Keras MaxPooling2D gives ValueError: Negative dimension size caused by subtracting 2 from 1 for 'MaxPool_x' #3945

pranay360 · 2016-10-03T09:44:39Z

Please look into the following link.
http://stackoverflow.com/questions/39815518/keras-maxpooling2d-layer-gives-valueerror

I have been facing this issue on the following config Ubuntu 16.04, Keras, TensorFlow GPU and Cuda 7.5.
On my windows config with GPU enabled Theano there is no issue in executing the script.

What might be the reason for this?

cdalbergue · 2016-10-03T11:57:49Z

Hello, I think taht you need to specify the dimension ordering:

Either change image_dim_ordering: 'tf' to image_dim_ordering: 'th' in your ~/.keras/keras.json
Specify the dim_ordering in your model like so:

model.add(MaxPooling2D(pool_size=(2, 2), dim_ordering="th"))

hope that helps 👍

burgalon · 2016-11-24T09:05:13Z

thanks @cdalbergue That actually resolves the issue however I'm not sure why it's needed as we are using the Tensorflow backend - shouldn't the dim_ordering be 'tf' ?

cdalbergue · 2016-11-24T09:26:19Z

@burgalon It depends of the version of tensorflow that you are using. 0.10 would use "tf", 0.11 would use "th"

Edit: or the opposite, can't really remember

jeffreynghm · 2016-12-16T01:27:52Z

That's weired. It's th for the latest version of keras... is there any reason? behind it @cdalbergue ?

PonyboyYbr · 2017-01-22T10:11:41Z

In fact, this problem results from your initial input.
input_shape=(3,224,224) is for "theano"
and your backend should be "tensorflow", so it should be
input_shape=(224,224,3).

arashmh · 2017-02-21T11:06:56Z

Just add this:

from keras import backend as K
K.set_image_dim_ordering('th')

sivakumar05 · 2017-03-19T05:18:50Z

Your suggestion really helped me,arashmh!

OmarMAmin · 2017-05-02T12:50:07Z

guys how am i supposed to set image dim ordering to theano while my backend is tensorflow?

I made your suggested fix, and it's somehow working

marektmn · 2018-05-05T19:56:46Z

tensorflow.python.framework.errors_impl.InvalidArgumentError: Negative dimension size caused by subtracting 3 from 1 for 'max_pooling2d_3/MaxPool' (op: 'MaxPool') with input shapes: [?,1,1,256].

guys I'm lost I have tried everything mentioned here and on stackoverflow. My dataset is grayscale 17x17x1, 3 classes, and I'm using Alexnet architecture.

This is output of x_train.shape & y_train.shape :

x_train shape: (18079, 17, 17, 1)
y_train_shape: (18079, 3)

Explicitly I have used :
from keras import backend as K K.set_image_dim_ordering('tf')
and my keras.json :
{ "floatx": "float32", "backend": "tensorflow", "epsilon": 1e-07, "image_data_format": "channels_last" }

Priyankk18k · 2018-09-28T10:19:24Z

In fact, this problem results from your initial input.
input_shape=(3,224,224) is for "theano"
and your backend should be "tensorflow", so it should be
input_shape=(224,224,3).

This helped me thanks.

sammilei · 2018-11-27T07:53:19Z

from keras import backend as K
if K.image_data_format() == 'channels_first':
input_shape = (IMG_CHANNELS, IMG_WIDTH, IMG_HEIGHT)
else:
input_shape = (IMG_WIDTH, IMG_HEIGHT, IMG_CHANNELS)

PallawiSinghal · 2018-12-26T08:57:45Z

Hi @sammilei how does this help. can u explain!

wtyn · 2019-08-07T08:15:28Z

In fact, this problem results from your initial input.
input_shape=(3,224,224) is for "theano"
and your backend should be "tensorflow", so it should be
input_shape=(224,224,3).

This help me, Thanks very much

mrgransky · 2020-04-27T06:47:04Z

in keras 2.2.5, I tried:
model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2), data_format='channels_first'))

knownstranger03 · 2020-05-13T09:01:35Z

Hello, I think taht you need to specify the dimension ordering:

Either change image_dim_ordering: 'tf' to image_dim_ordering: 'th' in your ~/.keras/keras.json

Specify the dim_ordering in your model like so:
model.add(MaxPooling2D(pool_size=(2, 2), dim_ordering="th"))
hope that helps

there is no dim_ordering parameter when using tensorflow.keras this line will throw another error

knownstranger03 · 2020-05-13T09:06:14Z

thanks @cdalbergue That actually resolves the issue however I'm not sure why it's needed as we are using the Tensorflow backend - shouldn't the dim_ordering be 'tf' ?

there is no parameter dim_ordering() when you're using functions from tensorflow.keras. The problem here has different reasons in keras and TF.

