Skip to content

Commit b8bfb7c

Browse files
author
Ye Shaokai
committed
updates
1 parent 616a056 commit b8bfb7c

File tree

4 files changed

+142
-6
lines changed

4 files changed

+142
-6
lines changed

shaokai_train_avion_mc.sh renamed to shaokai_train_avion_mc_0.5b.sh

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -36,8 +36,8 @@ torchrun --nproc_per_node=4 \
3636
--image_grid_pinpoints "(1x1),...,(6x6)" \
3737
--mm_patch_merge_type spatial_unpad \
3838
--bf16 True \
39-
--run_name shaokai_llama_ov_0.5b_avion_mc \
40-
--output_dir experiments/shaokai_llama_ov_0.5b_avion_mc \
39+
--run_name shaokai_llava_ov_0.5b_avion_mc \
40+
--output_dir experiments/shaokai_llava_ov_0.5b_avion_mc \
4141
--num_train_epochs 1 \
4242
--per_device_train_batch_size 1 \
4343
--per_device_eval_batch_size 4 \
@@ -62,7 +62,7 @@ torchrun --nproc_per_node=4 \
6262
--dataloader_drop_last True \
6363
--frames_upbound 32 \
6464
--root /data/shaokai/EK100 \
65-
--action_predictions /data/shaokai/avaion_predictions_test.json \
65+
--action_predictions /data/shaokai/avion_predictions_test.json \
6666
--val_metadata /data/shaokai/epic-kitchens-100-annotations/EPIC_100_validation.csv \
6767
--llava_num_frames 16 \
6868
--topk_predictions 5 > train_kitchen_0.5b_avion_mc.out 2>&1

‎shaokai_train_avion_mc_7b.sh‎

Lines changed: 68 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,68 @@
1+
#!/bin/bash
2+
3+
# Export environment variables
4+
export CUDA_VISIBLE_DEVICES="0,1,2,3"
5+
export OMP_NUM_THREADS="8"
6+
export NCCL_IB_DISABLE="0"
7+
export NCCL_IB_GID_INDEX="3"
8+
export NCCL_SOCKET_IFNAME="eth0"
9+
export NCCL_DEBUG="INFO"
10+
export ACCELERATE_CPU_AFFINITY="1"
11+
export WANDB_API_KEY="4474ec79de023b0c3ffb43588ab6163264f875db"
12+
export HF_HOME=/data/shaokai
13+
14+
15+
# Run the command using torchrun
16+
torchrun --nproc_per_node=4 \
17+
--nnodes=1 \
18+
--node_rank=0 \
19+
--master_addr=127.0.0.1 \
20+
--master_port=29500 \
21+
llava/train/train_mem.py \
22+
--deepspeed scripts/zero3.json \
23+
--model_name_or_path lmms-lab/llava-onevision-qwen2-7b-ov \
24+
--version qwen_1_5 \
25+
--data_path scripts/train/EK100_avion_mc.yaml \
26+
--video_folder /data/shaokai/\
27+
--mm_tunable_parts mm_vision_tower,mm_mlp_adapter,mm_language_model \
28+
--mm_vision_tower_lr 2e-6 \
29+
--vision_tower google/siglip-so400m-patch14-384 \
30+
--mm_projector_type mlp2x_gelu \
31+
--mm_vision_select_layer -2 \
32+
--mm_use_im_start_end False \
33+
--mm_use_im_patch_token False \
34+
--group_by_modality_length True \
35+
--image_aspect_ratio anyres_max_9 \
36+
--image_grid_pinpoints "(1x1),...,(6x6)" \
37+
--mm_patch_merge_type spatial_unpad \
38+
--bf16 True \
39+
--run_name shaokai_llava_ov_7b_avion_mc \
40+
--output_dir experiments/shaokai_llava_ov_7b_avion_mc \
41+
--num_train_epochs 1 \
42+
--per_device_train_batch_size 1 \
43+
--per_device_eval_batch_size 4 \
44+
--gradient_accumulation_steps 2 \
45+
--evaluation_strategy steps \
46+
--eval_steps 500\
47+
--save_strategy steps \
48+
--save_steps 1000 \
49+
--learning_rate 1e-5 \
50+
--weight_decay 0. \
51+
--warmup_ratio 0.03 \
52+
--lr_scheduler_type cosine \
53+
--logging_steps 1 \
54+
--tf32 True \
55+
--model_max_length 32768 \
56+
--gradient_checkpointing True \
57+
--dataloader_num_workers 4 \
58+
--lazy_preprocess True \
59+
--report_to wandb \
60+
--torch_compile True \
61+
--torch_compile_backend inductor \
62+
--dataloader_drop_last True \
63+
--frames_upbound 32 \
64+
--root /data/shaokai/EK100 \
65+
--action_predictions /data/shaokai/avion_predictions_test.json \
66+
--val_metadata /data/shaokai/epic-kitchens-100-annotations/EPIC_100_validation.csv \
67+
--llava_num_frames 16 \
68+
--topk_predictions 5 > train_kitchen_7b_avion_mc.out 2>&1

