MGM
MGM copied to clipboard
dvlab-research
怎么从你们的检查点启动呢
Hi, I'm not sure what's the meaning of "start from checkpoint". Did you mean continue training from the pre-trained checkpoint?
Hi, I'm not sure what's the meaning of "start from checkpoint". Did you mean continue training from the pre-trained checkpoint?
Hi,没错QAQ 我可能不太专业哈,同样的基础模型，你们微调了2个阶段后，我只想微调你们用数据预训练后的基础模型，同样的型号，只要用同样的bash文件吗？QAQ 我是这么理解的。 比如LLAMA和miniGemini-Llama，我想微调的是miniGemini的-llama
I have the same question. Could author write a document to explain it?
` from minigemini.model.builder import load_pretrained_model
loal Mini-Gemini-2B
tokenizer, model, image_processor, context_len = load_pretrained_model(model_path="work_dirs/Mini-Gemini-2B", model_base=None, model_name='YanweiLi/Mini-Gemini-2B', load_8bit=False, load_4bit=False, device_map="auto", device="cuda", use_flash_attn=False) `
` from minigemini.model.builder import load_pretrained_model
loal Mini-Gemini-2B
tokenizer, model, image_processor, context_len = load_pretrained_model(model_path="work_dirs/Mini-Gemini-2B", model_base=None, model_name='YanweiLi/Mini-Gemini-2B', load_8bit=False, load_4bit=False, device_map="auto", device="cuda", use_flash_attn=False) `
https://huggingface.co/YanweiLi/Mini-Gemini-8x7B/blob/main/config.json elif "mixtral" in model_args.model_name_or_path:
Hi, I guess you can try this
FINETUNE_NAME=Mini-Gemini-7B
STAGE3_NAME=Your_prefer_name
AUX_SIZE=768
deepspeed minigemini/train/train_mem.py \
--deepspeed ./scripts/zero2_offload.json \
--model_name_or_path ./work_dirs/$FINETUNE_NAME \
--version v1 \
--data_path ./data/MiniGemini-Finetune/minigemini_instruction.json \
--image_folder ./data/MiniGemini-Finetune \
--vision_tower model_zoo/OpenAI/clip-vit-large-patch14-336 \
--vision_tower_aux model_zoo/OpenAI/openclip-convnext-large-d-320-laion2B-s29B-b131K-ft-soup \
--mm_projector_type mlp2x_gelu \
--mm_vision_select_layer -2 \
--mm_use_im_start_end False \
--mm_use_im_patch_token False \
--image_aspect_ratio pad \
--group_by_modality_length True \
--image_size_aux $AUX_SIZE \
--bf16 True \
--output_dir ./work_dirs/$STAGE3_NAME \
--num_train_epochs 1 \
--per_device_train_batch_size 8 \
--per_device_eval_batch_size 4 \
--gradient_accumulation_steps 2 \
--evaluation_strategy "no" \
--save_strategy "steps" \
--save_steps 1000 \
--save_total_limit 1 \
--learning_rate 2e-5 \
--weight_decay 0. \
--warmup_ratio 0.03 \
--lr_scheduler_type "cosine" \
--logging_steps 1 \
--tf32 True \
--model_max_length 2048 \
--gradient_checkpointing True \
--dataloader_num_workers 4 \
--lazy_preprocess True \
--report_to wandb
Hi, I guess you can try this
FINETUNE_NAME=Mini-Gemini-7B STAGE3_NAME=Your_prefer_name AUX_SIZE=768 deepspeed minigemini/train/train_mem.py \ --deepspeed ./scripts/zero2_offload.json \ --model_name_or_path ./work_dirs/$FINETUNE_NAME \ --version v1 \ --data_path ./data/MiniGemini-Finetune/minigemini_instruction.json \ --image_folder ./data/MiniGemini-Finetune \ --vision_tower model_zoo/OpenAI/clip-vit-large-patch14-336 \ --vision_tower_aux model_zoo/OpenAI/openclip-convnext-large-d-320-laion2B-s29B-b131K-ft-soup \ --mm_projector_type mlp2x_gelu \ --mm_vision_select_layer -2 \ --mm_use_im_start_end False \ --mm_use_im_patch_token False \ --image_aspect_ratio pad \ --group_by_modality_length True \ --image_size_aux $AUX_SIZE \ --bf16 True \ --output_dir ./work_dirs/$STAGE3_NAME \ --num_train_epochs 1 \ --per_device_train_batch_size 8 \ --per_device_eval_batch_size 4 \ --gradient_accumulation_steps 2 \ --evaluation_strategy "no" \ --save_strategy "steps" \ --save_steps 1000 \ --save_total_limit 1 \ --learning_rate 2e-5 \ --weight_decay 0. \ --warmup_ratio 0.03 \ --lr_scheduler_type "cosine" \ --logging_steps 1 \ --tf32 True \ --model_max_length 2048 \ --gradient_checkpointing True \ --dataloader_num_workers 4 \ --lazy_preprocess True \ --report_to wandb
