8-10 DeepSeek VLM モデル

git clone https://github.com/deepseek-ai/Janus.git

cd Janus

pip install -e . --index-url https://pypi.jetson-ai-lab.io/jp6/cu126

cd ..

import os
import csv
import cv2
import time
import torch
from PIL import Image
from transformers import AutoModelForCausalLM

from janus.models import MultiModalityCausalLM, VLChatProcessor
from janus.utils.io import load_pil_images

# Specify the path of the model
model_path = "deepseek-ai/Janus-Pro-1B"
# Load the processor
vl_chat_processor: VLChatProcessor = VLChatProcessor.from_pretrained(model_path, cache_dir="./model")
tokenizer = vl_chat_processor.tokenizer

# Load the model
vl_gpt: MultiModalityCausalLM = AutoModelForCausalLM.from_pretrained(
    model_path, trust_remote_code=True, torch_dtype=torch.bfloat16, cache_dir="./model",
)
vl_gpt = vl_gpt.to(torch.bfloat16).cuda().eval()

# Set up the conversation messages, roles, and content
img_path = "Janus/images/doge.png"
conversation = [
    {
        "role": "<|User|>",
        "content": "\nDescribe the image.",
        "images": [img_path],
    },
    {"role": "<|Assistant|>", "content": ""},
]

# Load the image and prepare input
frame = cv2.cvtColor(cv2.imread(img_path), cv2.COLOR_BGR2RGB)
pil_images = [Image.fromarray(frame)]
prepare_inputs = vl_chat_processor(
    conversations=conversation, images=pil_images, force_batchify=True
).to(vl_gpt.device)

# Run the image encoder to get image embedding data
inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs)

# Run the model to get a response
outputs = vl_gpt.language_model.generate(
    inputs_embeds=inputs_embeds,
    attention_mask=prepare_inputs.attention_mask,
    pad_token_id=tokenizer.eos_token_id,
    bos_token_id=tokenizer.bos_token_id,
    eos_token_id=tokenizer.eos_token_id,
    max_new_tokens=512,
    do_sample=False,
    use_cache=True,
)

# Decode the generated text and output the result
answer = tokenizer.decode(outputs[0].cpu().tolist(), skip_special_tokens=True)
print(answer)

import os
import PIL.Image
import torch
import numpy as np
from transformers import AutoModelForCausalLM
from janus.models import MultiModalityCausalLM, VLChatProcessor

# Specify the model path
model_path = "deepseek-ai/Janus-Pro-1B"
# Load the multimodal chat processor and specify cache directory
vl_chat_processor: VLChatProcessor = VLChatProcessor.from_pretrained(
    model_path, cache_dir="./model"
)
# Get the tokenizer
tokenizer = vl_chat_processor.tokenizer

# Load the multimodal causal language model, allow remote code, 
# specify cache directory
vl_gpt: MultiModalityCausalLM = AutoModelForCausalLM.from_pretrained(
    model_path, trust_remote_code=True, cache_dir="./model"
)
# Convert model to bfloat16 type, move to CUDA GPU, and set to evaluation mode
vl_gpt = vl_gpt.to(torch.bfloat16).cuda().eval()

# Define the conversation content: first line is user input, 
# second line is assistant's placeholder
conversation = [
    {
        "role": "<|User|>",
        "content": "A stunning princess from kabul in red, "
                    "white traditional clothing, blue eyes, brown hair",
    },
    {"role": "<|Assistant|>", "content": ""},
]

# Apply the SFT (Supervised Fine-Tuning) template on conversation to 
# generate a training format prompt
sft_format = vl_chat_processor.apply_sft_template_for_multi_turn_prompts(
    conversations=conversation,
    sft_format=vl_chat_processor.sft_format,
    system_prompt="",
)
# Append the image start tag after the prompt
prompt = sft_format + vl_chat_processor.image_start_tag

# Define the image generation function, 
# using torch.inference_mode to ensure no gradient calculation
@torch.inference_mode()
def generate(
    mmgpt: MultiModalityCausalLM,
    vl_chat_processor: VLChatProcessor,
    prompt: str,
    temperature: float = 1,
    parallel_size: int = 1,
    cfg_weight: float = 5,
    image_token_num_per_image: int = 576,
    img_size: int = 384,
    patch_size: int = 16,
):
    # Encode prompt into token IDs
    input_ids = vl_chat_processor.tokenizer.encode(prompt)
    input_ids = torch.LongTensor(input_ids)

    # Create a tokens tensor with length parallel_size*2, repeat input_ids
    tokens = torch.zeros((parallel_size*2, len(input_ids)), dtype=torch.int).cuda()
    for i in range(parallel_size*2):
        tokens[i, :] = input_ids
        # Keep original prompt in even rows, 
        # pad with pad_id in odd rows for contrastive guidance
        if i % 2 != 0:
            tokens[i, 1:-1] = vl_chat_processor.pad_id

    # Convert tokens to embeddings
    inputs_embeds = mmgpt.language_model.get_input_embeddings()(tokens)

    # Prepare tensor to store generated image tokens
    generated_tokens = torch.zeros(
        (parallel_size, image_token_num_per_image), 
        dtype=torch.int
    ).cuda()

    # Generate each image token step by step
    for i in range(image_token_num_per_image):
        # Call the model and pass past_key_values to speed up generation
        outputs = mmgpt.language_model.model(
            inputs_embeds=inputs_embeds,
            use_cache=True,
            past_key_values=outputs.past_key_values if i != 0 else None
        )
        hidden_states = outputs.last_hidden_state
        
        # Get logits split into conditional (cond) and unconditional (uncond) parts
        logits = mmgpt.gen_head(hidden_states[:, -1, :])
        logit_cond = logits[0::2, :]
        logit_uncond = logits[1::2, :]
        
        # Apply classifier-free guidance (CFG) weighting
        logits = logit_uncond + cfg_weight * (logit_cond - logit_uncond)
        probs = torch.softmax(logits / temperature, dim=-1)

        # Sample the next token according to probabilities
        next_token = torch.multinomial(probs, num_samples=1)
        generated_tokens[:, i] = next_token.squeeze(dim=-1)

        # Prepare input embeddings for next step; 
        # replicate tokens and convert to image embeddings
        next_token = torch.cat(
            [next_token.unsqueeze(dim=1), next_token.unsqueeze(dim=1)], dim=1
        ).view(-1)
        img_embeds = mmgpt.prepare_gen_img_embeds(next_token)
        inputs_embeds = img_embeds.unsqueeze(dim=1)

    # Use the vision decoder to decode tokens into images
    dec = mmgpt.gen_vision_model.decode_code(
        generated_tokens.to(dtype=torch.int),
        shape=[parallel_size, 8, img_size//patch_size, img_size//patch_size]
    )
    # Convert values to uint8 image format
    dec = dec.to(torch.float32).cpu().numpy().transpose(0, 2, 3, 1)
    dec = np.clip((dec + 1) / 2 * 255, 0, 255)

    # Create blank array and fill with final images
    visual_img = np.zeros((parallel_size, img_size, img_size, 3), dtype=np.uint8)
    visual_img[:, :, :] = dec

    # Create output folder and save images
    os.makedirs('generated_samples', exist_ok=True)
    for i in range(parallel_size):
        save_path = os.path.join('generated_samples', f"img_{i}.jpg")
        PIL.Image.fromarray(visual_img[i]).save(save_path)

# Call the generate function to start image generation
generate(
    vl_gpt,
    vl_chat_processor,
    prompt,
)