`1.7.1` `RandomForestRegressor.fit()` hits `ValueError: Cannot cast object dtype to float32`

[khteh]

Describe the bug

I have converted the object dtype to category but it keeps hitting this error. I try LinearRegression and hits the same error.

Steps/Code to Reproduce

https://github.com/khteh/JupyterNotebooks/blob/master/ZTM/BulldozerPricingRegression.ipynb

data: https://github.com/mrdbourke/zero-to-mastery-ml/raw/refs/heads/master/data/bluebook-for-bulldozers.zip

Expected Results

No error, duh!

Actual Results

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/pandas/core/arrays/categorical.py:592, in Categorical.astype(self, dtype, copy)
    591 try:
--> 592     new_cats = new_cats.astype(dtype=dtype, copy=copy)
    593     fill_value = self.categories._na_value

ValueError: could not convert string to float: '100C'

During handling of the above exception, another exception occurred:

ValueError                                Traceback (most recent call last)
Cell In[205], line 1
----> 1 get_ipython().run_cell_magic('time', '', '#print("\\nX_train.info:")\n#X_train.info()\n#print("\\nY_train.info:")\n#Y_train.info()\nY_train.head()\nmodel = RandomForestRegressor(n_jobs=-1)\nmodel.fit(X=X_train, y=Y_train)\n')

File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/IPython/core/interactiveshell.py:2565, in InteractiveShell.run_cell_magic(self, magic_name, line, cell)
   2563 with self.builtin_trap:
   2564     args = (magic_arg_s, cell)
-> 2565     result = fn(*args, **kwargs)
   2567 # The code below prevents the output from being displayed
   2568 # when using magics with decorator @output_can_be_silenced
   2569 # when the last Python token in the expression is a ';'.
   2570 if getattr(fn, magic.MAGIC_OUTPUT_CAN_BE_SILENCED, False):

File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/IPython/core/magics/execution.py:1470, in ExecutionMagics.time(self, line, cell, local_ns)
   1468 if interrupt_occured:
   1469     if exit_on_interrupt and captured_exception:
-> 1470         raise captured_exception
   1471     return
   1472 return out

File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/IPython/core/magics/execution.py:1439, in ExecutionMagics.time(self, line, cell, local_ns)
   1437     if expr_val is not None:
   1438         code_2 = self.shell.compile(expr_val, source, 'eval')
-> 1439         out = eval(code_2, glob, local_ns)
   1440 except KeyboardInterrupt as e:
   1441     captured_exception = e

File <timed exec>:7

File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/sklearn/base.py:1365, in _fit_context.<locals>.decorator.<locals>.wrapper(estimator, *args, **kwargs)
   1358     estimator._validate_params()
   1360 with config_context(
   1361     skip_parameter_validation=(
   1362         prefer_skip_nested_validation or global_skip_validation
   1363     )
   1364 ):
-> 1365     return fit_method(estimator, *args, **kwargs)

File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/sklearn/ensemble/_forest.py:359, in BaseForest.fit(self, X, y, sample_weight)
    356 if issparse(y):
    357     raise ValueError("sparse multilabel-indicator for y is not supported.")
--> 359 X, y = validate_data(
    360     self,
    361     X,
    362     y,
    363     multi_output=True,
    364     accept_sparse="csc",
    365     dtype=DTYPE,
    366     ensure_all_finite=False,
    367 )
    368 # _compute_missing_values_in_feature_mask checks if X has missing values and
    369 # will raise an error if the underlying tree base estimator can't handle missing
    370 # values. Only the criterion is required to determine if the tree supports
    371 # missing values.
    372 estimator = type(self.estimator)(criterion=self.criterion)

File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/sklearn/utils/validation.py:2971, in validate_data(_estimator, X, y, reset, validate_separately, skip_check_array, **check_params)
   2969         y = check_array(y, input_name="y", **check_y_params)
   2970     else:
-> 2971         X, y = check_X_y(X, y, **check_params)
   2972     out = X, y
   2974 if not no_val_X and check_params.get("ensure_2d", True):

File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/sklearn/utils/validation.py:1368, in check_X_y(X, y, accept_sparse, accept_large_sparse, dtype, order, copy, force_writeable, force_all_finite, ensure_all_finite, ensure_2d, allow_nd, multi_output, ensure_min_samples, ensure_min_features, y_numeric, estimator)
   1362     raise ValueError(
   1363         f"{estimator_name} requires y to be passed, but the target y is None"
   1364     )
   1366 ensure_all_finite = _deprecate_force_all_finite(force_all_finite, ensure_all_finite)
-> 1368 X = check_array(
   1369     X,
   1370     accept_sparse=accept_sparse,
   1371     accept_large_sparse=accept_large_sparse,
   1372     dtype=dtype,
   1373     order=order,
   1374     copy=copy,
   1375     force_writeable=force_writeable,
   1376     ensure_all_finite=ensure_all_finite,
   1377     ensure_2d=ensure_2d,
   1378     allow_nd=allow_nd,
   1379     ensure_min_samples=ensure_min_samples,
   1380     ensure_min_features=ensure_min_features,
   1381     estimator=estimator,
   1382     input_name="X",
   1383 )
   1385 y = _check_y(y, multi_output=multi_output, y_numeric=y_numeric, estimator=estimator)
   1387 check_consistent_length(X, y)

