solarian-diffusion 0.1.0__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- diffusion_policy/__init__.py +0 -0
- diffusion_policy/codecs/__init__.py +0 -0
- diffusion_policy/codecs/imagecodecs_numcodecs.py +1408 -0
- diffusion_policy/common/__init__.py +0 -0
- diffusion_policy/common/checkpoint_util.py +64 -0
- diffusion_policy/common/json_logger.py +117 -0
- diffusion_policy/common/normalize_util.py +251 -0
- diffusion_policy/common/pose_repr_util.py +112 -0
- diffusion_policy/common/pytorch_util.py +82 -0
- diffusion_policy/common/q_critic_util.py +135 -0
- diffusion_policy/common/replay_buffer.py +672 -0
- diffusion_policy/common/sampler.py +204 -0
- diffusion_policy/config/legacy/task/blockpush_lowdim_seed.yaml +34 -0
- diffusion_policy/config/legacy/task/blockpush_lowdim_seed_abs.yaml +34 -0
- diffusion_policy/config/legacy/task/can_image.yaml +64 -0
- diffusion_policy/config/legacy/task/can_image_abs.yaml +64 -0
- diffusion_policy/config/legacy/task/can_lowdim.yaml +45 -0
- diffusion_policy/config/legacy/task/can_lowdim_abs.yaml +46 -0
- diffusion_policy/config/legacy/task/kitchen_lowdim.yaml +33 -0
- diffusion_policy/config/legacy/task/kitchen_lowdim_abs.yaml +38 -0
- diffusion_policy/config/legacy/task/lift_abs.yaml +72 -0
- diffusion_policy/config/legacy/task/lift_image.yaml +64 -0
- diffusion_policy/config/legacy/task/lift_image_abs.yaml +63 -0
- diffusion_policy/config/legacy/task/lift_lowdim.yaml +46 -0
- diffusion_policy/config/legacy/task/lift_lowdim_abs.yaml +47 -0
- diffusion_policy/config/legacy/task/pusht_image.yaml +40 -0
- diffusion_policy/config/legacy/task/pusht_lowdim.yaml +34 -0
- diffusion_policy/config/legacy/task/real_pusht_image.yaml +47 -0
- diffusion_policy/config/legacy/task/square_abs.yaml +72 -0
- diffusion_policy/config/legacy/task/square_image.yaml +64 -0
- diffusion_policy/config/legacy/task/square_image_abs.yaml +64 -0
- diffusion_policy/config/legacy/task/square_lowdim.yaml +46 -0
- diffusion_policy/config/legacy/task/square_lowdim_abs.yaml +47 -0
- diffusion_policy/config/legacy/task/tool_hang_abs.yaml +71 -0
- diffusion_policy/config/legacy/task/tool_hang_image.yaml +63 -0
- diffusion_policy/config/legacy/task/tool_hang_image_abs.yaml +63 -0
- diffusion_policy/config/legacy/task/tool_hang_lowdim.yaml +45 -0
- diffusion_policy/config/legacy/task/tool_hang_lowdim_abs.yaml +46 -0
- diffusion_policy/config/legacy/task/transport_image.yaml +75 -0
- diffusion_policy/config/legacy/task/transport_image_abs.yaml +75 -0
- diffusion_policy/config/legacy/task/transport_lowdim.yaml +49 -0
- diffusion_policy/config/legacy/task/transport_lowdim_abs.yaml +49 -0
- diffusion_policy/config/legacy/task/umi.yaml +45 -0
- diffusion_policy/config/legacy/task/umi_image.yaml +36 -0
- diffusion_policy/config/legacy/task/umi_teleop_image.yaml +41 -0
- diffusion_policy/config/legacy/train_bet_lowdim_workspace.yaml +132 -0
- diffusion_policy/config/legacy/train_diffusion_transformer_hybrid_workspace.yaml +141 -0
- diffusion_policy/config/legacy/train_diffusion_transformer_lowdim_kitchen_workspace.yaml +146 -0
- diffusion_policy/config/legacy/train_diffusion_transformer_lowdim_pusht_workspace.yaml +146 -0
- diffusion_policy/config/legacy/train_diffusion_transformer_lowdim_workspace.yaml +146 -0
- diffusion_policy/config/legacy/train_diffusion_transformer_real_hybrid_workspace.yaml +144 -0
- diffusion_policy/config/legacy/train_diffusion_transformer_umi_teleop_hybrid_workspace.yaml +156 -0
- diffusion_policy/config/legacy/train_diffusion_unet_ddim_hybrid_workspace.yaml +140 -0
- diffusion_policy/config/legacy/train_diffusion_unet_ddim_lowdim_workspace.yaml +146 -0
- diffusion_policy/config/legacy/train_diffusion_unet_hybrid_workspace.yaml +137 -0
- diffusion_policy/config/legacy/train_diffusion_unet_image_pretrained_workspace.yaml +149 -0
- diffusion_policy/config/legacy/train_diffusion_unet_image_workspace.yaml +150 -0
- diffusion_policy/config/legacy/train_diffusion_unet_lowdim_workspace.yaml +143 -0
- diffusion_policy/config/legacy/train_diffusion_unet_real_hybrid_workspace.yaml +141 -0
- diffusion_policy/config/legacy/train_diffusion_unet_real_image_workspace.yaml +152 -0
- diffusion_policy/config/legacy/train_diffusion_unet_real_pretrained_workspace.yaml +152 -0
- diffusion_policy/config/legacy/train_diffusion_unet_timm_workspace.yaml +145 -0
- diffusion_policy/config/legacy/train_diffusion_unet_umi_hybrid_workspace.yaml +141 -0
- diffusion_policy/config/legacy/train_diffusion_unet_umi_teleop_hybrid_workspace.yaml +153 -0
