Linear layer with domain-conditioned parameters (per-sample). Each domain has its own weight and bias vectors, stored in embeddings.

@brief exception indicating access out of the defined range @sa https://json.nlohmann.me/api/basic_json/out_of_range/

r""" This is the configuration class to store the configuration of a [`Florence2ForConditionalGeneration`]. It is used to instantiate an Flore

This module learns positional embeddings up to a fixed maximum size.

This module overrides nn.Embeddings' forward by multiplying with embeddings scale.

This module learns positional embeddings up to a fixed maximum size.

Helpers to convert between 6D rotation (Zhou et al.) and axis-angle.

This class implements a very simple positional encoding. It follows closely the encoder from the link below: https://pytorch.org/tutorial

! Input adapter for a (caching) istream. Ignores a UFT Byte Order Mark at beginning of input. Does not support changing the underlying std::streambuf

Multi-Head Self-Attention with optional fused SDPA fallback. If PyTorch provides `scaled_dot_product_attention`, it will be used (usuall

机械臂ros节点

Thin HTTP client that queries a remote policy server and returns actions.

Image to Patch Embedding

Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`Florence2DecoderLayer`] Args: config: Florence2L

Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a [`Florence2EncoderLayer`]. Args:

r""" This is the configuration class to store the configuration of a [`Florence2LanguagePreTrainedModel`]. It is used to instantiate a BART mo

Base class for Florence-2 model's outputs that also contains : pre-computed hidden states that can speed up sequential decoding. Args:

r""" This is the configuration class to store the configuration of a [`Florence2VisionModel`]. It is used to instantiate a Florence2VisionModel

Output sample: { 'domain_id': LongTensor[], # domain id 'language_instruction': str, 'image_input': FloatTensor[

MLP used in ViT-style blocks. Supports Linear or 1x1 Conv 'linear_layer' for token/channel mixing.

Multi-modal, domain-aware Transformer with optional soft prompts. See parameter and forward I/O descriptions inside the docstrings.

Standard Transformer block (pre-LN): LN → MHSA → residual, LN → MLP → residual.

Lightweight HTTP client that queries an XVLA FastAPI server for action predictions.

Lightweight HTTP client that queries an XVLA FastAPI server for action predictions.

Lightweight HTTP client that queries an XVLA FastAPI server for action predictions.

Lightweight HTTP client that queries an XVLA FastAPI server for action predictions.

Lightweight HTTP client that queries an XVLA FastAPI server for action predictions.

Lightweight HTTP client that queries an XVLA FastAPI server for action predictions.

Lightweight HTTP client that queries an XVLA FastAPI server for action predictions.

Lightweight HTTP client that queries an XVLA FastAPI server for action predictions.

Lightweight HTTP client that queries an XVLA FastAPI server for action predictions.

Lightweight HTTP client that queries an XVLA FastAPI server for action predictions.

Lightweight HTTP client that queries an XVLA FastAPI server for action predictions.

Lightweight HTTP client that queries an XVLA FastAPI server for action predictions.

@brief an iterator for primitive JSON types This class models an iterator for primitive JSON types (boolean, number, string). It's only purpose is to

AGI-bot variant of EE6DActionSpace using MSE for all components.

AIR-AGILEX-HQ. HDF5: /observations/eef_6d [T,20] -> L(10)+R(10) /observations/eef_left_time [T] /observations/eef_r

AIR-AGILEX (non-HQ). HDF5: /observations/eef_quaternion [T, 16] = L_xyz(3) L_quat(4) L_grip_raw(1) R_xyz(3) R_quat(4) R_grip_ra

AIRBOT. HDF5: /eef_6d [T,10] -> xyz(3)+rot6d(6)+grip_raw(1) Single arm (left), right is zeros. Grip <0.5 => closed.