File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/sklearn/utils/validation.py:971, in check_array(array, accept_sparse, accept_large_sparse, dtype, order, copy, force_writeable, force_all_finite, ensure_all_finite, ensure_non_negative, ensure_2d, allow_nd, ensure_min_samples, ensure_min_features, estimator, input_name)
    966 if pandas_requires_conversion:
    967     # pandas dataframe requires conversion earlier to handle extension dtypes with
    968     # nans
    969     # Use the original dtype for conversion if dtype is None
    970     new_dtype = dtype_orig if dtype is None else dtype
--> 971     array = array.astype(new_dtype)
    972     # Since we converted here, we do not need to convert again later
    973     dtype = None

File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/pandas/core/generic.py:6662, in NDFrame.astype(self, dtype, copy, errors)
   6656     results = [
   6657         ser.astype(dtype, copy=copy, errors=errors) for _, ser in self.items()
   6658     ]
   6660 else:
   6661     # else, only a single dtype is given
-> 6662     new_data = self._mgr.astype(dtype=dtype, copy=copy, errors=errors)
   6663     res = self._constructor_from_mgr(new_data, axes=new_data.axes)
   6664     return res.__finalize__(self, method="astype")

File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/pandas/core/internals/managers.py:430, in BaseBlockManager.astype(self, dtype, copy, errors)
    427 elif using_copy_on_write():
    428     copy = False
--> 430 return self.apply(
    431     "astype",
    432     dtype=dtype,
    433     copy=copy,
    434     errors=errors,
    435     using_cow=using_copy_on_write(),
    436 )

File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/pandas/core/internals/managers.py:363, in BaseBlockManager.apply(self, f, align_keys, **kwargs)
    361         applied = b.apply(f, **kwargs)
    362     else:
--> 363         applied = getattr(b, f)(**kwargs)
    364     result_blocks = extend_blocks(applied, result_blocks)
    366 out = type(self).from_blocks(result_blocks, self.axes)

File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/pandas/core/internals/blocks.py:784, in Block.astype(self, dtype, copy, errors, using_cow, squeeze)
    781         raise ValueError("Can not squeeze with more than one column.")
    782     values = values[0, :]  # type: ignore[call-overload]
--> 784 new_values = astype_array_safe(values, dtype, copy=copy, errors=errors)
    786 new_values = maybe_coerce_values(new_values)
    788 refs = None

File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/pandas/core/dtypes/astype.py:237, in astype_array_safe(values, dtype, copy, errors)
    234     dtype = dtype.numpy_dtype
    236 try:
--> 237     new_values = astype_array(values, dtype, copy=copy)
    238 except (ValueError, TypeError):
    239     # e.g. _astype_nansafe can fail on object-dtype of strings
    240     #  trying to convert to float
    241     if errors == "ignore":

File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/pandas/core/dtypes/astype.py:179, in astype_array(values, dtype, copy)
    175     return values
    177 if not isinstance(values, np.ndarray):
    178     # i.e. ExtensionArray
--> 179     values = values.astype(dtype, copy=copy)
    181 else:
    182     values = _astype_nansafe(values, dtype, copy=copy)

File ~/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/pandas/core/arrays/categorical.py:603, in Categorical.astype(self, dtype, copy)
    598     except (
    599         TypeError,  # downstream error msg for CategoricalIndex is misleading
    600         ValueError,
    601     ):
    602         msg = f"Cannot cast {self.categories.dtype} dtype to {dtype}"
--> 603         raise ValueError(msg)
    605     result = take_nd(
    606         new_cats, ensure_platform_int(self._codes), fill_value=fill_value
    607     )
    609 return result

ValueError: Cannot cast object dtype to float32

Versions

System:
    python: 3.13.3 (main, Aug 14 2025, 11:53:40) [GCC 14.2.0]
executable: /home/khteh/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/bin/python
   machine: Linux-6.14.0-28-generic-x86_64-with-glibc2.41

Python dependencies:
      sklearn: 1.7.1
          pip: 25.0
   setuptools: 80.9.0
        numpy: 2.3.2
        scipy: 1.16.1
       Cython: None
       pandas: 2.3.1
   matplotlib: 3.10.5
       joblib: 1.5.1
threadpoolctl: 3.6.0