- diffusion_policy/config/legacy/train_diffusion_unet_umi_timm_workspace.yaml +145 -0
- diffusion_policy/config/legacy/train_diffusion_unet_umi_workspace.yaml +150 -0
- diffusion_policy/config/legacy/train_diffusion_unet_video_workspace.yaml +128 -0
- diffusion_policy/config/legacy/train_ibc_dfo_hybrid_workspace.yaml +111 -0
- diffusion_policy/config/legacy/train_ibc_dfo_lowdim_workspace.yaml +109 -0
- diffusion_policy/config/legacy/train_ibc_dfo_real_hybrid_workspace.yaml +111 -0
- diffusion_policy/config/legacy/train_robomimic_image_workspace.yaml +92 -0
- diffusion_policy/config/legacy/train_robomimic_lowdim_workspace.yaml +92 -0
- diffusion_policy/config/legacy/train_robomimic_real_image_workspace.yaml +92 -0
- diffusion_policy/config/rl_annotate.yaml +265 -0
- diffusion_policy/config/rl_annotate_per_frame.yaml +162 -0
- diffusion_policy/config/rl_annotate_per_frame_v4.yaml +162 -0
- diffusion_policy/config/rl_annotate_v2.yaml +141 -0
- diffusion_policy/config/rl_annotate_v3.yaml +165 -0
- diffusion_policy/config/task/lift.yaml +89 -0
- diffusion_policy/config/task/square.yaml +92 -0
- diffusion_policy/config/task/tool_hang.yaml +93 -0
- diffusion_policy/config/task/tool_hang_abs.yaml +71 -0
- diffusion_policy/config/task/umi.yaml +105 -0
- diffusion_policy/config/task/umi_bimanual.yaml +135 -0
- diffusion_policy/config/task/umi_cup_wild.yaml +106 -0
- diffusion_policy/config/task/umi_image.yaml +36 -0
- diffusion_policy/config/task/umi_pick_lego.yaml +106 -0
- diffusion_policy/config/task/umi_pick_single_lego_175.yaml +106 -0
- diffusion_policy/config/task/umi_pick_single_lego_bbox.yaml +135 -0
- diffusion_policy/config/task/umi_teleop_image.yaml +41 -0
- diffusion_policy/config/train_diffusion_transformer_umi_bimanual_workspace.yaml +185 -0
- diffusion_policy/config/train_diffusion_transformer_umi_workspace.yaml +185 -0
- diffusion_policy/config/train_diffusion_unet_image_workspace.yaml +152 -0
- diffusion_policy/config/train_diffusion_unet_timm_robomimic_workspace.yaml +152 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_cup_wild.yaml +260 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_lego_pile_bbox.yaml +265 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_lego_pile_bbox_awr.yaml +324 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_lego_pile_bbox_parl.yaml +399 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_single_lego.yaml +258 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_single_lego_125.yaml +260 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_single_lego_175.yaml +260 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_single_lego_bbox.yaml +265 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_two_legos_bbox.yaml +265 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_workspace.yaml +224 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_workspace_lego.yaml +213 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_workspace_lego_resume.yaml +213 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_workspace_pick_lego.yaml +254 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_workspace_red_cup.yaml +219 -0
- diffusion_policy/config/train_diffusion_unet_umi_bimanual_workspace.yaml +203 -0
- diffusion_policy/config/train_iql_umi.yaml +186 -0
- diffusion_policy/config/train_iql_umi_v2.yaml +212 -0
- diffusion_policy/config/train_iql_umi_v3.yaml +212 -0
- diffusion_policy/dataset/__init__.py +0 -0
- diffusion_policy/dataset/base_dataset.py +76 -0
- diffusion_policy/dataset/umi_dataset.py +652 -0
- diffusion_policy/env_runner/__init__.py +0 -0
- diffusion_policy/env_runner/base_image_runner.py +9 -0
- diffusion_policy/env_runner/real_pusht_image_runner.py +10 -0
- diffusion_policy/model/__init__.py +0 -0
- diffusion_policy/model/common/__init__.py +0 -0
- diffusion_policy/model/common/dict_of_tensor_mixin.py +38 -0
- diffusion_policy/model/common/lr_scheduler.py +46 -0
- diffusion_policy/model/common/module_attr_mixin.py +14 -0
- diffusion_policy/model/common/normalizer.py +353 -0
- diffusion_policy/model/common/tensor_util.py +960 -0
- diffusion_policy/model/diffusion/__init__.py +0 -0
- diffusion_policy/model/diffusion/conditional_unet1d.py +238 -0
- diffusion_policy/model/diffusion/conv1d_components.py +46 -0
- diffusion_policy/model/diffusion/ema_model.py +88 -0
- diffusion_policy/model/diffusion/mask_generator.py +221 -0
- diffusion_policy/model/diffusion/positional_embedding.py +17 -0
- diffusion_policy/model/vision/__init__.py +0 -0
