IHP-24 - wip doesn't work

.gitignore

@@ -2,6 +2,7 @@ params.conf
 .idea
 .bsp
 converted-to-torchscript.pt
+nsfw_model.pt
 
 target/
 !.mvn/wrapper/maven-wrapper.jar

alias.sbt

@@ -1 +1,2 @@
 addCommandAlias("buildResizer", "project resizer;assembly;")
+addCommandAlias("buildRecognizer", "project recognizer;assembly;")

docker-compose.yml

@@ -15,30 +15,52 @@ services:
   #      MINIO_USER: minioadmin
   #      MINIO_PASSWORD: minioadmin
 
-  resizer1:
-    image: ghcr.io/baklanov-soft/image-hosting-processing-resizer:master
-    container_name: resizer1
-    depends_on:
-      - kafka-init
-      - minio
-    environment:
-      KAFKA_BOOTSTRAP_SERVERS: kafka:9092
-      CONSUMER_GROUP_ID: resizer-local-test
-      MINIO_HOST: "http://minio:9000"
-      MINIO_USER: minioadmin
-      MINIO_PASSWORD: minioadmin
+  #  resizer1:
+  #    image: ghcr.io/baklanov-soft/image-hosting-processing-resizer:master
+  #    container_name: resizer1
+  #    depends_on:
+  #      - kafka-init
+  #      - minio
+  #    environment:
+  #      KAFKA_BOOTSTRAP_SERVERS: kafka:9092
+  #      CONSUMER_GROUP_ID: resizer-local-test
+  #      MINIO_HOST: "http://minio:9000"
+  #      MINIO_USER: minioadmin
+  #      MINIO_PASSWORD: minioadmin
+  #      NEW_IMAGES_TOPIC: "new-images.v1"
+  #
+  #  resizer2:
+  #    image: ghcr.io/baklanov-soft/image-hosting-processing-resizer:master
+  #    container_name: resizer2
+  #    depends_on:
+  #      - kafka-init-new-images
+  #    environment:
+  #      KAFKA_BOOTSTRAP_SERVERS: kafka:9092
+  #      CONSUMER_GROUP_ID: resizer-local-test
+  #      MINIO_HOST: "http://minio:9000"
+  #      MINIO_USER: minioadmin
+  #      MINIO_PASSWORD: minioadmin
+  #      NEW_IMAGES_TOPIC: "new-images.v1"
 
-  resizer2:
-    image: ghcr.io/baklanov-soft/image-hosting-processing-resizer:master
-    container_name: resizer2
+  recognizer1:
+    image: test/recognizer:latest
+    container_name: recognizer1
     depends_on:
-      - kafka-init
+      - kafka-init-new-images
+      - kafka-init-categories
+    volumes:
+      - recognizer1-djl-cache:/root/.djl.ai
     environment:
       KAFKA_BOOTSTRAP_SERVERS: kafka:9092
-      CONSUMER_GROUP_ID: resizer-local-test
+      CONSUMER_GROUP_ID: recognizer-local-test
+      NEW_IMAGES_TOPIC: "new-images.v1"
+      CATEGORIES_TOPIC: "categories.v1"
+      NSFW_SYNSET_PATH: "synset.txt"
+      NSFW_MODEL_PATH: "nsfw_model.pt"
       MINIO_HOST: "http://minio:9000"
       MINIO_USER: minioadmin
       MINIO_PASSWORD: minioadmin
+      DEBUG_CATEGORIES: false
 
   kafka:
     container_name: kafka
@@ -138,19 +160,20 @@ services:
       "--topic", "categories.v1"
     ]
 
-  #  kafka-ui:
-  #    image: provectuslabs/kafka-ui
-  #    container_name: kafka-ui
-  #    ports:
-  #      - "8000:8000"
-  #    environment:
-  #      SERVER_PORT: 8000
-  #      KAFKA_CLUSTERS_0_NAME: image-hosting
-  #      KAFKA_CLUSTERS_0_BOOTSTRAPSERVERS: kafka:9092
-  #      KAFKA_CLUSTERS_0_READONLY: true
-  #    depends_on:
-  #      - kafka
+#  kafka-ui:
+#    image: provectuslabs/kafka-ui
+#    container_name: kafka-ui
+#    ports:
+#      - "8000:8000"
+#    environment:
+#      SERVER_PORT: 8000
+#      KAFKA_CLUSTERS_0_NAME: image-hosting
+#      KAFKA_CLUSTERS_0_BOOTSTRAPSERVERS: kafka:9092
+#      KAFKA_CLUSTERS_0_READONLY: true
+#    depends_on:
+#      - kafka
 
