kevinkke233-maker commited on
Commit
1f593d0
·
1 Parent(s): ed60b2c

Add Stage 1 SWD-H checkpoints

Browse files
README.md CHANGED
@@ -22,7 +22,9 @@ GitHub: https://github.com/kevinkke233-maker/Light-MER
22
  | File | Description | Status |
23
  |---|---|---|
24
  | `light-mer-teacher-qwen3-8b.pth` | Qwen3-8B teacher checkpoint for Stage 1 SWD-H distillation | Released |
25
- | `light-mer-stage1-swdh-qwen3-0.6b.pth` | Qwen3-0.6B student after Stage 1 SWD-H distillation | COMING SOON |
 
 
26
  | `light-mer-stage2-mgrpo-qwen3-0.6b.pth` | Qwen3-0.6B student after Stage 2 M-GRPO refinement | COMING SOON |
27
 
28
  ## Usage
@@ -41,6 +43,18 @@ TEACHER_CKPT=checkpoints/light-mer-teacher-qwen3-8b.pth \
41
  bash scripts/train_stage1_swdh.sh
42
  ```
43
 
 
 
 
 
 
 
 
 
 
 
 
 
44
  ## License
45
 
46
  This checkpoint repository is released under the Apache License 2.0. Please also follow the licenses and usage terms of the external datasets and pretrained models used with Light-MER.
 
22
  | File | Description | Status |
23
  |---|---|---|
24
  | `light-mer-teacher-qwen3-8b.pth` | Qwen3-8B teacher checkpoint for Stage 1 SWD-H distillation | Released |
25
+ | `stage1-swdh-qwen3-0.6b/checkpoint_000060_loss_1.291.pth` | Recommended Qwen3-0.6B student checkpoint after Stage 1 SWD-H distillation | Released |
26
+ | `stage1-swdh-qwen3-0.6b/checkpoint_*.pth` | Stage 1 SWD-H checkpoint trajectory from epoch 5 to epoch 60 | Released |
27
+ | `stage1-swdh-qwen3-0.6b/config.yaml` | Stage 1 SWD-H training config used for the released checkpoints | Released |
28
  | `light-mer-stage2-mgrpo-qwen3-0.6b.pth` | Qwen3-0.6B student after Stage 2 M-GRPO refinement | COMING SOON |
29
 
30
  ## Usage
 
43
  bash scripts/train_stage1_swdh.sh
44
  ```
45
 
46
+ For Stage 1 inference or evaluation, download one of the released student checkpoints, for example:
47
+
48
+ ```text
49
+ checkpoints/light-mer-stage1-swdh-qwen3-0.6b.pth
50
+ ```
51
+
52
+ The recommended Stage 1 checkpoint is:
53
+
54
+ ```text
55
+ stage1-swdh-qwen3-0.6b/checkpoint_000060_loss_1.291.pth
56
+ ```
57
+
58
  ## License
59
 
