This work presents an instruction-conditioned multimodal fine-tuning framework that adapts pretrained diffusion models to robotic visual prediction tasks by dynamically aligning textual instructions with visual regions in input frames via cross-attention mechanisms.
- computing platform：AutoDL
- mirror：pytorch2.10, cuda12.1
follow [RoboTwin installing instruction](RoboTwin/INSTALLATION.md at main · TianxingChen/RoboTwin) to download and install RoboTwin in corresponding folder which we have created.
unzip project.zip file, run following command to install RoboTwin virtual environment.
conda activate RoboTwin
pip install -r requirements.txtMake sure your generated data are put in the ./auto-tmpfolder.
Run following command to activate virtual render.
sudo apt-get update && sudo apt-get install -y xvfb x11-utils
Xvfb :99 -screen 0 1024x768x16 &
export DISPLAY=:99modify RoboTwin config settings in
task_name: block_hammer_beat
render_freq: 0
eval_video_log: false
use_seed: false
collect_data: true
save_path: /root/autodl_tmp
dual_arm: true
st_episode: 0
head_camera_type: D435
wrist_camera_type: D435
front_camera_type: D435
pcd_crop: true
pcd_down_sample_num: 1024
episode_num: 100
save_freq: 15
save_type:
raw_data: false
pkl: true
data_type:
rgb: true
observer: false
depth: false
pointcloud: false
endpose: true
qpos: true
mesh_segmentation: false
actor_segmentation: falseblock_handover
task_name: block_handover
render_freq: 0
eval_video_log: false
use_seed: false
collect_data: true
save_path: /root/autodl_tmp
dual_arm: true
st_episode: 0
head_camera_type: D435
wrist_camera_type: D435
front_camera_type: D435
pcd_crop: true
pcd_down_sample_num: 1024
episode_num: 100
save_freq: 15
save_type:
raw_data: false
pkl: true
data_type:
rgb: true
observer: false
depth: false
pointcloud: false
endpose: true
qpos: true
mesh_segmentation: false
actor_segmentation: falseblocks_stack_easy
task_name: blocks_stack_easy
render_freq: 0
eval_video_log: false
use_seed: false
collect_data: true
save_path: /root/autodl_tmp
dual_arm: true
st_episode: 0
head_camera_type: D435
wrist_camera_type: D435
front_camera_type: D435
pcd_crop: true
pcd_down_sample_num: 1024
episode_num: 100
save_freq: 15
save_type:
raw_data: false
pkl: true
data_type:
rgb: true
observer: false
depth: false
pointcloud: false
endpose: true
qpos: true
mesh_segmentation: false
actor_segmentation: falsemake sure you are in the project folder.
following [RoboTwin installing instruction](RoboTwin/INSTALLATION.md at main · TianxingChen/RoboTwin) to generate following .pkl data:
"/autodl-tmp/block_hammer_beat_D435_pkl"
"/autodl-tmp/block_handover_D435_pkl"
"/autodl-tmp/blocks_stack_easy_D435_pkl"or you can just run our bash script generate all the data we need.
cd project/path
./run_all_task.sh(we have already provide the instruct-pix2pix folder, so you can go Part.3 directly if you don't want to download yourself)
Run following command:
# make sure you are in the path/to/project/folder/autodl-tmp
git clone https://github.com/timothybrooks/instruct-pix2pix.git # clone the original repository
conda env create -f environment.yaml # set up conda environment
cd instruct-pix2pixmodify config/train.yaml:
model:
base_learning_rate: 1.0e-04
target: ldm.models.diffusion.ddpm_edit.LatentDiffusion
params:
ckpt_path: stable_diffusion/models/ldm/stable-diffusion-v1/v1-5-pruned-emaonly.ckpt
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: edited
cond_stage_key: edit
image_size: 32
channels: 4
cond_stage_trainable: false # Note: different from the one we trained before
conditioning_key: hybrid
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: true
load_ema: false
scheduler_config: # 10000 warmup steps
target: ldm.lr_scheduler.LambdaLinearScheduler
params:
warm_up_steps: [ 0 ]
cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
f_start: [ 1.e-6 ]
f_max: [ 1. ]
f_min: [ 1. ]
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 32 # unused
in_channels: 8
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
data:
target: main.DataModuleFromConfig
params:
batch_size: 32
num_workers: 2
train:
target: edit_dataset.EditDataset
params:
path: data/clip-filtered-dataset
split: train
min_resize_res: 256
max_resize_res: 256
crop_res: 256
flip_prob: 0.5
validation:
target: edit_dataset.EditDataset
params:
path: data/clip-filtered-dataset
split: val
min_resize_res: 256
max_resize_res: 256
crop_res: 256
lightning:
callbacks:
image_logger:
target: main.ImageLogger
params:
batch_frequency: 2000
max_images: 2
increase_log_steps: False
trainer:
max_epochs: 100 # 2000 to 100
benchmark: True
accumulate_grad_batches: 8 # 4 to 8
check_val_every_n_epoch: 4run following command to generate data pairs and corresponding dataset structure.
cd project/path
python dataset_gen.pydataset path tree:
autodl-tmp/instruct-pix2pix/data/instruct-pix2pix-dataset-000/
├── 0000000/
│ ├── prompt.json
│ ├── 000000_0.jpg
│ └── 000000_1.jpg
├── 0000001/
│ ├── prompt.json
│ └── 000001_0.jpg
│ └── 000001_1.jpg
└── ...generate seeds.json for training:
cd autodl-tmp/instruct-pix2pix
python dataset_creation/prepare_dataset.py data/instruct-pix2pix-dataset-000our code only runs on dual GPU settings, before continue, make sure your device have dual GPU available.
run following command to start training:
python main.py --name default --base configs/train.yaml --train --gpus 0,1
run following command to generate evalutaion results and predictions.
python eval.py --ckpt logs/train_default/checkpoints/last.ckpt