Refine CLIP embedding microservice by leveraging the third-party CLIP (#1298)

lvliang-intel · web-flow · commit 77272351b845 · 2025-02-18T20:55:20.000+08:00
* Refine CLI embedding microservice using dependency
Signed-off-by: lvliang-intel &lt;liang1.lv@intel.com&gt;
diff --git a/comps/embeddings/deployment/docker_compose/compose.yaml b/comps/embeddings/deployment/docker_compose/compose.yaml
@@ -4,6 +4,7 @@
 include:
   - ../../../third_parties/tei/deployment/docker_compose/compose.yaml
   - ../../../third_parties/bridgetower/deployment/docker_compose/compose.yaml
+  - ../../../third_parties/clip/deployment/docker_compose/compose_intel_cpu.yaml
 
 x-multimodal-bridgetower-embedding-config: &multimodal-bridgetower-embedding-config
   image: ${REGISTRY:-opea}/embedding:${TAG:-latest}
@@ -64,7 +65,11 @@ services:
       no_proxy: ${no_proxy}
       http_proxy: ${http_proxy}
       https_proxy: ${https_proxy}
+      CLIP_EMBEDDING_ENDPOINT: ${CLIP_EMBEDDING_ENDPOINT}
       EMBEDDING_COMPONENT_NAME: "OPEA_CLIP_EMBEDDING"
+    depends_on:
+      multimodal-clip-embedding:
+        condition: service_healthy
     restart: unless-stopped
 
   multimodal-bridgetower-embedding-server:
diff --git a/comps/embeddings/src/integrations/clip.py b/comps/embeddings/src/integrations/clip.py
@@ -1,13 +1,10 @@
 # Copyright (C) 2024 Intel Corporation
 # SPDX-License-Identifier: Apache-2.0
 
+import asyncio
 import os
-from typing import List, Union
 
-import torch
-import torch.nn as nn
-from einops import rearrange
-from transformers import AutoProcessor, AutoTokenizer, CLIPModel
+import requests
 
 from comps import CustomLogger, OpeaComponent, OpeaComponentRegistry, ServiceType
 from comps.cores.proto.api_protocol import EmbeddingRequest, EmbeddingResponse, EmbeddingResponseData
@@ -16,51 +13,6 @@
 logflag = os.getenv("LOGFLAG", False)
 
 
-model_name = "openai/clip-vit-base-patch32"
-
-clip = CLIPModel.from_pretrained(model_name)
-processor = AutoProcessor.from_pretrained(model_name)
-tokenizer = AutoTokenizer.from_pretrained(model_name)
-
-
-class vCLIP(nn.Module):
-    def __init__(self, cfg):
-        super().__init__()
-
-        self.num_frm = cfg["num_frm"]
-        self.model_name = cfg["model_name"]
-
-    def embed_query(self, texts):
-        """Input is list of texts."""
-        text_inputs = tokenizer(texts, padding=True, return_tensors="pt")
-        text_features = clip.get_text_features(**text_inputs)
-        return text_features
-
-    def get_embedding_length(self):
-        text_features = self.embed_query("sample_text")
-        return text_features.shape[1]
-
-    def get_image_embeddings(self, images):
-        """Input is list of images."""
-        image_inputs = processor(images=images, return_tensors="pt")
-        image_features = clip.get_image_features(**image_inputs)
-        return image_features
-
-    def get_video_embeddings(self, frames_batch):
-        """Input is list of list of frames in video."""
-        self.batch_size = len(frames_batch)
-        vid_embs = []
-        for frames in frames_batch:
-            frame_embeddings = self.get_image_embeddings(frames)
-            frame_embeddings = rearrange(frame_embeddings, "(b n) d -> b n d", b=len(frames_batch))
-            # Normalize, mean aggregate and return normalized video_embeddings
-            frame_embeddings = frame_embeddings / frame_embeddings.norm(dim=-1, keepdim=True)
-            video_embeddings = frame_embeddings.mean(dim=1)
-            video_embeddings = video_embeddings / video_embeddings.norm(dim=-1, keepdim=True)
-            vid_embs.append(video_embeddings)
-        return torch.cat(vid_embs, dim=0)
-
-
 @OpeaComponentRegistry.register("OPEA_CLIP_EMBEDDING")
 class OpeaClipEmbedding(OpeaComponent):
     """A specialized embedding component derived from OpeaComponent for CLIP embedding services.
@@ -74,7 +26,7 @@ class OpeaClipEmbedding(OpeaComponent):
 
     def __init__(self, name: str, description: str, config: dict = None):
         super().__init__(name, ServiceType.EMBEDDING.name.lower(), description, config)
-        self.embeddings = vCLIP({"model_name": "openai/clip-vit-base-patch32", "num_frm": 4})
+        self.base_url = os.getenv("CLIP_EMBEDDING_ENDPOINT", "http://localhost:6990")
 
