Upload 3 files

app.py

@@ -0,0 +1,59 @@
+import numpy as np
+from safetensors import safe_open
+from datasets import load_dataset
+import torch
+from multilingual_clip import pt_multilingual_clip
+import transformers
+import gradio as gr
+import clip
+
+
+def load_embeddings(file_path, key="vectors"):
+    with safe_open(file_path, framework="numpy") as f:
+        embeddings = f.get_tensor(key)
+    return embeddings
+
+
+image_embeddings = load_embeddings("clothes_desc.safetensors")
+
+
+image_embeddings = image_embeddings / np.linalg.norm(
+    image_embeddings, axis=1, keepdims=True
+)
+
+
+ds = load_dataset("wbensvage/clothes_desc")["train"]
+
+model_name = "M-CLIP/LABSE-Vit-L-14"
+model = pt_multilingual_clip.MultilingualCLIP.from_pretrained(model_name)
+tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)
+
+
+def encode_text(texts, model, tokenizer):
+    with torch.no_grad():
+        embs = model.forward(texts, tokenizer)
+    embs = embs.detach().cpu().numpy()
+    embs = embs / np.linalg.norm(embs, axis=1, keepdims=True)
+    return embs
+
+
+def find_images(query, top_k):
+    query_embedding = encode_text([query], model, tokenizer)
+    similarity = np.dot(query_embedding, image_embeddings.T)
+    top_k_indices = np.argsort(-similarity[0])[:top_k]
+    images = [ds[int(i)]["image"] for i in top_k_indices]
+    return images
+
+
+iface = gr.Interface(
+    fn=find_images,
+    inputs=[
+        gr.Textbox(lines=2, placeholder="Enter search text here...", label="Query"),
+        gr.Slider(10, 50, step=10, value=20, label="Number of images"),
+    ],
+    outputs=gr.Gallery(label="Search Results", columns=5, height="auto"),
+    title="Multilingual CLIP Image Search",
+    description="Enter a text query",
+)
+
+iface.launch()

clothes_desc.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2d70d7a406ceb193d93bacc36f1e9b83b8c0008ce478cce9826f3cccc702c79
+size 1536088

requirements.txt

@@ -0,0 +1,7 @@
+torch
+transformers
+multilingual-clip
+safetensors
+gradio
+numpy
+datasets