CoreML MLModel 无法获取 .modelDescription 参数 SqueezeNetInt8LUT

Question

I am trying to use SqueezeNet to make a prediction on a test image (that is just an asset in my project) using Objective-C. I am struggling quite a bit, in part because the documentation examples are all in Swift.

The issue I am encountering is that while I can instantiate a squeezeNet model, I cannot get the .modelDescription parameter, neither in the debugger nor runtime.

The code to causing this is similar to this:

@interface TestViewController ()
@property (strong, nonatomic) MLModel* model;
@end

@implementation TestViewController

- (void)viewDidLoad {
    [super viewDidLoad];
    NSURL* modelURL = [[NSBundle mainBundle] URLForResource:@"SqueezeNetInt8LUT" withExtension:@"mlmodelc"];
    MLModelConfiguration* config = [MLModelConfiguration new];
    self.model = [[SqueezeNetInt8LUT alloc] initWithContentsOfURL:modelURL error:nil];

}

- (IBAction)didTapTest:(id)sender {

    NSLog(@"%@", self.model.modelDescription);

    UIImage* testImage = [UIImage imageNamed:@"mountain"];
    MLImageConstraint* constraint = self.model.modelDescription.inputDescriptionsByName[@"image"].imageConstraint;

    MLFeatureValue* imageFeature = [MLFeatureValue featureValueWithCGImage:cgtest constraint:constraint options:nil error:nil];
    NSMutableDictionary* featureDict = [[NSMutableDictionary alloc] init];
    featureDict[@"mountain"] = imageFeature;
    MLDictionaryFeatureProvider* featureProv = [[MLDictionaryFeatureProvider new] initWithDictionary:featureDict error:nil];
    MLDictionaryFeatureProvider* pred = [self.model predictionFromFeatures:featureProv error:nil];
}
@end

Where the exception is thrown as soon as I do the NSLog. Any ideas? Did I initialize the model wrong? Despite po self.model giving <SqueezeNetInt8LUT: 0x600001d90040>

Answer 1

There were several issues with the above code that are not apparent from spending a whole day reading CoreML documentation and examples:

1. Do NOT do this: @property (strong, nonatomic) MLModel* model; instead do @property (strong, nonatomic) SqueezeNetInt8LUT* model;

2. I named the entry in the MLDictionaryFeatureProvider @"mountain", but unlike a normal dictionary, the key name for FeatureProviders MUST match the name of the input parameter expected by the model, as shown in the image attached.

so the line
featureDict[@"mountain"] = imageFeature; needed to be featureDict[@"image"] = imageFeature;