ducido's picture
Initial commit
987ed1b verified
Raw
History Blame Contribute Delete
25.3 kB
# MIT License
# Copyright (c) 2025 ReinFlow Authors
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
"""
MLP models for flow matching with learnable stochastic interpolate noise.
"""
import torch
import torch.nn as nn
import logging
import numpy as np
from copy import deepcopy
from typing import Tuple
from torch import Tensor
from diffusion_policy.model.common.mlp import MLP, ResidualMLP
from diffusion_policy.model.diffusion.positional_embedding import SinusoidalPosEmb
from diffusion_policy.model.common.modules import SpatialEmb, RandomShiftsAug
from diffusion_policy.model.common.vit import VitEncoder
log = logging.getLogger(__name__)
import einops
from typing import List
class FlowMLP(nn.Module):
def __init__(
self,
horizon_steps,
action_dim,
cond_dim,
time_dim=16,
mlp_dims=[256, 256],
cond_mlp_dims=None,
activation_type="Mish",
out_activation_type="Identity",
use_layernorm=False,
residual_style=False,
):
super().__init__()
self.time_dim = time_dim
self.act_dim_total = action_dim * horizon_steps
self.horizon_steps = horizon_steps
self.action_dim=action_dim
self.cond_dim=cond_dim
self.mlp_dims=mlp_dims
self.activation_type=activation_type
self.out_activation_type=out_activation_type
self.use_layernorm=use_layernorm
self.residual_style=residual_style
self.time_embedding = nn.Sequential(
SinusoidalPosEmb(time_dim),
nn.Linear(time_dim, time_dim * 2),
nn.Mish(),
nn.Linear(time_dim * 2, time_dim),
)
model = ResidualMLP if residual_style else MLP
# obs encoder
if cond_mlp_dims:
self.cond_mlp = MLP(
[cond_dim] + cond_mlp_dims,
activation_type=activation_type,
out_activation_type="Identity",
)
self.cond_enc_dim = cond_mlp_dims[-1]
else:
self.cond_enc_dim = cond_dim
input_dim = time_dim + action_dim * horizon_steps + self.cond_enc_dim
# velocity head
self.mlp_mean = model(
[input_dim] + mlp_dims + [self.act_dim_total],
activation_type=activation_type,
out_activation_type=out_activation_type,
use_layernorm=use_layernorm,
)
def forward(
self,
action,
time,
cond,
output_embedding=False,
):
"""
**Args**:
action: (B, Ta, Da)
time: (B,) or int, diffusion step
cond: dict with key state/rgb; more recent obs at the end
state: (B, To, Do)
**Outpus**:
velocity.
vel: (B, Ta, Da) when output_embedding==False
vel,time_emb, cond_emb: when output_embedding==False
"""
B, Ta, Da = action.shape
# flatten action chunk
action = action.view(B, -1)
# flatten obs history
state = cond["state"].view(B, -1)
# obs encoder
cond_emb = self.cond_mlp(state) if hasattr(self, "cond_mlp") else state
# time encoder
if isinstance(time, int) or isinstance(time, float):
time=torch.ones((B,1), device=action.device)* time
time_emb = self.time_embedding(time.view(B, 1)).view(B, self.time_dim)
# velocity head
vel_feature = torch.cat([action, time_emb, cond_emb], dim=-1)
vel = self.mlp_mean(vel_feature)
if output_embedding:
return vel.view(B, Ta, Da), time_emb, cond_emb
return vel.view(B, Ta, Da)
def sample_action(self,cond:dict,inference_steps:int,clip_intermediate_actions:bool,act_range:List[float], z:Tensor=None,save_chains:bool=False):
"""
simply return action via integration (Euler's method). the initial noise could be specified.
when `save_chains` is True, also return the denoising trajectory.
