CODEWITHSUNDEEP
pythontensorflowconv-neural-network
Nick Rizzolo
Nick Rizzolo

Reputation: 125

How do I set up a image dataset in tensorflow for a CNN?

I have a folder of images on my computer but I do not know how to set up the input layer. This is the code they provided on there sample page. I am not sure how to implement my folder into this format.

 def main(unused_argv):
  # Load training and eval data
  mnist = tf.contrib.learn.datasets.load_dataset("mnist")
  train_data = mnist.train.images # Returns np.array
  train_labels = np.asarray(mnist.train.labels, dtype=np.int32)
  eval_data = mnist.test.images # Returns np.array
  eval_labels = np.asarray(mnist.test.labels, dtype=np.int32)

Upvotes: 0

Views: 3107

Answers (1)

coder3101
coder3101

Reputation: 4165

You need to convert your images in the folders to vectors (set of numbers) to feed into the network.

In the above example of mnist you actually use a compressed binary version of the images. So no preprocessing is required from your end (except scaling pixels to be between 0-1).

Suppose you have a image that is RGB named orig.jpg in current working directory.

You will load that image as 

import numpy as np 
from PIL import Image 
img = Image.open('orig.jpg').convert('RGB')
img = np.array(img)

This will give you all the image content in the form of numbers to a numpy array variable. img in this case.

From your question, you are implementing a Convolution Neural network, so you do not need to flatten the img variable to one dimensional vector.

You can check the shape of the variable by running

np.shape(img)

Typically the shape will be a (lenght, width, 3)

One last step before you train your model with the image you may need to normalize the RGB values. You can do so by dividing img with 255. This usually improves your optimization algorithm

If you want to iterate over the complete image in a directory. You could do so by 

for filename in os.listdir(directory_name):
    if filename.endswith(".jpg"):  
         # code to extract the values using os.path.join(directory_name,filename)

The complete code to iterate and save all the images with .jpg to np.array is as follows.

from PIL import Image
import numpy as np
import os.path

length = 128 # pixels in length
width = 128 # pixels in width


imgs = np.empty((0,length, width, 3)) #empty dummy array, we will append to this array all the images
for filename in os.listdir(directory_name):
    if filename.endswith(".jpg"):
        img = Image.open(os.path.join(directory_name,filename)).convert('RGB')
        imgs = np.append(imgs, np.array(img).reshape((1, length, width, 3)), axis=0)

After the loop completes, the imgs will have a shape of (num_image, length, width, 3)

Now you can use this imgs and build some input pipeline with tf.data.Dataset.from_tensor_slices(....) or you can directly use the imgs

Upvotes: 2

Related Questions