add dexvla
This commit is contained in:
lesjie-wen
parent
fe483b1d0d
commit
1aa4f0c086
|
|
@ -0,0 +1,270 @@
|
|||
# coding=utf-8
|
||||
# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. 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.
|
||||
"""Qwen2VL model configuration"""
|
||||
|
||||
import os
|
||||
from typing import Union
|
||||
|
||||
from transformers.configuration_utils import PretrainedConfig
|
||||
from transformers.modeling_rope_utils import rope_config_validation
|
||||
from transformers.utils import logging
|
||||
from transformers import AutoModel, AutoConfig
|
||||
|
||||
logger = logging.get_logger(__name__)
|
||||
|
||||
|
||||
class Qwen2VLAVisionConfig(PretrainedConfig):
|
||||
model_type = "dex_vla"
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
depth=32,
|
||||
embed_dim=1280,
|
||||
hidden_size=3584,
|
||||
hidden_act="quick_gelu",
|
||||
mlp_ratio=4,
|
||||
num_heads=16,
|
||||
in_channels=3,
|
||||
patch_size=14,
|
||||
spatial_merge_size=2,
|
||||
temporal_patch_size=2,
|
||||
**kwargs,
|
||||
):
|
||||
super().__init__(**kwargs)
|
||||
|
||||
self.depth = depth
|
||||
self.embed_dim = embed_dim
|
||||
self.hidden_size = hidden_size
|
||||
self.hidden_act = hidden_act
|
||||
self.mlp_ratio = mlp_ratio
|
||||
self.num_heads = num_heads
|
||||
self.in_channels = in_channels
|
||||
self.patch_size = patch_size
|
||||
self.spatial_merge_size = spatial_merge_size
|
||||
self.temporal_patch_size = temporal_patch_size
|
||||
|
||||
@classmethod
|
||||
def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
|
||||
cls._set_token_in_kwargs(kwargs)
|
||||
|
||||
config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
|
||||
|
||||
if config_dict.get("model_type") == "qwen2_vl":
|
||||
config_dict = config_dict["vision_config"]
|
||||
|
||||
if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
|
||||
logger.warning(
|
||||
f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
|
||||
f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
|
||||
)
|
||||
|
||||
return cls.from_dict(config_dict, **kwargs)
|
||||
|
||||
|
||||
class DexVLAConfig(PretrainedConfig):
|
||||
r"""
|
||||
This is the configuration class to store the configuration of a [`Qwen2VLModel`]. It is used to instantiate a
|
||||
Qwen2-VL model according to the specified arguments, defining the model architecture. Instantiating a configuration
|
||||
with the defaults will yield a similar configuration to that of
|
||||
Qwen2-VL-7B-Instruct [Qwen/Qwen2-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct).
|
||||
|
||||
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
||||
documentation from [`PretrainedConfig`] for more information.
|
||||
|
||||
|
||||
Args:
|
||||
vocab_size (`int`, *optional*, defaults to 152064):
|
||||
Vocabulary size of the Qwen2VL model. Defines the number of different tokens that can be represented by the
|
||||
`inputs_ids` passed when calling [`Qwen2VLModel`]
|
||||
hidden_size (`int`, *optional*, defaults to 8192):
|
||||
Dimension of the hidden representations.
|
||||
intermediate_size (`int`, *optional*, defaults to 29568):
|
||||
Dimension of the MLP representations.
|
||||
num_hidden_layers (`int`, *optional*, defaults to 80):
|
||||
Number of hidden layers in the Transformer encoder.
|
||||
num_attention_heads (`int`, *optional*, defaults to 64):
|
||||
Number of attention heads for each attention layer in the Transformer encoder.
|
||||
num_key_value_heads (`int`, *optional*, defaults to 8):
|
||||
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
|
||||
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
|
||||
`num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
|
||||
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
|
||||
by meanpooling all the original heads within that group. For more details checkout [this
|
||||
paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
|
||||
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
|
||||
The non-linear activation function (function or string) in the decoder.
|
||||
max_position_embeddings (`int`, *optional*, defaults to 32768):
|
||||
The maximum sequence length that this model might ever be used with.
