checkpoint_D_D/config.yaml

fusion_mode: UVformer
Linear: true
load_from_checkpoint: 
from_scratch: false
resume_from_checkpoint: null
train_uvformer: false
train_uvformer_gripper: true
colour_aug: [0, 0, 0, 0]
occ_loss: false
occ_loss_weight: [0.5,0.25,0.25,0.25] # occ, r, g, b 
train_action: true
eval: true
rotate: true
data_path_list: []
loss_weight: {occ: 0, action: 1}
validation: false
data_tasks_groups: None
num_epochs: 20
window_size: 12
learning_rate: 0.0001
lr_scheduler: constant
calvin_dataset: task_D_D
batch_size_calvin: 5
save_dir: 

original_image_shapes:
  rgb_static: [200, 200, 3]
  rgb_gripper: [84, 84, 3]
cam_trained: ["rgb_static","rgb_gripper"]
image_feature_out: [0]
extrinsic_jitter: 0
UVformer:
  cam_trained: ["rgb_static","rgb_gripper"]
  original_image_shapes:
    rgb_static: [200, 200, 3]
    rgb_gripper: [84, 84, 3]
  type: DeformableTransformer
  transformer_config:
    task_name: LLD
    hidden_dim: 1024
    nhead: 8
    num_decoder_layers: 2
    dim_feedforward: 1024
    dropout: 0.1
    PreProcessing: {left: 0.5, right: 0.5, front: 0.5 , back: 0.5, dilate: 1}
    grid_resolution: [0.05, 0.05, 0.1]
    ref_z_range: [0.3, 0.8]
    rotate: true
    upsample: 2
    self_attn_type: cosformer
    reuse_ref_point: false
    norm: None
    use_ffn: false
PETR:
  hidden_dim: 1024
  depth_step: 0.05
  depth_num: 4
  depth_start: 0.05
  position_range: [-0.2, -0.2, -0.05, 0.2, 0.2, 0.3]


vision_encoder_path: ViT-L-14
replan: -1
visualize: false
vision_encoder_pretrained: openai
lm_path: 
calvin_conf_path: 
batch_size_sim: 1
calvin_seq_path: 
tokenizer_path: 
reset: false
diverse_inst: false
future_act_len: -1
cross_attn_every_n_layers: 4
run_name: RobotFlamingoDBG
use_media_placement_augmentation: false
offline: false
logging_steps: 100
clip_cache_dir: 
gradient_accumulation_steps: 1
openflamingo_checkpoint: 
delete_previous_checkpoint: false
seed: 42
loss_multiplier_calvin: 1.0
warmup_steps: 5000
local_rank: 0
weight_decay: 0.1
precision: fp32
workers: 4
train_num_samples_calvin: 100
dataset_resampled: true
dist_url: env://
dist_backend: nccl
horovod: false
no_set_device_rank: false
report_to_wandb: false
wandb_project: RoboFlamingo_D2D
wandb_entity: null
save_checkpoints_to_wandb: false
freeze_embed: false
use_gripper: true
use_state: true
hist_window: 1
eval_hist_size: -1
sep_resampler: false
train_params: -1
rgb_pad: 10
gripper_pad: 4
n_timesteps: 150
predict_epsilon: false
head_type: lstm
n_obs_steps: 6
diff_horizon: 32
last_action: false
use_hist: false
traj_cons: true
debug: false
sep_lm_head: false
clip_state: false
unfreeze_vit: false
text_aug: false
residual: false
tcp_rel: false
dif_ws: false
partial_data: false
freeze_sampler: false
fwd_pred: false
fwd_pred_hand: false
no_pretrain: false
real_data: false
no_image_patch: false
cotrain: false
batch_size_vl: 20
vl_task_weights: 0.005
global_latent: 1
save_every_iter: 10000
pad_length: -1
hidden_size: 768
decoder_type: lstm
data_type: calvin
min_window_size: 12
max_window_size: 24
llm_name: mpt_dolly_3b
pooling: max
multi_step_action: 1