60
  This checkpoint repository is released under the Apache License 2.0. Please also follow the licenses and usage terms of the external datasets and pretrained models used with Light-MER.
stage1-swdh-qwen3-0.6b/checkpoint_000005_loss_2.097.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9c3a889b52211fcca62587b17e81c0ebb4ac2846543a6d4ca7b0f50f6c1ce6ff
3
+ size 499014241
stage1-swdh-qwen3-0.6b/checkpoint_000010_loss_1.904.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:98ea9d9a61f5ce3b14259fd3f4088099350e7a654e918c1eda2cc05c95e52393
3
+ size 499014241
stage1-swdh-qwen3-0.6b/checkpoint_000015_loss_1.759.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:92a4bf0aacbaf0c7b3978bd3fac9741c84063a259ac849ce1a2eb435f578cced
3
+ size 499014241
stage1-swdh-qwen3-0.6b/checkpoint_000020_loss_1.635.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:193bb449def5caf06ce7de55d24b666e9fc648a90843904a4d61b8a5c6dd7130
3
+ size 499014241
stage1-swdh-qwen3-0.6b/checkpoint_000025_loss_1.549.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d8b56153385529e1448797d3f13e36e7fa1c1c5823a7d7dd687dcc587dcb5468
3
+ size 499014241
stage1-swdh-qwen3-0.6b/checkpoint_000030_loss_1.465.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:55c83cbb0e45b278bc58b0472315aa9ab70265d977dfe59cecc10d55cff21ea0
3
+ size 499014241
stage1-swdh-qwen3-0.6b/checkpoint_000035_loss_1.415.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9361f2336b12fbcec403cd6b968df782807a61d96639b5337ccd7bd85f78e1a2
3
+ size 499014241
stage1-swdh-qwen3-0.6b/checkpoint_000040_loss_1.373.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:29cb00584d0aa2584be0feebca991ced28dbea17631c361b6c19876996246c45
3
+ size 499014241
stage1-swdh-qwen3-0.6b/checkpoint_000045_loss_1.337.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:60ad10494a6b258926d06377ff35bf86229e2453f384873e1469e61ca9063976
3
+ size 499014241
stage1-swdh-qwen3-0.6b/checkpoint_000050_loss_1.317.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cf6777c33d368ce5669aee096f3443731962b8c7328ccdc9e4265ded65781b11
3
+ size 499014241
stage1-swdh-qwen3-0.6b/checkpoint_000055_loss_1.300.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cb7a09215935e4d484099c967c64b7f2182241c57bedbc071e9dbadf5303e3ca
3
+ size 499014241
stage1-swdh-qwen3-0.6b/checkpoint_000060_loss_1.291.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dba3ea44640b460e1705a129ee9b90c2d4008c4a7b2026f5fcc455950873cab0
3
+ size 499014241
stage1-swdh-qwen3-0.6b/config.yaml ADDED
@@ -0,0 +1,128 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ## SWD-H sweep candidate: p=1 robust sliced Wasserstein
2
+ ## Validation target: three independent inference+score repeats with avg >= 74.16
3
+
4
+ model:
5
+ arch: affectgpt
6
+ model_type: pretrain_vicuna
7
+
8
+ frozen_video_proj: False
9
+ frozen_video_Qformer: False
10
+ frozen_audio_Qformer: False
11
+ frozen_audio_proj: False
12
+ frozen_multi_Qformer: False
13
+ frozen_multi_llama_proj: False
14
+ frozen_llm: False
15
+
16
+ multi_fusion_type: attention
17
+ video_fusion_type: attention
18
+ audio_fusion_type: attention
19
+ image_fusion_type: mean
20
+
21
+ ckpt: ""
22
+ ckpt_2: ""
23
+
24
+ llama_model: "Qwen3_0.6B"
25
+ acoustic_encoder: "HUBERT_BASE"
26
+ visual_encoder: "CLIP_VIT_BASE"
27
+
28
+ teacher:
29
+ llama_model: "Qwen3"
30
+ visual_encoder: "CLIP_VIT_LARGE"
31
+ acoustic_encoder: "HUBERT_LARGE"
32
+ ckpt: "/users/2780718z/sharedscratch/Year_3/OT/AffectGPT-master/AffectGPT/output/qwen3_emercoarse_highlevelfilter4_outputhybird_bestsetup_bestfusion_lz/qwen3_emercoarse_highlevelfilter4_outputhybird_bestsetup_bestfusion_lz_20260209171/checkpoint_000060_loss_0.609.pth"
33
+
34
+ use_swd: True
35
+ swd_n_projections: 200
36
+ swd_p: 1
37
+
38
+ ot_weight: 1.0
39
+ ot_ramp_steps: 5000
40
+ ot_epsilon: 0.1
41
+ ot_common_dim: 256
42
+ ot_num_iters: 20
43
+
44
+ kl_weight: 0.0
45
+ kl_temperature: 2.0
46
+
47
+ num_audio_query_token: 1
48
+ num_video_query_token: 1
49
+ num_multi_query_token: 1
50
+ num_image_query_token: 1
51
+ max_length: 1024
52
+
53
+ vis_processor:
54
+ train:
55
+ name: "alpro_video_train"
56
+ n_frms: 8
57
+ image_size: 224
58
+ text_processor:
59
+ train:
60
+ name: "blip_caption"
61
+ img_processor:
62
+ train:
63
+ name: "blip2_image_train"
64
+ image_size: 224
65
+
66
+ datasets:
67
+ mercaptionplus:
68
+ data_type: video
69
+ face_or_frame: "multiface_audio_face_text"
70
+ label_type: "hybird"
71
+
72
+ run:
73
+ task: video_text_pretrain
74
+
75
+ lr_sched: "linear_warmup_cosine_lr"
76
+ init_lr: 5e-5
77
+ min_lr: 5e-6
78
+ warmup_lr: 5e-7
79
+ weight_decay: 0.01
80
+
81
+ max_epoch: 60
82
+ iters_per_epoch: 5000
83
+ warmup_steps: 5000
84
+
85
+ batch_size_train: 3
86
+ batch_size_eval: 3
87
+ accum_grad_iters: 8
88
+
89
+ seed: 42
90
+ num_workers: 4
91
+
92
+ amp: True
93
+ resume_ckpt_path: null
94
+
95
+ evaluate: False
96
+ train_splits: ["train"]
97
+
98
+ device: "cuda"
99
+ world_size: 1
100
+ dist_url: "env://"
101
+ distributed: False
102
+
103
+ output_dir: "output/final_swd_h_qwen3_06b_swd_p1"
104
+
105
+ inference:
106
+ task: video_text_pretrain
107
+
108
+ vis_processor:
109
+ train:
110
+ name: "alpro_video_eval"
111
+ n_frms: 8
112
+ image_size: 224
113
+ text_processor:
114
+ train:
115
+ name: "blip_caption"
116
+ img_processor:
117
+ train:
118
+ name: "blip2_image_eval"
119
+ image_size: 224
120
+
121
+ face_or_frame: "xxx"
122
+ base_root: "output_final_swd_h_qwen3_06b_swd_p1/repeat1/results"
123
+ ckpt_root: xxx
124
+ ckpt_name: xxx
125
+ test_epoch: xxx
126
+ test_epochs: xxx-xxx
127
+ skip_epoch: 1
128
+ gpu: 0