|
||||
initializer_range (`float`, *optional*, defaults to 0.02):
|
||||
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
||||
rms_norm_eps (`float`, *optional*, defaults to 1e-05):
|
||||
The epsilon used by the rms normalization layers.
|
||||
use_cache (`bool`, *optional*, defaults to `True`):
|
||||
Whether or not the model should return the last key/values attentions (not used by all models). Only
|
||||
relevant if `config.is_decoder=True`.
|
||||
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
||||
Whether the model's input and output word embeddings should be tied.
|
||||
rope_theta (`float`, *optional*, defaults to 1000000.0):
|
||||
The base period of the RoPE embeddings.
|
||||
use_sliding_window (`bool`, *optional*, defaults to `False`):
|
||||
Whether to use sliding window attention.
|
||||
sliding_window (`int`, *optional*, defaults to 4096):
|
||||
Sliding window attention (SWA) window size. If not specified, will default to `4096`.
|
||||
max_window_layers (`int`, *optional*, defaults to 80):
|
||||
The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
|
||||
attention_dropout (`float`, *optional*, defaults to 0.0):
|
||||
The dropout ratio for the attention probabilities.
|
||||
vision_config (`Dict`, *optional*):
|
||||
The config for the visual encoder initialization.
|
||||
rope_scaling (`Dict`, *optional*):
|
||||
Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
|
||||
and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
|
||||
accordingly.
|
||||
Expected contents:
|
||||
`rope_type` (`str`):
|
||||
The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
|
||||
'llama3'], with 'default' being the original RoPE implementation.
|
||||
`factor` (`float`, *optional*):
|
||||
Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
|
||||
most scaling types, a `factor` of x will enable the model to handle sequences of length x *
|
||||
original maximum pre-trained length.
|
||||
`original_max_position_embeddings` (`int`, *optional*):
|
||||
Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
|
||||
pretraining.
|
||||
`attention_factor` (`float`, *optional*):
|
||||
Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
|
||||
computation. If unspecified, it defaults to value recommended by the implementation, using the
|
||||
`factor` field to infer the suggested value.
|
||||
`beta_fast` (`float`, *optional*):
|
||||
Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
|
||||
ramp function. If unspecified, it defaults to 32.
|
||||
`beta_slow` (`float`, *optional*):
|
||||
Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
|
||||
ramp function. If unspecified, it defaults to 1.
|
||||
`short_factor` (`List[float]`, *optional*):
|
||||
Only used with 'longrope'. The scaling factor to be applied to short contexts (<
|
||||
`original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
|
||||
size divided by the number of attention heads divided by 2
|
||||
`long_factor` (`List[float]`, *optional*):
|
||||
Only used with 'longrope'. The scaling factor to be applied to long contexts (<
|
||||
`original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
|