- diffusion_policy/model/vision/crop_randomizer.py +288 -0
- diffusion_policy/model/vision/multi_image_obs_encoder.py +195 -0
- diffusion_policy/model/vision/timm_obs_encoder.py +525 -0
- diffusion_policy/policy/__init__.py +0 -0
- diffusion_policy/policy/base_image_policy.py +27 -0
- diffusion_policy/policy/diffusion_unet_image_policy.py +264 -0
- diffusion_policy/policy/diffusion_unet_timm_policy.py +607 -0
- diffusion_policy/policy/parl_action_optimizer.py +675 -0
- diffusion_policy/workspace/__init__.py +0 -0
- diffusion_policy/workspace/base_workspace.py +146 -0
- diffusion_policy/workspace/train_diffusion_unet_image_workspace.py +1585 -0
- solarian_diffusion/__init__.py +32 -0
- solarian_diffusion/config.py +31 -0
- solarian_diffusion/policy_server.py +316 -0
- solarian_diffusion/rtc.py +272 -0
- solarian_diffusion/service.py +16 -0
- solarian_diffusion/trainer.py +213 -0
- solarian_diffusion-0.1.0.dist-info/METADATA +126 -0
- solarian_diffusion-0.1.0.dist-info/RECORD +159 -0
- solarian_diffusion-0.1.0.dist-info/WHEEL +5 -0
- solarian_diffusion-0.1.0.dist-info/entry_points.txt +8 -0
- solarian_diffusion-0.1.0.dist-info/licenses/LICENSE +201 -0
- solarian_diffusion-0.1.0.dist-info/licenses/NOTICE +45 -0
- solarian_diffusion-0.1.0.dist-info/top_level.txt +3 -0
- umi/__init__.py +0 -0
- umi/common/__init__.py +0 -0
- umi/common/pose_util.py +115 -0
|
@@ -0,0 +1,152 @@
|
|
|
1
|
+
defaults:
|
|
2
|
+
- _self_
|
|
3
|
+
- task: lift_image_abs
|
|
4
|
+
|
|
5
|
+
name: train_diffusion_unet_timm
|
|
6
|
+
_target_: diffusion_policy.workspace.train_diffusion_unet_image_workspace.TrainDiffusionUnetImageWorkspace
|
|
7
|
+
|
|
8
|
+
task_name: ${task.name}
|
|
9
|
+
shape_meta: ${task.shape_meta}
|
|
10
|
+
exp_name: "default"
|
|
11
|
+
|
|
12
|
+
n_action_steps: 8
|
|
13
|
+
|
|
14
|
+
policy:
|
|
15
|
+
_target_: diffusion_policy.policy.diffusion_unet_timm_policy.DiffusionUnetTimmPolicy
|
|
16
|
+
|
|
17
|
+
shape_meta: ${shape_meta}
|
|
18
|
+
|
|
19
|
+
noise_scheduler:
|
|
20
|
+
_target_: diffusers.DDIMScheduler
|
|
21
|
+
num_train_timesteps: 50
|
|
22
|
+
beta_start: 0.0001
|
|
23
|
+
beta_end: 0.02
|
|
24
|
+
# beta_schedule is important
|
|
25
|
+
# this is the best we found
|
|
26
|
+
beta_schedule: squaredcos_cap_v2
|
|
27
|
+
clip_sample: True
|
|
28
|
+
set_alpha_to_one: True
|
|
29
|
+
steps_offset: 0
|
|
30
|
+
prediction_type: epsilon # or sample
|
|
31
|
+
|
|
32
|
+
obs_encoder:
|
|
33
|
+
_target_: diffusion_policy.model.vision.timm_obs_encoder.TimmObsEncoder
|
|
34
|
+
shape_meta: ${shape_meta}
|
|
35
|
+
model_name: 'resnet18'
|
|
36
|
+
# model_name: 'vit_base_patch16_224'
|
|
37
|
+
pretrained: False
|
|
38
|
+
global_pool: 'avg'
|
|
39
|
+
transforms:
|
|
40
|
+
# - _target_: torchvision.transforms.Resize
|
|
41
|
+
# size: 224
|
|
42
|
+
# antialias: True
|
|
43
|
+
- _target_: torchvision.transforms.RandomRotation
|
|
44
|
+
degrees:
|
|
45
|
+
- -5.0
|
|
46
|
+
- 5.0
|
|
47
|
+
expand: false
|
|
48
|
+
- _target_: torchvision.transforms.ColorJitter
|
|
49
|
+
brightness: 0.2
|
|
50
|
+
contrast: 0.5
|
|
51
|
+
saturation: 0.2
|
|
52
|
+
hue: 0.2
|
|
53
|
+
|
|
54
|
+
use_group_norm: True
|
|
55
|
+
share_rgb_model: False
|
|
56
|
+
imagenet_norm: True
|
|
57
|
+
|
|
58
|
+
num_inference_steps: 16
|
|
59
|
+
obs_as_global_cond: True
|
|
60
|
+
diffusion_step_embed_dim: 128
|
|
61
|
+
down_dims: [256,512,1024]
|
|
62
|
+
kernel_size: 5
|
|
63
|
+
n_groups: 8
|
|
64
|
+
cond_predict_scale: True
|
|
65
|
+
input_pertub: 0.1 # reference: https://github.com/forever208/DDPM-IP
|
|
66
|
+
inpaint_fixed_action_prefix: False
|
|
67
|
+
|
|
68
|
+
ema:
|
|
69
|
+
_target_: diffusion_policy.model.diffusion.ema_model.EMAModel
|
|
70
|
+
update_after_step: 0
|
|
71
|
+
inv_gamma: 1.0
|
|
72
|
+
power: 0.75
|
|
73
|
+
min_value: 0.0
|
|
74
|
+
max_value: 0.9999
|
|
75
|
+
|
|
76
|
+
dataloader:
|
|
77
|
+
batch_size: ${eval:'128 if "${task.name}" == "tool_hang" else 256'}
|
|
78
|
+
num_workers: 16
|
|
79
|
+
shuffle: True
|
|
80
|
+
pin_memory: True
|
|
81
|
+
persistent_workers: True
|
|
82
|
+
|
|
83
|
+
val_dataloader:
|
|
84
|
+
batch_size: ${eval:'128 if "${task.name}" == "tool_hang" else 256'}
|
|
85
|
+
num_workers: 16
|
|
86
|
+
shuffle: False
|
|
87
|
+
pin_memory: True
|
|
88
|
+
persistent_workers: True
|
|
89
|
+
|
|
90
|
+
optimizer:
|
|
91
|
+
_target_: torch.optim.AdamW
|
|
92
|
+
lr: 1.0e-4
|
|
93
|
+
betas: [0.95, 0.999]
|
|
94
|
+
eps: 1.0e-8
|
|
95
|
+
weight_decay: 1.0e-6
|
|
96
|
+
|
|
97
|
+
training:
|
|
98
|
+
device: "cuda:0"
