From aec5c3ddda74956f6ffa84bd0453d5567d7e17bb Mon Sep 17 00:00:00 2001 From: Kajud Date: Sun, 27 Sep 2026 15:25:47 +0200 Subject: [PATCH 1/5] docs: diagnoza padnietego treningu ACT (torch w sandboxie), wznowienie do 4000 - LOG: run2 zabity przy 3528, checkpoint 3000 caly; import torch pada tylko w sandboxie narzedzia (WinError 1114 shm.dll), poza sandboxem CUDA dziala; resume bez symlinku last; kolizja dwoch sesji na jednym GPU i pliku logu; checkpoint 4000 (loss 0.575) tylko na laptopie - STATUS: akapit ACT uzupelniony o stan 15:24 i pulapki Co-Authored-By: Claude Fable 5.1 --- docs/LOG.md | 23 +++++++++++++++++++++++ docs/STATUS.md | 6 ++++++ 2 files changed, 29 insertions(+) diff --git a/docs/LOG.md b/docs/LOG.md index d7d9320..a003042 100644 --- a/docs/LOG.md +++ b/docs/LOG.md @@ -890,3 +890,26 @@ linii polecen trafilo wlasny push (README zawieral te slowa). LFS: 630 MB z 1 GB **Nastepny krok:** rollout checkpointu koncowego na Pi z wylacznikiem; jesli ruch w zla strone - ARM_FRAMES krok 1. Wiecej epizodow (`--resume=true`) i druga statyczna kamera, jesli polityka nie generalizuje po polozeniu szyszki. **Sprzet:** dotkniety zdalnie (odczyt kamery `lerobot-find-cameras` na Pi, kopiowanie plikow; ramie nie ruszane) + +## 2026-09-27 - frane + Claude - diagnoza padnietego treningu ACT, wznowienie +**Zrobione:** Pytanie frane "co sie dzieje z trenowaniem, wznowic?". Stan: run2 (`C:/Users/frane/outputs/act_so101_grasp2_run2`) +zabity przy kroku 3528 (restart sesji ~14:55), checkpointy 1000/2000/3000 z `training_state` cale. Blad `import torch` +(WinError 1114 `shm.dll`) z poprzedniej sesji to NIE uszkodzony torch: pliki torch 2.11+cu128 zgodne z RECORD (11821 +plikow), reczne ladowanie wszystkich DLL z `torch/lib` przechodzi. Import pada TYLKO wewnatrz sandboxa narzedzia Bash/ +PowerShell w Claude Code; z wylaczonym sandboxem (`dangerouslyDisableSandbox`) `import torch` + CUDA + matmul dzialaja. +Przeinstalowanie torcha (15:12, `torch_reinstall.log`) bylo niepotrzebne. Wznowienie: `python tools/train_win.py +--config_path=/checkpoints/003000/pretrained_model/train_config.json --resume=true` dziala BEZ symlinku `last` +(lerobot 0.6.1 bierze katalog checkpointu z `config_path`; "Resuming data order at epoch 1, sample 1544"). Odpalone +15:17:10 jako proces odlaczony (`Start-Process`, przezywa restart sesji), do 7000 krokow. +**Nie dziala / otwarte:** 20 s pozniej DRUGA sesja Claude (worktree `robot-pinecone-test-plan`) wznowila ten sam +checkpoint do tego samego katalogu z `--steps=4000` i do tego samego pliku `act_grasp2_run4.log` (ta sama numeracja) - +oba procesy dzielily GPU (1.7 zamiast 3.3 kroku/s) i obydwa zapisalyby `checkpoints/004000` w tej samej chwili. +Moj proces zabity (duplikat), trening drugiej sesji zostawiony: 15:24:25 checkpoint 4000 (loss 0.575 (l1 0.255)), koniec. +Kroki 4000 -> 7000 NIE trenowane; 4000 lezy tylko na laptopie (`.../act_so101_grasp2_run2/checkpoints/004000`). +Commity z checkpointami 2000/3000 (`models/`, LFS, branch `frane/act-training` w tamtym worktree) NIE sa na origin +(push nie doszedl; 2 x 207 MB przy 630 MB z 1 GB limitu LFS - kolejne checkpointy do LFS sie nie zmieszcza, wagi +na Pi przez scp jak w SETUP.md). Pulapka: dwie sesje przy jednym GPU/katalogu `outputs` - przed startem treningu +`Get-CimInstance Win32_Process | ? Name -eq python.exe` i wlasna nazwa logu. +**Nastepny krok:** po checkpoincie 4000 (druga sesja) rollout na Pi z wylacznikiem; jesli ma byc 7000 krokow, ta sama +komenda `--config_path=.../004000/...