shaokai_train_random_mc.sh renamed to shaokai_train_random_mc_0.5b.sh

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -36,8 +36,8 @@ torchrun --nproc_per_node=4 \
3636
--image_grid_pinpoints "(1x1),...,(6x6)" \
3737
--mm_patch_merge_type spatial_unpad \
3838
--bf16 True \
39-
--run_name shaokai_llama_ov_0.5b_random_mc \
40-
--output_dir experiments/shaokai_llama_ov_0.5b_random_mc \
39+
--run_name shaokai_llava_ov_0.5b_random_mc \
40+
--output_dir experiments/shaokai_llava_ov_0.5b_random_mc \
4141
--num_train_epochs 1 \
4242
--per_device_train_batch_size 1 \
4343
--per_device_eval_batch_size 4 \
@@ -62,7 +62,7 @@ torchrun --nproc_per_node=4 \
6262
--dataloader_drop_last True \
6363
--frames_upbound 32 \
6464
--root /data/shaokai/EK100 \
65-
--action_predictions /data/shaokai/avaion_predictions_test.json \
65+
--action_predictions /data/shaokai/avion_predictions_test.json \
6666
--val_metadata /data/shaokai/epic-kitchens-100-annotations/EPIC_100_validation.csv \
6767
--llava_num_frames 16 \
6868
--topk_predictions 5 > train_kitchen_0.5b_random_mc.out 2>&1

‎shaokai_train_random_mc_7b.sh‎

Lines changed: 68 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,68 @@
1+
#!/bin/bash
2+
3+
# Export environment variables
4+
export CUDA_VISIBLE_DEVICES="0,1,2,3"
5+
export OMP_NUM_THREADS="8"
6+
export NCCL_IB_DISABLE="0"
7+
export NCCL_IB_GID_INDEX="3"
8+
export NCCL_SOCKET_IFNAME="eth0"
9+
export NCCL_DEBUG="INFO"
10+
export ACCELERATE_CPU_AFFINITY="1"
11+
export WANDB_API_KEY="4474ec79de023b0c3ffb43588ab6163264f875db"
12+
export HF_HOME=/data/shaokai
13+
14+
15+
# Run the command using torchrun
16+
torchrun --nproc_per_node=4 \
17+
--nnodes=1 \
18+
--node_rank=0 \
19+
--master_addr=127.0.0.1 \
20+
--master_port=29500 \
21+
llava/train/train_mem.py \
22+
--deepspeed scripts/zero3.json \
23+
--model_name_or_path lmms-lab/llava-onevision-qwen2-7b-ov \
24+
--version qwen_1_5 \
25+
--data_path scripts/train/EK100_random_mc.yaml \
26+
--video_folder /data/shaokai/\
27+
--mm_tunable_parts mm_vision_tower,mm_mlp_adapter,mm_language_model \
28+
--mm_vision_tower_lr 2e-6 \
29+
--vision_tower google/siglip-so400m-patch14-384 \
30+
--mm_projector_type mlp2x_gelu \
31+
--mm_vision_select_layer -2 \
32+
--mm_use_im_start_end False \
33+
--mm_use_im_patch_token False \
34+
--group_by_modality_length True \
35+
--image_aspect_ratio anyres_max_9 \
36+
--image_grid_pinpoints "(1x1),...,(6x6)" \
37+
--mm_patch_merge_type spatial_unpad \
38+
--bf16 True \
39+
--run_name shaokai_llava_ov_7b_random_mc \
40+
--output_dir experiments/shaokai_llava_ov_7b_random_mc \
41+
--num_train_epochs 1 \
42+
--per_device_train_batch_size 1 \
43+
--per_device_eval_batch_size 4 \
44+
--gradient_accumulation_steps 2 \
45+
--evaluation_strategy steps \
46+
--eval_steps 500\
47+
--save_strategy steps \
48+
--save_steps 1000 \
49+
--learning_rate 1e-5 \
50+
--weight_decay 0. \
51+
--warmup_ratio 0.03 \
52+
--lr_scheduler_type cosine \
53+
--logging_steps 1 \
54+
--tf32 True \
55+
--model_max_length 32768 \
56+
--gradient_checkpointing True \
57+
--dataloader_num_workers 4 \
58+
--lazy_preprocess True \
59+
--report_to wandb \
60+
--torch_compile True \
61+
--torch_compile_backend inductor \
62+
--dataloader_drop_last True \
63+
--frames_upbound 32 \
64+
--root /data/shaokai/EK100 \
65+
--action_predictions /data/shaokai/avion_predictions_test.json \
66+
--val_metadata /data/shaokai/epic-kitchens-100-annotations/EPIC_100_validation.csv \
67+
--llava_num_frames 16 \
68+
--topk_predictions 5 > train_kitchen_7b_random_mc.out 2>&1

0 commit comments

Comments
 (0)