||||
size divided by the number of attention heads divided by 2
|
||||
`low_freq_factor` (`float`, *optional*):
|
||||
Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
|
||||
`high_freq_factor` (`float`, *optional*):
|
||||
Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
|
||||
|
||||
```python
|
||||
>>> from transformers import Qwen2VLForConditionalGeneration, Qwen2VLConfig
|
||||
|
||||
>>> # Initializing a Qwen2VL style configuration
|
||||
>>> configuration = Qwen2VLConfig()
|
||||
|
||||
>>> # Initializing a model from the Qwen2-VL-7B style configuration
|
||||
>>> model = Qwen2VLForConditionalGeneration(configuration)
|
||||
|
||||
>>> # Accessing the model configuration
|
||||
>>> configuration = model.config
|
||||
```"""
|
||||
|
||||
model_type = "qwen2_vla"
|
||||
keys_to_ignore_at_inference = ["past_key_values"]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocab_size=152064,
|
||||
hidden_size=8192,
|
||||
intermediate_size=29568,
|
||||
num_hidden_layers=80,
|
||||
num_attention_heads=64,
|
||||
num_key_value_heads=8,
|
||||
hidden_act="silu",
|
||||
max_position_embeddings=32768,
|
||||
initializer_range=0.02,
|
||||
rms_norm_eps=1e-05,
|
||||
use_cache=True,
|
||||
tie_word_embeddings=False,
|
||||
rope_theta=1000000.0,
|
||||
use_sliding_window=False,
|
||||
sliding_window=4096,
|
||||
max_window_layers=80,
|
||||
attention_dropout=0.0,
|
||||
vision_config=None,
|
||||
rope_scaling=None,
|
||||
# For loading policy head
|
||||
policy_head_type='dit_diffusion_policy', # dit_diffusion_policy
|
||||
policy_head_size='DiT_L',
|
||||
action_dim=10,
|
||||
state_dim=7,
|
||||
non_lora_lr=1e-4,
|
||||
**kwargs,
|
||||
):
|
||||
if isinstance(vision_config, dict):
|
||||
self.vision_config = Qwen2VLAVisionConfig(**vision_config)
|
||||
elif vision_config is None:
|
||||
self.vision_config = Qwen2VLAVisionConfig()
|
||||
|
||||
self.vocab_size = vocab_size
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.hidden_size = hidden_size
|
||||
self.intermediate_size = intermediate_size
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.use_sliding_window = use_sliding_window
|
||||
self.sliding_window = sliding_window
|
||||
self.max_window_layers = max_window_layers
|
||||
|
||||
# for loading policy head
|
||||
self.policy_head_type = policy_head_type
|
||||
# if policy_head_type == 'dit_diffusion_policy':
|
||||
# # self.policy_head_size = kwargs.get("policy_head_size", "none")
|
||||
# self.policy_head_size = policy_head_size
|
||||
# # self.policy_head_config = register_configuration_class(self.policy_head_type, model_size=policy_head_size)
|
||||
# self.policy_head_config = AutoConfig.for_model(model_type=self.policy_head_type,
|
||||
# model_size=self.policy_head_size,
|
||||
# global_cond_dim=hidden_size, action_dim=action_dim,
|
||||
# state_dim=state_dim)
|
||||
# elif policy_head_type == 'unet_diffusion_policy':
|
||||
# self.policy_head_config = AutoConfig.for_model(model_type=self.policy_head_type,
|
||||
# global_cond_dim=hidden_size, action_dim=action_dim,
|
||||
# state_dim=state_dim)
|
||||
# for backward compatibility
|
||||
if num_key_value_heads is None:
