ValueError: Negative dimension size caused by subtracting 2 from 1 for 'max_pooling2d_6/MaxPool' (op: 'MaxPool') with input shapes: [?,1,1,64]

Question

I am getting an error of Negative dimension size when I am keeping height and width of the input image anything below 362X362. I am surprised because this error is generally caused because of wrong input dimensions. I did not find any reason why number or rows and columns can cause an error. Below is my code-

batch_size = 32
num_classes = 7
epochs=50
height = 362
width = 362

train_datagen = ImageDataGenerator(
        rotation_range=40,
        width_shift_range=0.2,
        height_shift_range=0.2,
        rescale=1./255,
        shear_range=0.2,
        zoom_range=0.2,
        horizontal_flip=True,
        fill_mode='nearest')

test_datagen = ImageDataGenerator(rescale=1./255)

train_generator = train_datagen.flow_from_directory(
    'train',
        target_size=(height, width),
        batch_size=batch_size,
        class_mode='categorical')

validation_generator = test_datagen.flow_from_directory(
     'validation',
        target_size=(height, width),
        batch_size=batch_size,
        class_mode='categorical')

base_model = InceptionV3(weights='imagenet', include_top=False, input_shape=
(height,width,3))

x = base_model.output
x = Conv2D(32, (3, 3), use_bias=True, activation='relu') (x) #line2
x = MaxPooling2D(pool_size=(2, 2))(x)
x = Conv2D(64, (3, 3), activation='relu') (x) #line3
x = MaxPooling2D(pool_size=(2, 2))(x)
x = Flatten()(x)
x = Dense(batch_size, activation='relu')(x) #line1
x = (Dropout(0.5))(x)
predictions = Dense(num_classes, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)

for layer in base_model.layers:
    layer.trainable = False

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

model.fit_generator(
        train_generator,
        samples_per_epoch=128,
        nb_epoch=epochs,
        validation_data=validation_generator,
        verbose=2)

for i, layer in enumerate(base_model.layers):
    print(i, layer.name)

for layer in model.layers[:309]:
    layer.trainable = False
for layer in model.layers[309:]:
    layer.trainable = True

from keras.optimizers import SGD
model.compile(optimizer=SGD(lr=0.0001, momentum=0.9), 
loss='categorical_crossentropy', metrics=['accuracy'])

model.save('my_model.h5')
model.fit_generator(
        train_generator,
        samples_per_epoch=512,
        nb_epoch=epochs,
        validation_data=validation_generator,
        verbose=2)

Answer 1

Replace this:

x = MaxPooling2D(pool_size=(2, 2))(x)

with this:

x = MaxPooling2D((2,2), padding='same')(x)

to prevent dimension during downsampling.

Answer 2

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

Answer 3

InceptionV3 downsamples the input image very aggressively. For the input 362x362 image, base_model.output tensor is (?, 9, 9, 2048) - that is easy to see if you write

base_model.summary()

After that, your model downsamples the (?, 9, 9, 2048) tensor even further (like in this question):

(?, 9, 9, 2048)  # input
(?, 7, 7, 32)    # after 1st conv-2d
(?, 3, 3, 32)    # after 1st max-pool-2d
(?, 1, 1, 64)    # after 2nd conv-2d
error: can't downsample further!

You can prevent the conv layer from reducing the tensor size by adding padding='same' parameter, even that will make the error disappear. Or by simply reducing the number of downsamples.