[python] dont format string values with precision

[JasonTam]

I was running into this:
ValueError: Unknown format code 'f' for object of type 'str'
for string values when setting precision in create_tree_digraph

[StrikerRUS]

@JasonTam Hi! Thanks for creating PR.

Can you please provide the example in which you got described error?

UPD:
... or even better enhance the test here, so that it could be passed only with your modification.

[JasonTam]

Sure, here is a minimal example:

import pandas as pd
import numpy as np
import lightgbm as lgb
np.random.seed(0)

x = pd.concat([
    pd.Series(np.random.choice(['a', 'b'], size=(100)), dtype='category'),
    pd.Series(np.random.randint(0, 2, size=100)),
], axis=1)
y = np.random.randint(0, 2, size=100)
ds = lgb.Dataset(x, y)

params = {'min_child_samples': 2}
gbm = lgb.train(params, ds)
lgb.create_tree_digraph(gbm, precision=2)

Not sure of the best way to modify the test:, the cheapest thing I could do is to convert all 30 columns to categorical. This guarantees that the test only passes with the modification. But I feel like it's not good for the test overall.

gbm.fit(self.X_train, self.y_train, categorical_feature=list(range(30)), verbose=False)

If I only convert a couple of columns to categorical, the model will just try to ignore those and split on the other columns.

[StrikerRUS]

Thanks for the example!

The error is caused by the scientific notation, e.g. 1.0000000180025095e-35. However, we are sill able to convert this value to float with specified precision instead of silent ignoring precision argument.

I think that float2str could be rewritten as follows:

    def float2str(value, precision=None):
        if precision is None:
            return str(value)
        else:
            try:
                return "{0:.{1}f}".format(float(value), precision)
            except TypeError:
                warnings.warn('Some notification that precison cannot be set to some values')
                return str(value)

Now it works correctly with scientific notation and notifies user that precision cannot be set in case of strange values.

[JasonTam]

I thought the error is caused by categorical values? The split threshold would be on a set of categories ie: "0||1||2" which is a string. In my fix, I ignore precision if the value is of type string. Not sure if scientific notation is related to this.

[StrikerRUS]

LightGBM cannot handle strings at all (see #789).

[JasonTam]

Im not talking about the categorical values or codes of the dataset itself. I'm talking about the threshold that will be shown in the plot. It's a string if the split is based on set membership.

[StrikerRUS]

Threshold cannot be string as well.

[JasonTam]

Idk man, when I use categorical data, value becomes a string when plotting. You can try my example with a debugger and take a look at value and see what I'm talking about

[StrikerRUS]

Just run your example completely without float2str function:

[StrikerRUS]

Please provide the example and the resulted graph with strings you're talking about.

[JasonTam]

Not sure exactly what you're asking for.
With my example on master branch, I can't create the graph -- it's going to error out with the following:

ValueError: Unknown format code 'f' for object of type 'str'
> /Users/jtam/tools/LightGBM/python-package/lightgbm/plotting.py(271)float2str()
    269 
    270     def float2str(value, precision=None):
--> 271         return "{0:.{1}f}".format(value, precision) if precision is not None else str(value)
    272 
    273     def add(root, parent=None, decision=None):

ipdb> value
'1'

If I apply my PR, I will get the following:

The 1 in the top node is actually a string (note how it's been skipped in the precision formatting , where all other values actually have precision=2).

Could it perhaps be my python or pandas version? -- not sure
(Python 3.6.0 pandas 0.22.0)

[StrikerRUS]

Oh, I see. I thought you're having the threshold value like this one "0||1||2", but you're talking about the one element from this set "1".

[JasonTam]

I actually do have thresholds like "0|1|2" , but for a private dataset. If I have time, I'll try to write a small example with it too.

(^ graph rendered after applying PR)

[StrikerRUS]

Please remove the brackets and use string_type from compat.py module instead of str in isinstance().

[StrikerRUS]

If I have time, I'll try to write a small example with it too.

OK, thank you! No need to do this.

[StrikerRUS]

Please remove these brackets too and I'll merge.

(precision is not None)

[JasonTam]

I think the travis error is unrelated?

[StrikerRUS]

Thank you very much for your contribution and patience!

I think the travis error is unrelated?

Yeah, I've restart the test.