add policy head
This commit is contained in:
lesjie-wen
parent
1aa4f0c086
commit
20f346956a
|
|
@ -0,0 +1,95 @@
|
|||
import os
|
||||
from typing import Union, List
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
from transformers.utils import logging
|
||||
from transformers import AutoConfig, AutoModelForCausalLM
|
||||
logger = logging.get_logger(__name__)
|
||||
|
||||
MODEL_STRUCTURE = {
|
||||
'ScaleDP_H': {'depth': 32, 'n_emb': 1280, 'num_heads': 16, },
|
||||
'ScaleDP_L': {'depth': 24, 'n_emb': 1024, 'num_heads': 16, }, # 400M
|
||||
}
|
||||
|
||||
class ScaleDPPolicyConfig(PretrainedConfig):
|
||||
'''
|
||||
Configuration for ScaleDP policy head
|
||||
'''
|
||||
model_type = "scale_dp_policy"
|
||||
def __init__(
|
||||
self,
|
||||
eval: bool = False,
|
||||
action_dim: int = 14, # action dim
|
||||
# output_dim: int = 14, # action dim
|
||||
cond_dim: int = 1536, # the input dim of the condition
|
||||
state_dim: int = 14, # the input dim of the state
|
||||
prediction_horizon: int = 16, # horizon
|
||||
n_obs_steps: int = 2, # number of observation steps
|
||||
depth: int = 28, # number of DiT blocks
|
||||
n_emb: int = 256, # embedding size
|
||||
num_heads: int = 16,
|
||||
mlp_ratio: int = 4.0,
|
||||
time_as_cond: bool = True,
|
||||
obs_as_cond: bool = True,
|
||||
learn_sigma: bool = False,
|
||||
model_size: str = "none",
|
||||
num_inference_timesteps: int = 10,
|
||||
num_queries: int = 16,
|
||||
noise_samples: int = 1,
|
||||
num_train_timesteps: int = 100,
|
||||
is_tinyvla: bool = False,
|
||||
**kwargs
|
||||
):
|
||||
if model_size != "none":
|
||||
depth = MODEL_STRUCTURE[model_size]['depth']
|
||||
n_emb = MODEL_STRUCTURE[model_size]['n_emb']
|
||||
num_heads = MODEL_STRUCTURE[model_size]['num_heads']
|
||||
else:
|
||||
# raise ValueError("model_size show not be 'none'")
|
||||
pass
|
||||
# print("model_size should not be 'none'")
|
||||
self.eval = eval
|
||||
|
||||
self.input_dim = action_dim
|
||||
self.output_dim = action_dim
|
||||
self.prediction_horizon = prediction_horizon
|
||||
|
||||
self.is_tinyvla = is_tinyvla
|
||||
|
||||
self.cond_dim = cond_dim
|
||||
self.state_dim = state_dim
|
||||
|
||||
self.n_obs_steps = n_obs_steps
|
||||
self.depth = depth
|
||||
self.n_emb = n_emb
|
||||
self.num_heads = num_heads
|
||||
self.mlp_ratio = mlp_ratio
|
||||
self.time_as_cond = time_as_cond
|
||||
self.obs_as_cond = obs_as_cond
|
||||
self.learn_sigma = learn_sigma
|
||||
|
||||
self.num_inference_timesteps = num_inference_timesteps
|
||||
self.num_queries = prediction_horizon
|
||||
self.noise_samples = noise_samples
|
||||
self.num_train_timesteps = num_train_timesteps
|
||||
super().__init__(**kwargs)
|
||||
|
||||
@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)
|
||||
|
||||
# get the vision config dict if we are loading from CLIPConfig
|
||||
if config_dict.get("model_type") == "llava_pythia":
|
||||
config_dict = config_dict["action_head"]
|
||||
|
||||
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)
|
||||
|
||||
AutoConfig.register("scale_dp_policy", ScaleDPPolicyConfig)
