Janus B：多模態理解與生成任務的統一模型

劍鋒1.3B

Janus 是一個新的自回歸框架，整合了多模態理解和生成。與先前的模型使用單一視覺編碼器來執行理解和生成任務不同，Janus 為這些功能引入了兩個獨立的視覺編碼路徑。

理解和產生編碼的差異

• 在多模態理解任務中，視覺編碼器會擷取高階語意訊息，例如物件類別和視覺屬性。此編碼器專注於推斷複雜的含義，強調高維語義元素。
• 另一方面，在視覺生成任務中，重點放在生成精細細節並保持整體一致性。因此，需要能夠捕捉空間結構和紋理的低維編碼。

設定環境

以下是在 Google Colab 中執行 Janus 的步驟：

git clone https://github.com/deepseek-ai/Janus
cd Janus
pip install -e .
# If needed, install the following as well
# pip install wheel
# pip install flash-attn --no-build-isolation

願景任務

載入模型

使用以下程式碼載入視覺任務所需的模型：

import torch
from transformers import AutoModelForCausalLM
from janus.models import MultiModalityCausalLM, VLChatProcessor
from janus.utils.io import load_pil_images

# Specify the model path
model_path = "deepseek-ai/Janus-1.3B"
vl_chat_processor = VLChatProcessor.from_pretrained(model_path)
tokenizer = vl_chat_processor.tokenizer

vl_gpt = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
vl_gpt = vl_gpt.to(torch.bfloat16).cuda().eval()

載入和準備圖像以進行編碼

接下來，載入圖片並將其轉換為模型可以理解的格式：

conversation = [
    {
        "role": "User",
        "content": "<image_placeholder>\nDescribe this chart.",
        "images": ["images/pie_chart.png"],
    },
    {"role": "Assistant", "content": ""},
]

# Load the image and prepare input
pil_images = load_pil_images(conversation)
prepare_inputs = vl_chat_processor(
    conversations=conversation, images=pil_images, force_batchify=True
).to(vl_gpt.device)

# Run the image encoder and obtain image embeddings
inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs)
</image_placeholder>

產生回應

最後，運行模型以產生回應：

# Run the model and generate 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,
)

answer = tokenizer.decode(outputs[0].cpu().tolist(), skip_special_tokens=True)
print(f"{prepare_inputs['sft_format'][0]}", answer)

範例輸出

The image depicts a pie chart that illustrates the distribution of four different categories among four distinct groups. The chart is divided into four segments, each representing a category with a specific percentage. The categories and their corresponding percentages are as follows:

1. **Hogs**: This segment is colored in orange and represents 30.0% of the total.
2. **Frog**: This segment is colored in blue and represents 15.0% of the total.
3. **Logs**: This segment is colored in red and represents 10.0% of the total.
4. **Dogs**: This segment is colored in green and represents 45.0% of the total.

The pie chart is visually divided into four segments, each with a different color and corresponding percentage. The segments are arranged in a clockwise manner starting from the top-left, moving clockwise. The percentages are clearly labeled next to each segment.

The chart is a simple visual representation of data, where the size of each segment corresponds to the percentage of the total category it represents. This type of chart is commonly used to compare the proportions of different categories in a dataset.

To summarize, the pie chart shows the following:
- Hogs: 30.0%
- Frog: 15.0%
- Logs: 10.0%
- Dogs: 45.0%

This chart can be used to understand the relative proportions of each category in the given dataset.

輸出展示了對圖像的適當理解，包括其顏色和文字。

影像生成任務

載入模型

使用以下程式碼載入影像產生任務所需的模型：

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-1.3B"
vl_chat_processor = VLChatProcessor.from_pretrained(model_path)
tokenizer = vl_chat_processor.tokenizer

vl_gpt = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
vl_gpt = vl_gpt.to(torch.bfloat16).cuda().eval()

準備提示

接下來，依照使用者的要求準備提示：

# Set up the prompt
conversation = [
    {
        "role": "User",
        "content": "cute japanese girl, wearing a bikini, in a beach",
    },
    {"role": "Assistant", "content": ""},
]

# Convert the prompt into the appropriate format
sft_format = vl_chat_processor.apply_sft_template_for_multi_turn_prompts(
    conversations=conversation,
    sft_format=vl_chat_processor.sft_format,
    system_prompt="",
)

prompt = sft_format + vl_chat_processor.image_start_tag

產生影像

以下函數用於產生影像。預設情況下，產生 16 張影像：

@torch.inference_mode()
def generate(
    mmgpt: MultiModalityCausalLM,
    vl_chat_processor: VLChatProcessor,
    prompt: str,
    temperature: float = 1,
    parallel_size: int = 16,
    cfg_weight: float = 5,
    image_token_num_per_image: int = 576,
    img_size: int = 384,
    patch_size: int = 16,
):
    input_ids = vl_chat_processor.tokenizer.encode(prompt)
    input_ids = torch.LongTensor(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
        if i % 2 != 0:
            tokens[i, 1:-1] = vl_chat_processor.pad_id

    inputs_embeds = mmgpt.language_model.get_input_embeddings()(tokens)

    generated_tokens = torch.zeros((parallel_size, image_token_num_per_image), dtype=torch.int).cuda()

    for i in range(image_token_num_per_image):
        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

        logits = mmgpt.gen_head(hidden_states[:, -1, :])
        logit_cond = logits[0::2, :]
        logit_uncond = logits[1::2, :]

        logits = logit_uncond + cfg_weight * (logit_cond - logit_uncond)
        probs = torch.softmax(logits / temperature, dim=-1)

        next_token = torch.multinomial(probs, num_samples=1)
        generated_tokens[:, i] = next_token.squeeze(dim=-1)

        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)

    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],
    )
    dec = dec.to(torch.float32).cpu().numpy().transpose(0, 2, 3, 1)
    dec = np.clip((dec + 1) / 2 * 255, 0, 255)

    visual_img = np.zeros((parallel_size, img_size, img_size, 3), dtype=np.uint8)
    visual_img[:, :, :] = dec

    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)

# Run the image generation
generate(vl_gpt, vl_chat_processor, prompt)

產生的影像將保存在 generated_samples 資料夾中。

產生結果範例

以下是產生影像的範例：

• 狗的描繪相對較好。
• 建築物保持整體形狀，但某些細節（例如窗戶）可能顯得不切實際。
• 人類，然而，要很好地生成是很有挑戰性的，在真實感和類似動畫的風格中都存在明顯的扭曲。

以上是Janus B：多模態理解與生成任務的統一模型的詳細內容。更多資訊請關注PHP中文網其他相關文章！