` `--resume=true` poza sandboxem, jeden proces naraz. +**Sprzet:** nie diff --git a/docs/STATUS.md b/docs/STATUS.md index bd573a0..af88608 100644 --- a/docs/STATUS.md +++ b/docs/STATUS.md @@ -91,6 +91,12 @@ zadania i przypisania na tablicy Projects (link nizej). Czego nie ma tutaj albo kroku 2000 (l1 0.31). Checkpointy 1000/2000 na Pi w `~/models/act_so101_grasp2//pretrained_model`, 1000 tez na masterze (`models/`, LFS). Rollout lokalnie na Pi (0.65 s na 100 akcji): komenda w SETUP.md "Rollout polityki ACT na Pi". NIE sprawdzone na robocie w chwili pisania. + 15:24: run2 padl przy 3528 (restart sesji); wznowienie z 3000 (`--config_path=.../003000/pretrained_model/ + train_config.json --resume=true`, bez `last`) doszlo do 4000 (loss 0.575 (l1 0.255)), checkpoint `C:/Users/frane/outputs/ + act_so101_grasp2_run2/checkpoints/004000` (laptop, NIE na Pi, NIE w LFS). Kroki 4000 -> 7000 NIE liczone. + Pulapka: `import torch` pada w sandboxie narzedzia Claude Code (WinError 1114 shm.dll) - trening poza sandboxem, + JEDEN proces naraz (dwie sesje naraz dzielily GPU i katalog, LOG 15:20). Checkpointy 2000/3000 w LFS tylko + lokalnie (branch `frane/act-training`, push nie doszedl; limit LFS 1 GB). - Zera stawow vs URDF i kamera na ramieniu: `tools/frame_check.py` (FK placo + werdykt operatora) i `tools/hand_eye_calib.py` (marker/collect/solve/predict, AX=XB) gotowe z testami (28), instrukcja `docs/ARM_FRAMES.md`. NIE uruchomione na sprzecie - do zrobienia przez sesje przy Pi. To warunek wstepny dla IK/GraspGenX. From 438e78fc165c065b630f9be0a748d17d73b9dc2f Mon Sep 17 00:00:00 2001 From: Kajud Date: Sun, 27 Sep 2026 15:28:32 +0200 Subject: [PATCH 2/5] docs/LOG: przywrocona linia Sprzet w wpisie o treningu ACT (zgubiona w merge) Co-Authored-By: Claude Fable 5.1 --- docs/LOG.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/docs/LOG.md b/docs/LOG.md index 2d34049..596cd4b 100644 --- a/docs/LOG.md +++ b/docs/LOG.md @@ -937,6 +937,8 @@ na Pi przez scp jak w SETUP.md). Pulapka: dwie sesje przy jednym GPU/katalogu `o `Get-CimInstance Win32_Process | ? Name -eq python.exe` i wlasna nazwa logu. **Nastepny krok:** po checkpoincie 4000 (druga sesja) rollout na Pi z wylacznikiem; jesli ma byc 7000 krokow, ta sama komenda `--config_path=.../004000/...` `--resume=true` poza sandboxem, jeden proces naraz. +**Sprzet:** nie + ## 2026-09-27 - pawel120 + Claude - panel zbiorczy robota (wizytowka) **Zrobione:** `tools/robot_panel.py` (:8090) + `robot_panel.html`: jeden panel zamiast trzech okien SSH (panel.md). Nadzorca uslug `pinecone_bot/supervisor.py` (start/stop SIGINT -> terminate -> kill, logi w pamieci i `logs/.log`, From 17cb7ed7096c735c7a37206d125102c37b2525d3 Mon Sep 17 00:00:00 2001 From: Kajud Date: Sun, 27 Sep 2026 15:33:45 +0200 Subject: [PATCH 3/5] docs: proba treningu 4000 -> 7000 zatrzymana na prosbe (przed checkpointem 5000) Co-Authored-By: Claude Fable 5.1 --- docs/LOG.md | 3 +++ docs/STATUS.md | 3 ++- 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/docs/LOG.md b/docs/LOG.md index 596cd4b..b4ea277 100644 --- a/docs/LOG.md +++ b/docs/LOG.md @@ -935,6 +935,9 @@ Commity z checkpointami 2000/3000 (`models/`, LFS, branch `frane/act-training` w (push nie doszedl; 2 x 207 MB przy 630 MB z 1 GB