 volumes:
   minio-data:
   db-data:
+  recognizer1-djl-cache:

recognizer/Dockerfile

@@ -0,0 +1,9 @@
+FROM eclipse-temurin:17.0.6_10-jre-jammy
+
+WORKDIR /opt/app
+
+COPY ./target/scala-2.13/image-hosting-processing-recognizer-assembly-0.1.0-SNAPSHOT.jar ./
+COPY synset.txt ./
+COPY nsfw_model.pt ./
+
+ENTRYPOINT ["java", "-cp", "image-hosting-processing-recognizer-assembly-0.1.0-SNAPSHOT.jar", "com.github.baklanovsoft.imagehosting.recognizer.Main"]

recognizer/build_local.sh

@@ -0,0 +1,3 @@
+docker buildx build --platform linux/amd64 -t test/recognizer .
+
+docker image ls | grep test/recognizer

recognizer/convert.py

@@ -0,0 +1,18 @@
+from transformers import AutoImageProcessor, AutoModelForImageClassification
+import torch
+from PIL import Image
+from transformers import AutoTokenizer
+
+model_name = "DenisNovac/nsfw_image_detection"
+
+model = AutoModelForImageClassification.from_pretrained(model_name, torchscript=True, return_dict=False)
+
+processor = AutoImageProcessor.from_pretrained(model_name)
+
+image = Image.open("images/hentai.jpg")
+image_inputs = processor(images=image, return_tensors="pt")
+
+config = {'forward': [image_inputs['pixel_values']]}
+converted = torch.jit.trace_module(model,  config)
+
+torch.jit.save(converted, "nsfw_model.pt")

recognizer/download-model.sh

@@ -1,3 +1,3 @@
 # https://huggingface.co/DenisNovac/nsfw_image_detection
 # fork of https://huggingface.co/Falconsai/nsfw_image_detection
-wget -O converted-to-torchscript.pt https://huggingface.co/DenisNovac/nsfw_image_detection/resolve/main/converted-to-torchscript.pt?download=true
+wget -O nsfw_model.pt https://huggingface.co/DenisNovac/nsfw_image_detection/resolve/main/converted-to-torchscript.pt?download=true

recognizer/src/main/resources/application.conf

@@ -16,9 +16,9 @@ new-images-topic = ${?NEW_IMAGES_TOPIC}
 debug-categories = false
 debug-categories = ${?DEBUG_CATEGORIES}
 
-nsfw-synset = "synset.txt"
-nsfw-synset = ${?NSFW_SYNSET}
-nsfw-model-path = "recognizer/converted-to-torchscript.pt"
+nsfw-synset-path = "recognizer/synset.txt"
+nsfw-synset-path = ${?NSFW_SYNSET_PATH}
+nsfw-model-path = "recognizer/nsfw_model.pt"
 nsfw-model-path = ${?NSFW_MODEL_PATH}
 
 minio {

recognizer/src/main/scala/com/github/baklanovsoft/imagehosting/recognizer/AppConfig.scala

@@ -9,7 +9,7 @@ final case class AppConfig(
     newImagesTopic: String,
     categoriesTopic: String,
     debugCategories: Boolean,
-    nsfwSynset: String,
+    nsfwSynsetPath: String,
     nsfwModelPath: String,
     minio: MinioCreds
 )

recognizer/src/main/scala/com/github/baklanovsoft/imagehosting/recognizer/Main.scala

@@ -32,7 +32,7 @@ object Main extends IOApp with KafkaJsonDeserializer {
     resources = for {
                   detection      <- if (config.debugCategories) ObjectDetection.debug[IO](minioClient)
                                     else ObjectDetection.production[IO]
-                  nsfw           <- NsfwDetection.of[IO](config.nsfwModelPath, config.nsfwSynset)
+                  nsfw           <- NsfwDetection.of[IO](config.nsfwModelPath, config.nsfwSynsetPath)
                   categorization <- Resource.eval(
                                       CategorizationStream
                                         .of[IO](

recognizer/src/main/scala/com/github/baklanovsoft/imagehosting/recognizer/NsfwDetection.scala

