Accuracy Stuck at 50% Keras

Question

Code

import numpy as np
from keras.preprocessing.image import ImageDataGenerator
from keras.models import Sequential,Model
from keras.layers import Dropout, Flatten, Dense,Input
from keras import applications
from keras.preprocessing import image
from keras import backend as K
K.set_image_dim_ordering('tf')


# dimensions of our images.
img_width, img_height = 150,150

top_model_weights_path = 'bottleneck_fc_model.h5'
train_data_dir = 'Cats and Dogs Dataset/train'
validation_data_dir = 'Cats and Dogs Dataset/validation'
nb_train_samples = 20000
nb_validation_samples = 5000
epochs = 50
batch_size = 16
input_tensor = Input(shape=(150,150,3))

base_model=applications.VGG16(include_top=False, weights='imagenet',input_tensor=input_tensor)
for layer in base_model.layers:
    layer.trainable = False

top_model=Sequential()
top_model.add(Flatten(input_shape=base_model.output_shape[1:]))
top_model.add(Dense(256,activation="relu"))
top_model.add(Dropout(0.5))
top_model.add(Dense(1,activation='softmax'))
top_model.load_weights(top_model_weights_path)
model = Model(inputs=base_model.input,outputs=top_model(base_model.output))


datagen = ImageDataGenerator(rescale=1. / 255)

train_data = datagen.flow_from_directory(train_data_dir,target_size=(img_width, img_height),batch_size=batch_size,classes=['dogs', 'cats'],class_mode="binary",shuffle=False)


validation_data = datagen.flow_from_directory(validation_data_dir,target_size=(img_width, img_height),classes=['dogs', 'cats'], batch_size=batch_size,class_mode="binary",shuffle=False)


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

model.fit_generator(train_data, steps_per_epoch=nb_train_samples//batch_size, epochs=epochs,validation_data=validation_data, shuffle=False,verbose=

I have implemented a Image Classifier on the cats and dogs Dataset(https://www.kaggle.com/c/dogs-vs-cats/data) using keras(transfer learned using the VGG16 network). The code runs without errors but the accuracy is stuck at 0.0 % for about half of the epoch and after half it increases to an of accuracy of 50%. I am using Atom with hydrogen.

How do I fix this.I really don't think I have a bias problem with such a dataset with VGG16(although i am relatively new to this field).

Sparkie

Change the last activation layer size from 1 to 2. – Christianna

Answer 1

Change your activation at your output layer to sigmoid

from

top_model.add(Dense(1,activation='softmax')) 

to

top_model.add(Dense(1,activation='sigmoid'))

Acinaciform

thanks a lot it worked. I thought softmax was identical to sigmoid for binary classification. (stats.stackexchange.com/questions/233658/…) Can you help me as to why softmax didnt work? – Sparkie

A 2D soft max is equivalent to a 1D sigmoid. If you want to use a softmax you need to use Dense(2) so that there are actually two classes the softmax can differentiate. A softmax run on n logits gives n numbers whose sum is 1. So if you run it on a single logit you always get a 1. – Calderon

@janbernlöhr Yes this is the solution I offered above. Using a softmax is an overkill for a binary classification task though. – Christianna

I fully agree to prefer the sigmoid for binary classification - it is computationally less demanding and should be numerically more stable. My comment was just to answer Niteya’s answer. – Calderon

Thanks @janbernlöhr , i tried both solutions and a sigmoid actually gave a better accuracy till 3 epochs after that both just kind of saturated. – Sparkie

Answer 2

• Changing the activation function is a good step, along with this pay attention to the size of the dataset you are using.
• If small then reduce batch size, reducing batch size solved it for me :)

I guess my test set was too small to feed large batches into the CNN.

Publicspirited