Why does my Ml CNN Kaggle Cats / Dogs prection only spit out an error?

Question

I´m pretty new to ML and Computer Vision.I am trying to do a categotrial prediction for Cats/ Dogs 0 as the Cats and 1 as the Dogs. But my model.fit() function spits out this error.
...

    ValueError: Input 0 of layer sequential_5 is incompatible with the layer: : expected min_ndim=4, found ndim=2. Full shape received: [None, 10000]

This is my ML model:

import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
import cv2

#the images are stored in the Folders 'Cat/' and 'Dog/'
animal = ['Cat/','Dog/']
images_cat= []
images_dog=[]

# reads in the images 
for x in animal:
    for i in range(1,12500): # the are images from '1.jpg' till '12499.jpg' for each Cats and Dogs
        try:
            image_path = x+ str(i) +'.jpg'# this gets the path of the images for example 'Cat/1.jpg'
            #print(image_path)
            img = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2GRAY)
            img_resized = cv2.resize(img,dsize=(100,100))
            if x == 'Cat/':
                images_cat.append(img_resized)
            elif x == 'Dog/':
                images_dog.append(img_resized)
                
        except cv2.error as e:
            #some images spit out an errer and the apprently can't be read so therefore I just give them the first image to add to the list
            if x == 'Cat/':
                images_cat.append(images_cat[1])
            elif x == 'Dog/':
                images_dog.append(images_dog[1])

# assign targets to values

y_cat = np.zeros(len(images_cat)) # Cat == 0
y_dog = np.ones(len(images_dog)) # Dog == 1


# trainig_images = 80%   test_images= 20%
training_sample_count = round(0.8* len(y_cat))

#list slicing the images to get 80% of the images as calculated above
X_cat_train = images_cat [:training_sample_count]
y_cat_train_fin = y_cat[:training_sample_count]

X_dog_train = images_dog [:training_sample_count]
y_dog_train_fin = y_dog[:training_sample_count]

# create the final training list
X_train = X_cat_train + X_dog_train
y_train=[]

y_train.append(y_cat_train_fin.data)
y_train.append(y_dog_train_fin.data)

y_train = np.reshape(y_train,(19998,))
np.shape(y_train)# output: (19998,)

#normalizing the data
X_train = [x / 255.0 for x in X_train]
X_train = np.reshape(X_train,(19998,10000))
np.shape(X_train) #output: (19998, 10000)

from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense, Dropout, MaxPooling2D, Conv2D, Flatten 

model = Sequential()
model.add(Conv2D(32,kernel_size=(5,5),padding='same', activation ='relu'))
model.add(MaxPooling2D((3,3)))

model.add(Conv2D(32,kernel_size=(5,5),padding='same', activation ='relu'))
model.add(MaxPooling2D((3,3)))

model.add(Dropout(0.25))
model.add(Flatten())

model.add(Dense(1, activation='softmax'))
model.compile(optimizer='adam', loss="sparse_categorical_crossentropy", metrics=["accuracy"])

model.fit(
    X_train,
    y_train,
    epochs=10,
    batch_size=10000)

I haven't gotten to the test Images yet, but I basicly try to train this model for future Data (like own Images of Cats or Dogs to be then predicted). I would be happy if anyone could help me with my problem as I am stuck atm. Cheers :)

Answer 1

There are a few issues with your model -

1. You are doing a binary classification yet using a configuration for multi-class single-label classification. Changing your loss and last layer activation to get correct results. Please check the table below for reference.

2. You are passing a 1D array for each sample to the Conv2D layer when it needs a 3D tensor. That's the reason for the error expected min_ndim=4, found ndim=2. The expected dimension was (batch, height, width, channels) and instead it got (batch, pixels). I have added a model.add(Reshape((100,100,1), input_shape=(10000,))) which reshapes the 10000 pixels to (100,100,1) to be able to be passed into the conv2d layer properly.

3. Lastly, you have 19998 sample images. Though possible, it doesn't make sense to have a batch size of 10000. Batch size is the number of samples that will result in gradient updates. In your case, each epoch will only have 2 gradient updates because 19998/10000 ~ 2. I would advise having batch size as something like 128 or 64 or 32. I have set it to 128 in the model.fit

Find the updated code below.

from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense, Dropout, MaxPooling2D, Conv2D, Flatten 

X_train = np.random.random((500, 10000))
Y_train = np.random.randint(0,2,(500,)) #0, 0, 1, 0, 1...

model = Sequential()
model.add(Reshape((100,100,1), input_shape=(10000,)))
model.add(Conv2D(32,kernel_size=(5,5), padding='same', activation ='relu'))
model.add(MaxPooling2D((3,3)))

model.add(Conv2D(32,kernel_size=(5,5),padding='same', activation ='relu'))
model.add(MaxPooling2D((3,3)))

model.add(Dropout(0.25))
model.add(Flatten())

model.add(Dense(1, activation='sigmoid'))
model.compile(optimizer='adam', loss="binary_crossentropy", metrics=["accuracy"])

model.fit(
    X_train,
    Y_train,
    epochs=3,
    batch_size=128)

Epoch 1/3
4/4 [==============================] - 2s 498ms/step - loss: 0.7019 - accuracy: 0.4680
Epoch 2/3
4/4 [==============================] - 2s 534ms/step - loss: 0.6939 - accuracy: 0.5260
Epoch 3/3
4/4 [==============================] - 2s 524ms/step - loss: 0.6922 - accuracy: 0.5240

Answer 2

The Conv2D layer expects inputs of shape (batch_size, x, y, depth). Your X_train is being reshaped to only have size (batch_size, x*y) which is not what the Conv2D expects.

It may work to just take out this reshape: X_train = np.reshape(X_train,(19998,10000)). If not, you could reshape to (19998, 100, 100, 1).