"""
B = cond['state'].shape[0]
device=cond['state'].device
x_hat:Tensor=z if z is not None else torch.randn(B, self.horizon_steps, self.action_dim, device=device)
if save_chains:
x_chain=torch.zeros((B, inference_steps+1, self.horizon_steps, self.action_dim), device=device)
dt = (1 / inference_steps) * torch.ones_like(x_hat, device=device)
steps = torch.linspace(0, 1-1 / inference_steps, inference_steps, device=device).repeat(B, 1)
for i in range(inference_steps):
t = steps[:, i]
vt = self.forward(x_hat, t, cond)
x_hat += vt * dt
if clip_intermediate_actions or i == inference_steps-1: # always clip the output action. appended by ReinFlow Authors on 04/25/2025
x_hat = x_hat.clamp(*act_range)
if save_chains:
x_chain[:, i+1] = x_hat
if save_chains:
return x_hat, x_chain
return x_hat
class ExploreNoiseNet(nn.Module):
'''
Neural network to generate learnable exploration noise, conditioned on time embeddings and or state embeddings.
\sigma(s,t) or \sigma(s)
'''
def __init__(self,
in_dim:int,
out_dim:int,
logprob_denoising_std_range:list, #[min_std, max_std]
device,
hidden_dims=[16], #[8] [32],
activation_type='Tanh'
):
super().__init__()
self.device = device
self.mlp_logvar = MLP(
[in_dim] + hidden_dims +[out_dim],
activation_type=activation_type,
out_activation_type="Identity",
).to(self.device)
self.set_noise_range(logprob_denoising_std_range)
def set_noise_range(self, logprob_denoising_std_range:list):
self.logprob_denoising_std_range=logprob_denoising_std_range
min_logprob_denoising_std = self.logprob_denoising_std_range[0]
max_logprob_denoising_std = self.logprob_denoising_std_range[1]
self.logvar_min = torch.nn.Parameter(torch.log(torch.tensor(min_logprob_denoising_std**2, dtype=torch.float32, device=self.device)), requires_grad=False)
self.logvar_max = torch.nn.Parameter(torch.log(torch.tensor(max_logprob_denoising_std**2, dtype=torch.float32, device=self.device)), requires_grad=False)
def forward(self, noise_feature:torch.Tensor):
'''
'''
noise_logvar = self.mlp_logvar(noise_feature)
noise_std = self.process_noise(noise_logvar)
return noise_std
def process_noise(self, noise_logvar):
'''
input:
torch.Tensor([B, Ta , Da]) log \sigma^2
output:
torch.Tensor([B, 1, Ta * Da]), sigma, floating point values, bounded in [min_logprob_denoising_std, max_logprob_denoising_std]
'''
noise_logvar = noise_logvar
noise_logvar = torch.tanh(noise_logvar)
noise_logvar = self.logvar_min + (self.logvar_max - self.logvar_min) * (noise_logvar + 1)/2.0
noise_std = torch.exp(0.5 * noise_logvar)
return noise_std
class NoisyFlowMLP(nn.Module):
def __init__(
self,
policy:FlowMLP,
denoising_steps,
learn_explore_noise_from,
inital_noise_scheduler_type,
min_logprob_denoising_std,
max_logprob_denoising_std,
learn_explore_time_embedding,
time_dim_explore,
use_time_independent_noise,
device,
noise_hidden_dims=None,
activation_type='Tanh'
):
super().__init__()
self.device=device
self.policy:FlowMLP = policy.to(self.device)
"""
input: [batchsize, time_dim + cond_enc_dim]
output: positive tensor of shape [batchsize, self.denoising_steps, self.horizon_steps x self.act_dim]
"""
self.denoising_steps: int = denoising_steps
self.learn_explore_noise_from: int = learn_explore_noise_from
self.initial_noise_scheduler_type: str = inital_noise_scheduler_type
if min_logprob_denoising_std > max_logprob_denoising_std:
raise ValueError(f"min_logprob_denoising_std must not exceed max_logprob_denoising_std, but received min_logprob_denoising_std={min_logprob_denoising_std} > max_logprob_denoising_std={max_logprob_denoising_std}. Revise your configuration file!")
