ValueError: Error when checking input: expected conv2d_25_input to have shape (144, 256, 3) but got array with shape (256, 144, 3)

Question

I have built a CNN with images of size 720 (height) and 1280 (width). I attempt to re scale images to 144x256. However, I receive this error:

ValueError: Error when checking input: expected conv2d_25_input to have shape (144, 256, 3) but got array with shape (256, 144, 3)

I do the obvious thing and swap the height and width, but the error still persists:

ValueError: Error when checking input: expected conv2d_33_input to have shape (256, 144, 3) but got array with shape (144, 256, 3)

The error is the same but reversed. Here is my code, ran on google colab. I have embedded the data from github into the project so you should have no problem running the entire program. Can someone explain where the error originates from? Thank you!

# Get data
! git clone https://github.com/finn-williams/cachedetector.git

# Import dependancies
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import cv2
import os
import random
%tensorflow_version 1.x
import tensorflow
#from keras.models import Sequential
from keras.layers import Conv2D,Dense,Flatten,Dropout,MaxPooling2D
#from keras.preprocessing.image import ImageDataGenerator
#from keras.callbacks import ReduceLROnPlateau
#from keras import *
from keras.preprocessing.image import ImageDataGenerator
#from keras.preprocessing import image
from keras.callbacks import ModelCheckpoint
from keras import *

# Get data paths
total_images_train_cache = os.listdir('/content/cachedetector/Data/train/Cache')
total_images_train_no_cache = os.listdir('/content/cachedetector/Data/train/NoCache')

# Image processing with generators
image_train_transformer = ImageDataGenerator(rescale = 1./255,
                                   rotation_range = 5,
                                   shear_range = 0.2,
                                   width_shift_range = 0.1,  # Shifts width of each image up to second parameter
                                   height_shift_range = 0.1, # Shifts width of each image up to second parameter
                                   zoom_range = 0.2,  # Zooms into image
                                   brightness_range = [0.2, 1.0], # Changes brightness between parameters
                                   channel_shift_range = 70 # Changes brightness
                                   )

image_test_transformer = ImageDataGenerator(rescale = 1./255)

# Hyperparameters
image_height = 144
image_width = 256
batch_size = 10 # High batch = poor generalization
no_of_epochs  = 200

# Reshape Images
training_set = image_train_transformer.flow_from_directory('/content/cachedetector/Data/train',
                                                 target_size = (image_width, image_height),
                                                 batch_size = batch_size,
                                                 class_mode = 'binary')

test_set = image_test_transformer.flow_from_directory('/content/cachedetector/Data/test',
                                            target_size = (image_width, image_height),
                                            batch_size = batch_size,
                                            class_mode = 'binary')

val_set = image_test_transformer.flow_from_directory('/content/cachedetector/Data/val',
                                            target_size = (image_width, image_height),
                                            batch_size = 1,
                                            shuffle = False,
                                            class_mode = 'binary')

# Build model
model = Sequential([Conv2D(32, (3, 3),input_shape = (image_height, image_width, 3), activation = 'relu'),
                    # Keep filters low at beginning/iamges close to output
                    Conv2D(32,(3,3),activation='relu'),
                    MaxPooling2D(pool_size=(2,2)),
                    Dropout(0.2),
                    Conv2D(64,(3,3),activation='relu'),
                    # Ramp up filters in hidden layer
                    Conv2D(64,(3,3),activation='relu'),
                    MaxPooling2D(pool_size=(2,2)),
                    # Max pooling to reduce dimenstions caused by filters
                    Dropout(0.2),
                    Conv2D(128,(3,3),activation='relu'),
                    Conv2D(128,(3,3),activation='relu'),
                    MaxPooling2D(pool_size=(2,2)),
                    Dropout(0.2),
                    Conv2D(256,(3,3),activation='relu'),
                    Conv2D(256,(3,3),activation='relu'),
                    MaxPooling2D(pool_size=(2,2)),
                    Flatten(),
                    Dense(units=256,activation='relu'),
                    Dropout(0.2),
                    Dense(units=1,activation='sigmoid')])

# Model settings
model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

# Train
history = model.fit_generator(training_set,
                    steps_per_epoch=1319//batch_size, # When to conduct new epoch (calculated using ceil(num_samples / batch_size) )
                    epochs=no_of_epochs,  
                    validation_data=test_set,
                    validation_steps=145//batch_size,
                    callbacks=callbacks_list
                   )

Answer 1

Clearly, there is this inconsistency between the shape of input images:

target_size = (image_width, image_height)

and the input shape of the model:

input_shape = (image_height, image_width, 3)

These two should be the same. So, in order to save effort and make the smallest changes, just fix the input shape of the model as follow:

input_shape = (image_width, image_height, 3)