|
|
99
|
+
seed: 42
|
|
100
|
+
debug: False
|
|
101
|
+
resume: True
|
|
102
|
+
# optimization
|
|
103
|
+
lr_scheduler: cosine
|
|
104
|
+
lr_warmup_steps: 500
|
|
105
|
+
num_epochs: 600
|
|
106
|
+
gradient_accumulate_every: 1
|
|
107
|
+
# EMA destroys performance when used with BatchNorm
|
|
108
|
+
# replace BatchNorm with GroupNorm.
|
|
109
|
+
use_ema: True
|
|
110
|
+
freeze_encoder: False
|
|
111
|
+
# training loop control
|
|
112
|
+
# in epochs
|
|
113
|
+
rollout_every: 50
|
|
114
|
+
checkpoint_every: 50
|
|
115
|
+
val_every: 1
|
|
116
|
+
sample_every: 5
|
|
117
|
+
# steps per epoch
|
|
118
|
+
max_train_steps: null
|
|
119
|
+
max_val_steps: null
|
|
120
|
+
# misc
|
|
121
|
+
tqdm_interval_sec: 1.0
|
|
122
|
+
|
|
123
|
+
logging:
|
|
124
|
+
project: diffusion_policy_debug
|
|
125
|
+
resume: True
|
|
126
|
+
mode: online
|
|
127
|
+
name: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
|
|
128
|
+
tags: ["${name}", "${task_name}", "${exp_name}"]
|
|
129
|
+
id: null
|
|
130
|
+
group: null
|
|
131
|
+
|
|
132
|
+
checkpoint:
|
|
133
|
+
topk:
|
|
134
|
+
monitor_key: test_mean_score
|
|
135
|
+
mode: max
|
|
136
|
+
k: 5
|
|
137
|
+
format_str: 'epoch_{epoch:04d}-test_mean_score_{test_mean_score:.3f}.ckpt'
|
|
138
|
+
save_last_ckpt: True
|
|
139
|
+
save_last_snapshot: False
|
|
140
|
+
|
|
141
|
+
multi_run:
|
|
142
|
+
run_dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
|
|
143
|
+
wandb_name_base: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
|
|
144
|
+
|
|
145
|
+
hydra:
|
|
146
|
+
job:
|
|
147
|
+
override_dirname: ${name}
|
|
148
|
+
run:
|
|
149
|
+
dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
|
|
150
|
+
sweep:
|
|
151
|
+
dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
|
|
152
|
+
subdir: ${hydra.job.num}
|
|
@@ -0,0 +1,260 @@
|
|
|
1
|
+
defaults:
|
|
2
|
+
- _self_
|
|
3
|
+
- task: umi_cup_wild
|
|
4
|
+
|
|
5
|
+
name: train_diffusion_unet_timm
|
|
6
|
+
_target_: diffusion_policy.workspace.train_diffusion_unet_image_workspace.TrainDiffusionUnetImageWorkspace
|
|
7
|
+
|
|
8
|
+
# task_name: ${task.name}
|
|
9
|
+
shape_meta: ${task.shape_meta}
|
|
10
|
+
exp_name: "default"
|
|
11
|
+
batch_size: 64
|
|
12
|
+
# visual_encoder: "vit_base_patch16_clip_224.openai"
|
|
13
|
+
# visual_encoder_tag: "vitb"
|
|
14
|
+
visual_encoder: "vit_large_patch14_clip_224"
|
|
15
|
+
visual_encoder_tag: "vitl"
|
|
16
|
+
# visual_encoder: "resnet34.a1_in1k"
|
|
17
|
+
# visual_encoder_tag: "resnet34"
|
|
18
|
+
model_type: "unet"
|
|
19
|
+
# this will be overridden at call time
|
|
20
|
+
num_gpus: 1
|
|
21
|
+
|
|
22
|
+
|
|
23
|
+
# Parsed from the dataset file name's hyphen-separated tokens.
|
|
24
|
+
# eg: pick_lego-black_mirror-real_200-sim_200-no_bbox.zarr.zip
|
|
25
|
+
# Resolver: train.py::parse_dataset_field. Last arg is the fallback default.
|
|
26
|
+
# mirror options: black_mirror, mirror, swap_mirror, no_mirror
|
|
27
|
+
mirror: ${parse_dataset_field:${task.dataset_path},mirror,no_mirror}
|
|
28
|
+
# real_data_size / sim_data_size are null when the filename has no real_<N>/sim_<N>
|
|
29
|
+
# token; the wandb tags below skip themselves in that case.
|
|
30
|
+
real_data_size: ${parse_dataset_field:${task.dataset_path},real_data_size,null}
|
|
31
|
+
sim_data_size: ${parse_dataset_field:${task.dataset_path},sim_data_size,null}
|
|
32
|
+
bbox: ${parse_dataset_field:${task.dataset_path},bbox,no_bbox}
|
|
33
|
+
task_name: ${parse_dataset_field:${task.dataset_path},data_task,pick_lego}
|
|
34
|
+
|
|
35
|
+
# Override dataset path for fine-tuning
|
|
36
|
+
# This will override the dataset_path in task/umi.yaml
|
|
37
|
+
# Uncomment and set to your new dataset path:
|
|
38
|
+
# task.dataset_path: "path/to/your/new/dataset.zarr.zip"
|
|
39
|
+
|
|
40
|
+
n_action_steps: 8
|
|
41
|
+
|
|
42
|
+
policy:
|
|
43
|
+
_target_: diffusion_policy.policy.diffusion_unet_timm_policy.DiffusionUnetTimmPolicy
|
|
44
|
+
|
|
45
|
+
shape_meta: ${shape_meta}
|
|
46
|
+
|
|
47
|
+
noise_scheduler:
|
|
48
|
+
_target_: diffusers.DDIMScheduler
|
|
49
|
+
num_train_timesteps: 50
|
|
50
|
+
beta_start: 0.0001
|
|
51
|
+
beta_end: 0.02
|
|
52
|
+
# beta_schedule is important
|
|
53
|
+
# this is the best we found
|
|
54
|
+
beta_schedule: squaredcos_cap_v2
|
|
55
|
+
clip_sample: True
|
|
56
|
+
set_alpha_to_one: True
|
|
57
|
+
steps_offset: 0
|