self.min_logprob_denoising_std: float = min_logprob_denoising_std
self.max_logprob_denoising_std: float = max_logprob_denoising_std
self.learn_explore_time_embedding: bool = learn_explore_time_embedding
self.set_logprob_noise_levels()
self.noise_hidden_dims=noise_hidden_dims
self.use_time_independent_noise = use_time_independent_noise
self.time_dim_explore =time_dim_explore
self.noise_activation_type=activation_type
self.init_exploration_noise_net()
def init_exploration_noise_net(self):
if self.use_time_independent_noise:
noise_input_dim = self.policy.cond_enc_dim
if not self.noise_hidden_dims:
self.noise_hidden_dims = [16]
else:
if self.learn_explore_time_embedding:
noise_input_dim = self.time_dim_explore + self.policy.cond_enc_dim
self.time_embedding_explore = nn.Embedding(num_embeddings=self.denoising_steps,
embedding_dim = self.time_dim_explore,
device=self.device)
else:
noise_input_dim = self.policy.time_dim + self.policy.cond_enc_dim
if not self.noise_hidden_dims:
self.noise_hidden_dims = [int(np.sqrt(noise_input_dim**2 + self.policy.act_dim_total**2))]
self.explore_noise_net=ExploreNoiseNet(in_dim=noise_input_dim,
out_dim=self.policy.act_dim_total,
logprob_denoising_std_range=[self.min_logprob_denoising_std, self.max_logprob_denoising_std],
device=self.device,
hidden_dims=self.noise_hidden_dims,
activation_type=self.noise_activation_type)
def forward(
self,
action,
time,
cond,
learn_exploration_noise=False,
step=-1,
verbose=False,
**kwargs,
)->Tuple[Tensor, Tensor]:
"""
inputs:
x: (B, Ta, Da)
time: (B,) floating point in [0,1) flow matching time
cond: dict with key state/rgb; more recent obs at the end
state: (B, To, Do)
step: (B,) torch.tensor, optional, flow matching inference step, from 0 to denoising_steps-1
*here, B is the n_envs
outputs:
vel [B, Ta, Da]
noise_std [B, Ta x Da]
"""
B = action.shape[0]
vel, time_emb, cond_emb = self.policy.forward(action, time, cond, output_embedding=True)
# noise head (for exploration). allow gradient flow.
if self.initial_noise_scheduler_type=='const' or step < self.learn_explore_noise_from:
noise_std = self.logprob_noise_levels[:, step].repeat(B,1)
else:
if self.use_time_independent_noise:
noise_feature = cond_emb
else:
if self.learn_explore_time_embedding:
step_ts = torch.tensor(step, device = self.device).repeat(B)
time_emb_explore = self.time_embedding_explore(step_ts)
noise_feature = torch.cat([time_emb_explore, cond_emb], dim=-1)
else:
noise_feature = torch.cat([time_emb.detach(), cond_emb], dim=-1)
noise_std = self.explore_noise_net.forward(noise_feature=noise_feature)
if verbose:
log.info(f"step={step}, learnable noise = {noise_std.mean()}")
if verbose:
log.info(f"step={step}, set to learn from {self.learn_explore_noise_from}, will learn exploration noise ? {step >= self.learn_explore_noise_from}, noise_std={noise_std.mean()}require_grad={noise_std.requires_grad}")
return vel, noise_std if learn_exploration_noise else noise_std.detach()
@torch.no_grad()
def stochastic_interpolate(self,t):
valid_noise_schedulers=['vp', 'lin', 'const', 'const_schedule_itr', 'learn_decay']
if self.initial_noise_scheduler_type == 'vp':
a = 0.2 #2.0
std = torch.sqrt(a * t * (1 - t))
elif self.initial_noise_scheduler_type == 'lin':
k=0.1
b=0.0
std = k*t+b
elif self.initial_noise_scheduler_type == 'const' or 'const_schedule_itr':
std = torch.ones_like(t) * self.min_logprob_denoising_std
else:
raise ValueError(f"Invalid noise scheduler type {self.initial_noise_scheduler_type}, must be in the following: {valid_noise_schedulers}")
return std
@torch.no_grad()
def set_logprob_noise_levels(self, force_level=None, verbose=False):
'''
create noise std for logrporbability calcualion.