limitu LFS - kolejne checkpointy do LFS sie nie zmieszcza, wagi na Pi przez scp jak w SETUP.md). Pulapka: dwie sesje przy jednym GPU/katalogu `outputs` - przed startem treningu `Get-CimInstance Win32_Process | ? Name -eq python.exe` i wlasna nazwa logu. +15:30 wznowienie 4000 -> 7000 (`--config_path=.../004000/... --resume=true --steps=7000`, 3.4 kroku/s, loss 0.52 przy +~4300) ZATRZYMANE 15:34 na prosbe frane (nie obciazac laptopa) przy kroku ~4413, przed checkpointem 5000 - te ~400 +krokow przepadlo, stan nadal = checkpoint 4000. Log `act_grasp2_run5_to7000.err`. **Nastepny krok:** po checkpoincie 4000 (druga sesja) rollout na Pi z wylacznikiem; jesli ma byc 7000 krokow, ta sama komenda `--config_path=.../004000/...` `--resume=true` poza sandboxem, jeden proces naraz. **Sprzet:** nie diff --git a/docs/STATUS.md b/docs/STATUS.md index 9b7b254..8505c62 100644 --- a/docs/STATUS.md +++ b/docs/STATUS.md @@ -118,7 +118,8 @@ zadania i przypisania na tablicy Projects (link nizej). Czego nie ma tutaj albo ACT na Pi". NIE sprawdzone na robocie w chwili pisania. 15:24: run2 padl przy 3528 (restart sesji); wznowienie z 3000 (`--config_path=.../003000/pretrained_model/ train_config.json --resume=true`, bez `last`) doszlo do 4000 (loss 0.575 (l1 0.255)), checkpoint `C:/Users/frane/outputs/ - act_so101_grasp2_run2/checkpoints/004000` (laptop, NIE na Pi, NIE w LFS). Kroki 4000 -> 7000 NIE liczone. + act_so101_grasp2_run2/checkpoints/004000` (laptop, NIE na Pi, NIE w LFS). Kroki 4000 -> 7000 NIE liczone + (proba 15:30 zatrzymana na prosbe frane przy ~4413, przed checkpointem 5000; wznowic ta sama komenda od 004000). Pulapka: `import torch` pada w sandboxie narzedzia Claude Code (WinError 1114 shm.dll) - trening poza sandboxem, JEDEN proces naraz (dwie sesje naraz dzielily GPU i katalog, LOG 15:20). Checkpointy 2000/3000 w LFS tylko lokalnie (branch `frane/act-training`, push nie doszedl; limit LFS 1 GB). From d36a1fb544603e2ea689d20aadf74c5695ac8ab6 Mon Sep 17 00:00:00 2001 From: Kajud Date: Sun, 27 Sep 2026 16:05:32 +0200 Subject: [PATCH 4/5] docs: trening ACT dokonczony do 7000 (loss 0.301), checkpointy 5000-7000 na laptopie Co-Authored-By: Claude Fable 5.1 --- docs/LOG.md | 8 ++++++-- docs/STATUS.md | 4 ++-- 2 files changed, 8 insertions(+), 4 deletions(-) diff --git a/docs/LOG.md b/docs/LOG.md index b4ea277..9a1090c 100644 --- a/docs/LOG.md +++ b/docs/LOG.md @@ -938,8 +938,12 @@ na Pi przez scp jak w SETUP.md). Pulapka: dwie sesje przy jednym GPU/katalogu `o 15:30 wznowienie 4000 -> 7000 (`--config_path=.../004000/... --resume=true --steps=7000`, 3.4 kroku/s, loss 0.52 przy ~4300) ZATRZYMANE 15:34 na prosbe frane (nie obciazac laptopa) przy kroku ~4413, przed checkpointem 5000 - te ~400 krokow przepadlo, stan nadal = checkpoint 4000. Log `act_grasp2_run5_to7000.err`. -**Nastepny krok:** po checkpoincie 4000 (druga sesja) rollout na Pi z wylacznikiem; jesli ma byc 7000 krokow, ta sama -komenda `--config_path=.../004000/...