Vision-CAIR commited on
Commit
51fc377
Β·
1 Parent(s): dd78ee5

Delete train_configs

Browse files
train_configs/minigpt4_llama2_stage1_pretrain.yaml DELETED
@@ -1,55 +0,0 @@
1
- model:
2
- arch: minigpt4
3
- model_type: pretrain_llama2
4
-
5
-
6
- datasets:
7
- laion:
8
- vis_processor:
9
- train:
10
- name: "blip2_image_train"
11
- image_size: 224
12
- text_processor:
13
- train:
14
- name: "blip_caption"
15
- sample_ratio: 115
16
- cc_sbu:
17
- vis_processor:
18
- train:
19
- name: "blip2_image_train"
20
- image_size: 224
21
- text_processor:
22
- train:
23
- name: "blip_caption"
24
- sample_ratio: 14
25
-
26
-
27
- run:
28
- task: image_text_pretrain
29
- # optimizer
30
- lr_sched: "linear_warmup_cosine_lr"
31
- init_lr: 1e-4
32
- min_lr: 8e-5
33
- warmup_lr: 1e-6
34
-
35
- weight_decay: 0.05
36
- max_epoch: 4
37
- batch_size_train: 64
38
- batch_size_eval: 64
39
- num_workers: 4
40
- warmup_steps: 5000
41
- iters_per_epoch: 5000
42
-
43
- seed: 42
44
- output_dir: "output/minigpt4_stage1_pretrain"
45
-
46
- amp: True
47
- resume_ckpt_path: null
48
-
49
- evaluate: False
50
- train_splits: ["train"]
51
-
52
- device: "cuda"
53
- world_size: 1
54
- dist_url: "env://"
55
- distributed: True
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
train_configs/minigpt4_llama2_stage2_finetune.yaml DELETED
@@ -1,50 +0,0 @@
1
- model:
2
- arch: minigpt4
3
- model_type: pretrain_llama2
4
-
5
- max_txt_len: 160
6
- end_sym: "</s>"
7
- prompt_path: "prompts/alignment.txt"
8
- prompt_template: '[INST] {} [/INST] '
9
- ckpt: '/path/to/stage1/checkpoint/'
10
-
11
-
12
- datasets:
13
- cc_sbu_align:
14
- vis_processor:
15
- train:
16
- name: "blip2_image_train"
17
- image_size: 224
18
- text_processor:
19
- train:
20
- name: "blip_caption"
21
-
22
- run:
23
- task: image_text_pretrain
24
- # optimizer
25
- lr_sched: "linear_warmup_cosine_lr"
26
- init_lr: 3e-5
27
- min_lr: 1e-5
28
- warmup_lr: 1e-6
29
-
30
- weight_decay: 0.05
31
- max_epoch: 5
32
- iters_per_epoch: 200
33
- batch_size_train: 12
34
- batch_size_eval: 12
35
- num_workers: 4
36
- warmup_steps: 200
37
-
38
- seed: 42
39
- output_dir: "output/minigpt4_stage2_finetune"
40
-
41
- amp: True
42
- resume_ckpt_path: null
43
-
44
- evaluate: False
45
- train_splits: ["train"]
46
-
47
- device: "cuda"
48
- world_size: 1
49
- dist_url: "env://"
50
- distributed: True
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
train_configs/minigpt4_stage1_pretrain.yaml DELETED
@@ -1,55 +0,0 @@
1
- model:
2
- arch: minigpt4
3
- model_type: pretrain_vicuna0
4
-
5
-
6
- datasets:
7
- laion:
8
- vis_processor:
9
- train:
10
- name: "blip2_image_train"
11
- image_size: 224
12
- text_processor:
13
- train:
14
- name: "blip_caption"
15
- sample_ratio: 115
16
- cc_sbu:
17
- vis_processor:
18
- train:
19
- name: "blip2_image_train"
20
- image_size: 224
21
- text_processor:
22
- train:
23
- name: "blip_caption"
24
- sample_ratio: 14
25
-
26
-
27
- run:
28
- task: image_text_pretrain
29
- # optimizer
30
- lr_sched: "linear_warmup_cosine_lr"
31
- init_lr: 1e-4
32
- min_lr: 8e-5
33
- warmup_lr: 1e-6
34
-
35
- weight_decay: 0.05
36
- max_epoch: 4
37
- batch_size_train: 64
38
- batch_size_eval: 64
39
- num_workers: 4
40
- warmup_steps: 5000
41
- iters_per_epoch: 5000
42
-
43
- seed: 42
44
- output_dir: "output/minigpt4_stage1_pretrain"
45
-
46
- amp: True
47
- resume_ckpt_path: null
48
-
49
- evaluate: False
50
- train_splits: ["train"]
51
-
52
- device: "cuda"
53
- world_size: 1
54
- dist_url: "env://"
55
- distributed: True
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
train_configs/minigpt4_stage2_finetune.yaml DELETED
@@ -1,50 +0,0 @@
1
- model:
2
- arch: minigpt4
3
- model_type: pretrain_vicuna0
4
-
5
- max_txt_len: 160
6
- end_sym: "###"
7
- prompt_path: "prompts/alignment.txt"
8
- prompt_template: '###Human: {} ###Assistant: '
9
- ckpt: '/path/to/stage1/checkpoint/'
10
-
11
-
12
- datasets:
13
- cc_sbu_align:
14
- vis_processor:
15
- train:
16
- name: "blip2_image_train"
17
- image_size: 224
18
- text_processor:
19
- train:
20
- name: "blip_caption"
21
-
22
- run:
23
- task: image_text_pretrain
24
- # optimizer
25
- lr_sched: "linear_warmup_cosine_lr"
26
- init_lr: 3e-5
27
- min_lr: 1e-5
28
- warmup_lr: 1e-6
29
-
30
- weight_decay: 0.05
31
- max_epoch: 5
32
- iters_per_epoch: 200
33
- batch_size_train: 12
34
- batch_size_eval: 12
35
- num_workers: 4
36
- warmup_steps: 200
37
-
38
- seed: 42
39
- output_dir: "output/minigpt4_stage2_finetune"
40
-
41
- amp: True
42
- resume_ckpt_path: null
43
-
44
- evaluate: False
45
- train_splits: ["train"]
46
-
47
- device: "cuda"
48
- world_size: 1
49
- dist_url: "env://"
50
- distributed: True