Update README.md
Browse files
README.md
CHANGED
@@ -1,5 +1,116 @@
|
|
1 |
-
---
|
2 |
-
license: other
|
3 |
-
license_name: meta
|
4 |
-
license_link: https://ai.meta.com/llama/licence
|
5 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: other
|
3 |
+
license_name: meta
|
4 |
+
license_link: https://ai.meta.com/llama/licence
|
5 |
+
datasets:
|
6 |
+
- toshi456/llava_pretrain_blip_laion_cc_sbu_558k_ja
|
7 |
+
base_model: mylesgoose/Meta-Llama-3.1-8B-Instruct-goose-abliterated
|
8 |
+
---
|
9 |
+
Install https://github.com/LLaVA-VL/LLaVA-NeXT/tree/main prior to running below. Thanks to that team for their fantastic work.
|
10 |
+
you can test with something like this.
|
11 |
+

|
12 |
+
|
13 |
+
""""
|
14 |
+
from llava.model.builder import load_pretrained_model
|
15 |
+
from llava.mm_utils import get_model_name_from_path, process_images, tokenizer_image_token
|
16 |
+
from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, IGNORE_INDEX
|
17 |
+
from llava.conversation import conv_templates, SeparatorStyle
|
18 |
+
|
19 |
+
from PIL import Image
|
20 |
+
import requests
|
21 |
+
import copy
|
22 |
+
import torch
|
23 |
+
|
24 |
+
pretrained = "mylesgoose/Meta-Llama-3.1-8B-Instruct-goose-abliterated-pre-llava"
|
25 |
+
model_name = "llava_llama3"
|
26 |
+
device = "cuda"
|
27 |
+
device_map = "auto"
|
28 |
+
tokenizer, model, image_processor, max_length = load_pretrained_model(pretrained, None, model_name, device_map=device_map) # Add any other thing you want to pass in llava_model_args
|
29 |
+
|
30 |
+
model.eval()
|
31 |
+
model.tie_weights()
|
32 |
+
|
33 |
+
image = Image.open("https://cdn-uploads.huggingface.co/production/uploads/65069ffda7ba30bf62aea321/XJRK1McipixmNVUyiL5v1.jpeg")
|
34 |
+
image_tensor = process_images([image], image_processor, model.config)
|
35 |
+
image_tensor = [_image.to(dtype=torch.float16, device=device) for _image in image_tensor]
|
36 |
+
conv_template = "llava_llama_3"
|
37 |
+
question = DEFAULT_IMAGE_TOKEN + "\nWhat is shown in this image? Is there anything strange about this image? Is this normal behaviour"
|
38 |
+
conv = copy.deepcopy(conv_templates[conv_template])
|
39 |
+
conv.append_message(conv.roles[0], question)
|
40 |
+
conv.append_message(conv.roles[1], None)
|
41 |
+
prompt_question = conv.get_prompt()
|
42 |
+
|
43 |
+
# Check if tokenizer_image_token returns the attention mask
|
44 |
+
input_ids, attention_mask = tokenizer_image_token(
|
45 |
+
prompt_question, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt"
|
46 |
+
)
|
47 |
+
input_ids = input_ids.unsqueeze(0).to(device)
|
48 |
+
image_sizes = [image.size]
|
49 |
+
|
50 |
+
# If attention_mask is not returned, create it manually (adjust as needed)
|
51 |
+
if attention_mask is None:
|
52 |
+
attention_mask = torch.ones_like(input_ids)
|
53 |
+
attention_mask[:, :IMAGE_TOKEN_INDEX] = 1
|
54 |
+
attention_mask[:, IMAGE_TOKEN_INDEX+1:] = 1
|
55 |
+
|
56 |
+
cont = model.generate(
|
57 |
+
input_ids,
|
58 |
+
images=image_tensor,
|
59 |
+
image_sizes=image_sizes,
|
60 |
+
attention_mask=attention_mask,
|
61 |
+
do_sample=True,
|
62 |
+
temperature=0.9,
|
63 |
+
max_new_tokens=256,
|
64 |
+
)
|
65 |
+
text_outputs = tokenizer.batch_decode(cont, skip_special_tokens=True)
|
66 |
+
print(text_outputs)
|
67 |
+
|
68 |
+
"""
|
69 |
+
|
70 |
+
I Trained the llama 3.1 model integratign the google vison encoder. This is a base model . It has only the encoder integrated into it. It has not been trained on any closed datasets. Other than what is listed.
|
71 |
+
LLM_VERSION="mylesgoose/Meta-Llama-3.1-8B-Instruct-goose-abliterated"
|
72 |
+
LLM_VERSION_CLEAN="${LLM_VERSION//\//_}"
|
73 |
+
VISION_MODEL_VERSION="google/siglip-so400m-patch14-384"
|
74 |
+
VISION_MODEL_VERSION_CLEAN="${VISION_MODEL_VERSION//\//_}"
|
75 |
+
|
76 |
+
############### Pretrain ################
|
77 |
+
|
78 |
+
PROMPT_VERSION=plain
|
79 |
+
|
80 |
+
BASE_RUN_NAME="llavanext-${VISION_MODEL_VERSION_CLEAN}-${LLM_VERSION_CLEAN}-mlp2x_gelu-pretrain_blip558k_plain"
|
81 |
+
echo "BASE_RUN_NAME: ${BASE_RUN_NAME}"
|
82 |
+
|
83 |
+
deepspeed llava/train/train_mem.py \
|
84 |
+
--deepspeed scripts/zero3.json \
|
85 |
+
--model_name_or_path ${LLM_VERSION} \
|
86 |
+
--version ${PROMPT_VERSION} \
|
87 |
+
--data_path ./data/llava_data/LLaVA-Pretrain/blip_laion_cc_sbu_558k.json \
|
88 |
+
--image_folder ./data/llava_data/LLaVA-Pretrain/images \
|
89 |
+
--vision_tower ${VISION_MODEL_VERSION} \
|
90 |
+
--mm_tunable_parts="mm_mlp_adapter" \
|
91 |
+
--mm_vision_select_layer -2 \
|
92 |
+
--mm_projector_type mlp2x_gelu \
|
93 |
+
--mm_use_im_start_end False \
|
94 |
+
--mm_use_im_patch_token False \
|
95 |
+
--bf16 True \
|
96 |
+
--output_dir ./checkpoints/projectors/${BASE_RUN_NAME} \
|
97 |
+
--num_train_epochs 1 \
|
98 |
+
--per_device_train_batch_size 6 \
|
99 |
+
--per_device_eval_batch_size 6 \
|
100 |
+
--gradient_accumulation_steps 6 \
|
101 |
+
--evaluation_strategy "no" \
|
102 |
+
--save_strategy "steps" \
|
103 |
+
--save_steps 500 \
|
104 |
+
--learning_rate 1e-3 \
|
105 |
+
--weight_decay 0. \
|
106 |
+
--warmup_ratio 0.03 \
|
107 |
+
--lr_scheduler_type "cosine" \
|
108 |
+
--logging_steps 1 \
|
109 |
+
--tf32 True \
|
110 |
+
--model_max_length 131072 \
|
111 |
+
--gradient_checkpointing True \
|
112 |
+
--dataloader_num_workers 6 \
|
113 |
+
--lazy_preprocess True \
|
114 |
+
--report_to wandb \
|
115 |
+
--run_name $BASE_RUN_NAME \
|
116 |
+
--attn_implementation flash_attention_2
|