Problems with Python's Numpy array dimensions

Question

I am struggling with a problem that I just can't get to work! I am currently working with Tensorflow and worked through the basic tutorials.

As you can see in the tutorial, the neural-network-model expects a train_images Numpy array of shape (60000, 28, 28) as there are 60,000 images of size 28x28 in the training set of the tutorial. I am reading in the Flavia-dataset, which is a set of pictures of leaves. The traning set consists of 1588 pictures which have 300x300px resolution. Here's my code:

for root, dirs, files in os.walk(pathname_data):
for name in files:
    img = keras.preprocessing.image.load_img(os.path.join(root,name), color_mode="grayscale",target_size=(300,300))      #get image in 300x300 grayscale
    array = keras.preprocessing.image.img_to_array(img)                                                                 #convert to numpy array
    array = array.squeeze(axis=2)                                                                                       #convert to 300x300 2d array
    array = array / 255.0                                                                                               #preprocess data
    pathSegments = os.path.normpath(os.path.join(root,name)).split(os.sep)                                              #split path
    if pathSegments[len(pathSegments)-3] == "Train":                                                                    #assign to training- or testSet
        #TODO: how to store the 2x2 arrays ?? 
        #store in training set 
    elif pathSegments[len(pathSegments)-3] == "Test":
        #store in test set 

My question is now, how do I store "array" so that I end up with a (1588, 300, 300)-shaped Numpy array that I can feed to my model? I already tried to experiment with reshape, append and transpose, but as of yet to no avail :( Any help greatly appreciated!

Answer 1

I assume that every 'array' you generate from a file is a (300, 300) shape

You can either pregenerate the array and use a counter

all_img = np.empty((1588, 300, 300))
count = 0
for root, dirs, files in os.walk(pathname_data):
    for name in files:
        ...
        all_img[count] = array.copy()
        count += 1
        ...

or you could append all the images to a list and change it into an array later

all_img = []
for root, dirs, files in os.walk(pathname_data):
    for name in files:
        ...
        all_img.append(array)
        ...
all_img = np.array(all_img)

Both this method will give you a (1588, 300, 300) array, I've no experience with Tensorflow, but this is the shape you required.