Upload 4 files

app.py

@@ -1,70 +1,19 @@
-import gradio as gr
-from huggingface_hub import InferenceClient
-
-
-def respond(
-    message,
-    history: list[dict[str, str]],
-    system_message,
-    max_tokens,
-    temperature,
-    top_p,
-    hf_token: gr.OAuthToken,
-):
-    """
-    For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
-    """
-    client = InferenceClient(token=hf_token.token, model="openai/gpt-oss-20b")
-
-    messages = [{"role": "system", "content": system_message}]
-
-    messages.extend(history)
-
-    messages.append({"role": "user", "content": message})
-
-    response = ""
-
-    for message in client.chat_completion(
-        messages,
-        max_tokens=max_tokens,
-        stream=True,
-        temperature=temperature,
-        top_p=top_p,
-    ):
-        choices = message.choices
-        token = ""
-        if len(choices) and choices[0].delta.content:
-            token = choices[0].delta.content
-
-        response += token
-        yield response
-
-
-"""
-For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
-"""
-chatbot = gr.ChatInterface(
-    respond,
-    type="messages",
-    additional_inputs=[
-        gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
-        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
-        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
-        gr.Slider(
-            minimum=0.1,
-            maximum=1.0,
-            value=0.95,
-            step=0.05,
-            label="Top-p (nucleus sampling)",
-        ),
-    ],
-)
-
-with gr.Blocks() as demo:
-    with gr.Sidebar():
-        gr.LoginButton()
-    chatbot.render()
-
-
-if __name__ == "__main__":
-    demo.launch()
+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.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