Hey
We are testing a project on xcode 14 beta 5 and we have an issue with a model that is simply Apple's Vision Feature Print (embeddings). The model has the input 299x299, then a visionFeaturePrint layer and the output is float64[2048]. The model is in Core ML Package v3 and was created using CoreML Tools, cutting the layer added by Create ML into a classification model. The result depends solely on the interaction that invokes the prediction despite the input image (simulator/Apple M1 chip). On the device works as expected.
let config = MLModelConfiguration()
#if targetEnvironment(simulator)
config.computeUnits = .cpuOnly
#else
config.computeUnits = .all
#endif
model = try! ImageSemanticInfo_iOS(configuration: config)
let buffer = thumb!.toCVPixelBuffer()!
for _ in 0..<3{
let results = try! model!.prediction(image: buffer).sceneprint
}
For example, if we take just the first entry of the embedding, we will always get the following results, regardless of the input image used:
0.474750816822052 - First call
0.3231460750102997 - Second call
0.37376347184181213 - Third call