACTION_PLACEHOLDER_TOKEN,
ACTION_END_TOKEN,
CRL_GOAL_REPR_TOKEN,
CRL_OBS_REPR_TOKEN,
]
num_new_tokens = tokenizer.add_tokens(prts_special_tokens, special_tokens=True)
logger.info(f"Added {num_new_tokens} new special tokens to the tokenizer.")
self.action_token = getattr(tokenizer, "action_token", ACTION_PLACEHOLDER_TOKEN)
self.action_token_id = tokenizer.convert_tokens_to_ids(self.action_token)
token_dict = {
"action_start_token_id": ACTION_START_TOKEN,
"action_token_id": ACTION_PLACEHOLDER_TOKEN,
"vision_start_token_id": VISION_START_TOKEN,
"image_token_id": IMAGE_PLACEHOLDER_TOKEN,
"video_token_id": VIDEO_PLACEHOLDER_TOKEN,
"crl_goal_repr_token_id": CRL_GOAL_REPR_TOKEN,
"crl_obs_repr_token_id": CRL_OBS_REPR_TOKEN,
}
self.token_ids = {key: tokenizer.convert_tokens_to_ids(value) for key, value in token_dict.items()}
def __call__(
self,
images: Optional[ImageInput] = None,
text: Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]] = None,
videos: Optional[VideoInput] = None,
actions: Union[torch.Tensor] = None,
**kwargs: Unpack[Qwen3VLProcessorKwargs],
) -> BatchFeature:
output_kwargs = self._merge_kwargs(
Qwen3VLProcessorKwargs,
tokenizer_init_kwargs=self.tokenizer.init_kwargs,
**kwargs,
)
image_inputs = {}
if images is not None:
image_inputs = self.image_processor(images=images, **output_kwargs["images_kwargs"])
image_grid_thw = image_inputs["image_grid_thw"]
else:
image_grid_thw = None
videos_inputs = {}
if videos is not None:
videos_inputs = self.video_processor(videos=videos, **output_kwargs["videos_kwargs"])
video_grid_thw = videos_inputs["video_grid_thw"]
if "return_metadata" not in kwargs:
video_metadata = videos_inputs.pop("video_metadata", None)
else:
video_metadata = videos_inputs.get("video_metadata", None)
else:
video_grid_thw = None
video_metadata = None
if not isinstance(text, list):
text = [text]
text = text.copy()
if image_grid_thw is not None:
merge_length = self.image_processor.merge_size**2
index = 0
for i in range(len(text)):
while self.image_token in text[i]:
num_image_tokens = image_grid_thw[index].prod() // merge_length
text[i] = text[i].replace(self.image_token, "<|placeholder|>" * num_image_tokens, 1)
index += 1
text[i] = text[i].replace("<|placeholder|>", self.image_token)
if video_grid_thw is not None:
merge_length = self.video_processor.merge_size**2
index = 0
for i in range(len(text)):
while self.video_token in text[i]:
if video_metadata is not None and index < len(video_metadata):
metadata = video_metadata[index]
if metadata.fps is None:
logger.warning_once(
"Qwen3VL requires frame timestamps to construct prompts, but the `fps` of the input video could not be inferred. "
"Probably `video_metadata` was missing from inputs and you passed pre-sampled frames. "
"Defaulting to `fps=24`. Please provide `video_metadata` for more accurate results."
