Conghoan branch #1

CongHoanCoder · 2026-01-20T21:47:37Z

I make a few changes from your code. Please help me review it. Thanks

…a_sim_insertion_human

mdaiter · 2026-01-21T02:34:14Z

.ipynb_checkpoints/modeling_act-checkpoint.ipynb

Why is this being deleted entirely?

It's file is too heavy to commit so I shifted it to LFS. but I didn't change this file. so just reject this change

mdaiter · 2026-01-21T02:35:19Z

.gitattributes

 outputs/train/aloha_sim_transfer_cube_human/model_10000.safetensors filter=lfs diff=lfs merge=lfs -text
 resnet18-f37072fd.safetensors filter=lfs diff=lfs merge=lfs -text
 outputs/train/act_aloha_sim_transfer_cube_human/model_10000.safetensors filter=lfs diff=lfs merge=lfs -text
+*.ipynb filter=lfs diff=lfs merge=lfs -text


Why is this being added? Are you saying you want to keep .iypnb files as lfs hosts?

mdaiter · 2026-01-21T02:35:43Z

01. Test/test.py

This is good - agreed

Although we need to work on the naming. No spaces in the names. I'd create a directory called tests/ and put test.py under there

mdaiter · 2026-01-21T02:36:40Z

modeling_act.ipynb

Are you deleting this file entirely, or are you shifting it to LFS? If so, where's the LFS upload?

It's file is too heavy to commit so I shifted it to LFS. but I didn't change this file. so just reject this change

mdaiter · 2026-01-21T02:37:08Z

networks.py

 import tinygrad
 from tinygrad import Tensor, nn, dtypes
-from tinygrad.ops import Variable
+# from tinygrad.ops import Variable


No dangling comments. Why is this commented out? What shifted?

it has been shifted to "from tinygrad import Variable"

Where's that import? Is it needed now?

mdaiter · 2026-01-21T02:38:12Z

normalize.py

+                mean_val = stats[key]["mean"]
+                std_val = stats[key]["std"]
+                # Convert to numpy if needed (handle both numpy arrays and torch tensors)
+                if hasattr(mean_val, 'numpy'):
+                    mean_val = mean_val.numpy()
+                if hasattr(std_val, 'numpy'):
+                    std_val = std_val.numpy()
+                buffer["mean"].assign(mean_val)


This is okay, but it's just verbose. Is there any way to make this into a one or two liner?

mdaiter · 2026-01-21T02:38:26Z

normalize.py

+                min_val = stats[key]["min"]
+                max_val = stats[key]["max"]
+                # Convert to numpy if needed (handle both numpy arrays and torch tensors)
+                if hasattr(min_val, 'numpy'):
+                    min_val = min_val.numpy()
+                if hasattr(max_val, 'numpy'):
+                    max_val = max_val.numpy()
+                buffer["min"].assign(min_val)


(Read above)

mdaiter · 2026-01-21T02:38:50Z

outputs/train/act_aloha_sim_transfer_cube_human/model_10000.safetensors

Why is this deleted?

I back up your result and trained it again to compare results. You can reject the change.

mdaiter · 2026-01-21T02:39:48Z

outputs/train/aloha_sim_insertion_human/model_30000.safetensors

What changed?

I back up your result and trained it again to compare results. You can reject the change.

mdaiter · 2026-01-21T02:40:04Z

resnet.py

+        import pathlib
        resnet18_IMAGENET1K = ResNet(Block, [2, 2, 2, 2], num_classes=1000)
-        state_dict = nn.state.safe_load("resnet18-f37072fd.safetensors")
+        model_path = pathlib.Path(__file__).parent / "resnet18-f37072fd.safetensors"


mdaiter · 2026-01-21T02:40:46Z

test.py

+# parser.add_argument("--env_name", type=str, choices=['AlohaTransferCube-v0', 'AlohaInsertion-v0'], default='AlohaTransferCube-v0')
+parser.add_argument("--env_name", type=str, choices=['AlohaTransferCube-v0', 'AlohaInsertion-v0'], default='AlohaInsertion-v0')
+# parser.add_argument("--model_path", type=str, default='outputs/train/aloha_sim_transfer_cube_human/model_final.safetensors')
+# parser.add_argument("--model_path", type=str, default='outputs/train/aloha_sim_transfer_cube_human/model_final.safetensors')
+parser.add_argument("--model_path", type=str, default='outputs/train/aloha_sim_insertion_human/model_30000_original.safetensors')


