The custom prediction result always points to one result.

[woshichuanqilz]

1. To make the train quicker I reduce the num_objects to 2 and batch_size(which solve this issue,

My train code:

    # -*- coding: utf-8 -*-
    from imageai.Prediction.Custom import ModelTraining
    model_trainer = ModelTraining()
    model_trainer.setModelTypeAsResNet()
    model_trainer.setDataDirectory("heros_img")
    model_trainer.trainModel(num_objects=2, num_experiments=50, enhance_data=True, batch_size=7, show_network_summary=True)

test code:

    # -*- coding: utf-8 -*-
    from imageai.Prediction.Custom import CustomImagePrediction
    import os
    execution_path = os.getcwd()
    prediction = CustomImagePrediction()
    prediction.setModelTypeAsResNet()
    prediction.setModelPath(os.path.join(execution_path, "model_ex-050_acc-0.476190.h5"))
    prediction.setJsonPath(os.path.join(execution_path, "model_class.json"))
    prediction.loadModel(num_objects=2)
    predictions, probabilities = prediction.predictImage(os.path.join(execution_path, "test.png"), result_count=5)
    for eachPrediction, eachProbability in zip(predictions, probabilities):
        print(eachPrediction + " : " + eachProbability)

Whatever image I put in always give one test of one result, like, apple always be 99%+: apple : 99.97028708457947 banana : 0.02971168141812086

2. I just crop the image with the apple, is it can predict with large background?
3. After training I get so many h5 files which one should I use?

My goal is to use the imageai to recognize the heros in dota. Because dota images is much simpler than the real world, I guess less train will be work. One hero just has several images from some visual angles. Am I right, hope can get some advice from you, thank you ~

[LittleOrangeC]

For your question #3 I believe you should use the .h5 file with the highest accuracy (acc) score.

[OlafenwaMoses]

How many images did you use to train your model? It is paramount that your training images for each type of objects should be up to 500 numbers or more to ensure your generated models can predict new images correctly.

[andersffs]

Getting the same error with >900 train images.