How to convert tf.int64 to tf.float32?

Question

I tried:

test_image = tf.convert_to_tensor(img, dtype=tf.float32)

Then following error appears:

ValueError: Tensor conversion requested dtype float32 for Tensor with dtype int64: 'Tensor("test/ArgMax:0", shape=TensorShape([Dimension(None)]), dtype=int64)'

Mark McDonald · Answer 1 · 2016-11-10T23:31:34.440

57

You can cast generally using:

tf.cast(my_tensor, tf.float32)

Replace tf.float32 with your desired type.

Edit: It seems at the moment at least, that tf.cast won't cast to an unsigned dtype (e.g. tf.uint8). To work around this, you can cast to the signed equivalent and used tf.bitcast to get all the way. e.g.

tf.bitcast(tf.cast(my_tensor, tf.int8), tf.uint8)

edited Nov 10 '16 at 23:31

answered Jun 14 '16 at 10:38

Mark McDonald

7,571
6
46
53

1

It should be noted that tf cannot compute gradients for these operations, so they cannot be used to simulate quantized weights. – oarfish Nov 23 '17 at 13:46

score 12 · Accepted Answer · answered Feb 24 '16 at 08:15

12

Oops, I find the function in the API...

 tf.to_float(x, name='ToFloat')

answered Feb 24 '16 at 08:15

Yee Liu

1,417
3
15
17

4

`tf.to_float()` is now deprecated and `tf.cast()` should be used instead. – Richard Nov 21 '19 at 19:39

score 5 · Answer 3 · answered Oct 01 '16 at 23:44

You can use either tf.cast(x, tf.float32) or tf.to_float(x), both of which cast to float32.

Example:

sess = tf.Session()

# Create an integer tensor.
tensor = tf.convert_to_tensor(np.array([0, 1, 2, 3, 4]), dtype=tf.int64)
sess.run(tensor)
# array([0, 1, 2, 3, 4])

# Use tf.cast()
tensor_float = tf.cast(tensor, tf.float32)
sess.run(tensor_float)
# array([ 0.,  1.,  2.,  3.,  4.], dtype=float32)

# Use tf.to_float() to cast to float32
tensor_float = tf.to_float(tensor)
sess.run(tensor_float)
# array([ 0.,  1.,  2.,  3.,  4.], dtype=float32)

when I cast an mage with type of `tf.uint8` to `tf.float32`, and used `matplotlib` to show them, `tf.float32` change. How can show the main image? — Tavakoli, Feb 07 '17 at 18:29

score 1 · Answer 4 · answered Jul 31 '18 at 03:50

imagetype cast you can use tf.image.convert_image_dtype() which convert image range [0 255] to [0 1]:

img_uint8 = tf.constant([1,2,3], dtype=tf.uint8)
img_float = tf.image.convert_image_dtype(img_uint8, dtype=tf.float32)
with tf.Session() as sess:
    _img= sess.run([img_float])
    print(_img, _img.dtype)

output:

[0.00392157 0.00784314 0.01176471] float32

if you only want to cast type and keep value range use tf.cast or tf.to_float as @stackoverflowuser2010 and @Mark McDonald answered

score 1 · Answer 5 · answered Mar 31 '20 at 04:25

In case your data is actually a Pandas dataframe, we can first check for the datatype using:

print(dataset.dtypes)

To cast all the entries into float32 (for e.g.),

# Typecast
dataset = dataset.astype('float32')

#print them to verify
print(dataset.dtypes)

How to convert tf.int64 to tf.float32?

5 Answers5

Linked