|
58
|
+
prediction_type: epsilon # or sample
|
|
59
|
+
|
|
60
|
+
obs_encoder:
|
|
61
|
+
_target_: diffusion_policy.model.vision.timm_obs_encoder.TimmObsEncoder
|
|
62
|
+
shape_meta: ${shape_meta}
|
|
63
|
+
|
|
64
|
+
##### from scratch #####
|
|
65
|
+
# model_name: 'vit_base_patch16_224'
|
|
66
|
+
# model_name: 'resnet34'
|
|
67
|
+
# model_name: 'vit_tiny_patch16_224'
|
|
68
|
+
# model_name: 'efficientnet_b0'
|
|
69
|
+
# model_name: 'efficientnet_b3'
|
|
70
|
+
# pretrained: False
|
|
71
|
+
# frozen: False
|
|
72
|
+
|
|
73
|
+
##### from scratch #####
|
|
74
|
+
# model_name: 'resnet34.a1_in1k'
|
|
75
|
+
model_name: ${visual_encoder}
|
|
76
|
+
# model_name: 'vit_large_patch14_clip_224' # Changed to match checkpoint cup_wild_vit_l_1img.ckpt (ViT-L with CLIP weights)
|
|
77
|
+
# model_name: 'convnext_base.clip_laion2b_augreg_ft_in12k'
|
|
78
|
+
pretrained: True
|
|
79
|
+
frozen: False
|
|
80
|
+
|
|
81
|
+
|
|
82
|
+
global_pool: ''
|
|
83
|
+
|
|
84
|
+
# 'avg' or 'max' or 'soft_attention' or 'spatial_embedding' or 'transformer' or 'attention_pool_2d'
|
|
85
|
+
feature_aggregation: 'attention_pool_2d'
|
|
86
|
+
# feature_aggregation: null
|
|
87
|
+
position_encording: 'sinusoidal' # 'learnable' or 'sinusoidal'. it only works for transformer
|
|
88
|
+
|
|
89
|
+
# it only works for resnet. 32 (7x7) or 16 (14x14)
|
|
90
|
+
downsample_ratio: 32
|
|
91
|
+
|
|
92
|
+
transforms:
|
|
93
|
+
- type: RandomCrop
|
|
94
|
+
ratio: 0.95
|
|
95
|
+
# RandomRotation is not compatible with BF16
|
|
96
|
+
# - _target_: torchvision.transforms.RandomRotation
|
|
97
|
+
# degrees:
|
|
98
|
+
# - -5.0
|
|
99
|
+
# - 5.0
|
|
100
|
+
# expand: false
|
|
101
|
+
- _target_: torchvision.transforms.ColorJitter
|
|
102
|
+
brightness: 0.3
|
|
103
|
+
contrast: 0.4
|
|
104
|
+
saturation: 0.5
|
|
105
|
+
hue: 0.08
|
|
106
|
+
# - _target_: torchvision.transforms.RandomPerspective
|
|
107
|
+
# distortion_scale: 0.5
|
|
108
|
+
# p: 1.0
|
|
109
|
+
# - _target_: torchvision.transforms.ElasticTransform
|
|
110
|
+
# alpha: 50.0
|
|
111
|
+
# sigma: 5.0
|
|
112
|
+
# - _target_: torchvision.transforms.RandomPerspective
|
|
113
|
+
# distortion_scale: 0.5
|
|
114
|
+
# p: 1.0
|
|
115
|
+
# - _target_: torchvision.transforms.ElasticTransform
|
|
116
|
+
# alpha: 50.0
|
|
117
|
+
# sigma: 5.0
|
|
118
|
+
# - _target_: torchvision.transforms.RandomCrop
|
|
119
|
+
# size: 192
|
|
120
|
+
|
|
121
|
+
|
|
122
|
+
use_group_norm: True
|
|
123
|
+
share_rgb_model: False
|
|
124
|
+
imagenet_norm: True
|
|
125
|
+
|
|
126
|
+
num_inference_steps: 16
|
|
127
|
+
obs_as_global_cond: True
|
|
128
|
+
diffusion_step_embed_dim: 128
|
|
129
|
+
down_dims: [256,512,1024]
|
|
130
|
+
kernel_size: 5
|
|
131
|
+
n_groups: 8
|
|
132
|
+
cond_predict_scale: True
|
|
133
|
+
input_pertub: 0.1 # reference: https://github.com/forever208/DDPM-IP
|
|
134
|
+
# train multiple diffusion steps per obs encoder forward pass
|
|
135
|
+
# doing this too much will cause instability!
|
|
136
|
+
train_diffusion_n_samples: 1
|
|
137
|
+
|
|
138
|
+
ema:
|
|
139
|
+
_target_: diffusion_policy.model.diffusion.ema_model.EMAModel
|
|
140
|
+
update_after_step: 0
|
|
141
|
+
inv_gamma: 1.0
|
|
142
|
+
power: 0.75
|
|
143
|
+
min_value: 0.0
|
|
144
|
+
max_value: 0.9999
|
|
145
|
+
|
|
146
|
+
dataloader:
|
|
147
|
+
batch_size: ${batch_size}
|
|
148
|
+
num_workers: 8
|
|
149
|
+
shuffle: True
|
|
150
|
+
pin_memory: True
|
|
151
|
+
persistent_workers: True
|
|
152
|
+
prefetch_factor: 4
|
|
153
|
+
|
|
154
|
+
val_dataloader:
|
|
155
|
+
batch_size: ${batch_size}
|
|
156
|
+
num_workers: 8
|
|
157
|
+
shuffle: False
|
|
158
|
+
pin_memory: True
|
|
159
|
+
persistent_workers: True
|
|
160
|
+
prefetch_factor: 4
|
|
161
|
+
|
|
162
|
+
# dataloader:
|
|
163
|
+
# batch_size: 64
|
|
164
|
+
# num_workers: 16
|
|
165
|
+
# shuffle: True
|
|
166
|
+
# pin_memory: True
|
|
167
|
+
# persistent_workers: True
|
|
168
|
+
|
|
169
|
+
# val_dataloader:
|
|
170
|
+
# batch_size: 64
|
|
171
|
+
# num_workers: 16
|
|
172
|
+
# shuffle: False
|
|
173
|
+
# pin_memory: True
|
|
174
|
+
# persistent_workers: True
|
|
175
|
+
|
|
176
|
+
optimizer:
|
|
177
|
+
_target_: torch.optim.AdamW
|
|
178
|
+
lr: 3.0e-4
|
|
179
|
+
betas: [0.95, 0.999]
|
|
180
|
+
eps: 1.0e-8
|
|
181
|
+
weight_decay: 1.0e-6
|
|
182
|
+
|
|
183
|
+
training:
|
|
184
|
+
device: "cuda:0"