|
||||
|
|
@ -0,0 +1,552 @@
|
|||
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
||||
# All rights reserved.
|
||||
from typing import Tuple
|
||||
|
||||
import timm
|
||||
import numpy as np
|
||||
import logging
|
||||
|
||||
import math
|
||||
from typing import Tuple
|
||||
|
||||
try:
|
||||
from typing import Literal
|
||||
except ImportError:
|
||||
from typing_extensions import Literal
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
import torch.utils.checkpoint
|
||||
from torch.jit import Final
|
||||
from timm.models.vision_transformer import Mlp, use_fused_attn
|
||||
from transformers.modeling_utils import PreTrainedModel
|
||||
from transformers import AutoModel, AutoModelForCausalLM
|
||||
|
||||
_logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class Attention(nn.Module):
|
||||
fused_attn: Final[bool]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
dim: int,
|
||||
num_heads: int = 8,
|
||||
qkv_bias: bool = False,
|
||||
qk_norm: bool = False,
|
||||
attn_drop: float = 0.,
|
||||
proj_drop: float = 0.,
|
||||
norm_layer: nn.Module = nn.LayerNorm,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
assert dim % num_heads == 0, 'dim should be divisible by num_heads'
|
||||
self.num_heads = num_heads
|
||||
self.head_dim = dim // num_heads
|
||||
self.scale = self.head_dim ** -0.5
|
||||
self.fused_attn = use_fused_attn()
|
||||
|
||||
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
|
||||
self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
|
||||
self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
|
||||
self.attn_drop = nn.Dropout(attn_drop)
|
||||
self.proj = nn.Linear(dim, dim)
|
||||
self.proj_drop = nn.Dropout(proj_drop)
|
||||
|
||||
def forward(self, x: torch.Tensor, attn_mask=None) -> torch.Tensor:
|
||||
B, N, C = x.shape
|
||||
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
|
||||
q, k, v = qkv.unbind(0)
|
||||
q, k = self.q_norm(q), self.k_norm(k)
|
||||
|
||||
if self.fused_attn:
|
||||
x = F.scaled_dot_product_attention(
|
||||
q, k, v, attn_mask=attn_mask,
|
||||
dropout_p=self.attn_drop.p if self.training else 0.,
|
||||
)
|
||||
else:
|
||||
q = q * self.scale
|
||||
# attn = q @ k.transpose(-2, -1)
|
||||
# if attn_mask is not None:
|
||||
# attn += attn_mask
|
||||
# attn = attn.softmax(dim=-1)
|
||||
# attn = self.attn_drop(attn)
|
||||
# x = attn @ v
|
||||
attn_scores = torch.matmul(q, k.transpose(-2, -1))
|
||||
|
||||
# Add attention mask if provided
|
||||
if attn_mask is not None:
|
||||
attn_scores += attn_mask
|
||||
|
||||
# Apply softmax to get attention weights (softmax is applied along the last dimension)
|
||||
attn_weights = F.softmax(attn_scores, dim=-1)
|
||||
|
||||
# Dropout on attention weights (if dropout is used)
|
||||
attn_weights = self.attn_drop(attn_weights)
|
||||
|
||||
# Apply attention weights to value tensor (V)
|
||||
x = torch.matmul(attn_weights, v)
|
||||
|
||||
x = x.transpose(1, 2).reshape(B, N, C)
|
||||
x = self.proj(x)
|
||||
x = self.proj_drop(x)
|
||||
return x
|
||||
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
def modulate(x, shift, scale):
|
||||
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
|
||||
|
||||
|
||||
#################################################################################
|
||||
# Embedding Layers for Timesteps and Class Labels #
|
||||
#################################################################################
|
||||
|
||||
class TimestepEmbedder(nn.Module):
|
||||
"""
|
||||
Embeds scalar timesteps into vector representations.
|
||||
"""
|
||||
|
||||
def __init__(self, hidden_size, frequency_embedding_size=256):
|
||||
super().__init__()
|
||||
self.mlp = nn.Sequential(
|
||||
nn.Linear(frequency_embedding_size, hidden_size, bias=True),
|
||||
nn.SiLU(),
|
||||
nn.Linear(hidden_size, hidden_size, bias=True),
|
||||
)
|
||||
self.frequency_embedding_size = frequency_embedding_size
|
||||
|
||||
@staticmethod
|
||||
def timestep_embedding(t, dim, max_period=10000):
|
||||
"""
|
||||
Create sinusoidal timestep embeddings.