Built with OpenMP: True

threadpoolctl info:
       user_api: blas
   internal_api: openblas
    num_threads: 16
         prefix: libscipy_openblas
       filepath: /home/khteh/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/numpy.libs/libscipy_openblas64_-8fb3d286.so
        version: 0.3.30
threading_layer: pthreads
   architecture: SkylakeX

       user_api: blas
   internal_api: openblas
    num_threads: 16
         prefix: libscipy_openblas
       filepath: /home/khteh/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/scipy.libs/libscipy_openblas-68440149.so
        version: 0.3.28
threading_layer: pthreads
   architecture: SkylakeX

       user_api: openmp
   internal_api: openmp
    num_threads: 16
         prefix: libgomp
       filepath: /home/khteh/.local/share/virtualenvs/JupyterNotebooks-uVG1pv5y/lib/python3.13/site-packages/scikit_learn.libs/libgomp-a34b3233.so.1.0.0
        version: None

[Shashank200702]

Hi, I’m new to open-source and interested in contributing. Could I try helping by writing a minimal reproducible example or adding a unit test to confirm this bug?

[Aishwarya0811]

hi @khteh after analyzing your .ipynb file code the problem is Categorical columns causing the error The error occurs because RandomForestRegressor requires all features to be numeric, but the bulldozer dataset contains numerous categorical columns

Sklearn internally attempts to cast all input data to numpy.float32,.RandomForestRegressor.html) which fails when encountering string values like "Large", "Alabama", or "Track Type Tractor, Dozer".

When model.fit() executes, X_train contains object columns that cause the casting error.

The exact fix needed
Solution 1: Complete preprocessing function
Solution 2: Pipeline-based approach
Solution 3: Emergency fix for immediate testing

[khteh]

Is there any recent change in your model implementation? Isn't there categorical codes that the models are supposed to use instead of their non-numerical labels? There is data preprocessing in my .ipynb:

# Convert object dtype to category dtype
for k,v in data.items():
    # v is Series
    if pd.api.types.is_object_dtype(v):
        data[k] = v.astype("category").cat.as_ordered() # This will assign a numerical value to each category
    if pd.isna(v).sum() or pd.isnull(v).sum():
        # Fix missing numerical values        
        if pd.api.types.is_numeric_dtype(v):
            # Add a flag to mark originally missing value
            data[f"{k}_missing"] = pd.isnull(v)
            # Fill missing values with media (NOT mean as mean is more supceptible to outliers)
            data[k] = v.fillna(v.median())
        # Fix missing categorical values
        else:
            # Add a flag to mark originally missing value
            data[f"{k}_missing"] = pd.isnull(v)
            # Turn categories into numbers and +1
            data[k] = pd.Categorical(v).codes + 1 # NULL / N.A will be assigned -1 by default. +1 here will offset all codes by +1. So, Null / N.A will have 0

Don't understand what you mean with your "Solution 3"

[virchan]

Thank you for reporting this issue, @khteh, and for helping to make scikit-learn better.

I haven't looked into your dataset yet, but from the error message you shared, it seems you may need to apply some preprocessing, such as using OneHotEncoder, to convert string-valued columns into numerical values before feeding them into RandomForestRegressor.

[khteh]

No, I expect it to use the categorical codes. Otherwise, what's the use case of Category dtype?

[virchan]

No, I expect it to use the categorical codes. Otherwise, what's the use case of Category dtype?

The category dtype is a Pandas construct (see Pandas' user guide).

Scikit-learn estimators only work with numerical features, so categorical values need to be explicitly transformed, as recommend in the FAQ page.

[khteh]

What about these lines of code?

        data[k] = v.astype("category").cat.as_ordered() # This will assign a numerical value to each category
...
            data[k] = pd.Categorical(v).codes + 1 # NULL / N.A will be assigned -1 by default. +1 here will offset all codes by +1. So, Null / N.A will have 0

[virchan]

What about these lines of code?

        data[k] = v.astype("category").cat.as_ordered() # This will assign a numerical value to each category
...
            data[k] = pd.Categorical(v).codes + 1 # NULL / N.A will be assigned -1 by default. +1 here will offset all codes by +1. So, Null / N.A will have 0

Here’s the situation:

pandas.Series.cat.as_ordered sets a categorical column as ordered, but does not convert it into numeric values.

The line

data[k] = pd.Categorical(v).codes + 1

only runs inside the if block:

if pd.isna(v).sum() or pd.isnull(v).sum():

This means that any categorical column without missing values never gets converted to numeric.

That explains why the training set still contains a string like '100C' when passed to RandomForestRegressor, as indicated by the ValueError.

Therefore, the issue lies in the data preprocessing pipeline, not in scikit-learn’s code or documentation.

With that said, I’m marking this issue as resolved.