cskrren commited on
Commit
3f4785f
·
verified ·
1 Parent(s): 8a5fa42

Keep v1 and EgoBench releases; simplify provenance and align training recipe

Browse files
README.md CHANGED
@@ -8,14 +8,21 @@ tags:
8
 
9
  Inference checkpoints for [InfiniHand](https://github.com/infinihand/InfiniHand).
10
 
11
- | Folder | File | MANO | Mask | Camera |
12
- | --- | --- | --- | --- | --- |
13
- | [v1](v1) | `infinihand_v1.pt` | Stage II 160k | 700k | 200k |
14
- | [egobench](egobench) | `infinihand_egobench.pt` | v3 Stage II 15k | v3 Stage II 30k | 200k |
15
- | [egobench_plus](egobench_plus) | `infinihand_egobench_plus.pt` | Same as egobench | Same as v1 | Same as v1 |
16
 
17
- The plus checkpoint keeps the egobench base and MANO weights and uses v1 mask and camera branches. Camera weights are identical across the three versions. Each folder contains its SHA256SUMS file.
 
 
 
18
 
19
- Files contain PyTorch tensor state dictionaries, without optimizer state or bundled MANO assets. Obtain MANO assets separately under their terms. Follow the code repository for environment setup and inference. The shared configuration uses `/path/to/...` placeholders. Third-party components retain their respective terms; see NOTICE.md in the code repository.
20
 
21
- Exports were checked for tensor names, shapes, finite values and exact serialized-weight readback. Version names do not imply independently validated benchmark rankings.
 
 
 
 
 
 
 
 
 
8
 
9
  Inference checkpoints for [InfiniHand](https://github.com/infinihand/InfiniHand).
10
 
11
+ ## Checkpoints
 
 
 
 
12
 
13
+ | Version | File | Training data |
14
+ | --- | --- | --- |
15
+ | [v1](v1) | `infinihand_v1.pt` | Our curated training corpus |
16
+ | [egobench](egobench) | `infinihand_egobench.pt` | EgoBench |
17
 
18
+ Each folder contains its SHA256SUMS file. Files contain PyTorch tensor state dictionaries, without optimizer state or bundled MANO assets. Obtain MANO assets separately under their terms.
19
 
20
+ ## Usage
21
+
22
+ Follow the [video and dataset inference guide](https://github.com/infinihand/InfiniHand#inference). Hand postprocessing is enabled by default; see [POST_PROCESS.md](https://github.com/infinihand/InfiniHand/blob/main/docs/POST_PROCESS.md). The shared model configuration uses `/path/to/...` placeholders.
23
+
24
+ ## Training configuration
25
+
26
+ The released training configurations use **Stage I: 300,000 steps** and **Stage II: 100,000 steps**. These describe the provided training recipe, not per-checkpoint iteration labels. See [DATA.md](https://github.com/infinihand/InfiniHand/blob/main/docs/DATA.md) and [training configurations](https://github.com/infinihand/InfiniHand/tree/main/configs/training).
27
+
28
+ Third-party components retain their respective terms; see [NOTICE.md](https://github.com/infinihand/InfiniHand/blob/main/NOTICE.md). Exports were checked for tensor names, shapes, finite values and exact serialized-weight readback. Version names do not imply independently validated benchmark rankings.
egobench_plus/SHA256SUMS DELETED
@@ -1 +0,0 @@
1
- 39e8c4cef3e07a17996850ad88d3bbc0ff3d448d27948d46829e92f6ab107533 infinihand_egobench_plus.pt
 
 
egobench_plus/infinihand_egobench_plus.pt DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:39e8c4cef3e07a17996850ad88d3bbc0ff3d448d27948d46829e92f6ab107533
3
- size 5585243331