File size: 8,035 Bytes
b8cd73d |
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 |
import copy
from typing import Any, Dict
from transformers import AutoConfig, PretrainedConfig
class ActionCodecConfig(PretrainedConfig):
model_type = "action_codec"
def __init__(
self,
embodiment_config: Dict[str, Any] = None,
n_tokens: int = 16,
n_quantizers: int = 1,
z_dim: int = 512,
vq_type: str = "vq",
vq_codebook_size: int = 2048,
vq_commitment_weight: float = 0.25,
vq_decay: float = 0.99,
vq_kmeans_init: bool = True,
vq_threshold_ema_dead_code: int = 2,
vq_quantizer_dropout: float = 0.25,
encoder_dim: int = 256,
encoder_n_layers: int = 6,
encoder_n_heads: int = 8,
encoder_add_self_attn: bool = False,
encoder_add_causal_mask: bool = False,
encoder_pos_encoding_type: str = "fourier",
decoder_dim: int = 256,
decoder_n_layers: int = 6,
decoder_n_heads: int = 8,
decoder_add_self_attn: bool = False,
decoder_add_causal_mask: bool = False,
decoder_pos_encoding_type: str = "fourier",
decoder_cls_size: int = 1,
**kwargs,
):
super().__init__(**kwargs)
if embodiment_config is None:
default_config = {
"franka_libero_20hz": {
"action_dim": 7,
"freq": 20,
"duration": 1,
"description": "20Hz 7-dim action for 1s. Delta eef position (xyz), orientation (rpy), and gripper position (1 open/0 close).",
},
"widowx_bridge_5hz": {
"action_dim": 7,
"freq": 5,
"duration": 1,
"description": "5Hz 7-dim action for 1s. Delta eef position (xyz), orientation (rpy), and gripper position (1 open/0 close).",
},
"franka_droid_15hz": {
"action_dim": 7,
"freq": 15,
"duration": 1,
"description": "15Hz 7-dim action for 1s. Delta eef position (xyz), orientation (rpy), and gripper position (1 open/0 close).",
},
}
self.embodiment_config = copy.deepcopy(default_config)
else:
self.embodiment_config = copy.deepcopy(embodiment_config)
self.n_tokens = n_tokens
self.n_quantizers = n_quantizers
self.z_dim = z_dim
self.encoder_dim = encoder_dim
self.encoder_n_layers = encoder_n_layers
self.encoder_n_heads = encoder_n_heads
self.encoder_add_self_attn = encoder_add_self_attn
self.encoder_add_causal_mask = encoder_add_causal_mask
self.encoder_pos_encoding_type = encoder_pos_encoding_type
self.decoder_dim = decoder_dim
self.decoder_n_layers = decoder_n_layers
self.decoder_n_heads = decoder_n_heads
self.decoder_add_self_attn = decoder_add_self_attn
self.decoder_add_causal_mask = decoder_add_causal_mask
self.decoder_pos_encoding_type = decoder_pos_encoding_type
self.decoder_cls_size = decoder_cls_size
self.vq_type = vq_type
self.vq_codebook_size = vq_codebook_size
self.vq_commitment_weight = vq_commitment_weight
self.vq_decay = vq_decay
self.vq_kmeans_init = vq_kmeans_init
self.vq_threshold_ema_dead_code = vq_threshold_ema_dead_code
self.vq_quantizer_dropout = vq_quantizer_dropout
class ActionCodecConfigOld(PretrainedConfig):
model_type = "action_codec"
def __init__(
self,
horizon: int = 20,
action_dim: int = 7,
action_encoding: str = "independent_v2",
horizon_patch_size: int = 1,
encoder_class: str = "action_codec.modules.perceiver.PerceiverEncoder",
decoder_class: str = "action_codec.modules.perceiver.PerceiverDecoder",
vq_class: str = "vector_quantize_pytorch.VectorQuantize",
encoder_kwargs: Dict[str, Any] = None,
decoder_kwargs: Dict[str, Any] = None,
vq_kwargs: Dict[str, Any] = None,
**kwargs,
):
super().__init__(**kwargs)
self.horizon = horizon
self.action_dim = action_dim
self.action_encoding = action_encoding
self.horizon_patch_size = horizon_patch_size
self.encoder_class = encoder_class
self.decoder_class = decoder_class
self.vq_class = vq_class
self.encoder_kwargs = (
dict(encoder_kwargs)
if encoder_kwargs is not None
else {
"dim": 384,
"in_len": horizon,
"out_len": 16,
"num_layers": 12,
"num_heads": 4,
"output_round": -1.0,
}
)
self.decoder_kwargs = (
dict(decoder_kwargs)
if decoder_kwargs is not None
else {
"dim": 384,
"in_len": 16,
"out_len": horizon,
"num_layers": 12,
"num_heads": 4,
}
)
self.vq_kwargs = (
dict(vq_kwargs)
if vq_kwargs is not None
else {
"dim": 512,
"codebook_size": 2048,
"kmeans_init": True,
"kmeans_iters": 10,
"decay": 0.99,
"commitment_weight": 0.25,
"rotation_trick": False,
"threshold_ema_dead_code": 2,
"use_cosine_sim": False,
"codebook_diversity_loss_weight": 0.0,
}
)
class BPEActionCodecConfig(PretrainedConfig):
model_type = "bpe_action_codec"
def __init__(
self,
horizon: int = 20,
action_dim: int = 7,
action_encoding: str = "independent_v2",
horizon_patch_size: int = 1,
encoder_class: str = "action_codec.modules.perceiver.PerceiverEncoder",
decoder_class: str = "action_codec.modules.perceiver.PerceiverDecoder",
vq_class: str = "vector_quantize_pytorch.VectorQuantize",
encoder_kwargs: Dict[str, Any] = None,
decoder_kwargs: Dict[str, Any] = None,
vq_kwargs: Dict[str, Any] = None,
**kwargs,
):
super().__init__(**kwargs)
self.horizon = horizon
self.action_dim = action_dim
self.action_encoding = action_encoding
self.horizon_patch_size = horizon_patch_size
self.encoder_class = encoder_class
self.decoder_class = decoder_class
self.vq_class = vq_class
self.encoder_kwargs = (
dict(encoder_kwargs)
if encoder_kwargs is not None
else {
"dim": 384,
"in_len": horizon,
"out_len": 16,
"num_layers": 12,
"num_heads": 4,
"output_round": -1.0,
}
)
self.decoder_kwargs = (
dict(decoder_kwargs)
if decoder_kwargs is not None
else {
"dim": 384,
"in_len": 16,
"out_len": horizon,
"num_layers": 12,
"num_heads": 4,
}
)
self.vq_kwargs = (
dict(vq_kwargs)
if vq_kwargs is not None
else {
"dim": 512,
"codebook_size": 2048,
"kmeans_init": True,
"kmeans_iters": 10,
"decay": 0.99,
"commitment_weight": 0.25,
"rotation_trick": False,
"threshold_ema_dead_code": 2,
"use_cosine_sim": False,
"codebook_diversity_loss_weight": 0.0,
}
)
AutoConfig.register("action_codec", ActionCodecConfig)
AutoConfig.register("bpe_action_codec", BPEActionCodecConfig)
ActionCodecConfig.register_for_auto_class()
__all__ = ["ActionCodecConfig", "BPEActionCodecConfig"]
|