UnboundLocalError("local variable 'values' referenced before assignment") about what-if-tool HOT 7 CLOSED

Nice to see that the function works with your manual testing. Are you able to log the input to the custom predict fn when call from the UI to see how it differs from your manual testing?

from what-if-tool.

it may be the case that if you pass a single example through the ui (by pressing predict button after making an edit), then your custom predict fn is returning a raw values instead of a list of length 1 of the value to return?

from what-if-tool.

The issue is most likely in your custom prediction fn. Perhaps you can write your predict function with logging to see exactly what is passed in, if its in the format your model expects, or if you need to preprocess the data sent to the function before you pass it to your model.

from what-if-tool.

Thanks @jameswex for the input, really appreciate it. Now, I have written another custom_function which takes a train.Example and performs the predict, however I'm getting a new error TypeError("object of type 'float' has no len()")

My custom predict function:

def custom_predict_fn(examples):
  if type(examples) == list:
    feats_all = []
    for example in examples:
      feats = []
      for col in features:
        #print(col)
        p = example.features.feature[col].float_list.value
        #print(p)
        feats.append(p[0])
      feats_all.append(feats)
  else:
    feats = []
    for col in features:
      #print(col)
      p = examples.features.feature[col].float_list.value
      #print(p)
      feats.append(p[0])
    feats_all = [feats]
  return model.predict(np.array(feats_all)).tolist()

I have tested the function with a slice of train.Example

test_examples = df_to_examples(test_features[0:10])
custom_predict_fn(test_examples[2:5])
Out: [2.229121702878904e-05, 6.0774786297023536e-05, 1.2393046192705309e-05]

custom_predict_fn(test_examples[2])
[2.229121702878904e-05]

It works correctly. However on UI, I get the error: TypeError("object of type 'float' has no len()")

This is how my train.Example sample looks like

features {
  feature {
    key: "a1"
    value {
      float_list {
        value: 1.0
      }
    }
  }
  feature {
    key: "a2"
    value {
      float_list {
        value: 1.0
      }
    }
  }
  feature {
    key: "a3"
    value {
      float_list {
        value: 1.0
      }
    }
  }
  feature {
    key: "cj_val"
    value {
      float_list {
        value: -0.9179026484489441
      }
    }
  }
  feature {
    key: "cos"
    value {
      float_list {
        value: -0.561825692653656
      }
    }
  }
  feature {
    key: "hr"
    value {
      float_list {
        value: 35.0
      }
    }
  }
  feature {
    key: "iscore"
    value {
      float_list {
        value: 8.0
      }
    }
  }
  feature {
    key: "label"
    value {
      float_list {
        value: 0.0
      }
    }
  }
  feature {
    key: "max_pct_2"
    value {
      float_list {
        value: 48.029998779296875
      }
    }
  }
  feature {
    key: "max_pct"
    value {
      float_list {
        value: 0.4039580523967743
      }
    }
  }

from what-if-tool.

Okay, interesting. I have tested it with single example, it still returns a list, something like [0.4]. I even added a check just before the return value, it seems no error is taking place there; the error is taking place after the value is returned most probably. Is there any specific format that the UI expects internally? I seems it's trying to get a length of a float value, maybe a feature?

from what-if-tool.

Okay, it's working now. From the comments, I found out the prediction should have shape like (batch, num_classes) for classification, but my model was predicting a single score for a binary classification problem. Now, it's working as expected. Thanks @jameswex

from what-if-tool.

Great to hear!

from what-if-tool.