How to build a CNN model for MNIST fashion and test it with a another set of image from web?

Tags: , ,

Importing the data and splitting it into 4 for test and train

c_trainX = trainX.reshape(x_train.shape[0],28,28,1)#x_train.shape[0] = 60
model3 = Sequential() # type of DNN
model3.add(Conv2D(28, kernel_size=(3,3), input_shape = (28,28,1)))  
model3.add(Dense(200, activation="relu"))                
model3.add(Dense(10, activation=tf.nn.softmax))         
model3.compile(optimizer='adam', loss='sparse_categorical_crossentropy', 
metrics=['accuracy']), y_train, epochs=15)

model3.evaluate(c_testX, y_test)
[0.5343353748321533, 0.9064000248908997]---- This is my validation loss and 
p = model3.predict(c_testX[:10])

import urllib
from PIL import Image
%matplotlib inline
img ="testing.jpg")
numpyimgdata = np.asarray(img)    
import cv2
load_img_rz = np.array("testing.jpg").resize((28,28)))
print("After resizing:",load_img_rz.shape)
numpyimgdata_reshaped_grey = cv2.cvtColor(load_img_rz, cv2.COLOR_BGR2GRAY)
your_new_array = np.expand_dims(numpyimgdata_reshaped_grey, axis=-1)
       numpyimgdata_reshaped = your_new_array.reshape(-1,28, 28, 1)     # this is done make the image in
                                                  # the same dimension of that of test and train data
image_predicted_array = model3.predict(numpyimgdata_reshaped)
test_pred = np.argmax(image_predicted_array, axis=1)     


This is actually wrong. It should be printed as a trouser which is denoted by 1 cause in mnist dataset 5 is sandal Label Description

  • 0 T-shirt/top
  • 1 Trouser
  • 2 Pullover
  • 3 Dress
  • 4 Coat
  • 5 Sandal
  • 6 Shirt
  • 7 Sneaker
  • 8 Bag
  • 9 Ankle boot

I tried with different images, I am getting number 5(sandals) when I try with some boot or canvas shoes even. What seems to be the actual mistake here?


The main problem is very simple. Here I will give you a complete implementation of your program. Please note that I may change the model definition and image preprocessing step. Ok, let get started.

Fashion MNIST

Get the data – Do some preprocessing – Visualize a sample.

from tensorflow.keras.datasets import fashion_mnist   

(x_train, y_train), (x_test, y_test) = fashion_mnist.load_data()

x_train = tf.expand_dims(x_train, -1)     # from 28 x 28 to 28 x 28 x 1 
x_train = tf.divide(x_train, 255)         # Normalize 
y_train = tf.one_hot(y_train , depth=10)  # Make target One-Hot

x_test = tf.expand_dims(x_test, -1)       # from 28 x 28 to 28 x 28 x 1 
x_test = tf.divide(x_test, 255)           # Normalize
y_test = tf.one_hot(y_test , depth=10)    # Make target One-Hot
x_train.shape, y_train.shape, x_test.shape, y_test.shape
(TensorShape([60000, 28, 28, 1]),
 TensorShape([60000, 10]),
 TensorShape([10000, 28, 28, 1]),
 TensorShape([10000, 10]))

[BONUS]: Look, these are 28 and grayscale images. Now, for any reason, if we want to resize and/or want to make it RGB (3 channel), we can do that too. Check my other answer here.

Now, let’s visualize one of a sample from our preprocessed data.

plt.imshow(x_train[0][:,:,0], cmap="gray")

enter image description here

Observe that, the main object right white and the background is black.

Model and Training

It’s better to use pretrained weight I think. However, here is a toy model to train.

model = Sequential() 
model.add(Conv2D(16, kernel_size=(3,3), input_shape = (28,28,1)))  
model.add(Conv2D(32, kernel_size=(3,3), activation="relu"))  
model.add(Conv2D(64, kernel_size=(3,3), activation="relu"))  
model.add(Conv2D(128, kernel_size=(3,3), activation="relu"))  
model.add(Dense(10, activation=tf.nn.softmax))       

# Unlike you I use categorical_crossentropy
# as because I one_hot encoded my y_train and y_test
              metrics=['accuracy']), y_train, batch_size=256, 
             epochs=15, validation_data=(x_test, y_test))
epoch:15: loss: 0.4552 - accuracy: 0.8370 - val_loss: 0.4008 - val_accuracy: 0.8606


Let’s make some predictions on the web-searched samples. Before that, let’s first define a function that will do the necessary preprocessing.

# a preprocess function 
def infer_prec(img, img_size):
    img = tf.expand_dims(img, -1)       # from 28 x 28 to 28 x 28 x 1 
    img = tf.divide(img, 255)           # normalize 
    img = tf.image.resize(img,          # resize acc to the input
             [img_size, img_size])
    img = tf.reshape(img,               # reshape to add batch dimension 
            [1, img_size, img_size, 1])
    return img 

Ok, I scrape some Fashion MNIST looking similar data, let’s open one of them.

import cv2
import matplotlib.pyplot as plt

img = cv2.imread('/content/a.jpg', 0)   # read image as gray scale    
print(img.shape)   # (300, 231)

plt.imshow(img, cmap="gray")

img = infer_prec(img, 28)  # call preprocess function 
print(img.shape)   # (1, 28, 28, 1)

enter image description here

All is good so far, except now we have a white background, which is not like our training sample on which our model is trained on. If I’m not wrong, all the samples of Fashion MNIST do have a black background. At this point, if we pass this sample to the model for prediction, it wouldn’t make accurate or close accurate predictions.

When we make an RGB sample to Grayscale, the white pixel remains white and the other colorful pixel gets black. For our case to handle this, we can use the bitwise_not operation on the grayscale image before passing it to the model for prediction. This bitwise_not simply makes 0 to 1 and vice-versa.

import cv2
import matplotlib.pyplot as plt

img = cv2.imread('/content/a.jpg', 0)  # read image as gray scale    
img = cv2.bitwise_not(img)             # < ----- bitwise_not
print(img.shape)   # (300, 231)

plt.imshow(img, cmap="gray")

img = infer_prec(img, 28)  # call preprocess function 
print(img.shape)   # (1, 28, 28, 1)

enter image description here

Now, pass it to the model for predicted probabilities.

y_pred = model.predict(img)

array([[3.1869055e-03, 5.6372599e-05, 1.1225128e-01, 2.2242602e-02,
        7.7411497e-01, 5.8861728e-11, 8.7906137e-02, 6.2964287e-12,
        2.4166984e-04, 2.0408438e-08]], dtype=float32)

Now we can get the predicted label and compare gt.

tf.argmax(y_pred, axis=-1).numpy() 
# array([4]) # Coat

Source: stackoverflow