)
metadata.fps = 24 if metadata.fps is None else metadata.fps
curr_timestamp = self._calculate_timestamps(
metadata.frames_indices,
metadata.fps,
self.video_processor.merge_size,
)
video_placeholder = ""
frame_seqlen = video_grid_thw[index][1:].prod() // merge_length
for frame_idx in range(video_grid_thw[index][0]):
curr_time = curr_timestamp[frame_idx]
video_placeholder += f"<{curr_time:.1f} seconds>"
video_placeholder += (
self.vision_start_token + "<|placeholder|>" * frame_seqlen + self.vision_end_token
)
if f"{self.vision_start_token}{self.video_token}{self.vision_end_token}" in text[i]:
text[i] = text[i].replace(
f"{self.vision_start_token}{self.video_token}{self.vision_end_token}",
video_placeholder,
1,
)
else:
text[i] = text[i].replace(self.video_token, video_placeholder, 1)
else:
num_video_tokens = video_grid_thw[index].prod() // merge_length
text[i] = text[i].replace(self.video_token, "<|placeholder|>" * num_video_tokens, 1)
index += 1
text[i] = text[i].replace("<|placeholder|>", self.video_token)
return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
return_mm_token_type_ids = output_kwargs["text_kwargs"].pop("return_mm_token_type_ids", None)
text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
self._check_special_mm_tokens(text, text_inputs, modalities=["image", "video"])
if return_mm_token_type_ids:
array_ids = np.array(text_inputs["input_ids"])
mm_token_type_ids = np.zeros_like(text_inputs["input_ids"])
mm_token_type_ids[array_ids == self.image_token_id] = 1
text_inputs["mm_token_type_ids"] = mm_token_type_ids.tolist()
output_data = {**text_inputs, **image_inputs, **videos_inputs}
if actions is not None:
output_data["actions"] = actions
return BatchFeature(data=output_data, tensor_type=return_tensors)
def _calculate_timestamps(self, indices: Union[list[int], np.ndarray], video_fps: float, merge_size: int = 2):
if not isinstance(indices, list):
indices = indices.tolist()
if len(indices) % merge_size != 0:
indices.extend(indices[-1] for _ in range(merge_size - len(indices) % merge_size))
timestamps = [idx / video_fps for idx in indices]
timestamps = [
(timestamps[i] + timestamps[i + merge_size - 1]) / 2 for i in range(0, len(timestamps), merge_size)
]
return timestamps
def _get_num_multimodal_tokens(self, image_sizes=None, video_sizes=None, **kwargs):
vision_data = {}
if image_sizes is not None:
images_kwargs = Qwen3VLProcessorKwargs._defaults.get("images_kwargs", {})
images_kwargs.update(kwargs)
merge_size = images_kwargs.get("merge_size", None) or self.image_processor.merge_size
num_image_patches = [
self.image_processor.get_number_of_image_patches(*image_size, images_kwargs)
for image_size in image_sizes
]
num_image_tokens = [(num_patches // merge_size**2) for num_patches in num_image_patches]
vision_data.update({"num_image_tokens": num_image_tokens, "num_image_patches": num_image_patches})
if video_sizes is not None:
videos_kwargs = Qwen3VLProcessorKwargs._defaults.get("videos_kwargs", {})
videos_kwargs.update(kwargs)
merge_size = videos_kwargs.get("merge_size", None) or self.video_processor.merge_size
num_video_patches = [
self.video_processor.get_number_of_video_patches(*video_size, videos_kwargs)
for video_size in video_sizes
]
num_video_tokens = [(num_patches // merge_size**2) for num_patches in num_video_patches]
vision_data["num_video_tokens"] = num_video_tokens
return MultiModalData(**vision_data)
def set_action_tokenizer(self, action_tokenizer):
self.action_tokenizer = action_tokenizer
prts_fast_action_tokens = [f"<|action_token_{i}|>" for i in range(action_tokenizer.vocab_size)]
num_new_tokens = self.tokenizer.add_tokens(prts_fast_action_tokens, special_tokens=True)
logger.info(f"Added {num_new_tokens} FAST action tokens to the tokenizer.")
self.action_token_start_index = self.tokenizer.convert_tokens_to_ids("<|action_token_0|>")
self.action_vocab_size = action_tokenizer.vocab_size
token_ids = self.tokenizer.convert_tokens_to_ids(prts_fast_action_tokens)
self.action_mapper = {k: v for k, v in zip(prts_fast_action_tokens, token_ids, strict=True)}
def preprocess_action(self, actions, **kwargs):
raise NotImplementedError
def post_process_image_text_to_text(
self, generated_outputs, skip_special_tokens=True, clean_up_tokenization_spaces=False, **kwargs
):
return self.tokenizer.batch_decode(
generated_outputs,
skip_special_tokens=skip_special_tokens,
clean_up_tokenization_spaces=clean_up_tokenization_spaces,
**kwargs,
)
@property
def model_input_names(self):
tokenizer_input_names = self.tokenizer.model_input_names
image_processor_input_names = self.image_processor.model_input_names
return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
PRTS_Qwen3VLProcessor.register_for_auto_class()
__all__ = ["PRTS_Qwen3VLProcessor"]
|