|
||||
:param t: a 1-D Tensor of N indices, one per batch element.
|
||||
These may be fractional.
|
||||
:param dim: the dimension of the output.
|
||||
:param max_period: controls the minimum frequency of the embeddings.
|
||||
:return: an (N, D) Tensor of positional embeddings.
|
||||
"""
|
||||
# https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
|
||||
half = dim // 2
|
||||
freqs = torch.exp(
|
||||
-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.bfloat16) / half
|
||||
).to(device=t.device)
|
||||
args = t[:, None].float() * freqs[None]
|
||||
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
||||
if dim % 2:
|
||||
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
|
||||
return embedding.to(dtype=torch.bfloat16)
|
||||
|
||||
def forward(self, t):
|
||||
t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
|
||||
t_emb = self.mlp(t_freq)
|
||||
return t_emb
|
||||
|
||||
|
||||
|
||||
#################################################################################
|
||||
# Core ScaleDP Model #
|
||||
#################################################################################
|
||||
|
||||
class ScaleDPBlock(nn.Module):
|
||||
"""
|
||||
A ScaleDP block with adaptive layer norm zero (adaLN-Zero) conScaleDPioning.
|
||||
"""
|
||||
|
||||
def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, **block_kwargs):
|
||||
super().__init__()
|
||||
self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
||||
self.attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, **block_kwargs)
|
||||
self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
||||
mlp_hidden_dim = int(hidden_size * mlp_ratio)
|
||||
approx_gelu = lambda: nn.GELU(approximate="tanh")
|
||||
self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0)
|
||||
self.adaLN_modulation = nn.Sequential(
|
||||
nn.SiLU(),
|
||||
nn.Linear(hidden_size, 6 * hidden_size, bias=True)
|
||||
)
|
||||
|
||||
def forward(self, x, c, attn_mask=None):
|
||||
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=1)
|
||||
x = x + gate_msa.unsqueeze(1) * self.attn(modulate(self.norm1(x), shift_msa, scale_msa), attn_mask=attn_mask) # norm, scale&shift, attn, scale,
|
||||
x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
|
||||
return x
|
||||
|
||||
|
||||
class FinalLayer(nn.Module):
|
||||
"""
|
||||
The final layer of ScaleDP.
|
||||
"""
|
||||
|
||||
def __init__(self, hidden_size, output_dim):
|
||||
super().__init__()
|
||||
self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
||||
self.linear = nn.Linear(hidden_size, output_dim, bias=True)
|
||||
self.adaLN_modulation = nn.Sequential(
|
||||
nn.SiLU(),
|
||||
nn.Linear(hidden_size, 2 * hidden_size, bias=True)
|
||||
)
|
||||
|
||||
def forward(self, x, c):
|
||||
shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
|
||||
x = modulate(self.norm_final(x), shift, scale)
|
||||
x = self.linear(x)
|
||||
return x
|
||||
|
||||
from .configuration_scaledp import ScaleDPPolicyConfig
|
||||
class ScaleDP(PreTrainedModel):
|
||||
"""
|
||||
Diffusion models with a Transformer backbone.