|
||||
num_key_value_heads = num_attention_heads
|
||||
|
||||
self.num_key_value_heads = num_key_value_heads
|
||||
self.hidden_act = hidden_act
|
||||
self.initializer_range = initializer_range
|
||||
self.rms_norm_eps = rms_norm_eps
|
||||
self.use_cache = use_cache
|
||||
self.rope_theta = rope_theta
|
||||
self.attention_dropout = attention_dropout
|
||||
self.rope_scaling = rope_scaling
|
||||
|
||||
# Validate the correctness of rotary position embeddings parameters
|
||||
# BC: if there is a 'type' field, move it to 'rope_type'.
|
||||
# and change type from 'mrope' to 'default' because `mrope` does defeault RoPE calculations
|
||||
# one can set it to "linear"/"dynamic" etc. to have scaled RoPE
|
||||
# TODO: @raushan update config in the hub
|
||||
if self.rope_scaling is not None and "type" in self.rope_scaling:
|
||||
if self.rope_scaling["type"] == "mrope":
|
||||
self.rope_scaling["type"] = "default"
|
||||
self.rope_scaling["rope_type"] = self.rope_scaling["type"]
|
||||
rope_config_validation(self, ignore_keys={"mrope_section"})
|
||||
|
||||
super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
|
||||
|
||||
from transformers import AutoConfig
|
||||
AutoConfig.register("dex_vla", DexVLAConfig)
|
||||
|
|
@ -0,0 +1,40 @@
|
|||
import torch.nn as nn
|
||||
|
||||
class ActionProjector(nn.Module):
|
||||
def __init__(self, in_dim, out_dim=1024):
|
||||
super(ActionProjector, self).__init__()
|
||||
self.global_1d_pool = nn.AdaptiveAvgPool1d(1)
|
||||
self.mlps = nn.ModuleList([
|
||||
# nn.LayerNorm(in_dim),
|
||||
nn.Linear(in_dim, in_dim),
|
||||
nn.GELU(),
|
||||
nn.Linear(in_dim, out_dim),
|
||||
nn.Dropout(0.0),
|
||||
]
|
||||
)
|
||||
|
||||
def forward(self, x):
|
||||
x = self.global_1d_pool(x.permute(1, 0)).permute(1, 0)
|
||||
for mlp in self.mlps:
|
||||
x = mlp(x)
|
||||
return x
|
||||
|
||||
|
||||
class FiLM(nn.Module):
|
||||
def __init__(self, feature_dim, condition_dim):
|
||||
super(FiLM, self).__init__()
|
||||
self.scale_fc = nn.Linear(condition_dim, feature_dim)
|
||||
self.shift_fc = nn.Linear(condition_dim, feature_dim)
|
||||
|
||||
nn.init.zeros_(self.scale_fc.weight)
|
||||
nn.init.zeros_(self.scale_fc.bias)
|
||||
nn.init.zeros_(self.shift_fc.weight)
|
||||
nn.init.zeros_(self.shift_fc.bias)
|
||||
|
||||
def forward(self, x, condition):
|
||||
# 计算缩放和偏移参数
|
||||
scale = self.scale_fc(condition)
|
||||
shift = self.shift_fc(condition)
|
||||
|
||||
# 应用 FiLM 调制
|
||||
return x * (1 + scale) + shift
|
||||
File diff suppressed because it is too large
Load Diff
|
|
@ -0,0 +1,151 @@
|
|||
from PIL import Image
|
||||
import numpy as np
|
||||
from torchvision.transforms.functional import to_pil_image, to_tensor
|
||||
import torchvision.transforms as transforms
|
||||
import torch
|
||||
from qwen_vl_utils import process_vision_info
|
||||
from qwen_vl_utils import fetch_image
|
||||
class Qwen2VLAProcess:
|
||||
def __init__(
|
||||
self,
|
||||
language=None,
|
||||
tokenizer=None,
|
||||
max_seq_len=512,
|
||||
multimodal_processor=None,
|
||||
camera_names=None,
|
||||
data_args=None,
|
||||
):
|
||||
super().__init__()
|
||||
self.tokenizer = tokenizer
|
||||
self.max_seq_len = max_seq_len
|
||||
self.camera_names = camera_names
|
||||
# self.language = language
|
||||
self.multimodal_processor = multimodal_processor
|
||||
self.data_args = data_args
|
||||
|
||||
def preprocess_image(self, image, size=224):
|
||||
# Model has been trained to handle images of different aspects ratios
|
||||
# resized to 224x224 in the range [-1, 1]. Bilinear and antialias resize
|
||||
# options are helpful to improve quality in some tasks.
|
||||
image = np.asarray(image)
|
||||
if image.ndim == 2: # Convert image without last channel into greyscale.
|
||||
image = np.stack((image,) * 3, axis=-1)
|
||||
image = image[..., :3] # Remove alpha layer.
