Duplicate from Mueris/TurkishVLMTAMGA

Co-authored-by: Mustafa Eren Işıktaşlı <Mueris@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

README.txt

@@ -0,0 +1,20 @@
+# CLIP2MT5 CrossAttention VQA Model
+
+This is a Vision-Language model combining **CLIP-ViT** and **mT5** using a custom cross-attention bridge.
+It supports Visual Question Answering (VQA) in Turkish.
+
+## Usage
+
+```python
+from PIL import Image
+from hf_clip2mt5 import load_for_inference, predict
+
+repo_id = "MUERIS/TurkishVLMTAMGA"
+
+model, tokenizer, device = load_for_inference(repo_id)
+
+image = Image.open("example.jpg")
+question = "Görselde kaç kişi var?"
+
+answer = predict(model, tokenizer, device, image, question)
+print(answer)

__init__.py

@@ -0,0 +1,2 @@
+from .model import CLIP2MT5_CrossAttention, load_model
+from .inference import load_for_inference, predict

app.py

@@ -0,0 +1,19 @@
+import gradio as gr
+from inference import load_for_inference, predict
+
+REPO_ID = "MUERIS/TurkishVLMTAMGA"
+
+model, tokenizer, device = load_for_inference(REPO_ID)
+
+def answer(image, question):
+    return predict(model, tokenizer, device, image, question)
+
+gr.Interface(
+    fn=answer,
+    inputs=[
+        gr.Image(type="pil"),
+        gr.Textbox(label="Question")
+    ],
+    outputs="text",
+    title="CLIP2MT5 Visual Question Answering"
+).launch()

config.json

@@ -0,0 +1,8 @@
+{
+  "model_type": "clip2mt5-crossattention",
+  "library": "pytorch",
+  "architectures": ["CLIP2MT5_CrossAttention"],
+  "pipeline_tag": "image-text-to-text",
+  "description": "CLIP + mT5 VQA Model using cross-attention.",
+  "author": "MUERIS"
+}

inference.py

@@ -0,0 +1,46 @@
+import torch
+from PIL import Image
+from torchvision import transforms
+
+from model import load_model
+
+
+# Preprocessing
+_transform = transforms.Compose([
+    transforms.Resize((224, 224)),
+    transforms.ToTensor(),
+    transforms.Normalize(
+        mean=[0.4815, 0.4578, 0.4082],
+        std=[0.2686, 0.2613, 0.2758]
+    )
+])
+
+
+def load_for_inference(repo_id, filename="model.pt"):
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    model = load_model(repo_id=repo_id, filename=filename, device=device)
+    tokenizer = model.tokenizer
+    return model, tokenizer, device
+
+
+def predict(model, tokenizer, device, image: Image.Image, question: str):
+    image_tensor = _transform(image).unsqueeze(0).to(device)
+
+    q = tokenizer(
+        question,
+        return_tensors='pt',
+        padding=True,
+        truncation=True,
+        max_length=64
+    ).to(device)
+
+    with torch.no_grad():
+        output_ids = model.generate(
+            images=image_tensor,
+            input_ids=q.input_ids,
+            attention_mask=q.attention_mask,
+            max_length=64,
+            num_beams=4
+        )
+
+    return tokenizer.decode(output_ids[0], skip_special_tokens=True)

model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7006cd53a99bc96da69f28839e99492d83fc73f55a6d429adb8cf7f8ccfa6d45
+size 2937292702

model.py

@@ -0,0 +1,91 @@
+import torch
+import torch.nn as nn
+from transformers import CLIPModel, AutoTokenizer, AutoModelForSeq2SeqLM
+from huggingface_hub import hf_hub_download
+
+
+class CLIP2MT5_CrossAttention(nn.Module):
+    def __init__(self, clip_name='openai/clip-vit-base-patch32',
+                 t5_name='mukayese/mt5-base-turkish-summarization'):
+        super().__init__()
+
+        self.clip = CLIPModel.from_pretrained(clip_name)
+        self.tokenizer = AutoTokenizer.from_pretrained(t5_name)
+        self.t5 = AutoModelForSeq2SeqLM.from_pretrained(t5_name)
+
+        self.vis_proj = nn.Linear(
+            self.clip.config.vision_config.hidden_size,
+            self.t5.config.d_model
+        )
+
+    def forward(self, images, input_ids, attention_mask, labels=None):
+        vision_outputs = self.clip.vision_model(pixel_values=images).last_hidden_state
+        vision_embeds = self.vis_proj(vision_outputs)
+
+        text_embeds = self.t5.encoder.embed_tokens(input_ids)
+
+        extended_input_embeds = torch.cat([vision_embeds, text_embeds], dim=1)
+
+        extended_attention_mask = torch.cat([
+            torch.ones(vision_embeds.size(0), vision_embeds.size(1),
+                       dtype=attention_mask.dtype, device=attention_mask.device),
+            attention_mask
+        ], dim=1)
+
+        if labels is not None:
+            labels = labels.clone()
+            labels[labels == self.tokenizer.pad_token_id] = -100
+
+        return self.t5(
+            inputs_embeds=extended_input_embeds,
+            attention_mask=extended_attention_mask,
+            labels=labels,
+            return_dict=True
+        )
+
+    @torch.no_grad()
+    def generate(self, images, input_ids, attention_mask, **gen_kwargs):
+        vision_outputs = self.clip.vision_model(pixel_values=images).last_hidden_state
+        vision_embeds = self.vis_proj(vision_outputs)
+
+        text_embeds = self.t5.encoder.embed_tokens(input_ids)
+
+        extended_input_embeds = torch.cat([vision_embeds, text_embeds], dim=1)
+
+        extended_attention_mask = torch.cat([
+            torch.ones(vision_embeds.size(0), vision_embeds.size(1),
+                       dtype=attention_mask.dtype, device=attention_mask.device),
+            attention_mask
+        ], dim=1)
+
+        return self.t5.generate(
+            inputs_embeds=extended_input_embeds,
+            attention_mask=extended_attention_mask,
+            **gen_kwargs
+        )
+
+
+
+# HF Loader for STATE_DICT
+
+
+def load_model(
+    repo_id: str,
+    filename: str = "model.pt",
+    clip_name="openai/clip-vit-base-patch32",
+    t5_name="mukayese/mt5-base-turkish-summarization",
+    device=None
+):
+    if device is None:
+        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+    model_path = hf_hub_download(repo_id=repo_id, filename=filename)
+
+    model = CLIP2MT5_CrossAttention(clip_name=clip_name, t5_name=t5_name)
+
+    state = torch.load(model_path, map_location=device)
+    model.load_state_dict(state)
+
+    model.to(device)
+    model.eval()
+    return model

requirements.txt

@@ -0,0 +1,6 @@
+torch
+torchvision
+transformers
+huggingface_hub
+gradio
+Pillow