Keras: After Flatten() both batch and output shape of a tensor become none

[rajatkoner08]

If I use Flatten() using functional API, both batch and output shape become none, where as in Sequential model its print correct output shape. I need both batch and output size later on and I have to use functional api because of my model complexity, is this is a issue of Keras?
@fchollet @farizrahman4u @Dref360

`input = Input(batch_shape=[64,224,224,3])

test = Conv2D(96,kernel_size=(11,11),strides=(4,4), activation='relu',name='conv1')(input)
print 'b4 flatten shape ', test.get_shape()

test = Flatten()(test)

print 'after flateen shape ',test.get_shape()

alexnet = Sequential()

alexnet.add(InputLayer(batch_input_shape=img_input))

alexnet.add(Conv2D(96, kernel_size=(11, 11), strides=(4, 4),
                   activation='relu', name='conv1'))

alexnet.add(Flatten())

print 'Sequential model shape ',alexnet.output_shape`

The output I got :

b4 flatten shape (64, 30, 30, 96)
after flatten shape (?, ?)
Sequential model shape (64, 86400)

[kylemcdonald]

My current workaround is to use Reshape((-1,)) instead, which doesn't have this problem. Here's a minimal example for reproducing this bug:

from keras.layers import Input, Flatten, Reshape

x = Input(batch_shape=(16, 10, 10))
print(x)

x = Input(batch_shape=(16, 10, 10))
x = Flatten()(x)
print(x)

x = Input(batch_shape=(16, 10, 10))
x = Reshape((-1,))(x)
print(x)

Prints:

Tensor("input_1:0", shape=(16, 10, 10), dtype=float32)
Tensor("flatten_1/Reshape:0", shape=(?, ?), dtype=float32) # from Flatten
Tensor("reshape_1/Reshape:0", shape=(16, 100), dtype=float32) # from Reshape

[demba]

Hi, I am using Keras 2.1.5 and am getting similar issues with Flatten() and Reshape not behaving as expected. The output of the same commands as above are

Tensor("input_5:0", shape=(16, 10, 10), dtype=float32)
Tensor("flatten_15/Reshape:0", shape=(?, ?), dtype=float32)
Tensor("reshape_27/Reshape:0", shape=(?, ?), dtype=float32)

The work around suggested in the previous comment doesn't work. My current workaround is

x = Input(batch_shape=(16, 10, 10))
x = keras.layers.Reshape((100,))(x)
print(x)

which outputs

Tensor("reshape_29/Reshape:0", shape=(16, 100), dtype=float32)

[demba]

Hi, I am using Keras 2.1.5 and am getting similar issues with Flatten() and Reshape not behaving as expected. The output of the same commands as above are

Tensor("input_5:0", shape=(16, 10, 10), dtype=float32)
Tensor("flatten_15/Reshape:0", shape=(?, ?), dtype=float32)
Tensor("reshape_27/Reshape:0", shape=(?, ?), dtype=float32)

The work around suggested in the previous comment doesn't work. My current workaround is

x = Input(batch_shape=(16, 10, 10))
x = keras.layers.Reshape((100,))(x)
print(x)

which outputs

Tensor("reshape_29/Reshape:0", shape=(16, 100), dtype=float32)

[ysyyork]

same issue here. No I need to explicitly take down the shape and use it in reshape

[thomasmooon]

+1
default_encoding: ANSI_X3.4-1968
ipython_version: 6.4.0,
os_name: posix,
platform: Linux-3.10.0-693.21.1.el7.x86_64-x86_64-with-Ubuntu-16.04-xenial,n
sys_platform: linux,
sys_version: 3.5.2
GCC 5.4.0 20160609
keras version: 2.1.6
python version: 3.5
tensorflow backend version: 1.8.0

[thomasmooon]

My current workaround is to query the shape before Flatten()is done this way:

# imports
import numpy as np
import keras
from keras.layers import *

# toy net
x = Input(shape = (12,100,10))
x = Dense(32)(x)
f = Flatten()(x)

# shape inference
shape_before_flatten = x.shape.as_list()[1:] # [1:] to skip None
shape_flatten = np.prod(shape_before_flatten) # value of shape in the non-batch dimension

# print
print("x = "+ str(x))
print("shape before Flatten() = " + str(shape_before_flatten))
print("shape after Flatten() = " + str(f.shape.as_list()))
print("shape_flatten in the non-batch dimension: " + str(shape_flatten))

x = Tensor("dense_17/BiasAdd:0", shape=(?, 12, 100, 32), dtype=float32)
shape before Flatten() = [12, 100, 32]
shape after Flatten() = [None, None]
shape_flatten in the non-batch dimension: 38400

[danaugrs]

I'm also experiencing this issue. Any updates?

[akashsengupta1997]

I am also experiencing the exact same issue.

[mohammadAbdolhosseiniMoghaddam]

does it affect the model structure or it is just a print issue?

[siddhartha-mukherjee-india]

Hi, I am using Keras 2.1.5 and am getting similar issues with Flatten() and Reshape not behaving as expected. The output of the same commands as above are

Tensor("input_5:0", shape=(16, 10, 10), dtype=float32)
Tensor("flatten_15/Reshape:0", shape=(?, ?), dtype=float32)
Tensor("reshape_27/Reshape:0", shape=(?, ?), dtype=float32)

The work around suggested in the previous comment doesn't work. My current workaround is

x = Input(batch_shape=(16, 10, 10))
x = keras.layers.Reshape((100,))(x)
print(x)

which outputs

Tensor("reshape_29/Reshape:0", shape=(16, 100), dtype=float32)

This helped me. Thanks a ton !!!

[zl2k]

Same problem. Reshape((-1,)) not work either.
Keras 2.2.4, Tensorflow 1.12

[FlashSnail]

Any updates? I have the same problem, after Flatten(), tensor's shape become (?,?)

[FlashSnail]

Any updates? I have the same problem, after Flatten(), tensor's shape become (?,?)

I have a workaround here:

x = Input(shape=(, 10, 10))
shape = np.prod(x.shape[1:])
x = Reshape((shape,))(x)

this code will return shape=(?,100)

[ustcjerry]

Any updates? I have the same problem, after Flatten(), tensor's shape become (?,?)

I have a workaround here:

x = Input(shape=(, 10, 10))
shape = np.prod(x.shape[1:])
x = Reshape((shape,))(x)

this code will return shape=(?,100)

Not working with:

x = Input(shape=(, 10, 10))
shape = np.prod(x.shape[1:])
x = Reshape((shape,))(x)
x = Dense(10, activation='relu'))(x)

[ustcjerry]

This works

x = Input(shape=(32, 10, 10))
shape = np.prod(x.shape[1:])
x = Reshape((shape,))(x)
print(x.shape)

output:
(?, 3200)

When adding some following layers, we can add an additional Reshape layer:

x = Input(shape=(, 10, 10))
shape = np.prod(x.shape[1:])
x = Reshape((shape,))(x)
x = Reshape((-1,))(x)
x = Dense(10, activation='relu'))(x)

[onkarkris]

This works

x = Input(shape=(32, 10, 10))
shape = np.prod(x.shape[1:])
x = Reshape((shape,))(x)
print(x.shape)

output:
(?, 3200)

When adding some following layers, we can add an additional Reshape layer:

x = Input(shape=(, 10, 10))
shape = np.prod(x.shape[1:])
x = Reshape((shape,))(x)
x = Reshape((-1,))(x)
x = Dense(10, activation='relu'))(x)

Above solution will not work! A quick fix for adding following layer is:
shape = np.prod(x.shape[1:])
reshaper=keras.layers.Lambda(lambda x: tf.keras.backend.reshape(x, shape=(self.batch_size, shape)))
x = reshaper(x)
x = Dense(10, activation='relu'))(x)