|
|
185
|
+
seed: 42
|
|
186
|
+
debug: False
|
|
187
|
+
resume: False
|
|
188
|
+
# optimization
|
|
189
|
+
lr_scheduler: cosine
|
|
190
|
+
lr_warmup_steps: 500
|
|
191
|
+
num_epochs: 50
|
|
192
|
+
gradient_accumulate_every: 1
|
|
193
|
+
# EMA destroys performance when used with BatchNorm
|
|
194
|
+
# replace BatchNorm with GroupNorm.
|
|
195
|
+
use_ema: True
|
|
196
|
+
freeze_encoder: False
|
|
197
|
+
# training loop control
|
|
198
|
+
# in epochs
|
|
199
|
+
rollout_every: 10
|
|
200
|
+
checkpoint_every: 10
|
|
201
|
+
val_every: 1
|
|
202
|
+
sample_every: 5
|
|
203
|
+
# steps per epoch
|
|
204
|
+
max_train_steps: null
|
|
205
|
+
max_val_steps: null
|
|
206
|
+
# misc
|
|
207
|
+
tqdm_interval_sec: 1.0
|
|
208
|
+
|
|
209
|
+
logging:
|
|
210
|
+
project: umi
|
|
211
|
+
resume: False
|
|
212
|
+
mode: online
|
|
213
|
+
name: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
|
|
214
|
+
tags:
|
|
215
|
+
- "${name}"
|
|
216
|
+
- "${task_name}"
|
|
217
|
+
- "${exp_name}"
|
|
218
|
+
# visual encoder could be vit_base_patch16_clip_224.openai, resnet34, etc.
|
|
219
|
+
- "${visual_encoder_tag}"
|
|
220
|
+
# batch size for each gpu
|
|
221
|
+
- "batch_size:${eval:'${batch_size} * ${num_gpus}'}"
|
|
222
|
+
- "num_gpus:${num_gpus}"
|
|
223
|
+
# model type could be unet or transformer
|
|
224
|
+
- "${model_type}"
|
|
225
|
+
# dataset file name (joined; substring-searchable on wandb)
|
|
226
|
+
# Note: full ${task.dataset_path} dropped — wandb tags max 64 chars.
|
|
227
|
+
- "${eval:'\"${task.dataset_path}\".split(\"/\")[-1].rsplit(\".\",1)[0]'}"
|
|
228
|
+
- "${mirror}"
|
|
229
|
+
# Only emit real_/sim_ tags when the dataset filename actually carries
|
|
230
|
+
# a real_<N> / sim_<N> token (otherwise the resolver returns None and we
|
|
231
|
+
# render an empty string here, which is filtered out before wandb init).
|
|
232
|
+
- "${eval:'\"real_size:%d\" % ${real_data_size} if ${real_data_size} is not None else \"\"'}"
|
|
233
|
+
- "${eval:'\"sim_size:%d\" % ${sim_data_size} if ${sim_data_size} is not None else \"\"'}"
|
|
234
|
+
- "${eval:'(\"real\" if ${real_data_size} > 0 else \"no_real\") if ${real_data_size} is not None else \"\"'}"
|
|
235
|
+
- "${eval:'(\"sim\" if ${sim_data_size} > 0 else \"no_sim\") if ${sim_data_size} is not None else \"\"'}"
|
|
236
|
+
- "${bbox}"
|
|
237
|
+
id: null
|
|
238
|
+
group: ${task_name}_${model_type}
|
|
239
|
+
|
|
240
|
+
checkpoint:
|
|
241
|
+
topk:
|
|
242
|
+
monitor_key: train_loss
|
|
243
|
+
mode: min
|
|
244
|
+
k: 20
|
|
245
|
+
format_str: '{task_name}-{model_type}-{visual_encoder}-batch_{total_batch}-{dataset_file_name}-epoch_{epoch:04d}-train_loss_{train_loss:.3f}.ckpt'
|
|
246
|
+
save_last_ckpt: True
|
|
247
|
+
save_last_snapshot: False
|
|
248
|
+
|
|
249
|
+
multi_run:
|
|
250
|
+
run_dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
|
|
251
|
+
wandb_name_base: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
|
|
252
|
+
|
|
253
|
+
hydra:
|
|
254
|
+
job:
|
|
255
|
+
override_dirname: ${name}
|
|
256
|
+
run:
|
|
257
|
+
dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
|
|
258
|
+
sweep:
|
|
259
|
+
dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
|
|
260
|
+
subdir: ${hydra.job.num}
|
|
@@ -0,0 +1,265 @@
|
|
|
1
|
+
defaults:
|
|
2
|
+
- _self_
|
|
3
|
+
- task: umi_pick_single_lego_bbox
|
|
4
|
+
|
|
5
|
+
name: train_diffusion_unet_timm
|
|
6
|
+
_target_: diffusion_policy.workspace.train_diffusion_unet_image_workspace.TrainDiffusionUnetImageWorkspace
|
|
7
|
+
|
|
8
|
+
# task_name: ${task.name}
|
|
9
|
+
shape_meta: ${task.shape_meta}
|
|
10
|
+
exp_name: "default"
|
|
11
|
+
batch_size: 128
|
|
12
|
+
# visual_encoder: "vit_base_patch16_clip_224.openai"
|
|
13
|
+
# visual_encoder_tag: "vitb"
|
|
14
|
+
visual_encoder: "vit_large_patch14_clip_224"
|
|
15
|
+
visual_encoder_tag: "vitl"
|
|
16
|
+
# visual_encoder: "resnet34.a1_in1k"
|
|
17
|
+
# visual_encoder_tag: "resnet34"
|
|
18
|
+
model_type: "unet"
