问题
I keep getting this error in my linear model:
Cast string to float is not supported
Specifically, the error is on this line:
results = m.evaluate(input_fn=lambda: input_fn(df_test), steps=1)
If it helps, here's the stack trace:
File "tensorflowtest.py", line 164, in <module>
m.fit(input_fn=lambda: input_fn(df_train), steps=int(100))
File "/home/computer/.local/lib/python2.7/site-packages/tensorflow/contrib/learn/python/learn/estimators/linear.py", line 475, in fit
max_steps=max_steps)
File "/home/computer/.local/lib/python2.7/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 333, in fit
max_steps=max_steps)
File "/home/computer/.local/lib/python2.7/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 662, in _train_model
train_op, loss_op = self._get_train_ops(features, targets)
File "/home/computer/.local/lib/python2.7/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 963, in _get_train_ops
_, loss, train_op = self._call_model_fn(features, targets, ModeKeys.TRAIN)
File "/home/computer/.local/lib/python2.7/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 944, in _call_model_fn
return self._model_fn(features, targets, mode=mode, params=self.params)
File "/home/computer/.local/lib/python2.7/site-packages/tensorflow/contrib/learn/python/learn/estimators/linear.py", line 220, in _linear_classifier_model_fn
loss = loss_fn(logits, targets)
File "/home/computer/.local/lib/python2.7/site-packages/tensorflow/contrib/learn/python/learn/estimators/linear.py", line 141, in _log_loss_with_two_classes
logits, math_ops.to_float(target))
File "/home/computer/.local/lib/python2.7/site-packages/tensorflow/python/ops/math_ops.py", line 661, in to_float
return cast(x, dtypes.float32, name=name)
File "/home/computer/.local/lib/python2.7/site-packages/tensorflow/python/ops/math_ops.py", line 616, in cast
return gen_math_ops.cast(x, base_type, name=name)
File "/home/computer/.local/lib/python2.7/site-packages/tensorflow/python/ops/gen_math_ops.py", line 419, in cast
result = _op_def_lib.apply_op("Cast", x=x, DstT=DstT, name=name)
File "/home/computer/.local/lib/python2.7/site-packages/tensorflow/python/framework/op_def_library.py", line 749, in apply_op
op_def=op_def)
File "/home/computer/.local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 2380, in create_op
original_op=self._default_original_op, op_def=op_def)
File "/home/computer/.local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 1298, in __init__
self._traceback = _extract_stack()
UnimplementedError (see above for traceback): Cast string to float is not supported
[[Node: ToFloat = Cast[DstT=DT_FLOAT, SrcT=DT_STRING, _device="/job:localhost/replica:0/task:0/cpu:0"](Reshape_1)]]
The model is an adaptation of the tutorial from here and here. The tutorial code does run, so it's not a problem with my TensorFlow installation.
The input CSV is data in the form of many binary categorical columns (yes/no). Initially, I represented the data in each column as 0's and 1's, but I get the same error when I change it to ys and ns.
How do I fix this?
回答1:
I had the exact same problem, you need to make sure that the input data you are feeding the model is in the right format. ( not just the features but also the label column)
My problem was that i was not skipping the first row in the data file, so i was trying to convert the titles to float format.Something as simple as adding
skiprows=1
When reading the csv:
df_test = pd.read_csv(test_file, names=COLUMNS_TEST, skipinitialspace=True, skiprows=1, engine="python")
I would recommend you to check:
df_test.dtypes
You should get something like
Feature1 int64
Feature2 int64
Feature3 int64
Feature4 object
Feature5 object
Feature6 float64
dtype: object
If you are not getting the correct dtype then the model.fit is going to fail
回答2:
You can't literally cast a string to a number, particularly "y","n" to 1.0/0.0.
If you have numeric strings (e.g. "0") you could try tf.string_to_number(..)
回答3:
The problem is that you probably had indicated the feature like a real type but in your data frame is still string or when set in tf.constant you didn't cast to correct type.
Confirm the types of your columns. You can check just type (df is your dataframe):
df.info()
And you can see all columns and types, some like that:
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 178932 entries, 0 to 178931
Data columns (total 64 columns):
d_prcp 178932 non-null float64
d_stn 178932 non-null int64
ws_lat 178932 non-null float64
ws_lon 178932 non-null float64
d_year 178932 non-null int64
d_temp 178932 non-null float64
...
You can use this bellow function in order to convert your data in correct type in tensorflow. (this code is from a repo google/training-data-analyst: link here)
def make_input_fn(df):
def pandas_to_tf(pdcol):
# convert the pandas column values to float
t = tf.constant(pdcol.astype('float32').values)
# take the column which is of shape (N) and make it (N, 1)
return tf.expand_dims(t, -1)
def input_fn():
# create features, columns
features = {k: pandas_to_tf(df[k]) for k in FEATURES}
labels = tf.constant(df[TARGET].values)
return features, labels
return input_fn
def make_feature_cols():
input_columns = [tf.contrib.layers.real_valued_column(k) for k in FEATURES]
return input_columns
回答4:
I am using W10, Python3 and Tensorflow 1.9
The source of the error in my code was in the feature definition. I had a boolean feature with a default_value of -1 like this:
tf.feature_column.categorical_column_with_vocabulary_list(
key='partial_funding_indicator', vocabulary_list=['True', 'False'],
dtype=tf.string, **default_value=-1**, num_oov_buckets=None)
The issue did not arise when the default_value was changed to 0:
tf.feature_column.categorical_column_with_vocabulary_list(
key='partial_funding_indicator', vocabulary_list=['True', 'False'],
dtype=tf.string, **default_value=0**, num_oov_buckets=None)
default_value is the integer ID value to return for out-of-vocabulary feature values. For example, in a list/file of value like ['True', 'False'] to make default_value == True, it would be default_value=0; the list index.
回答5:
Your classes are probably in string forms and they need to be numeric (1 and 0 only for this specific tutorial)
回答6:
Normally this error is because m.evaluate is somehow empty.
Since you load your data from csv file, it is very likely that your data was stored as string instead of float or int inside the array. I suggest you check it manually to make sure.
回答7:
Somewhere in your code you're using tf.cast() to convert a string to number, but you cannot do this. Replace it with tf.strings.to_number():
tf.strings.to_number(x, out_type=tf.float32)
来源：https://stackoverflow.com/questions/40186722/cast-string-to-float-is-not-supported-in-linear-model