Using Keras to design a CNN: Understanding Tensor Shape

0
votes

just starting out with ML, and wanted to create my own CNN to detect orientation of images with faces. I followed a tutorial to accept input images of 64x64x1, and here is my code:

from keras.models import Model
from keras.layers import Input
from keras.layers import Dense
from keras.layers import Flatten
from keras.layers.convolutional import Conv2D
from keras.layers.pooling import MaxPooling2D
from keras.preprocessing.image import ImageDataGenerator

datagen = ImageDataGenerator()
train_it = datagen.flow_from_directory('firstThousandTransformed/', class_mode='categorical', batch_size=64, color_mode="grayscale")
val_it = datagen.flow_from_directory('validation/', class_mode='categorical', batch_size=64, color_mode="grayscale")

imageInput = Input(shape=(64,64,1))
conv1 = Conv2D(32, kernel_size=4, activation='relu')(imageInput)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
conv2 = Conv2D(16, kernel_size=4, activation='relu')(pool1)
pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
flat = Flatten()(pool2)
hidden1 = Dense(10, activation='relu')(flat)
output = Dense(4, activation='sigmoid')(hidden1)
model = Model(inputs=imageInput, outputs=output)
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
print(model.summary())
model.fit(train_it, steps_per_epoch=16, validation_data=val_it, validation_steps=8)

However, I get this error when I try to run:

Input to reshape is a tensor with 3810304 values, but the requested shape requires a multiple of 2704 [[node model/flatten/Reshape (defined at c:\Users\cdues\Desktop\kerasTutorial\orentationTry.py:33) ]] [Op:__inference_train_function_836]

Below is my model summary:

enter image description here

I need some help understanding what a Tensor shape is and where my code has gone wrong here. Just working through the tutorial with Keras, I didn't encounter Tensor shape and now I am sort of lost. Sorry for the basic question, can yall help a noobie out? Thanks!

2
pythontensorflowmachine-learningkerasdeep-learning

2 Answers

0
votes

First in ImageDataGenerator there is a parameter called rescale. Typically with pixel values in the range 0 to 255 rescale is set to 1/255 so pixel value fall in the range from 0 to 1. I recommend you use that. Documentation for ImageDataGenerator is here.. In flow from directory you can specify the image size with parameter target_size: Tuple of integers (height, width), default: (256, 256). Documentation is at location specified earlier. In your model you have 4 nodes in your output layer. This implies you are classifying images into one of 4 classes. If that is the case in model.compile you should use categorical cross entropy as the loss. Change the activation function in your output layer to softmax.

1
votes

Try using the target_size argument while calling flow_from_directory.

train_it = datagen.flow_from_directory('firstThousandTransformed/',
                                       class_mode='categorical', 
                                       batch_size=64,
                                       color_mode='grayscale',
                                       target_size=(64,64))

val_it = datagen.flow_from_directory('validation/',
                                      class_mode='categorical',
                                      batch_size=64,
                                      color_mode='grayscale',
                                      target_size=(64,64))

This way you can reshape the images from the directories before feeding to the model.