@@ -10,9 +10,9 @@ import ai.djl.translate.Translator
 import cats.effect.kernel.{Resource, Sync}
 import cats.implicits._
 import com.github.baklanovsoft.imagehosting.{BucketId, Category, ImageId, Score}
-import org.typelevel.log4cats.LoggerFactory
+import org.typelevel.log4cats.{Logger, LoggerFactory}
 
-import java.nio.file.Paths
+import java.nio.file.{Files, Paths}
 import scala.jdk.CollectionConverters._
 
 trait NsfwDetection[F[_]] {
@@ -24,37 +24,44 @@ trait NsfwDetection[F[_]] {
 
 object NsfwDetection {
 
-  private def buildTranslator[F[_]: Sync](synsetPath: String): F[Translator[Image, Classifications]] = Sync[F].delay {
-    // copypasted from here https://github.com/deepjavalibrary/djl/issues/1419
-    ImageClassificationTranslator
-      .builder()
-      .optSynsetArtifactName(synsetPath)
-      .addTransform(new Resize(256))
-      // from the model description it was trained on 224x224 images so looks like it fits
-      .addTransform(new CenterCrop(224, 224))
-      .addTransform(new ToTensor())
-      .addTransform(
-        new Normalize(
-          Array(
-            0.485f,
-            0.456f,
-            0.406f
-          ),
-          Array(
-            0.229f,
-            0.224f,
-            0.225f
+  private def buildTranslator[F[_]: Sync](synsetUrl: String): F[Translator[Image, Classifications]] =
+    Sync[F].delay {
+      // copypasted from here https://github.com/deepjavalibrary/djl/issues/1419
+      ImageClassificationTranslator
+        .builder()
+        .optSynsetUrl(synsetUrl)
+        .addTransform(new Resize(256))
+        // from the model description it was trained on 224x224 images so looks like it fits
+        .addTransform(new CenterCrop(224, 224))
+        .addTransform(new ToTensor())
+        .addTransform(
+          new Normalize(
+            Array(
+              0.485f,
+              0.456f,
+              0.406f
+            ),
+            Array(
+              0.229f,
+              0.224f,
+              0.225f
+            )
           )
         )
-      )
-      .optApplySoftmax(true)
-      .build()
-  }
+        .optApplySoftmax(true)
+        .build()
+    }
 
-  private def acquireModelPredictor[F[_]: Sync](modelPath: String, synsetPath: String) =
+  private def acquireModelPredictor[F[_]: Sync: Logger](modelPath: String, synsetPath: String) =
     Resource.make {
       for {
-        translator <- buildTranslator(synsetPath)
+        lookup    <- Sync[F].delay(Files.list(Paths.get("./")).toList)
+        _         <- Logger[F].info(s"Workdir absolute path: ${Paths.get("./").toAbsolutePath.toString}")
+        _         <- Logger[F].info(s"Lookup result: $lookup")
+        synsetUrl <- Sync[F].delay("file://" + Paths.get(synsetPath).toAbsolutePath.toString)
+        _         <- Logger[F].info(s"Synset constructed url: $synsetUrl")
+
+        translator <- buildTranslator(synsetUrl)
         criteria   <- Sync[F].delay {
                         Criteria
                           .builder()
@@ -77,8 +84,8 @@ object NsfwDetection {
 
   def of[F[_]: Sync: LoggerFactory](modelPath: String, synsetPath: String): Resource[F, NsfwDetection[F]] =
     for {
-      logger         <- Resource.eval(LoggerFactory[F].create)
-      (_, predictor) <- acquireModelPredictor[F](modelPath, synsetPath)
+      implicit0(logger: Logger[F]) <- Resource.eval(LoggerFactory[F].create)
+      (_, predictor)               <- acquireModelPredictor[F](modelPath, synsetPath)
     } yield new NsfwDetection[F] {
 
       override def detect(image: Image, bucketId: BucketId, imageId: ImageId): F[Option[(Category, Score)]] =