Merge pull request #24 from smartscanapp/staging

SmartScan SDK V1.2.0

Author: dev-diaries41

core/src/androidTest/kotlin/com/fpf/smartscansdk/core/data/images/Entity.kt

@@ -1,7 +1,7 @@
 package com.fpf.smartscansdk.core.data.images
 
 import androidx.room.*
-import com.fpf.smartscansdk.core.data.Embedding
+import com.fpf.smartscansdk.core.embeddings.Embedding
 
 @Entity(tableName = "image_embeddings")
 data class ImageEmbeddingEntity(

core/src/main/java/com/fpf/smartscansdk/core/data/Classification.kt

@@ -1,13 +1,13 @@
-package com.fpf.smartscansdk.ml.models.providers.embeddings
+package com.fpf.smartscansdk.core.embeddings
 
-import com.fpf.smartscansdk.core.data.ClassificationError
-import com.fpf.smartscansdk.core.data.ClassificationResult
-import com.fpf.smartscansdk.core.data.PrototypeEmbedding
-import com.fpf.smartscansdk.core.embeddings.getSimilarities
-import com.fpf.smartscansdk.core.embeddings.getTopN
+sealed class ClassificationResult {
+    data class Success(val classId: String, val similarity: Float ): ClassificationResult()
+    data class Failure(val error: ClassificationError ): ClassificationResult()
+}
 
+enum class ClassificationError{MINIMUM_CLASS_SIZE, THRESHOLD, CONFIDENCE_MARGIN}
 
-fun classify(embedding: FloatArray, classPrototypes: List<PrototypeEmbedding>, threshold: Float = 0.4f, confidenceMargin: Float = 0.05f ): ClassificationResult{
+fun fewShotClassify(embedding: FloatArray, classPrototypes: List<PrototypeEmbedding>, threshold: Float = 0.4f, confidenceMargin: Float = 0.05f ): ClassificationResult{
     if(classPrototypes.size < 2) return ClassificationResult.Failure(error= ClassificationError.MINIMUM_CLASS_SIZE) // Using a single class prototype leads to many false positives
 
     // No threshold filter applied here to allow confidence check by comparing top 2 matches
@@ -24,6 +24,4 @@ fun classify(embedding: FloatArray, classPrototypes: List<PrototypeEmbedding>, t
 
     val classId = classPrototypes[bestIndex].id
     return ClassificationResult.Success(classId=classId, similarity = bestSim)
-}
-
-
+}

core/src/main/java/com/fpf/smartscansdk/core/data/Processors.kt

@@ -56,3 +56,7 @@ fun unflattenEmbeddings(flattened: FloatArray, embeddingDim: Int): List<FloatArr
     }
     return embeddings
 }
+
+
+
+

ml/src/main/java/com/fpf/smartscansdk/ml/models/providers/embeddings/FewShotClassifier.kt renamed to core/src/main/java/com/fpf/smartscansdk/core/embeddings/Classification.kt

@@ -1,5 +1,4 @@
-package com.fpf.smartscansdk.core.data
-
+package com.fpf.smartscansdk.core.embeddings
 
 // `Embedding` represents a raw vector for a single media item, with `id` corresponding to its `MediaStoreId`.
 data class Embedding(
@@ -14,7 +13,3 @@ data class PrototypeEmbedding(
     val date: Long,
     val embeddings: FloatArray
 )
-
-
-
-

core/src/main/java/com/fpf/smartscansdk/core/embeddings/EmbeddingUtils.kt

@@ -1,7 +1,6 @@
 package com.fpf.smartscansdk.core.embeddings
 
 import android.util.Log
-import com.fpf.smartscansdk.core.data.Embedding
 import kotlinx.coroutines.Dispatchers
 import kotlinx.coroutines.withContext
 import java.io.File
@@ -12,6 +11,7 @@ import java.io.RandomAccessFile
 import java.nio.ByteBuffer
 import java.nio.ByteOrder
 import java.nio.channels.FileChannel
+import kotlin.collections.map
 
 class FileEmbeddingStore(
     private val file: File,
@@ -25,13 +25,10 @@ class FileEmbeddingStore(
     }
 
     private var cache: LinkedHashMap<Long, Embedding>? = null
+    private var cachedIds: List<Long>? = null
 
     override val exists: Boolean get() = file.exists()
 
