在 Python 中使用生成器输入 Keras model.fit_generator

Question

I am learning how to use generator in Python and feed it into Keras model.fit_generator.

from keras.utils import to_categorical
from keras.models import Sequential
from keras.layers import Dense, Conv2D, Flatten
import pandas as pd
import os
import cv2

class Generator:
    def __init__(self,path):
        self.path = path

    def gen(self, feat, labels):
        i=0
        while (True):
            im = cv2.imread(feat[i],0)
            im = im.reshape(28,28,1)
            yield im,labels[i]
            i+=1

if __name__ == "__main__":
    input_dir = './mnist'
    output_file = 'dataset.csv'

    filename = []
    label = []
    for root,dirs,files in os.walk(input_dir):
        for file in files:
            full_path = os.path.join(root,file)
            filename.append(full_path)
            label.append(os.path.basename(os.path.dirname(full_path)))

    data = pd.DataFrame(data={'filename': filename, 'label':label})
    data.to_csv(output_file,index=False)

    feat = data.iloc[:,0]
    labels = pd.get_dummies(data.iloc[:,1]).as_matrix()

    image_gen = Generator(input_dir)

    # #create model
    model = Sequential()
    model.add(Conv2D(64, kernel_size=3, activation="relu", input_shape=(28,28,1)))
    model.add(Conv2D(32, kernel_size=3, activation="relu"))
    model.add(Flatten())
    model.add(Dense(2, activation="softmax"))

    model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
    model.fit_generator(image_gen.gen(filename,labels), steps_per_epoch=5 ,epochs=5, verbose=1)

I have 2 subfolders inside ./mnist folder, corresponding to each class in my dataset. I created a Dataframe that contains the path of each image and the label (which is the name of the corresponding subfolder).

I created Generator class that loads the image whose path is written in the DataFrame.

It gave me error: ValueError: Error when checking input: expected conv2d_1_input to have 4 dimensions, but got array with shape (28, 28, 1)

Could anyone please help? And also, is it the correct way to implement generator in general?

Thanks!

Answer 1

I think answers to your questions can be found in Keras documentation.

In terms of the input shape, Conv2D layers expects 4-dimensional input, but you explicitly reshape to (28,28,1) in your generator, so 3 dimensions. On the Conv2D info and the input format, see this documentation .

In terms of the generator itself, Keras documentation provides an example with generator being a function, the same is discussed in Python Wiki . But your particular implementation seem to work, at least for the first iteration, if you get to the point of feeding the data into the convolution layer.