In keras: change the input shape accoring to the backend framework in use or change the dim_ordering=(tf/th)

In tensorflow: goto the the code line where error is raised and change the padding='valid' parameter to padding='same'.If the parameter doesn't exist then add as in the example below.

model.add(MaxPooling2D((2,2), strides=(2,2), padding='same'))

more info on the topic can be found here - https://www.tensorflow.org/api_docs/python/tf/keras/layers/MaxPool2D

Lotus-Blue · 2020-05-26T04:12:36Z

In fact, this problem results from your initial input.
input_shape=(3,224,224) is for "theano"
and your backend should be "tensorflow", so it should be
input_shape=(224,224,3).

thank you

jashshah999 · 2020-07-10T19:20:18Z

ValueError: Negative dimension size caused by subtracting 2 from 1 for '{{node max_pooling2d_2/MaxPool}} = MaxPoolT=DT_FLOAT, data_format="NHWC", ksize=[1, 2, 2, 1], padding="VALID", strides=[1, 2, 2, 1]' with input shapes: [?,1,1,64].

Tried all of the above solutions but nothing has worked. Please help!

CODE is below for reference:

import os
import numpy as np
from keras.models import Sequential
from keras.layers import Activation, Dropout, Flatten, Dense
from keras.preprocessing.image import ImageDataGenerator
from keras.layers import Convolution2D, MaxPooling2D, ZeroPadding2D
from keras import optimizers
import scipy
import keras
import pylab as pl
import matplotlib.cm as cm

img_width, img_height = 150, 150

train_data_dir = '/home/jashshah999/Desktop/data/train/'

datagen = ImageDataGenerator(rescale=1./255)

train_generator = datagen.flow_from_directory(
train_data_dir,
target_size=(img_width, img_height),
batch_size=16,
class_mode='binary')

model = Sequential()
model.add(Convolution2D(32, 3, 3, input_shape=(img_width,img_height,3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2),strides=(2,2),data_format="channels_last"))

model.add(Convolution2D(32, 3, 3))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2),strides=(2,2),data_format="channels_last"))

model.add(Convolution2D(64, 3, 3))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2),strides=(2,2),data_format="channels_last"))

model.add(Flatten())
model.add(Dense(64))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(1))
model.add(Activation('sigmoid'))

model.compile(loss='binary_crossentropy',
optimizer='rmsprop',
metrics=['accuracy'])

nb_epoch = 10
nb_train_samples = 2500

model.fit_generator(
train_generator,
samples_per_epoch=nb_train_samples,
nb_epoch=nb_epoch,
validation_data=validation_generator,
nb_val_samples=nb_validation_samples)

Light-- · 2020-08-21T07:45:49Z

@burgalon It depends of the version of tensorflow that you are using. 0.10 would use "tf", 0.11 would use "th"

Edit: or the opposite, can't really remember

what about tf2?
@burgalon

HiroshanG · 2020-09-20T12:31:15Z

thanks @cdalbergue That actually resolves the issue however I'm not sure why it's needed as we are using the Tensorflow backend - shouldn't the dim_ordering be 'tf' ?

there is no parameter dim_ordering() when you're using functions from tensorflow.keras. The problem here has different reasons in keras and TF.

In keras: change the input shape accoring to the backend framework in use or change the dim_ordering=(tf/th)

In tensorflow: goto the the code line where error is raised and change the padding='valid' parameter to padding='same'.If the parameter doesn't exist then add as in the example below.

model.add(MaxPooling2D((2,2), strides=(2,2), padding='same'))

more info on the topic can be found here - https://www.tensorflow.org/api_docs/python/tf/keras/layers/MaxPool2D

this worked for me... Thanks

ronghui19 · 2020-12-31T09:51:25Z

Just add this:

from keras import backend as K
K.set_image_dim_ordering('th')

save my life

fatemeshahbazi · 2021-01-02T19:29:10Z

from keras import backend as K
if K.image_data_format() == 'channels_first':
input_shape = (IMG_CHANNELS, IMG_WIDTH, IMG_HEIGHT)
else:
input_shape = (IMG_WIDTH, IMG_HEIGHT, IMG_CHANNELS)

thanks a lot , it helped me:))

kangzhiq · 2021-01-21T17:45:10Z

For keras 2.0, it becomes:

import tensorflow.keras.backend as K
# to set the image order
K.set_image_data_format('channels_first')

See the link below:
Keras 2.0 release note

dolaameng mentioned this issue Oct 5, 2016

Sharing Ideas on Continuous Expansion of FAQ #3966

Closed

pranay360 closed this as completed Mar 19, 2017

wzell mentioned this issue Jul 7, 2017

Run error wzell/cmd#5

Closed

jeroenclaes mentioned this issue Aug 23, 2017

Negative dimension size caused by subtracting 35 from 27 for 'max_pooling1d_12/MaxPool' (op: 'MaxPool') with input shapes: [?,27,1,128] rstudio/keras3#110