         health_status = self.check_health()
         if not health_status:
@@ -89,46 +41,38 @@ async def invoke(self, input: EmbeddingRequest) -> EmbeddingResponse:
         Returns:
             EmbeddingResponse: The response in OpenAI embedding format, including embeddings, model, and usage information.
         """
-        # Parse input according to the EmbeddingRequest format
-        if isinstance(input.input, str):
-            texts = [input.input.replace("\n", " ")]
-        elif isinstance(input.input, list):
-            if all(isinstance(item, str) for item in input.input):
-                texts = [text.replace("\n", " ") for text in input.input]
-            else:
-                raise ValueError("Invalid input format: Only string or list of strings are supported.")
-        else:
-            raise TypeError("Unsupported input type: input must be a string or list of strings.")
-        embed_vector = self.get_embeddings(texts)
-        if input.dimensions is not None:
-            embed_vector = [embed_vector[i][: input.dimensions] for i in range(len(embed_vector))]
-
-        # for standard openai embedding format
-        res = EmbeddingResponse(
-            data=[EmbeddingResponseData(index=i, embedding=embed_vector[i]) for i in range(len(embed_vector))]
-        )
-        return res
+        json_payload = input.model_dump()
+        try:
+            response = await asyncio.to_thread(
+                requests.post,
+                f"{self.base_url}/v1/embeddings",
+                headers={"Content-Type": "application/json"},
+                json=json_payload,
+            )
+            response.raise_for_status()
+            response_json = response.json()
+
+            return EmbeddingResponse(
+                data=[EmbeddingResponseData(**item) for item in response_json.get("data", [])],
+                model=response_json.get("model", input.model),
+                usage=response_json.get("usage", {}),
+            )
+        except requests.RequestException as e:
+            raise RuntimeError(f"Failed to invoke embedding service: {str(e)}")
 
     def check_health(self) -> bool:
         """Checks if the embedding model is healthy.
 
         Returns:
             bool: True if the embedding model is initialized, False otherwise.
         """
-        if self.embeddings:
+        try:
+            _ = requests.post(
+                f"{self.base_url}/v1/embeddings",
+                headers={"Content-Type": "application/json"},
+                json={"input": "health check"},
+            )
+
             return True
-        else:
+        except requests.RequestException as e:
             return False
-
-    def get_embeddings(self, text: Union[str, List[str]]) -> List[List[float]]:
-        """Generates embeddings for input text.
-
-        Args:
-            text (Union[str, List[str]]): Input text or list of texts.
-
-        Returns:
-            List[List[float]]: List of embedding vectors.
-        """
-        texts = [text] if isinstance(text, str) else text
-        embed_vector = self.embeddings.embed_query(texts).tolist()
-        return embed_vector
diff --git a/comps/embeddings/src/requirements.txt b/comps/embeddings/src/requirements.txt
@@ -1,9 +1,7 @@
 aiohttp
 docarray
-einops
 fastapi
 huggingface_hub
-open-clip-torch
 openai
 opentelemetry-api
 opentelemetry-exporter-otlp
@@ -13,5 +11,4 @@ predictionguard==2.2.1
 prometheus-fastapi-instrumentator
 PyYAML
 shortuuid
-transformers
 uvicorn
diff --git a/comps/third_parties/clip/deployment/docker_compose/compose_intel_cpu.yaml b/comps/third_parties/clip/deployment/docker_compose/compose_intel_cpu.yaml
@@ -1,19 +1,22 @@
 # Copyright (C) 2024 Intel Corporation
 # SPDX-License-Identifier: Apache-2.0
 
-version: "3.8"
-
 services:
-  embedding:
+  multimodal-clip-embedding:
     image: opea/embedding-multimodal-clip:latest
-    container_name: embedding-multimodal-server
+    container_name: multimodal-clip-embedding-server
     ports:
-      - "6000:6000"
+      - "${MULTIMODAL_CLIP_EMBEDDER_PORT:-6990}:6990"
     ipc: host
     environment:
       no_proxy: ${no_proxy}
       http_proxy: ${http_proxy}
       https_proxy: ${https_proxy}
+    healthcheck:
+      test: ["CMD-SHELL", "sleep 30 && exit 0"]
+      interval: 1s
+      timeout: 35s
+      retries: 1
     restart: unless-stopped
 
 networks:
diff --git a/tests/embeddings/test_embeddings_clip.sh b/tests/embeddings/test_embeddings_clip.sh
@@ -17,12 +17,25 @@ function build_docker_images() {
     else
         echo "opea/embedding built successful"
     fi
+
+    cd $WORKPATH
+    echo $(pwd)
+    docker build --no-cache -t opea/embedding-multimodal-clip:comps --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f comps/third_parties/clip/src/Dockerfile .
+
+    if [ $? -ne 0 ]; then
+        echo "opea/embedding-multimodal-clip built fail"
+        exit 1
+    else
+        echo "opea/embedding-multimodal-clip built successful"
+    fi
 }
 
 function start_service() {
     export TAG=comps
     export host_ip=${ip_address}
     export EMBEDDER_PORT=10203
+    export MULTIMODAL_CLIP_EMBEDDER_PORT=10204
+    export CLIP_EMBEDDING_ENDPOINT=http://${host_ip}:${MULTIMODAL_CLIP_EMBEDDER_PORT}
     service_name="clip-embedding-server"
     cd $WORKPATH
     cd comps/embeddings/deployment/docker_compose/
@@ -42,6 +55,7 @@ function validate_service() {
     else
         echo "Result wrong. Received was $result"
         docker logs clip-embedding-server
+        docker logs multimodal-clip-embedding-server
         exit 1
     fi
 }
@@ -57,7 +71,7 @@ function validate_microservice() {
 }
 
 function stop_docker() {
-    cid=$(docker ps -aq --filter "name=clip-embedding-server*")
+    cid=$(docker ps -aq --filter "name=clip-embedding-server*" --filter "name=multimodal-clip-embedding-server*")
     if [[ ! -z "$cid" ]]; then docker stop $cid && docker rm $cid && sleep 1s; fi
 }
 