-    override val isCached: Boolean
-        get() = cache != null
-
-
     // prevent OOM in FileEmbeddingStore.save() by batching writes
     private suspend fun save(embeddingsList: List<Embedding>): Unit = withContext(Dispatchers.IO) {
         if (embeddingsList.isEmpty()) return@withContext
@@ -99,6 +96,19 @@ class FileEmbeddingStore(
         }
     }
 
+    suspend fun get(ids: List<Long>): List<Embedding> = withContext(Dispatchers.IO) {
+        val map = cache ?: run {
+            val all = get()
+            LinkedHashMap(all.associateBy { it.id })
+        }
+        val embeddings = mutableListOf<Embedding>()
+
+        for (id in ids) {
+            map.get(id)?.let { embeddings.add(it) }
+        }
+        embeddings
+    }
+
     override suspend fun add(newEmbeddings: List<Embedding>): Unit = withContext(Dispatchers.IO) {
         if (newEmbeddings.isEmpty()) return@withContext
 
@@ -185,29 +195,42 @@ class FileEmbeddingStore(
         }
     }
 
-    suspend fun get(ids: List<Long>): List<Embedding> = withContext(Dispatchers.IO) {
-        val map = cache ?: run {
-            val all = get()
-            LinkedHashMap(all.associateBy { it.id })
-        }
-        val embeddings = mutableListOf<Embedding>()
 
-        for (id in ids) {
-            map.get(id)?.let { embeddings.add(it) }
-        }
-        embeddings
+    override fun clear(){
+        cache = null
     }
 
-    suspend fun get(id: Long): Embedding? = withContext(Dispatchers.IO) {
-        val map = cache ?: run {
-            val all = get()
-            LinkedHashMap(all.associateBy { it.id })
+
+    override suspend fun query(embedding: FloatArray, topK: Int, threshold: Float): List<Embedding> {
+
+        cachedIds = null // clear on new search
+
+        val storedEmbeddings = get()
+
+        if (storedEmbeddings.isEmpty()) return emptyList()
+
+        val similarities = getSimilarities(embedding, storedEmbeddings.map { it.embeddings })
+        val resultIndices = getTopN(similarities, topK, threshold)
+
+        if (resultIndices.isEmpty()) return emptyList()
+
+        val idsToCache = mutableListOf<Long>()
+        val results = resultIndices.map{idx ->
+            idsToCache.add( storedEmbeddings[idx].id)
+            storedEmbeddings[idx]
         }
-        map.get(id)
+        cachedIds = idsToCache
+        return results
     }
 
-    override fun clear(){
-        cache = null
+    suspend fun query(start: Int, end: Int): List<Embedding> {
+        val ids = cachedIds ?: return emptyList()
+        val s = start.coerceAtLeast(0)
+        val e = end.coerceAtMost(ids.size)
+        if (s >= e) return emptyList()
+
+        val batch = get(ids.subList(s, e))
+        return batch
     }
 
 }

core/src/main/java/com/fpf/smartscansdk/core/data/Embedings.kt renamed to core/src/main/java/com/fpf/smartscansdk/core/embeddings/Embeddings.kt

@@ -0,0 +1,8 @@
+package com.fpf.smartscansdk.core.embeddings
+
+interface IDetectorProvider<T> {
+    fun closeSession() = Unit
+    suspend fun initialize()
+    fun isInitialized(): Boolean
+    suspend fun detect(data: T): Pair<List<Float>, List<FloatArray>>
+}

core/src/main/java/com/fpf/smartscansdk/core/embeddings/FileEmbeddingRetriever.kt

@@ -1,12 +1,15 @@
 package com.fpf.smartscansdk.core.embeddings
 
-import com.fpf.smartscansdk.core.data.Embedding
-
 interface IEmbeddingStore {
-    val isCached: Boolean
     val exists: Boolean
     suspend fun add(newEmbeddings: List<Embedding>)
     suspend fun remove(ids: List<Long>)
     suspend fun get(): List<Embedding>
     fun clear()
+
+    suspend fun query(
+        embedding: FloatArray,
+        topK: Int,
+        threshold: Float
+    ): List<Embedding>
 }