generate a tensor `self.logprob_noise_levels` of shape `[1, self.denoising_steps, self.policy.horizion_steps x self.policy.act_dim]`
'''
self.logprob_noise_levels = torch.zeros(self.denoising_steps, device=self.device, requires_grad=False)
steps = torch.linspace(0, 1-1 /self.denoising_steps, self.denoising_steps, device=self.device)
for i, t in enumerate(steps):
if force_level:
self.logprob_noise_levels[i] = torch.tensor(force_level, device=self.device)
else:
self.logprob_noise_levels[i] = self.stochastic_interpolate(t)
self.logprob_noise_levels = self.logprob_noise_levels.clamp(min=self.min_logprob_denoising_std, max=self.max_logprob_denoising_std)
self.logprob_noise_levels = self.logprob_noise_levels.unsqueeze(0).unsqueeze(-1).repeat(1, 1, self.policy.horizon_steps * self.policy.action_dim)
if verbose:
log.info(f"Set logprob noise levels. self.logprob_noise_levels={self.logprob_noise_levels}")
class VisionFlowMLP(nn.Module):
"""With ViT backbone"""
def __init__(
self,
backbone: VitEncoder,
action_dim,
horizon_steps,
cond_dim, # proprioception only
img_cond_steps=1,
time_dim=16,
mlp_dims=[256, 256],
activation_type="Mish",
out_activation_type="Identity",
use_layernorm=False,
residual_style=False,
spatial_emb=0,
visual_feature_dim=128, # visual feature dim
dropout=0,
num_img=1, # currently only supports 1 or 2
augment=False,
):
super().__init__()
# action chunk
self.action_dim = action_dim
self.horizon_steps = horizon_steps
self.act_dim_total = action_dim * horizon_steps
# historical proprioception and visual inputs
self.prop_dim = cond_dim
self.img_cond_steps = img_cond_steps
# time
self.time_dim = time_dim
self.backbone = backbone
self.mlp_dims = mlp_dims
self.activation_type = activation_type
self.out_activation_type = out_activation_type
self.use_layernorm = use_layernorm
self.residual_style = residual_style
self.spatial_emb = spatial_emb
self.dropout = dropout
self.num_img = num_img
self.augment = augment
# vision
self.backbone = backbone
if augment:
self.aug = RandomShiftsAug(pad=4)
if spatial_emb > 0:
assert spatial_emb > 1, "this is the dimension"
if num_img == 2:
self.compress1 = SpatialEmb(
num_patch=self.backbone.num_patch,
patch_dim=self.backbone.patch_repr_dim,
prop_dim=cond_dim,
proj_dim=spatial_emb,
dropout=dropout,
)
self.compress2 = deepcopy(self.compress1)
elif num_img == 1: # TODO: clean up
self.compress = SpatialEmb(
num_patch=self.backbone.num_patch,
patch_dim=self.backbone.patch_repr_dim,
prop_dim=cond_dim,
proj_dim=spatial_emb,
dropout=dropout,
)
else:
raise NotImplementedError(f"num_img={num_img} Currently we only support 1 or 2 image inputs")
visual_feature_dim = spatial_emb * num_img
else: # spatial embedding not specified, use default value 128
self.compress = nn.Sequential(
nn.Linear(self.backbone.repr_dim, visual_feature_dim),
nn.LayerNorm(visual_feature_dim),
nn.Dropout(dropout),
nn.ReLU(),
)
self.cond_enc_dim = visual_feature_dim + self.prop_dim # rgb and proprioception
self.time_embedding = nn.Sequential(
SinusoidalPosEmb(time_dim),
nn.Linear(time_dim, time_dim * 2),
nn.Mish(),
nn.Linear(time_dim * 2, time_dim),
)
# Flow
input_dim = (
time_dim + \
action_dim * horizon_steps + \
self.cond_enc_dim
)
# output action chunk
output_dim = action_dim * horizon_steps
# velocity head
model = ResidualMLP if residual_style else MLP
self.mlp_mean = model(
[input_dim] + mlp_dims + [output_dim],
activation_type=activation_type,
out_activation_type=out_activation_type,
use_layernorm=use_layernorm,
)
def forward(
self,
action,
time,
cond: dict,
output_embedding=False,
**kwargs,
):
"""
inputs:
action: (B, Ta, Da) action chunk
time: (B,) or float within [0,1), flow time
cond: dict with key state/rgb; more recent obs at the end
state: (B, To, Do)
rgb: (B, To, C, H, W)
outputs:
TODO long term: more flexible handling of cond
"""
B, Ta, Da = action.shape
_, T_rgb, C, H, W = cond["rgb"].shape
