unum-cloud

/

uform-gen2-qwen-500m

like 71

Follow

Unum 28

Hello. Thank you for your interest in our model.

This model has a special prompt formatting rule. You should put " <image>" text (space is intended) in front of prompt text.

Thank you very much for the quick answer.

following the instructions of a " <image>"

I attached an image of the prompt:

txt = f"""<im_start>system\nYou are a helpful assistant.<im_end>\n<im_start>user\n <image> describe image<im_end>\n<im_start>assistant\n{item["text"]}{self.processor.tokenizer.eos_token}"""

I attach the error when training:

code fine tune:

https://colab.research.google.com/drive/1w5hTTzHp52KlGJdVrBpCaMpAfdMolAsT?usp=sharing

Hello. Processor tokenizes texts using chat_template, so there is no need to pass formatted text.
Generally speaking, current Processor doesn't suit well for training. You can tokenize dialog using processor.tokenizer and image using processor.feature_extractor. I think code of the gradio demo can be very useful for you

Following your example of 'space', I organized the code this way but I see that it doesn't work either, what am I doing wrong?

%%time
from transformers import AutoModel, AutoProcessor
import torch
from PIL import Image
model = AutoModel.from_pretrained("unum-cloud/uform-gen2-qwen-500m",
                                  trust_remote_code=True,
                                  device_map="auto",
                                  # torch_dtype=torch.bfloat16, #error
                                  low_cpu_mem_usage=True,
                                  )
processor = AutoProcessor.from_pretrained("unum-cloud/uform-gen2-qwen-500m",
                                          trust_remote_code=True)

from datasets import load_dataset

dataset = load_dataset("ybelkada/football-dataset", split="train")


from datasets import load_dataset
from transformers import AutoProcessor
import torch
from PIL import Image
from torch.utils.data import Dataset, DataLoader
# from transformers import AutoModel, AutoProcessor

device = "cuda:0" if torch.cuda.is_available() else "cpu"


class ImageCaptioningDataset(Dataset):
    def __init__(self, dataset, processor):
        self.dataset = dataset
        self.processor = processor

    def __len__(self):
        return len(self.dataset)

    def __getitem__(self, idx):
        item = self.dataset[idx]
        messages = [{"role": "system", "content": "You are a helpful assistant."}]        
        
        # item["text"]' msg assistant describe imagen
        assistant_msg = item["text"]

        user_msg= " <image>describe image" # example message user
        
        messages.append({"role": "user", "content": user_msg})
        messages.append({"role": "assistant", "content": assistant_msg})

        # Asumiendo que 'item["image"]' es un objeto PIL.Image o una ruta a la imagen
        if isinstance(item["image"], str):
            image = Image.open(item["image"])  # Cargar la imagen desde una ruta de archivo
        else:
            image = item["image"]  # Si ya es un objeto PIL.Image
        # image = image.convert("RGB")  # Convert la image to RGB

        # Procesamiento de la imagen
        image = (self.processor.feature_extractor(image).unsqueeze(0))
        # print("image:",image)

        print(messages)

        model_inputs = self.processor.tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
        input_ids = model_inputs
        # print("input_ids:",input_ids)
        
        
        total_length = input_ids.shape[1] + self.processor.num_image_latents - 1
        # print("input_ids.shape[1]",input_ids.shape[1],"self.processor.num_image_latents",self.processor.num_image_latents,"total_length:",total_length)

        # attention_mask = torch.ones(1, model_inputs.shape[1] + processor.num_image_latents - 1)
        attention_mask = torch.ones(1, total_length)
        print("--->>attention_mask",attention_mask.shape) # torch.Size([1, 294])


        # Preparar las entradas del modelo
        model_inputs = {
            "input_ids": input_ids,
            "images": image,
            "attention_mask": attention_mask
        }
        print(model_inputs)
        # model_inputs = {k: v.to(device) for k, v in model_inputs.items()}
        return model_inputs

train_dataset = ImageCaptioningDataset(dataset, processor)
train_dataloader = DataLoader(train_dataset, shuffle=True, batch_size=1)


import torch
from tqdm.auto import tqdm  # Importa tqdm para la barra de progreso

optimizer = torch.optim.AdamW(model.parameters(), lr=5e-5)

device = "cuda:0" if torch.cuda.is_available() else "cpu"

model.train()

for epoch in range(5):
    print(f"Epoch: {epoch}")

    # Inicializa la barra de progreso
    progress_bar = tqdm(enumerate(train_dataloader), total=len(train_dataloader))

    for idx, batch in progress_bar:
        input_ids =  batch['input_ids'].to(device)  # batch.pop("input_ids")
        images =  batch['images'].to(device)  # batch.pop("images")
        attention_mask= batch['attention_mask'].to(device)  # batch.pop("attention_mask")

        print(f"Input IDs shape: {input_ids.shape}, Images shape: {images.shape}, Attention Mask shape: {attention_mask.shape}")
        outputs = model(input_ids=input_ids, images=images, labels=input_ids, attention_mask=attention_mask)

        loss = outputs.loss

        # Actualización de la barra de progreso con la información de la pérdida
        progress_bar.set_description(f"Loss: {loss.item():.4f}")

        loss.backward()

        optimizer.step()
        optimizer.zero_grad(set_to_none=True)  # Optimiza el reseteo de gradientes

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-18-1cd4f0a661d2> in <cell line: 10>()
     20 
     21         print(f"Input IDs shape: {input_ids.shape}, Images shape: {images.shape}, Attention Mask shape: {attention_mask.shape}")
---> 22         outputs = model(input_ids=input_ids, images=images, labels=input_ids, attention_mask=attention_mask)
     23 
     24         loss = outputs.loss

6 frames
~/.cache/huggingface/modules/transformers_modules/unum-cloud/uform-gen2-qwen-500m/3912572ad204f82a2b0f875d3a1700faaebab719/vision_encoder.py in forward(self, x)
    175 
    176     def forward(self, x: Tensor) -> Tensor:
--> 177         h, w = x.shape[2:]
    178         x = self.patch_embed(x).flatten(start_dim=2).transpose(2, 1)
    179         x = x + self.interpolate_pos_encoding(x, h, w)

ValueError: too many values to unpack (expected 2)

url space

https://huggingface.co/spaces/unum-cloud/uform-gen2-qwen-500m-demo

https://colab.research.google.com/drive/1w5hTTzHp52KlGJdVrBpCaMpAfdMolAsT?usp=sharing

Are there any updates?

hello. I am going to prepare a colab notebook with fine-tuning example, stay tuned

good news, thanks. compatible with peft , sft ?

Any news on this colab notebook?