walaa2022 commited on
Commit
b0a246e
·
verified ·
1 Parent(s): 97f3940

Upload 2 files

Browse files
Files changed (2) hide show
  1. run.sh +53 -0
  2. setup.py +7 -0
run.sh ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+
3
+ conda activate blip3o
4
+
5
+
6
+ export HF_HOME=/HF/Home/
7
+ export OUTPUT_FOLDER=/Your/Model/Output/
8
+ export IMG_FOLDER=/Your/Image/Folder
9
+
10
+
11
+ torchrun --nproc_per_node=8 \
12
+ blip3o/train/train_mem.py \
13
+ --deepspeed ./deepspeed_scripts/zero1.json \
14
+ --model_name_or_path Qwen/Qwen2.5-VL-7B-Instruct \
15
+ --version qwen \
16
+ --data_type "mix" \
17
+ --image_folder ${IMG_FOLDER} \
18
+ --gen_vision_tower eva-clip-E-14-plus \
19
+ --gen_projector_type mlp2x_gelu \
20
+ --mm_projector_type mlp2x_gelu \
21
+ --mm_vision_select_layer -2 \
22
+ --mm_use_im_start_end False \
23
+ --mm_use_im_patch_token False \
24
+ --bf16 True \
25
+ --output_dir ${OUTPUT_FOLDER} \
26
+ --num_train_epochs 1 \
27
+ --per_device_train_batch_size 16 \
28
+ --per_device_eval_batch_size 4 \
29
+ --gradient_accumulation_steps 1 \
30
+ --eval_strategy "no" \
31
+ --save_strategy "steps" \
32
+ --save_steps 1000 \
33
+ --save_total_limit 1 \
34
+ --learning_rate 1e-4 \
35
+ --weight_decay 0. \
36
+ --warmup_ratio 0.003 \
37
+ --lr_scheduler_type "cosine_with_min_lr" \
38
+ --lr_scheduler_kwargs '{"min_lr":1e-5}' \
39
+ --model_max_length 512 \
40
+ --logging_steps 1 \
41
+ --tf32 True \
42
+ --gradient_checkpointing True \
43
+ --dataloader_num_workers 4 \
44
+ --lazy_preprocess True \
45
+ --gen_pooling early_pool2d_4 \
46
+ --n_query 64 \
47
+ --n_und_query 0 \
48
+ --report_to none \
49
+ --run_name blip3o_qwen_vl_7b
50
+
51
+
52
+
53
+
setup.py ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ from setuptools import setup, find_packages
2
+
3
+ setup(
4
+ name="blip3o",
5
+ version="0.1.0",
6
+ packages=find_packages(),
7
+ )