|
||||
"""
|
||||
config_class = ScaleDPPolicyConfig
|
||||
def __init__(
|
||||
self,
|
||||
config: ScaleDPPolicyConfig,
|
||||
):
|
||||
super().__init__(config)
|
||||
# compute number of tokens for main trunk and conScaleDPion encoder
|
||||
if config.n_obs_steps is None:
|
||||
config.n_obs_steps = config.prediction_horizon
|
||||
T = config.prediction_horizon
|
||||
T_cond = 1
|
||||
if not config.time_as_cond:
|
||||
T += 1
|
||||
T_cond -= 1
|
||||
obs_as_cond = config.cond_dim > 0
|
||||
if obs_as_cond:
|
||||
assert config.time_as_cond
|
||||
T_cond += config.n_obs_steps
|
||||
|
||||
self.is_tinyvla = config.is_tinyvla
|
||||
if config.is_tinyvla:
|
||||
self.global_1d_pool = nn.AdaptiveAvgPool1d(1)
|
||||
self.norm_after_pool = nn.LayerNorm(config.cond_dim)
|
||||
# self.combine = nn.Linear(cond_dim+state_dim, cond_dim)
|
||||
self.combine = nn.Sequential(
|
||||
nn.Linear(config.cond_dim+config.state_dim, 1024),
|
||||
nn.ReLU(),
|
||||
nn.Linear(1024, 1024),
|
||||
nn.ReLU(),
|
||||
nn.Linear(1024, config.cond_dim)
|
||||
)
|
||||
self.learn_sigma = config.learn_sigma
|
||||
self.input_dim = config.input_dim
|
||||
self.output_dim = config.output_dim * 2 if config.learn_sigma else config.output_dim
|
||||
self.num_heads = config.num_heads
|
||||
|
||||
self.x_embedder = nn.Linear(config.input_dim, config.n_emb)
|
||||
self.t_embedder = TimestepEmbedder(config.n_emb)
|
||||
self.cond_obs_emb = None
|
||||
if obs_as_cond:
|
||||
self.cond_obs_emb = nn.Linear(config.cond_dim, config.n_emb)
|
||||
|
||||
# Will use fixed sin-cos embedding:
|
||||
self.pos_embed = nn.Parameter(torch.zeros(1, config.prediction_horizon, config.n_emb))
|
||||
|
||||
self.blocks = nn.ModuleList([
|
||||
ScaleDPBlock(config.n_emb, config.num_heads, mlp_ratio=config.mlp_ratio) for _ in range(config.depth)
|
||||
])
|
||||
self.final_layer = FinalLayer(config.n_emb, output_dim=config.output_dim)
|
||||
# self.initialize_weights()
|
||||
# constants
|
||||
self.T = T
|
||||
self.T_cond = T_cond
|
||||
self.prediction_horizon = config.prediction_horizon
|
||||
self.time_as_cond = config.time_as_cond
|
||||
self.action_dim = config.output_dim
|
||||
self.obs_as_cond = obs_as_cond
|
||||
logger.info(
|
||||
"number of parameters in ScaleDP: %e", sum(p.numel() for p in self.parameters())
|
||||
)
|
||||
|
||||
from diffusers.schedulers.scheduling_ddim import DDIMScheduler
|
||||
self.num_inference_timesteps = config.num_inference_timesteps
|
||||
# self.proj_to_action = nn.Identity()
|
||||
self.noise_scheduler = DDIMScheduler(
|
||||
num_train_timesteps=config.num_train_timesteps, # 100
|
||||
beta_schedule='squaredcos_cap_v2',
|
||||
clip_sample=True,
|
||||
set_alpha_to_one=True,
|
||||
steps_offset=0,
|
||||
prediction_type='epsilon'
|
||||
)
|
||||
self.num_queries = config.num_queries #16
|
||||
self.noise_samples = config.noise_samples # 1
|
||||
# self.num_inference_timesteps = config.num_inference_timesteps # 100
|
||||
|
||||
def initialize_weights(self):
|
||||
# Initialize transformer layers:
|
||||
def _basic_init(module):
|
||||
if isinstance(module, nn.Linear):
|
||||
torch.nn.init.xavier_uniform_(module.weight)
|
||||
if module.bias is not None:
|
||||
nn.init.constant_(module.bias, 0)
|
||||
|
||||
self.apply(_basic_init)
|
||||
|
||||
nn.init.normal_(self.pos_embed, mean=0.0, std=0.02)
|
||||
|
||||
# Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
|
||||
w = self.x_embedder.weight.data
|
||||
nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
|
||||
nn.init.constant_(self.x_embedder.bias, 0)
|
||||
|
||||
# Initialize label embedding table:
|
||||
nn.init.normal_(self.cond_obs_emb.weight, mean=0.0, std=0.02)
|
||||
nn.init.constant_(self.cond_obs_emb.bias, 0)
|
||||
|
||||
# Initialize timestep embedding MLP:
|
||||
nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
|
||||
nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
|
||||
|
||||
# Zero-out adaLN modulation layers in ScaleDP blocks:
|
||||
for block in self.blocks:
|
||||
nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
|
||||
nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
|
||||
|
||||
# Zero-out output layers:
|
||||
nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0)
|
||||
nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0)
|
||||
nn.init.constant_(self.final_layer.linear.weight, 0)
|
||||
nn.init.constant_(self.final_layer.linear.bias, 0)
|
||||
|
||||
|
||||
def get_optim_groups(self, weight_decay: float = 1e-3):
|
||||
"""
|
||||
This long function is unfortunately doing something very simple and is being very defensive:
|
||||
We are separating out all parameters of the models into two buckets: those that will experience
|
||||
weight decay for regularization and those that won't (biases, and layernorm/embedding weights).