|
||||
assert image.shape[-1] == 3
|
||||
|
||||
image_pil = to_pil_image(image)
|
||||
|
||||
# Step 2: Define the resize transformation
|
||||
resize_transform = transforms.Resize((size, size), interpolation=transforms.InterpolationMode.BILINEAR)
|
||||
|
||||
# Step 3: Apply the resize transformation
|
||||
image_resized_pil = resize_transform(image_pil)
|
||||
|
||||
# Step 4: Convert back to tensor if needed
|
||||
image_resized = to_tensor(image_resized_pil)
|
||||
return image.numpy() / 127.5 - 1.0 # [0, 255]->[-1,1]
|
||||
|
||||
def qwen2_image_preprocess(self, each, camera_name):
|
||||
ele = {}
|
||||
each = Image.fromarray(each.squeeze(0).permute(1, 2, 0).numpy().astype(np.uint8))
|
||||
ele['image'] = each
|
||||
if 'wrist' in camera_name:
|
||||
w, h = eval(self.data_args.image_size_wrist)
|
||||
ele['resized_height'] = h
|
||||
ele['resized_width'] = w
|
||||
else:
|
||||
ele['resized_height'] = each.height
|
||||
ele['resized_width'] = each.width
|
||||
each = fetch_image(ele)
|
||||
return torch.from_numpy(np.array(each))
|
||||
|
||||
def forward_process(self, sample, use_reasoning=True):
|
||||
if sample['image'].ndim == 5 and sample['image'].shape[1] > 2:
|
||||
video = True
|
||||
else:
|
||||
video = False
|
||||
messages = self.datastruct_droid2llava(sample, video=video)
|
||||
|
||||
data_dict = dict(
|
||||
messages=messages,
|
||||
images=None
|
||||
)
|
||||
|
||||
image_data = torch.chunk(sample['image'], sample['image'].shape[0], 0)
|
||||
|
||||
images_list = []
|
||||
|
||||
for i, each in enumerate(image_data):
|
||||
if each.ndim == 4:
|
||||
img_pil = self.qwen2_image_preprocess(each, self.camera_names[i])
|
||||
else:
|
||||
img_pil = []
|
||||
for temp in each.squeeze(0):
|
||||
img_pil.append(self.qwen2_image_preprocess(temp, self.camera_names[i]))
|
||||
img_pil = torch.stack(img_pil, 0)
|
||||
images_list.append(img_pil)
|
||||
# TODO RESIZE
|
||||
# image_data = image_data / 255.0
|
||||
if video:
|
||||
image_data = None
|
||||
video_inputs = images_list
|
||||
else:
|
||||
image_data = images_list
|
||||
video_inputs = None
|
||||
|
||||
text = self.multimodal_processor.apply_chat_template(
|
||||
messages, tokenize=False, add_generation_prompt=True
|
||||
)
|
||||
|
||||
model_inputs = self.multimodal_processor(
|
||||
text=text,
|
||||
images=image_data,
|
||||
videos=video_inputs,
|
||||
padding=True,
|
||||
return_tensors="pt",
|
||||
)
|
||||
input_labels = torch.ones_like(model_inputs['input_ids']) * -100
|
||||
if use_reasoning:
|
||||
answer = sample['reasoning'] + "Next action:" + '<|im_end|>'
|
||||
else:
|
||||
answer = '' + '<|im_end|>'
|
||||
|
||||
output_text = self.tokenizer(answer, padding=True, return_tensors="pt")
|
||||
output_labels = output_text['input_ids']
|
||||
model_inputs['input_ids'] = torch.cat((model_inputs['input_ids'], output_text['input_ids']), dim=-1)
|
||||
model_inputs['attention_mask'] = torch.cat((model_inputs['attention_mask'], output_text['attention_mask']), dim=-1)
|
||||
labels = torch.cat((input_labels, output_labels), dim=-1)
|
||||
data_dict['state'] = sample['state']
|
||||
data_dict['action'] = sample['action']
|
||||
data_dict['is_pad'] = sample['is_pad']
|
||||
data_dict['labels'] = labels
|
||||
for k, v in model_inputs.items():
|
||||
data_dict[k] = v
|
||||
return data_dict
|
||||
|
||||
def datastruct_droid2llava(self, sample, video=False):
|
||||
len_image = sample['image'].shape[0]
|
||||
|
||||
messages = [
|
||||
{
|
||||
"role": "user",
|
||||
"content": [],
|
||||
},
|
||||
# {"role": "assistant", "content": f''},
|
||||
]
|
||||
|
||||
for i in range(len_image):
|
||||
if video:
|
||||
messages[0]['content'].append({
|
||||
"type": "video",
|
||||
"video": None,
|
||||
})
|
||||
else:
|
||||
messages[0]['content'].append({
|
||||
"type": "image",
|
||||
"image": None,
|
||||
})
|
||||
messages[0]['content'].append({"type": "text", "text": f""})
|
||||
messages[0]['content'][-1]['text'] = sample['raw_lang']
|
||||
|
||||
return messages
|
||||
Loading…
Reference in New Issue