No dangling comments. Also, why would we peg the model to open to 30,000 instead of final?

in the original version, there was only the model with 30000 steps. so I just run the train.py file again to get the models and compared the results. both 30000 steps and the final models are not good now. you can reject the change and help me to compare the results

mdaiter · 2026-01-21T02:41:03Z

test.py

    fps = env.metadata["render_fps"]

    # Encode all frames into a mp4 video.
-    video_path = output_directory / "rollout.mp4"


why rollout3 vs rollout?

it is just for testing to compare your result before

mdaiter · 2026-01-21T02:41:20Z

train.py

+# from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset


Didn't know it shifted.

Again, no dangling comments

mdaiter · 2026-01-21T02:41:50Z

train.py

 # Start of training code
 parser=argparse.ArgumentParser(description="Argument Parser for ACT training on simulated environments")
-parser.add_argument("env_name", type=str, choices=['aloha_sim_transfer_cube_human', 'aloha_sim_insertion_human'], default='aloha_sim_insertion_human')
+parser.add_argument("--env_name", type=str, choices=['aloha_sim_transfer_cube_human', 'aloha_sim_insertion_human'], default='aloha_sim_insertion_human')


mdaiter · 2026-01-21T02:42:25Z

train.py

+    ########################################################################
+    # Handle unused parameters by assigning zero gradients
+    optimizers_list = [opt]
+    if cfg.train_backbone_separately:
+        optimizers_list.append(opt_backbone)
+    for optimizer in optimizers_list:
+        for param in optimizer.params:
+            if param.grad is None:
+                # Create zero gradient with same shape and device as parameter
+                param.grad = Tensor.zeros(*param.shape, device=param.device, requires_grad=False)


can you explain this a bit?

There were some parameters not have gradients so when I run it, I got the error as the attachment. I searched and it said that the issue is that opt.step() is being called without first ensuring all parameters have gradients. The problem is that not all parameters are used in the forward pass, so they don't get gradients. then I changed the code like this

got it - okay!

mdaiter · 2026-01-21T02:43:05Z

train.py

+            batch_converted = {}
+            for k, v in batch.items():
+                if isinstance(v, torch.Tensor):
+                    batch_converted[k] = Tensor(v.detach().cpu().numpy(), requires_grad=False)
+                else:
+                    batch_converted[k] = v  # Keep strings, lists, etc. as-is
+
+            batch = batch_converted


So this makes sense in theory. But why are you mixing Torch and tinygrad tensors? Where's it coming from?

with the original code, I have encountered with the error below. The issue was that the batch contains mixed data types - some are torch tensors and some are lists or other non-tensor types. so I changed it.

mdaiter · 2026-01-21T12:00:39Z

cool - @CongHoanCoder , just clean up the pull request (e.g. no commented out imports / code, that's sloppy. also no randomly deleted files). I can check again and validate on my side when you're ready

CongHoanCoder added 2 commits January 20, 2026 14:00

updated normalize.py, train.py and the model 30000 steps for the aloh…

01a577e

…a_sim_insertion_human

modeling_act.ipynb updated

8a19821

mdaiter reviewed Jan 21, 2026

View reviewed changes

		# from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

		from lerobot.datasets.lerobot_dataset import LeRobotDataset

Conghoan branch #1

Are you sure you want to change the base?

Conghoan branch #1

Uh oh!

Conversation

CongHoanCoder commented Jan 20, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

CongHoanCoder Jan 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

CongHoanCoder Jan 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mdaiter commented Jan 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

CongHoanCoder Jan 21, 2026 •

edited

Loading

CongHoanCoder Jan 21, 2026 •

edited

Loading