` `--resume=true` poza sandboxem, jeden proces naraz. +15:42 wznowienie od 004000 raz jeszcze (run6), na prosbe frane pauza 15:52-15:57 przez NtSuspendProcess (GPU 0 %, +postep zachowany), potem laptop na baterii = 1.1 kroku/s (36 W), po zasilaczu 3.5 kroku/s. 16:05 KONIEC: krok 7000, +loss 0.301 (l1 0.198, kld 0.010); checkpointy 5000/6000/7000 w `.../act_so101_grasp2_run2/checkpoints/` (laptop). +Loss po krokach: 3000 0.79 -> 4000 0.575 -> 7000 0.301. Log `act_grasp2_run6_to7000.err`. +**Nastepny krok:** wagi 7000 na Pi przez scp (SETUP.md), rollout z wylacznikiem; jesli ruch w zla strone - ARM_FRAMES +krok 1. Checkpoint 7000 NIE do LFS (limit 1 GB). **Sprzet:** nie ## 2026-09-27 - pawel120 + Claude - panel zbiorczy robota (wizytowka) diff --git a/docs/STATUS.md b/docs/STATUS.md index 8505c62..d5573c6 100644 --- a/docs/STATUS.md +++ b/docs/STATUS.md @@ -118,8 +118,8 @@ zadania i przypisania na tablicy Projects (link nizej). Czego nie ma tutaj albo ACT na Pi". NIE sprawdzone na robocie w chwili pisania. 15:24: run2 padl przy 3528 (restart sesji); wznowienie z 3000 (`--config_path=.../003000/pretrained_model/ train_config.json --resume=true`, bez `last`) doszlo do 4000 (loss 0.575 (l1 0.255)), checkpoint `C:/Users/frane/outputs/ - act_so101_grasp2_run2/checkpoints/004000` (laptop, NIE na Pi, NIE w LFS). Kroki 4000 -> 7000 NIE liczone - (proba 15:30 zatrzymana na prosbe frane przy ~4413, przed checkpointem 5000; wznowic ta sama komenda od 004000). + act_so101_grasp2_run2/checkpoints/004000`. 16:05: trening DOKONCZONY do 7000 (loss 0.301, l1 0.198); checkpointy + 5000/6000/7000 w tym samym katalogu na laptopie, NIE na Pi, NIE w LFS (limit 1 GB) - na Pi przez scp (SETUP.md). Pulapka: `import torch` pada w sandboxie narzedzia Claude Code (WinError 1114 shm.dll) - trening poza sandboxem, JEDEN proces naraz (dwie sesje naraz dzielily GPU i katalog, LOG 15:20). Checkpointy 2000/3000 w LFS tylko lokalnie (branch `frane/act-training`, push nie doszedl; limit LFS 1 GB). From 497caed9856827e5f17e3994f0b24e71a71beeff Mon Sep 17 00:00:00 2001 From: Kajud Date: Sun, 27 Sep 2026 16:07:00 +0200 Subject: [PATCH 5/5] models: ACT checkpoint 7000 (so101_grasp2, loss 0.301) w Git LFS Co-Authored-By: Claude Fable 5.1 --- models/README.md | 7 +- .../007000/pretrained_model/config.json | 64 ++++++ .../007000/pretrained_model/model.safetensors | 3 + .../policy_postprocessor.json | 32 +++ ..._step_0_unnormalizer_processor.safetensors | 3 + .../pretrained_model/policy_preprocessor.json | 56 +++++ ...or_step_3_normalizer_processor.safetensors | 3 + .../007000/pretrained_model/train_config.json | 213 ++++++++++++++++++ 8 files changed, 379 insertions(+), 2 deletions(-) create mode 100644 models/act_so101_grasp2/007000/pretrained_model/config.json create mode 100644 models/act_so101_grasp2/007000/pretrained_model/model.safetensors create mode 100644 models/act_so101_grasp2/007000/pretrained_model/policy_postprocessor.json create mode 100644 models/act_so101_grasp2/007000/pretrained_model/policy_postprocessor_step_0_unnormalizer_processor.safetensors create mode 100644 models/act_so101_grasp2/007000/pretrained_model/policy_preprocessor.json create mode 100644 models/act_so101_grasp2/007000/pretrained_model/policy_preprocessor_step_3_normalizer_processor.safetensors create mode 100644 models/act_so101_grasp2/007000/pretrained_model/train_config.json diff --git a/models/README.md b/models/README.md index f3832bd..5ff2630 100644 --- a/models/README.md +++ b/models/README.md @@ -3,10 +3,13 @@ - `act_so101_grasp2/001000/pretrained_model/` - ACT po 1000 krokach (batch 8, AMP) na `datasets/so101_grasp2` (50 epizodow, kamera wrist), 2026-09-27 14:32, loss 2.28 przy kroku 800. Pierwszy uzywalny checkpoint do testu toru inferencji; kolejne (2000+) dochodza w miare