Closed

lowkeygit mentioned this issue Sep 21, 2017

Error during training ethanhe42/u-net#7

Closed

indiejoseph mentioned this issue Sep 28, 2017

Non-square images ml-hongkong/keras-transfer-learning-for-oxford102#3

Open

1 task

IcedDoggie mentioned this issue May 30, 2018

how to run IcedDoggie/Micro-Expression-with-Deep-Learning#1

Open

lmbak mentioned this issue Jul 29, 2019

Negative dimension size caused by subtracting 7 from 3 for 'avg_pool/AvgPool' (op: 'AvgPool') with input shapes: [?,3,3,2048]. azavea/raster-vision#808

Closed

YoungXueya mentioned this issue Dec 13, 2019

ValueError: Negative dimension size caused by subtracting 2 from 1 for 'fgcnn_layer_3/max_pooling2d_15/MaxPool' (op: 'MaxPool') with input shapes: [?,1,20,20]. shenweichen/DeepCTR#170

Closed

chunduriv mentioned this issue Oct 13, 2021

Negative dimension size caused by subtracting 2 from 1 for 'pool2/AvgPool' (op: 'AvgPool') with input shapes: [?,1,56,128]. #14095

Closed

tilakrayal mentioned this issue Jul 21, 2022

Forward AD threw error in tf.autodiff.ForwardAccumulator for tf.keras.layers.MaxPooling3D, but backward AD succeeded with same input tensorflow/tensorflow#56833

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Keras MaxPooling2D gives ValueError: Negative dimension size caused by subtracting 2 from 1 for 'MaxPool_x' #3945

Keras MaxPooling2D gives ValueError: Negative dimension size caused by subtracting 2 from 1 for 'MaxPool_x' #3945

pranay360 commented Oct 3, 2016

cdalbergue commented Oct 3, 2016 •

edited

Loading

burgalon commented Nov 24, 2016

cdalbergue commented Nov 24, 2016 •

edited

Loading

jeffreynghm commented Dec 16, 2016

PonyboyYbr commented Jan 22, 2017

arashmh commented Feb 21, 2017 •

edited

Loading

sivakumar05 commented Mar 19, 2017

OmarMAmin commented May 2, 2017

marektmn commented May 5, 2018

Priyankk18k commented Sep 28, 2018

sammilei commented Nov 27, 2018 •

edited

Loading

PallawiSinghal commented Dec 26, 2018

wtyn commented Aug 7, 2019

mrgransky commented Apr 27, 2020

knownstranger03 commented May 13, 2020

knownstranger03 commented May 13, 2020

Lotus-Blue commented May 26, 2020

jashshah999 commented Jul 10, 2020

Light-- commented Aug 21, 2020

HiroshanG commented Sep 20, 2020

ronghui19 commented Dec 31, 2020

fatemeshahbazi commented Jan 2, 2021

kangzhiq commented Jan 21, 2021

Keras MaxPooling2D gives ValueError: Negative dimension size caused by subtracting 2 from 1 for 'MaxPool_x' #3945

Keras MaxPooling2D gives ValueError: Negative dimension size caused by subtracting 2 from 1 for 'MaxPool_x' #3945

Comments

pranay360 commented Oct 3, 2016

cdalbergue commented Oct 3, 2016 • edited Loading

burgalon commented Nov 24, 2016

cdalbergue commented Nov 24, 2016 • edited Loading

jeffreynghm commented Dec 16, 2016

PonyboyYbr commented Jan 22, 2017

arashmh commented Feb 21, 2017 • edited Loading

sivakumar05 commented Mar 19, 2017

OmarMAmin commented May 2, 2017

marektmn commented May 5, 2018

Priyankk18k commented Sep 28, 2018

sammilei commented Nov 27, 2018 • edited Loading

PallawiSinghal commented Dec 26, 2018

wtyn commented Aug 7, 2019

mrgransky commented Apr 27, 2020

knownstranger03 commented May 13, 2020

knownstranger03 commented May 13, 2020

Lotus-Blue commented May 26, 2020

jashshah999 commented Jul 10, 2020

Light-- commented Aug 21, 2020

HiroshanG commented Sep 20, 2020

ronghui19 commented Dec 31, 2020

fatemeshahbazi commented Jan 2, 2021

kangzhiq commented Jan 21, 2021

cdalbergue commented Oct 3, 2016 •

edited

Loading

cdalbergue commented Nov 24, 2016 •

edited

Loading

arashmh commented Feb 21, 2017 •

edited

Loading

sammilei commented Nov 27, 2018 •

edited

Loading