|
||||
We are then returning the PyTorch optimizer object.
|
||||
"""
|
||||
|
||||
# separate out all parameters to those that will and won't experience regularizing weight decay
|
||||
decay = set()
|
||||
no_decay = set()
|
||||
whitelist_weight_modules = (torch.nn.Linear, Attention)
|
||||
blacklist_weight_modules = (torch.nn.LayerNorm, torch.nn.Embedding)
|
||||
for mn, m in self.named_modules():
|
||||
for pn, p in m.named_parameters():
|
||||
fpn = "%s.%s" % (mn, pn) if mn else pn # full param name
|
||||
|
||||
if pn.endswith("bias"):
|
||||
# all biases will not be decayed
|
||||
no_decay.add(fpn)
|
||||
elif pn.startswith("bias"):
|
||||
# MultiheadAttention bias starts with "bias"
|
||||
no_decay.add(fpn)
|
||||
elif pn.endswith("weight") and isinstance(m, whitelist_weight_modules):
|
||||
# weights of whitelist modules will be weight decayed
|
||||
decay.add(fpn)
|
||||
elif pn.endswith("weight") and isinstance(m, blacklist_weight_modules):
|
||||
# weights of blacklist modules will NOT be weight decayed
|
||||
no_decay.add(fpn)
|
||||
|
||||
# validate that we considered every parameter
|
||||
param_dict = {pn: p for pn, p in self.named_parameters()}
|
||||
inter_params = decay & no_decay
|
||||
union_params = decay | no_decay
|
||||
assert (
|
||||
len(inter_params) == 0
|
||||
), "parameters %s made it into both decay/no_decay sets!" % (str(inter_params),)
|
||||
assert (
|
||||
len(param_dict.keys() - union_params) == 0
|
||||
), "parameters %s were not separated into either decay/no_decay set!" % (
|
||||
str(param_dict.keys() - union_params),
|
||||
)
|
||||
|
||||
# create the pytorch optimizer object
|
||||
optim_groups = [
|
||||
{
|
||||
"params": [param_dict[pn] for pn in sorted(list(decay))],
|
||||
"weight_decay": weight_decay,
|
||||
},
|
||||
{
|
||||
"params": [param_dict[pn] for pn in sorted(list(no_decay))],
|
||||
"weight_decay": 0.0,
|
||||
},
|
||||
]
|
||||
return optim_groups
|
||||
|
||||
def configure_optimizers(self,
|
||||
learning_rate: float = 1e-4,
|
||||
weight_decay: float = 1e-3,
|
||||
betas: Tuple[float, float] = (0.9, 0.95)):
|
||||
optim_groups = self.get_optim_groups(weight_decay=weight_decay)
|
||||
optimizer = torch.optim.AdamW(
|
||||
optim_groups, lr=learning_rate, betas=betas
|
||||
)
|
||||
return optimizer
|
||||
|
||||
def forward(self, actions, hidden_states, states, is_pad):
|
||||
"""
|
||||
Forward pass for the diffusion head.
