use downsampling with efficientsam

src/main/java/ai/nets/samj/models/AbstractSamJ.java

@@ -71,7 +71,7 @@ public abstract class AbstractSamJ implements AutoCloseable {
 
 	protected static long ENCODE_MARGIN = 64;
 
-	protected static int MAX_IMG_SIZE = 2048;
+	protected static int MAX_IMG_SIZE = 2024;
 
 	/** Essentially, a syntactic-shortcut for a String consumer */
 	public interface DebugTextPrinter { void printText(String text); }
@@ -995,7 +995,7 @@ protected void recalculatePolys(List<Mask> masks, long[] encodeCoords) {
 				int newY = y * scale;
 				long newLen = pp.getRLEMask()[i + 1] * scale;
 				for (int j = 0; j < scale; j ++) {
-					upscaledRLE[i * scale + j * 2] = newX + encodeCoords[0] + (newY + j) * this.img.dimensionsAsLongArray()[0];
+					upscaledRLE[i * scale + j * 2] = newX + encodeCoords[0] + (encodeCoords[1] + newY + j) * this.img.dimensionsAsLongArray()[0];
 					upscaledRLE[i * scale + j * 2 + 1] = newLen;
 				}
 			}

src/main/java/ai/nets/samj/models/EfficientSamJ.java

@@ -201,13 +201,15 @@ protected void createEncodeImageScript() {
 		code += "im_shm = shared_memory.SharedMemory(name='"
 							+ shma.getNameForPython() + "', size=" + shma.getSize() 
 							+ ")" + System.lineSeparator();
-		int size = 1;
-		for (long l : targetDims) {size *= l;}
+		int size = (int) targetDims[2];
+		for (int i = 0; i < targetDims.length - 1; i ++) {
+			size *= Math.ceil(targetDims[i] / (double) scale);
+			}
 		code += "im = np.ndarray(" + size + ", dtype='" + CommonUtils.getDataTypeFromRAI(Cast.unchecked(shma.getSharedRAI()))
 			  + "', buffer=im_shm.buf).reshape([";
-		for (long ll : targetDims)
-			code += ll + ", ";
-		code = code.substring(0, code.length() - 2);
+		for (int i = 0; i < targetDims.length - 1; i ++)
+			code += (int) Math.ceil(targetDims[i] / (double) scale) + ", ";
+		code += targetDims[2];
 		code += "])" + System.lineSeparator();
 		//code += "np.save('/home/carlos/git/crop.npy', im)" + System.lineSeparator();
 		code += "input_h = im.shape[1]" + System.lineSeparator();