|
|
19
|
+
# this will be overridden at call time
|
|
20
|
+
num_gpus: 1
|
|
21
|
+
|
|
22
|
+
|
|
23
|
+
# Parsed from the dataset file name's hyphen-separated tokens.
|
|
24
|
+
# eg: pick_lego-black_mirror-real_200-sim_200-no_bbox.zarr.zip
|
|
25
|
+
# Resolver: train.py::parse_dataset_field. Last arg is the fallback default.
|
|
26
|
+
# mirror options: black_mirror, mirror, swap_mirror, no_mirror
|
|
27
|
+
mirror: ${parse_dataset_field:${task.dataset_path},mirror,no_mirror}
|
|
28
|
+
# real_data_size / sim_data_size are null when the filename has no real_<N>/sim_<N>
|
|
29
|
+
# token; the wandb tags below skip themselves in that case.
|
|
30
|
+
real_data_size: ${parse_dataset_field:${task.dataset_path},real_data_size,null}
|
|
31
|
+
sim_data_size: ${parse_dataset_field:${task.dataset_path},sim_data_size,null}
|
|
32
|
+
bbox: ${parse_dataset_field:${task.dataset_path},bbox,no_bbox}
|
|
33
|
+
# Surfaces the bbox integration mode (low_dim vs mask_channel) onto the
|
|
34
|
+
# wandb tag set so runs are filterable by it. Sourced from the task
|
|
35
|
+
# shape_meta so it tracks the actual model wiring, not a separate flag.
|
|
36
|
+
bbox_mode: ${task.shape_meta.obs.bbox.bbox_mode}
|
|
37
|
+
task_name: ${parse_dataset_field:${task.dataset_path},data_task,pick_lego}
|
|
38
|
+
|
|
39
|
+
# Override dataset path for fine-tuning
|
|
40
|
+
# This will override the dataset_path in task/umi.yaml
|
|
41
|
+
# Uncomment and set to your new dataset path:
|
|
42
|
+
# task.dataset_path: "path/to/your/new/dataset.zarr.zip"
|
|
43
|
+
|
|
44
|
+
n_action_steps: 8
|
|
45
|
+
|
|
46
|
+
policy:
|
|
47
|
+
_target_: diffusion_policy.policy.diffusion_unet_timm_policy.DiffusionUnetTimmPolicy
|
|
48
|
+
|
|
49
|
+
shape_meta: ${shape_meta}
|
|
50
|
+
|
|
51
|
+
noise_scheduler:
|
|
52
|
+
_target_: diffusers.DDIMScheduler
|
|
53
|
+
num_train_timesteps: 50
|
|
54
|
+
beta_start: 0.0001
|
|
55
|
+
beta_end: 0.02
|
|
56
|
+
# beta_schedule is important
|
|
57
|
+
# this is the best we found
|
|
58
|
+
beta_schedule: squaredcos_cap_v2
|
|
59
|
+
clip_sample: True
|
|
60
|
+
set_alpha_to_one: True
|
|
61
|
+
steps_offset: 0
|
|
62
|
+
prediction_type: epsilon # or sample
|
|
63
|
+
|
|
64
|
+
obs_encoder:
|
|
65
|
+
_target_: diffusion_policy.model.vision.timm_obs_encoder.TimmObsEncoder
|
|
66
|
+
shape_meta: ${shape_meta}
|
|
67
|
+
|
|
68
|
+
##### from scratch #####
|
|
69
|
+
# model_name: 'vit_base_patch16_224'
|
|
70
|
+
# model_name: 'resnet34'
|
|
71
|
+
# model_name: 'vit_tiny_patch16_224'
|
|
72
|
+
# model_name: 'efficientnet_b0'
|
|
73
|
+
# model_name: 'efficientnet_b3'
|
|
74
|
+
# pretrained: False
|
|
75
|
+
# frozen: False
|
|
76
|
+
|
|
77
|
+
##### from scratch #####
|
|
78
|
+
# model_name: 'resnet34.a1_in1k'
|
|
79
|
+
model_name: ${visual_encoder}
|
|
80
|
+
# model_name: 'vit_large_patch14_clip_224' # Changed to match checkpoint cup_wild_vit_l_1img.ckpt (ViT-L with CLIP weights)
|
|
81
|
+
# model_name: 'convnext_base.clip_laion2b_augreg_ft_in12k'
|
|
82
|
+
pretrained: True
|
|
83
|
+
frozen: False
|
|
84
|
+
|
|
85
|
+
|
|
86
|
+
global_pool: ''
|
|
87
|
+
|
|
88
|
+
# 'avg' or 'max' or 'soft_attention' or 'spatial_embedding' or 'transformer' or 'attention_pool_2d'
|
|
89
|
+
feature_aggregation: 'attention_pool_2d'
|
|
90
|
+
# feature_aggregation: null
|
|
91
|
+
position_encording: 'sinusoidal' # 'learnable' or 'sinusoidal'. it only works for transformer
|
|
92
|
+
|
|
93
|
+
# it only works for resnet. 32 (7x7) or 16 (14x14)
|
|
94
|
+
downsample_ratio: 32
|
|
95
|
+
|
|
96
|
+
transforms:
|
|
97
|
+
- type: RandomCrop
|
|
98
|
+
ratio: 0.95
|
|
99
|
+
# RandomRotation is not compatible with BF16
|
|
100
|
+
# - _target_: torchvision.transforms.RandomRotation
|
|
101
|
+
# degrees:
|
|
102
|
+
# - -5.0
|
|
103
|
+
# - 5.0
|
|
104
|
+
# expand: false
|
|
105
|
+
- _target_: torchvision.transforms.ColorJitter
|
|
106
|
+
brightness: 0.3
|
|
107
|
+
contrast: 0.4
|
|
108
|
+
saturation: 0.5
|
|
109
|
+
hue: 0.08
|
|
110
|
+
# - _target_: torchvision.transforms.RandomPerspective
|
|
111
|
+
# distortion_scale: 0.5
|
|
112
|
+
# p: 1.0
|
|
113
|
+
# - _target_: torchvision.transforms.ElasticTransform
|
|
114
|
+
# alpha: 50.0
|
|
115
|
+
# sigma: 5.0
|
|
116
|
+
# - _target_: torchvision.transforms.RandomPerspective
|
|
117
|
+
# distortion_scale: 0.5
|
|
118
|
+
# p: 1.0
|
|
119
|
+
# - _target_: torchvision.transforms.ElasticTransform
|
|
120
|
+
# alpha: 50.0
|
|
121
|
+
# sigma: 5.0
|
|
122
|
+
# - _target_: torchvision.transforms.RandomCrop
|
|
123
|
+
# size: 192
|
|
124
|
+
|
|
125
|
+
|
|
126
|
+
use_group_norm: True
|
|
127
|
+
share_rgb_model: False
|
|
128
|
+
imagenet_norm: True
|
|
129
|
+
|
|
130
|
+
num_inference_steps: 16
|
|
131
|
+
obs_as_global_cond: True
|
|
132
|
+
diffusion_step_embed_dim: 128
|
|
133
|
+
down_dims: [256,512,1024]