|
||||
:param actions: target actions, shape [B, Ta, D] D:10 = 3+6+1
|
||||
:param hidden_states: hidden states from the llava_pythia, as the conScaleDPion for the diffusion, shape [B,Tokens, D] 8 1200 1024
|
||||
:param states: robot states, shape [B, D]
|
||||
:return: loss
|
||||
"""
|
||||
if actions is not None: # training time
|
||||
B = actions.size(0)
|
||||
actions = actions[:, :self.num_queries]
|
||||
is_pad = is_pad[:, :self.num_queries]
|
||||
num_noise_samples = self.noise_samples
|
||||
# sample noise to add to actions
|
||||
noise = torch.randn([num_noise_samples] + list(actions.shape), device=actions.device,
|
||||
dtype=actions.dtype) # num_noise, B, Ta, D(1, 2, 16, 14)
|
||||
# sample a diffusion iteration for each data point
|
||||
timesteps = torch.randint(
|
||||
0, self.noise_scheduler.config.num_train_timesteps,
|
||||
(B,), device=actions.device
|
||||
).long()
|
||||
|
||||
timesteps, noise = timesteps.to(actions.device), noise.to(actions.device)
|
||||
|
||||
# add noise to the clean actions according to the noise magnitude at each diffusion iteration
|
||||
# (this is the forward diffusion process)
|
||||
noisy_actions = torch.cat([self.noise_scheduler.add_noise(
|
||||
actions, noise[i], timesteps)
|
||||
for i in range(len(noise))], dim=0) # [num_noise_samples * B, Ta, action_dim]
|
||||
|
||||
noisy_actions = noisy_actions.to(dtype=actions.dtype)
|
||||
assert hidden_states.ndim == 3
|
||||
|
||||
hidden_states = hidden_states.repeat(num_noise_samples, 1, 1)
|
||||
timesteps = timesteps.repeat(num_noise_samples)
|
||||
is_pad = is_pad.repeat(num_noise_samples, 1)
|
||||
states = states.repeat(num_noise_samples, 1)
|
||||
|
||||
noise_pred = self.model_forward(noisy_actions, timesteps, global_cond=hidden_states, states=states)
|
||||
noise = noise.view(noise.size(0) * noise.size(1), *noise.size()[2:])
|
||||
loss = torch.nn.functional.mse_loss(noise_pred, noise, reduction='none')
|
||||
loss = (loss * ~is_pad.unsqueeze(-1)).mean()
|
||||
# loss_dict['loss'] = loss
|
||||
return {'loss': loss}
|
||||
# return loss
|
||||
else: # inference time
|
||||
B = 1
|
||||
Tp = self.num_queries
|
||||
action_dim = self.action_dim
|
||||
|
||||
# initialize action from Guassian noise
|
||||
noisy_action = torch.randn((B, Tp, action_dim)).cuda()
|
||||
|
||||
naction = noisy_action.to(dtype=hidden_states.dtype)
|
||||
# init scheduler
|
||||
self.noise_scheduler.set_timesteps(self.num_inference_timesteps)
|
||||
|
||||
for k in self.noise_scheduler.timesteps:
|
||||
# predict noise
|
||||
noise_pred = self.model_forward(naction, k, global_cond=hidden_states, states=states)
|
||||
|
||||
# inverse diffusion step (remove noise)
|
||||
naction = self.noise_scheduler.step(
|
||||
model_output=noise_pred,
|
||||
timestep=k,
|
||||
sample=naction
|
||||
).prev_sample
|
||||
|
||||
return naction
|
||||
|
||||
def model_forward(self, x, t, global_cond, states):
|
||||
"""
|
||||
Forward pass of ScaleDP.
