Copy merv-base to merv-frozen
Browse files
merv-frozen/checkpoints/latest-checkpoint.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:4a7a574a43de24b167e5c8cf99a779bceb0b6036ceeb9cdaeae15d9ab81f54a7
|
3 |
+
size 27190883468
|
merv-frozen/config.json
ADDED
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"dataset": {
|
3 |
+
"align_stage_components": [
|
4 |
+
"download/videollava/valley_llavaimage.json",
|
5 |
+
"download/videollava"
|
6 |
+
],
|
7 |
+
"dataset_root_dir": "data",
|
8 |
+
"finetune_stage_components": [
|
9 |
+
"download/videollava/videochatgpt_llavaimage_tune.json",
|
10 |
+
"download/videollava"
|
11 |
+
],
|
12 |
+
"type": "videollava"
|
13 |
+
},
|
14 |
+
"hf_token": ".hf_token",
|
15 |
+
"model": {
|
16 |
+
"feature_fusion": "cross_attention_avg_lq",
|
17 |
+
"align_epochs": 1,
|
18 |
+
"align_global_batch_size": 256,
|
19 |
+
"align_learning_rate": 0.001,
|
20 |
+
"align_lr_scheduler_type": "linear-warmup+cosine-decay",
|
21 |
+
"align_max_grad_norm": 1.0,
|
22 |
+
"align_max_steps": null,
|
23 |
+
"align_per_device_batch_size": 16,
|
24 |
+
"align_train_strategy": "fsdp-shard-grad-op",
|
25 |
+
"align_warmup_ratio": 0.03,
|
26 |
+
"align_weight_decay": 0.0,
|
27 |
+
"arch_specifier": "no-align+3davg+linear",
|
28 |
+
"enable_gradient_checkpointing": true,
|
29 |
+
"enable_mixed_precision_training": true,
|
30 |
+
"finetune_epochs": 1,
|
31 |
+
"finetune_global_batch_size": 128,
|
32 |
+
"finetune_learning_rate": 2e-05,
|
33 |
+
"finetune_lr_scheduler_type": "linear-warmup+cosine-decay",
|
34 |
+
"finetune_max_grad_norm": 1.0,
|
35 |
+
"finetune_max_steps": null,
|
36 |
+
"finetune_per_device_batch_size": 8,
|
37 |
+
"finetune_train_strategy": "fsdp-full-shard",
|
38 |
+
"finetune_warmup_ratio": 0.03,
|
39 |
+
"finetune_weight_decay": 0.1,
|
40 |
+
"image_resize_strategy": "resize-naive",
|
41 |
+
"llm_backbone_id": "llama2-7b-pure",
|
42 |
+
"llm_max_length": 2048,
|
43 |
+
"model_id": "merv-base",
|
44 |
+
"num_frames": [
|
45 |
+
16,
|
46 |
+
16,
|
47 |
+
32,
|
48 |
+
16
|
49 |
+
],
|
50 |
+
"projector_token_length": 64,
|
51 |
+
"reduce_in_full_precision": false,
|
52 |
+
"type": "merv-base",
|
53 |
+
"video_backbone_ids": [
|
54 |
+
"languagebind-video-noclass",
|
55 |
+
"dinov2-video-all-tokens",
|
56 |
+
"vivit-google-b-all-no-cls-16frames",
|
57 |
+
"siglip-vit-b16-224px-all-no-cls"
|
58 |
+
],
|
59 |
+
"visual_feature_length": 1024
|
60 |
+
},
|
61 |
+
"pretrained_checkpoint": null,
|
62 |
+
"run_id": "merv-base",
|
63 |
+
"run_root_dir": "runs",
|
64 |
+
"seed": 7,
|
65 |
+
"stage": "finetune",
|
66 |
+
"trackers": [
|
67 |
+
"jsonl",
|
68 |
+
"wandb"
|
69 |
+
],
|
70 |
+
"type": "merv-base"
|
71 |
+
}
|
merv-frozen/config.yaml
ADDED
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
dataset:
|
2 |
+
align_stage_components:
|
3 |
+
- download/videollava/valley_llavaimage.json
|
4 |
+
- download/videollava
|
5 |
+
dataset_root_dir: data
|
6 |
+
finetune_stage_components:
|
7 |
+
- download/videollava/videochatgpt_llavaimage_tune.json
|
8 |
+
- download/videollava
|
9 |
+
type: videollava
|
10 |
+
hf_token: .hf_token
|
11 |
+
model:
|
12 |
+
align_epochs: 1
|
13 |
+
align_global_batch_size: 256
|
14 |
+
align_learning_rate: 0.001
|
15 |
+
align_lr_scheduler_type: linear-warmup+cosine-decay
|
16 |
+
align_max_grad_norm: 1.0
|
17 |
+
align_max_steps: null
|
18 |
+
align_per_device_batch_size: 16
|
19 |
+
align_train_strategy: fsdp-shard-grad-op
|
20 |
+
align_warmup_ratio: 0.03
|
21 |
+
align_weight_decay: 0.0
|
22 |
+
arch_specifier: no-align+3davg+linear
|
23 |
+
enable_gradient_checkpointing: true
|
24 |
+
enable_mixed_precision_training: true
|
25 |
+
feature_fusion: cross_attention_avg_lq
|
26 |
+
finetune_epochs: 1
|
27 |
+
finetune_global_batch_size: 128
|
28 |
+
finetune_learning_rate: 2.0e-05
|
29 |
+
finetune_lr_scheduler_type: linear-warmup+cosine-decay
|
30 |
+
finetune_max_grad_norm: 1.0
|
31 |
+
finetune_max_steps: null
|
32 |
+
finetune_per_device_batch_size: 8
|
33 |
+
finetune_train_strategy: fsdp-full-shard
|
34 |
+
finetune_warmup_ratio: 0.03
|
35 |
+
finetune_weight_decay: 0.1
|
36 |
+
image_resize_strategy: resize-naive
|
37 |
+
llm_backbone_id: llama2-7b-pure
|
38 |
+
llm_max_length: 2048
|
39 |
+
model_id: merv-base
|
40 |
+
num_frames:
|
41 |
+
- 16
|
42 |
+
- 16
|
43 |
+
- 32
|
44 |
+
- 16
|
45 |
+
projector_token_length: 64
|
46 |
+
reduce_in_full_precision: false
|
47 |
+
type: merv-base
|
48 |
+
video_backbone_ids:
|
49 |
+
- languagebind-video-noclass
|
50 |
+
- dinov2-video-all-tokens
|
51 |
+
- vivit-google-b-all-no-cls-16frames
|
52 |
+
- siglip-vit-b16-224px-all-no-cls
|
53 |
+
visual_feature_length: 1024
|
54 |
+
pretrained_checkpoint: null
|
55 |
+
run_id: merv-base
|
56 |
+
run_root_dir: runs
|
57 |
+
seed: 7
|
58 |
+
stage: finetune
|
59 |
+
trackers:
|
60 |
+
- jsonl
|
61 |
+
- wandb
|
62 |
+
type: merv-base
|