|
|
134
|
+
kernel_size: 5
|
|
135
|
+
n_groups: 8
|
|
136
|
+
cond_predict_scale: True
|
|
137
|
+
input_pertub: 0.1 # reference: https://github.com/forever208/DDPM-IP
|
|
138
|
+
# train multiple diffusion steps per obs encoder forward pass
|
|
139
|
+
# doing this too much will cause instability!
|
|
140
|
+
train_diffusion_n_samples: 1
|
|
141
|
+
|
|
142
|
+
ema:
|
|
143
|
+
_target_: diffusion_policy.model.diffusion.ema_model.EMAModel
|
|
144
|
+
update_after_step: 0
|
|
145
|
+
inv_gamma: 1.0
|
|
146
|
+
power: 0.75
|
|
147
|
+
min_value: 0.0
|
|
148
|
+
max_value: 0.9999
|
|
149
|
+
|
|
150
|
+
dataloader:
|
|
151
|
+
batch_size: ${batch_size}
|
|
152
|
+
num_workers: 8
|
|
153
|
+
shuffle: True
|
|
154
|
+
pin_memory: True
|
|
155
|
+
persistent_workers: True
|
|
156
|
+
prefetch_factor: 4
|
|
157
|
+
|
|
158
|
+
val_dataloader:
|
|
159
|
+
batch_size: ${batch_size}
|
|
160
|
+
num_workers: 8
|
|
161
|
+
shuffle: False
|
|
162
|
+
pin_memory: True
|
|
163
|
+
persistent_workers: True
|
|
164
|
+
prefetch_factor: 4
|
|
165
|
+
|
|
166
|
+
# dataloader:
|
|
167
|
+
# batch_size: 64
|
|
168
|
+
# num_workers: 16
|
|
169
|
+
# shuffle: True
|
|
170
|
+
# pin_memory: True
|
|
171
|
+
# persistent_workers: True
|
|
172
|
+
|
|
173
|
+
# val_dataloader:
|
|
174
|
+
# batch_size: 64
|
|
175
|
+
# num_workers: 16
|
|
176
|
+
# shuffle: False
|
|
177
|
+
# pin_memory: True
|
|
178
|
+
# persistent_workers: True
|
|
179
|
+
|
|
180
|
+
optimizer:
|
|
181
|
+
_target_: torch.optim.AdamW
|
|
182
|
+
lr: 3.0e-4
|
|
183
|
+
betas: [0.95, 0.999]
|
|
184
|
+
eps: 1.0e-8
|
|
185
|
+
weight_decay: 1.0e-6
|
|
186
|
+
|
|
187
|
+
training:
|
|
188
|
+
device: "cuda:0"
|
|
189
|
+
seed: 42
|
|
190
|
+
debug: False
|
|
191
|
+
resume: False
|
|
192
|
+
# optimization
|
|
193
|
+
lr_scheduler: cosine
|
|
194
|
+
lr_warmup_steps: 500
|
|
195
|
+
num_epochs: 200
|
|
196
|
+
gradient_accumulate_every: 1
|
|
197
|
+
# EMA destroys performance when used with BatchNorm
|
|
198
|
+
# replace BatchNorm with GroupNorm.
|
|
199
|
+
use_ema: True
|
|
200
|
+
freeze_encoder: False
|
|
201
|
+
# training loop control
|
|
202
|
+
# in epochs
|
|
203
|
+
rollout_every: 10
|
|
204
|
+
checkpoint_every: 10
|
|
205
|
+
val_every: 1
|
|
206
|
+
sample_every: 5
|
|
207
|
+
# steps per epoch
|
|
208
|
+
max_train_steps: null
|
|
209
|
+
max_val_steps: null
|
|
210
|
+
# misc
|
|
211
|
+
tqdm_interval_sec: 1.0
|
|
212
|
+
|
|
213
|
+
logging:
|
|
214
|
+
project: umi
|
|
215
|
+
resume: False
|
|
216
|
+
mode: online
|
|
217
|
+
name: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
|
|
218
|
+
tags:
|
|
219
|
+
- "${name}"
|
|
220
|
+
- "${task_name}"
|
|
221
|
+
- "${exp_name}"
|
|
222
|
+
# visual encoder could be vit_base_patch16_clip_224.openai, resnet34, etc.
|
|
223
|
+
- "${visual_encoder_tag}"
|
|
224
|
+
# batch size for each gpu
|
|
225
|
+
- "batch_size:${eval:'${batch_size} * ${num_gpus}'}"
|
|
226
|
+
- "num_gpus:${num_gpus}"
|
|
227
|
+
# model type could be unet or transformer
|
|
228
|
+
- "${model_type}"
|
|
229
|
+
# dataset file name (joined; substring-searchable on wandb)
|
|
230
|
+
# Note: full ${task.dataset_path} dropped — wandb tags max 64 chars.
|
|
231
|
+
- "${eval:'\"${task.dataset_path}\".split(\"/\")[-1].rsplit(\".\",1)[0]'}"
|
|
232
|
+
- "${mirror}"
|
|
233
|
+
# Only emit real_/sim_ tags when the dataset filename actually carries
|
|
234
|
+
# a real_<N> / sim_<N> token (otherwise the resolver returns None and we
|
|
235
|
+
# render an empty string here, which is filtered out before wandb init).
|
|
236
|
+
- "${eval:'\"real_size:%d\" % ${real_data_size} if ${real_data_size} is not None else \"\"'}"
|
|
237
|
+
- "${eval:'\"sim_size:%d\" % ${sim_data_size} if ${sim_data_size} is not None else \"\"'}"
|
|
238
|
+
- "${eval:'(\"real\" if ${real_data_size} > 0 else \"no_real\") if ${real_data_size} is not None else \"\"'}"
|
|
239
|
+
- "${eval:'(\"sim\" if ${sim_data_size} > 0 else \"no_sim\") if ${sim_data_size} is not None else \"\"'}"
|
|
240
|
+
- "${bbox}"
|
|
241
|
+
- "bbox_mode:${bbox_mode}"
|
|
242
|
+
id: null
|
|
243
|
+
group: ${task_name}_${model_type}
|
|
244
|
+
|
|
245
|
+
checkpoint:
|
|
246
|
+
topk:
|
|
247
|
+
monitor_key: train_loss
|
|
248
|
+
mode: min
|
|
249
|
+
k: 20
|
|
250
|
+
format_str: '{task_name}-{model_type}-{visual_encoder}-batch_{total_batch}-{dataset_file_name}-epoch_{epoch:04d}-train_loss_{train_loss:.3f}.ckpt'
|
|
251
|
+
save_last_ckpt: True
|
|
252
|
+
save_last_snapshot: False
|
|
253
|
+
|
|
254
|
+
multi_run:
|
|
255
|
+
run_dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
|
|
256
|
+
wandb_name_base: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
|
|
257
|
+
|
|
258
|
+
hydra:
|
|
259
|
+
job:
|
|
260
|
+
override_dirname: ${name}
|
|
261
|
+
run:
|
|
262
|
+
dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
|
|
263
|
+
sweep:
|
|
264
|
+
dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
|
|
265
|
+
subdir: ${hydra.job.num}
|