treningu. +- `act_so101_grasp2/007000/pretrained_model/` - ten sam trening dokonczony do 7000 krokow (wznawiany z 3000 i 4000), + 2026-09-27 16:05, loss 0.301 (l1 0.198). TEN checkpoint do rolloutu. Posrednie 2000-6000 tylko na laptopie frane + (`C:/Users/frane/outputs/act_so101_grasp2_run2/checkpoints/`) - nie w LFS, limit 1 GB. `*.safetensors` sa w Git LFS: po klonie `git lfs install` i `git lfs pull`. Na Pi: - tar -cf - -C models/act_so101_grasp2/001000 pretrained_model | ssh robot@172.20.10.4 \ - 'mkdir -p ~/models/act_so101_grasp2/001000 && tar -xf - -C ~/models/act_so101_grasp2/001000' + tar -cf - -C models/act_so101_grasp2/007000 pretrained_model | ssh robot@172.20.10.4 \ + 'mkdir -p ~/models/act_so101_grasp2/007000 && tar -xf - -C ~/models/act_so101_grasp2/007000' Rollout: docs/SETUP.md, sekcja "Rollout polityki ACT na Pi". diff --git a/models/act_so101_grasp2/007000/pretrained_model/config.json b/models/act_so101_grasp2/007000/pretrained_model/config.json new file mode 100644 index 0000000..56b8732 --- /dev/null +++ b/models/act_so101_grasp2/007000/pretrained_model/config.json @@ -0,0 +1,64 @@ +{ + "type": "act", + "n_obs_steps": 1, + "input_features": { + "observation.state": { + "type": "STATE", + "shape": [ + 6 + ] + }, + "observation.images.wrist": { + "type": "VISUAL", + "shape": [ + 3, + 480, + 640 + ] + } + }, + "output_features": { + "action": { + "type": "ACTION", + "shape": [ + 6 + ] + } + }, + "device": "cuda", + "use_amp": true, + "use_peft": false, + "push_to_hub": false, + "repo_id": null, + "private": null, + "tags": null, + "license": null, + "pretrained_path": "C:\\Users\\frane\\outputs\\act_so101_grasp2_run2\\checkpoints\\004000\\pretrained_model", + "pretrained_revision": null, + "chunk_size": 100, + "n_action_steps": 100, + "normalization_mapping": { + "VISUAL": "MEAN_STD", + "STATE": "MEAN_STD", + "ACTION": "MEAN_STD" + }, + "vision_backbone": "resnet18", + "pretrained_backbone_weights": "ResNet18_Weights.IMAGENET1K_V1", + "replace_final_stride_with_dilation": 0, + "pre_norm": false, + "dim_model": 512, + "n_heads": 8, + "dim_feedforward": 3200, + "feedforward_activation": "relu", + "n_encoder_layers": 4, + "n_decoder_layers": 1, + "use_vae": true, + "latent_dim": 32, + "n_vae_encoder_layers": 4, + "temporal_ensemble_coeff": null, + "dropout": 0.1, + "kl_weight": 10.0, + "optimizer_lr": 1e-05, + "optimizer_weight_decay": 0.0001, + "optimizer_lr_backbone": 1e-05 +} \ No newline at end of file diff --git a/models/act_so101_grasp2/007000/pretrained_model/model.safetensors b/models/act_so101_grasp2/007000/pretrained_model/model.safetensors new file mode 100644 index 0000000..6feb330 --- /dev/null +++ b/models/act_so101_grasp2/007000/pretrained_model/model.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dff38ea7468116e1207834dcb1625173664d378256bd2e4c28926f619f554849 +size 206699736 diff --git a/models/act_so101_grasp2/007000/pretrained_model/policy_postprocessor.json b/models/act_so101_grasp2/007000/pretrained_model/policy_postprocessor.json new file mode 100644 index 0000000..9ccb342 --- /dev/null +++ b/models/act_so101_grasp2/007000/pretrained_model/policy_postprocessor.json @@ -0,0 +1,32 @@ +{ + "name": "policy_postprocessor", + "steps": [ + { + "registry_name": "unnormalizer_processor", + "config": { + "eps": 1e-08, + "features": { + "action": { + "type": "ACTION", + "shape": [ + 6 + ] + } + }, + "norm_map": { + "VISUAL": "MEAN_STD", + "STATE": "MEAN_STD", + "ACTION": "MEAN_STD" + } + }, + "state_file": "policy_postprocessor_step_0_unnormalizer_processor.safetensors" + }, + { + "registry_name": "device_processor", + "config": { + "device": "cpu", + "float_dtype": null + } + } + ] +} \ No newline at end of file diff --git a/models/act_so101_grasp2/007000/pretrained_model/policy_postprocessor_step_0_unnormalizer_processor.safetensors b/models/act_so101_grasp2/007000/pretrained_model/policy_postprocessor_step_0_unnormalizer_processor.safetensors new file mode 100644 index 0000000..4f8010f --- /dev/null +++ b/models/act_so101_grasp2/007000/pretrained_model/policy_postprocessor_step_0_unnormalizer_processor.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd55aa8853b5330a197f8233da7d20089f531ca43441ef2b3e49b4c8b9054805 +size 6560 diff --git a/models/act_so101_grasp2/007000/pretrained_model/policy_preprocessor.json b/models/act_so101_grasp2/007000/pretrained_model/policy_preprocessor.json new file mode 100644 index 0000000..3bc0020 --- /dev/null +++ b/models/act_so101_grasp2/007000/pretrained_model/policy_preprocessor.json @@ -0,0 +1,56 @@ +{ + "name": "policy_preprocessor", + "steps": [ + { + "registry_name": "rename_observations_processor", + "config": { + "rename_map": {} + } + }, + { + "registry_name": "to_batch_processor", + "config": {} + }, + { + "registry_name": "device_processor", + "config": { + "device": "cuda", + "float_dtype": null + } + }, + { + "registry_name": "normalizer_processor", + "config": { + "eps": 1e-08, + "features": { + "observation.state": { + "type": "STATE", + "shape": [ + 6 + ] + }, + "observation.images.wrist": { + "type": "VISUAL", + "shape": [ + 3, + 480, + 640 + ] + }, + "action": { + "type": "ACTION", + "shape": [ + 6 + ] + } + }, + "norm_map": { + "VISUAL": "MEAN_STD", + "STATE": "MEAN_STD", + "ACTION": "MEAN_STD" + } + }, + "state_file": "policy_preprocessor_step_3_normalizer_processor.safetensors" + } + ] +} \ No newline at end of file diff --git a/models/act_so101_grasp2/007000/pretrained_model/policy_preprocessor_step_3_normalizer_processor.safetensors b/models/act_so101_grasp2/007000/pretrained_model/policy_preprocessor_step_3_normalizer_processor.safetensors new file mode 100644 index 0000000..730d698 --- /dev/null +++ b/models/act_so101_grasp2/007000/pretrained_model/policy_preprocessor_step_3_normalizer_processor.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0b0d2f8d8755a2df4b972ca293957c845a263b7e4a261bba9fb9100a443185f +size 6512 diff --git a/models/act_so101_grasp2/007000/pretrained_model/train_config.json b/models/act_so101_grasp2/007000/pretrained_model/train_config.json new file mode 100644 index 0000000..1bdc135 --- /dev/null +++ b/models/act_so101_grasp2/007000/pretrained_model/train_config.json @@ -0,0 +1,213 @@ +{ + "dataset": { + "repo_id": "local/so101_grasp2", + "repo_type": "dataset", + "root": "C:/Users/frane/datasets/so101_grasp2", + "episodes": null, + "image_transforms": { + "enable": false, + "max_num_transforms": 3, + "random_order": false, + "tfs": { + "brightness": { + "weight": 1.0, + "type": "ColorJitter", + "kwargs": { + "brightness": [ + 0.8, + 1.2 + ] + } + }, + "contrast": { + "weight": 1.0, + "type": "ColorJitter", + "kwargs": { + "contrast": [ + 0.8, + 1.2 + ] + } + }, + "saturation": { + "weight": 1.0, + "type": "ColorJitter", + "kwargs": { + "saturation": [ + 0.5, + 1.5 + ] + } + }, + "hue": { + "weight": 1.0, + "type": "ColorJitter", + "kwargs": { + "hue": [ + -0.05, + 0.05 + ] + } + }, + "sharpness": { + "weight": 1.0, + "type": "SharpnessJitter", + "kwargs": { + "sharpness": [ + 0.5, + 1.5 + ] + } + }, + "affine": { + "weight": 1.0, + "type": "RandomAffine", + "kwargs": { + "degrees": [ + -5.0, + 5.0 + ], + "translate": [ + 0.05, + 0.05 + ] + } + } + } + }, + "revision": null, + "use_imagenet_stats": true, + "video_backend": "pyav", + "return_uint8": false, + "depth_output_unit": "mm", + "streaming": false, + "eval_split": 0.0 + }, + "env": null, + "policy": { + "type": "act", + "n_obs_steps": 1, + "input_features": { + "observation.state": { + "type": "STATE", + "shape": [ + 6 + ] + }, + "observation.images.wrist": { + "type": "VISUAL", + "shape": [ + 3, + 480, + 640 + ] + } + }, + "output_features": { + "action": { + "type": "ACTION", + "shape": [ + 6 + ] + } + }, + "device": "cuda", + "use_amp": true, + "use_peft": false, + "push_to_hub": false, + "repo_id": null, + "private": null, + "tags": null, + "license": null, + "pretrained_path": "C:\\Users\\frane\\outputs\\act_so101_grasp2_run2\\checkpoints\\004000\\pretrained_model", + "pretrained_revision": null, + "chunk_size": 100, + "n_action_steps": 100, + "normalization_mapping": { + "VISUAL": "MEAN_STD", + "STATE": "MEAN_STD", + "ACTION": "MEAN_STD" + }, + "vision_backbone": "resnet18", + "pretrained_backbone_weights": "ResNet18_Weights.IMAGENET1K_V1", + "replace_final_stride_with_dilation": 0, + "pre_norm": false, + "dim_model": 512, + "n_heads": 8, + "dim_feedforward": 3200, + "feedforward_activation": "relu", + "n_encoder_layers": 4, + "n_decoder_layers": 1, + "use_vae": true, + "latent_dim": 32, + "n_vae_encoder_layers": 4, + "temporal_ensemble_coeff": null, + "dropout": 0.1, + "kl_weight": 10.0, + "optimizer_lr": 1e-05, + "optimizer_weight_decay": 0.0001, + "optimizer_lr_backbone": 1e-05 + }, + "reward_model": null, + "output_dir": "C:\\Users\\frane\\outputs\\act_so101_grasp2_run2", + "job_name": "act_so101_grasp2_run2", + "resume": true, + "seed": 1000, + "cudnn_deterministic": false, + "num_workers": 0, + "batch_size": 8, + "prefetch_factor": 4, + "persistent_workers": true, + "dataloader_multiprocessing_context": "spawn", + "steps": 7000, + "env_eval_freq": 20000, + "log_freq": 100, + "eval_steps": 0, + "max_eval_samples": 0, + "tolerance_s": 0.0001, + "save_checkpoint": true, + "save_freq": 1000, + "use_policy_training_preset": true, + "optimizer": { + "type": "adamw", + "lr": 1e-05, + "weight_decay": 0.0001, + "grad_clip_norm": 10.0, + "betas": [ + 0.9, + 0.999 + ], + "eps": 1e-08 + }, + "scheduler": null, + "eval": { + "n_episodes": 50, + "batch_size": 11, + "use_async_envs": true, + "recording": false, + "recording_repo_id": null, + "recording_private": false + }, + "wandb": { + "enable": false, + "disable_artifact": false, + "project": "lerobot", + "entity": null, + "notes": null, + "run_id": null, + "mode": null, + "add_tags": true + }, + "peft": null, + "job": { + "target": null, + "image": "huggingface/lerobot-gpu:latest", + "timeout": "2d", + "detach": false, + "tags": [] + }, + "save_checkpoint_to_hub": false, + "sample_weighting": null, + "rename_map": {}, + "checkpoint_path": "C:\\Users\\frane\\outputs\\act_so101_grasp2_run2\\checkpoints\\004000" +} \ No newline at end of file