How to configure predict pipeline for HomoNN & image dataset

[Drfarree]

Hi!

I'm trying to implement a predict pipeline for HomoNN using an image dataset (standard pytorch ImageFolder).

I attempted to follow the instructions in this documentation, specifically the part where I add the reader corresponding to the inference date:

predict_pipeline = PipeLine()
predict_pipeline.add_component(reader_0)
predict_pipeline.add_component(pipeline,
                               data=Data(predict_input={pipeline.data_transform_0.input.data: reader_0.output.data}))

However, the data transform component seems to be prepared for tabular data. When I try to adapt this to image data, and specify the reader accordingly, I fail to run the pipeline. I have tried several different configurations, unsuccessfully (see predict_pipeline.add_component(pipeline, data=...).

fate_project_path = os.path.abspath('../')
guest = 10000
host = 9999

# Load pipeline and deploy components
pipeline = PipeLine.load_model_from_file('pipeline_saved_metastasis_v3_pretrained_imagenet.pkl')
pipeline.deploy_component(["train_NN_0", "eval_0"])

# Define prediction pipeline
predict_pipeline = PipeLine()

# Define data reader
data_predict = {"name": "meta_center_2", "namespace": "experiment"}
data_path_predict = fate_project_path + '/examples/data/metastasis/center_2/inference'

reader_predict = Reader(name="reader_predict")
predict_pipeline.bind_table(name=data_predict['name'], namespace=data_predict['namespace'], path=data_path_predict)
reader_predict.get_party_instance(role='guest', party_id=guest).component_param(table=data_predict)
reader_predict.get_party_instance(role='host', party_id=host).component_param(table=data_predict)

# Add reader and train pipelines as components
predict_pipeline.add_component(reader_predict)
predict_pipeline.add_component(
    pipeline,
    data=Data(predict_input=reader_predict.output.data)
    #data=Data(predict_input={pipeline.get_component('reader_val.data'): reader_predict.output.data})
    #data=Data(validate_data=reader_predict.output.data)
    #data=Data(data=reader_predict.output.data)
)

predict_pipeline.predict()

The predict pipeline seems to fail to compile:

(The train pipeline runs smoothly and is correctly loaded through the pkl.)

Any insight would be greatly appreciated!