|
||||
x: (N, T, input_dim) noisy actions
|
||||
t: (N,) tensor of diffusion timesteps
|
||||
global_cond: (N, n_obs_steps, D) tensor of conScaleDPions: image embeddings
|
||||
"""
|
||||
if self.is_tinyvla:
|
||||
global_cond = self.global_1d_pool(global_cond.permute(0, 2, 1)).squeeze(-1)
|
||||
global_cond = self.norm_after_pool(global_cond)
|
||||
else:
|
||||
global_cond = global_cond.squeeze(1)
|
||||
global_cond = torch.cat([global_cond, states], dim=-1) if states is not None else global_cond
|
||||
global_cond = self.combine(global_cond)
|
||||
|
||||
if not torch.is_tensor(t):
|
||||
t = torch.tensor([t], dtype=torch.long, device=x.device)
|
||||
elif torch.is_tensor(t) and len(t.shape) == 0:
|
||||
t = t[None].to(x.device)
|
||||
t = t.expand(t.shape[0])
|
||||
|
||||
x = self.x_embedder(x) + self.pos_embed.to(device=x.device, dtype=x.dtype) # (N, T, D), where T = prediction_horizon
|
||||
t = self.t_embedder(t) # (N, D)
|
||||
if self.obs_as_cond:
|
||||
global_cond = self.cond_obs_emb(global_cond) # (N, D)
|
||||
# c = t + global_cond.sum(dim=1) # (N, D)
|
||||
c = t + global_cond # (N, D)
|
||||
for block in self.blocks:
|
||||
# x = block(x, c, attn_mask=self.mask) # (N, T, D)
|
||||
x = block(x, c, attn_mask=None) # (N, T, D)
|
||||
x = self.final_layer(x, c) # (N, T, output_dim)
|
||||
return x
|
||||
|
||||
#################################################################################
|
||||
# Sine/Cosine Positional Embedding Functions #
|
||||
#################################################################################
|
||||
# https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py
|
||||
|
||||
def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0):
|
||||
"""
|
||||
grid_size: int of the grid height and width
|
||||
return:
|
||||
pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
|
||||
"""
|
||||
grid_h = np.arange(grid_size, dtype=np.float32)
|
||||
grid_w = np.arange(grid_size, dtype=np.float32)
|
||||
grid = np.meshgrid(grid_w, grid_h) # here w goes first
|
||||
grid = np.stack(grid, axis=0)
|
||||
|
||||
grid = grid.reshape([2, 1, grid_size, grid_size])
|
||||
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
|
||||
if cls_token and extra_tokens > 0:
|
||||
pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
|
||||
return pos_embed
|
||||
|
||||
|
||||
def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
|
||||
assert embed_dim % 2 == 0
|
||||
|
||||
# use half of dimensions to encode grid_h
|
||||
emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2)
|
||||
emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2)
|
||||
|
||||
emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
|
||||
return emb
|
||||
|
||||
|
||||
def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
|
||||
"""
|
||||
embed_dim: output dimension for each position
|
||||
pos: a list of positions to be encoded: size (M,)
|
||||
out: (M, D)
|
||||
"""
|
||||
assert embed_dim % 2 == 0
|
||||
omega = np.arange(embed_dim // 2, dtype=np.float64)
|
||||
omega /= embed_dim / 2.
|
||||
omega = 1. / 10000 ** omega # (D/2,)
|
||||
|
||||
pos = pos.reshape(-1) # (M,)
|
||||
out = np.einsum('m,d->md', pos, omega) # (M, D/2), outer product
|
||||
|
||||
emb_sin = np.sin(out) # (M, D/2)
|
||||
emb_cos = np.cos(out) # (M, D/2)
|
||||
|
||||
emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D)
|
||||
return emb
|
||||
|
||||
|
||||
#################################################################################
|
||||
# ScaleDP Configs #
|
||||
#################################################################################
|
||||
|
||||
def ScaleDP_H(**kwargs):
|
||||
return ScaleDP(depth=32, n_emb=1280, num_heads=16, **kwargs)
|
||||
|
||||
def ScaleDP_L(**kwargs):
|
||||
return ScaleDP(depth=24, n_emb=1024, num_heads=16, **kwargs)
|
||||
|
||||
|
||||
|
||||
AutoModel.register(ScaleDPPolicyConfig, ScaleDP)
|
||||
Loading…
Reference in New Issue