# flatten chunk
action = action.view(B, -1)
# flatten history (proprioception, here we use the raw input without encoding)
state = cond["state"].view(B, -1)
# Take recent images --- sometimes we want to use fewer img_cond_steps than cond_steps (e.g., 1 image but 3 prio)
rgb = cond["rgb"][:, -self.img_cond_steps :]
# concatenate images in cond by channels
if self.num_img >1:
rgb = rgb.reshape(B, T_rgb, self.num_img, 3, H, W)
rgb = einops.rearrange(rgb, "b t n c h w -> b n (t c) h w")
elif self.num_img==1:
rgb = einops.rearrange(rgb, "b t c h w -> b (t c) h w")
else:
raise ValueError(f"self.num_img={self.num_img} <1. ")
# convert rgb to float32 for augmentation
rgb = rgb.float()
# visual and proprioceptive embeddings: get vit output - pass in two images separately
if self.num_img ==2: # TODO: properly handle multiple images
rgb1 = rgb[:, 0]
rgb2 = rgb[:, 1]
if self.augment:
rgb1 = self.aug(rgb1)
rgb2 = self.aug(rgb2)
feat1 = self.backbone.forward(rgb1)
feat1 = self.compress1.forward(feat1, state)
feat2 = self.backbone.forward(rgb2)
feat2 = self.compress2.forward(feat2, state)
feat = torch.cat([feat1, feat2], dim=-1)
elif self.num_img ==1: # single image
if self.augment:
rgb = self.aug(rgb)
feat = self.backbone.forward(rgb)
# compress
if isinstance(self.compress, SpatialEmb):
feat = self.compress.forward(feat, state)
else:
feat = feat.flatten(1, -1)
feat = self.compress(feat)
else:
raise NotImplementedError(f"num_img={self.num_img} Currently we only support 1 or 2 image inputs")
cond_encoded = torch.cat([feat, state], dim=-1) # visual and proprioception inputs.
# time embedding
time = time.view(B, 1)
time_emb = self.time_embedding(time).view(B, self.time_dim)
# all embeddings: time, visual-proprioceptive
emb = torch.cat([action, time_emb, cond_encoded], dim=-1)
# velocity head
vel = self.mlp_mean(emb)
if output_embedding:
return vel.view(B, Ta, Da), time_emb, cond_encoded
return vel.view(B, Ta, Da)
class NoisyVisionFlowMLP(NoisyFlowMLP):
def __init__(
self,
policy:VisionFlowMLP,
denoising_steps,
learn_explore_noise_from,
inital_noise_scheduler_type,
min_logprob_denoising_std,
max_logprob_denoising_std,
learn_explore_time_embedding,
time_dim_explore,
use_time_independent_noise,
device,
noise_hidden_dims=None,
activation_type='Tanh'
):
super().__init__(
policy,
denoising_steps,
learn_explore_noise_from,
inital_noise_scheduler_type,
min_logprob_denoising_std,
max_logprob_denoising_std,
learn_explore_time_embedding,
time_dim_explore,
use_time_independent_noise,
device,
noise_hidden_dims,
activation_type
)
def forward(
self,
action,
time,
cond,
learn_exploration_noise=False,
step=-1,
verbose=False,
**kwargs,
)->Tuple[Tensor, Tensor]:
"""
inputs:
x: (B, Ta, Da)
time: (B,) floating point in [0,1) flow matching time
cond: dict with key state/rgb; more recent obs at the end
state: (B, To, Do)
step: (B,) torch.tensor, optional, flow matching inference step, from 0 to denoising_steps-1
*here, B is the n_envs
outputs:
vel [B, Ta, Da]
noise_std [B, Ta x Da]
"""
B = action.shape[0]
self.policy: VisionFlowMLP
vel, time_emb, cond_emb = self.policy.forward(action, time, cond, output_embedding=True)
# noise head (for exploration). allow gradient flow.
if self.initial_noise_scheduler_type=='const' or step < self.learn_explore_noise_from:
noise_std = self.logprob_noise_levels[:, step].repeat(B,1)
else:
if self.use_time_independent_noise:
noise_feature = cond_emb
else:
if self.learn_explore_time_embedding:
step_ts = torch.tensor(step, device = self.device).repeat(B)
time_emb_explore = self.time_embedding_explore(step_ts)
noise_feature = torch.cat([time_emb_explore, cond_emb], dim=-1)
else:
noise_feature = torch.cat([time_emb.detach(), cond_emb], dim=-1)
# predict noise
noise_std = self.explore_noise_net.forward(noise_feature=noise_feature)
return vel, noise_std if learn_exploration_noise else noise_std.detach()