Following this operation, the following warning is reported. Is this normal? I see that the code does not load the weights of vision_tower and vision_tower_aux when loading MiniGeminiLlamaForCausalLM.from_trained, but the weights are reloaded from clip/openclip later. Isn’t that the same as the weights of vision_tower and vision_tower_aux in your pre-trained MiniGeminiLlamaForCausalLM? Are the weights inconsistent? (Of course, if you haven’t fine-tuned vision_tower and vision_tower_aux, it should be fine, but it doesn’t feel very standardized. Why not load the weights together with MiniGeminiLlamaForCausalLM.from_trained)
Some weights of the model checkpoint at model_zoo/YanweiLi/Mini-Gemini-34B-HD were not used when initializing MiniGeminiLlamaForCausalLM: ['model.vision_tower.vision_tower.vision_model.embeddings.class_embedding', 'model.vision_tower.vision_tower.vision_model.embeddings.patch_embedding.weight
Hi, I guess you can try this
FINETUNE_NAME=Mini-Gemini-7B STAGE3_NAME=Your_prefer_name AUX_SIZE=768 deepspeed minigemini/train/train_mem.py \ --deepspeed ./scripts/zero2_offload.json \ --model_name_or_path ./work_dirs/$FINETUNE_NAME \ --version v1 \ --data_path ./data/MiniGemini-Finetune/minigemini_instruction.json \ --image_folder ./data/MiniGemini-Finetune \ --vision_tower model_zoo/OpenAI/clip-vit-large-patch14-336 \ --vision_tower_aux model_zoo/OpenAI/openclip-convnext-large-d-320-laion2B-s29B-b131K-ft-soup \ --mm_projector_type mlp2x_gelu \ --mm_vision_select_layer -2 \ --mm_use_im_start_end False \ --mm_use_im_patch_token False \ --image_aspect_ratio pad \ --group_by_modality_length True \ --image_size_aux $AUX_SIZE \ --bf16 True \ --output_dir ./work_dirs/$STAGE3_NAME \ --num_train_epochs 1 \ --per_device_train_batch_size 8 \ --per_device_eval_batch_size 4 \ --gradient_accumulation_steps 2 \ --evaluation_strategy "no" \ --save_strategy "steps" \ --save_steps 1000 \ --save_total_limit 1 \ --learning_rate 2e-5 \ --weight_decay 0. \ --warmup_ratio 0.03 \ --lr_scheduler_type "cosine" \ --logging_steps 1 \ --tf32 True \ --model_max_length 2048 \ --gradient_checkpointing True \ --dataloader_num_workers 4 \ --lazy_preprocess True \ --report_to wandbFollowing this operation, the following warning is reported. Is this normal? I see that the code does not load the weights of vision_tower and vision_tower_aux when loading MiniGeminiLlamaForCausalLM.from_trained, but the weights are reloaded from clip/openclip later. Isn’t that the same as the weights of vision_tower and vision_tower_aux in your pre-trained MiniGeminiLlamaForCausalLM? Are the weights inconsistent? (Of course, if you haven’t fine-tuned vision_tower and vision_tower_aux, it should be fine, but it doesn’t feel very standardized. Why not load the weights together with MiniGeminiLlamaForCausalLM.from_trained)
Some weights of the model checkpoint at model_zoo/YanweiLi/Mini-Gemini-34B-HD were not used when initializing MiniGeminiLlamaForCausalLM: ['model.vision_tower.vision_tower.vision_model.embeddings.class_embedding', 'model.vision_tower.vision_tower.vision_model.embeddings.patch_embedding.weight
这个问题可以哦
而且我这边使用的是之前微调LLAVA的数据集，然后报错了，还没到加载模型这一步 [2024-04-19 15:17:04,563] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect) Traceback (most recent call last): File "/opt/conda/envs/minigemini/lib/python3.10/site-packages/deepspeed/launcher/runner.py", line 437, in main subprocess.check_call(ssh_check_cmd, stderr=subprocess.DEVNULL, stdout=subprocess.DEVNULL, shell=True) File "/opt/conda/envs/minigemini/lib/python3.10/subprocess.py", line 369, in check_call raise CalledProcessError(retcode, cmd) subprocess.CalledProcessError: Command 'ssh -o PasswordAuthentication=no your_ip_0 hostname' returned non-zero exit status 255.
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/opt/conda/envs/minigemini/bin/deepspeed", line 6, in