keras顺序模型中输入形状的错误

Question

All welcome. Trying to deal with keras. I have several images saved in .npy format, as well as their labels.

When training a model, I get an error:

ValueError: Error when checking input: expected dense_input to have shape (135, 240) but got array with shape (240, 3)

Which is very strange, because the shape of the submitted image is:

(135, 240, 3)

My class NeuralNetwork:

class NeuralNetwork():
    def __init__(self):

    self.model = keras.models.Sequential()
    self.model.add(keras.layers.Dense(1024, input_shape=(135, 240), activation="relu"))
    self.model.add(keras.layers.Dense(512, activation="relu"))
    self.model.add(keras.layers.Dense(9, activation="softmax"))

    opt = keras.optimizers.Adam()
    self.model.compile(loss="categorical_crossentropy", optimizer=opt,
                  metrics=["accuracy"])

    def FitModel(self, trainX, trainY):
        self.model.fit(trainX, trainY, epochs=30)


    def Predict(self, image):
        predictions = self.model.predict(image)

        choice = np.argmax(predictions[0])
        return choice

And main:

Data_Count = 7990

WIDTH = 240
HEIGHT = 135

nn = NeuralNetwork()

for i in range(1, DataCount+1):

    file_name = 'D:/TrainingData/training_data-{}.npy'.format(i)
    train_data = np.load(file_name)

    image = np.array([i[0] for i in train_data])[0]
    label = np.array([i[1] for i in train_data])[0]

    image = image / 255

    nn.FitModel(image, label)

Why she get only (240, 3), instead (135, 240)?

Thanks in advance for the answer!

Answer 1

Your input_shape for dense-layer is not correct. Your images, have shape (135, 240, 3) but you feed (135,240) which means you miss the image channels. In addition, you forget to Flatten the images before feeding them into dense-layer. Here is the example with some dummy data:

import numpy as np
from tensorflow.python import keras

model = keras.models.Sequential()
model.add(keras.layers.Flatten(input_shape=(135, 240, 3)))
model.add(keras.layers.Dense(1024, activation="relu"))
model.add(keras.layers.Dense(512, activation="relu"))
model.add(keras.layers.Dense(9, activation="softmax"))

model.compile(loss="categorical_crossentropy", optimizer='adam', metrics=["accuracy"])

# dummy data
images = np.zeros(shape=(7990, 135, 240, 3))
labels = np.zeros(shape=(7990, 9))

# train model
model.fit(x=images, y=labels, batch_size=128)