keras flow_from_dataframe 语义分割

Question

I'm trying to use keras' flow_from_dataframe for semantic segmentation (input is an image of dimension (height, width, 3), label is also an image of dimension (height, width), but can't get it to work.

As recommended here I (uninstalled the existing and) installed the latest keras preprocessing lib with

    pip install git+https://github.com/keras-team/keras-preprocessing.git

I'm getting following error for below mini example

ValueError: Error when checking target: expected conv1 to have 4 dimensions, but got array with shape (1, 1)

Using following versions in an anaconda virtual environment on Windows 7 in Pycharm

• tf 1.13.1
• keras-preprocessing 1.0.9
• keras 2.2.4
• keras-applications 1.0.7
• keras-base 2.2.4

I think the error lies in my usage of flow_from_dataframe, as I was able to write my own keras datagenerator following this blog .

Any advice, how to correctly setup flow_from_dataframe?

Fully working example which also generates the random training data.

import numpy as np
import pandas as pd
import os
import scipy.misc

import keras
from keras.models import Model
from keras.layers import Input, Conv2D
from keras_preprocessing.image import ImageDataGenerator


def get_file_list(root_path):
    """
    # Returns:
        file_list: _list_, list of full paths to all files found
    """
    file_list = []
    for root, dirs, files in os.walk(root_path):
        for name in files:
            file_list.append(os.path.join(root, name))
    return file_list


def gen_rand_img_labels(n_rand_imgs, path_img, path_label):
    for i in range(n_rand_imgs):
        img_rand = np.random.randint(0, 256, size=img_dim)
        scipy.misc.toimage(img_rand, cmin=0, cmax=255).save(os.path.join(path_img, 'img{}.png'.format(i)))

        label_rand = np.random.randint(0, n_classes, size=(img_dim[0], img_dim[1]))
        print('label_rand.shape: ', label_rand.shape)
        scipy.misc.toimage(label_rand, cmin=0, cmax=255).save(os.path.join(path_label, 'img{}.png'.format(i)))


if __name__ == "__main__":
    img_dim = (100, 200, 3)  # height, width, channels
    batch_size = 1
    nr_epochs = 1

    n_classes = 5
    n_rand_imgs = 10
    savepath_img = r'/path/to/img'
    savepath_label = r'/path/to/label'

    # --- generate random images and random labels and save them to disk
    gen_rand_img_labels(n_rand_imgs, savepath_img, savepath_label)


    # --- build Data Generator
    train_df = pd.DataFrame(columns=['path', 'label'])

    list_img_names = get_file_list(savepath_img)

    for fname in list_img_names:
        fname_pure = os.path.split(fname)[1]

        # read in png label file as numpy array
        y = scipy.misc.imread(os.path.join(savepath_label, fname_pure))
        y = keras.utils.to_categorical(y, n_classes)
        print('shape y: {}'.format(y.shape))
        train_df.loc[len(train_df)] = [fname, y]

    datagen = ImageDataGenerator(rescale=1/255.0, validation_split=0.25)
    train_generator = datagen.flow_from_dataframe(
        dataframe=train_df,
        x_col="path",
        y_col="label",
        subset="training",
        batch_size=batch_size,
        class_mode="raw",
        target_size=(img_dim[0], img_dim[1]))


    valid_generator = datagen.flow_from_dataframe(
        dataframe=train_df,
        x_col="path",
        y_col="label",
        subset="validation",
        batch_size=batch_size,
        class_mode="raw",
        target_size=(img_dim[0], img_dim[1]))

    # --- create the model and train it
    input_ = Input(shape=img_dim)
    x = Conv2D(n_classes, (3, 3), activation='relu', padding='same', name='conv1')(input_)
    model = Model(inputs=input_, outputs=[x])
    model.summary()

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

    # Train model
    STEP_SIZE_TRAIN = train_generator.n // train_generator.batch_size
    STEP_SIZE_VALID = valid_generator.n // valid_generator.batch_size
    model.fit_generator(generator=train_generator,
                        steps_per_epoch=STEP_SIZE_TRAIN,
                        validation_data=valid_generator,
                        validation_steps=STEP_SIZE_VALID,
                        epochs=nr_epochs)

Answer 1

I have updated the code for use with TensorFlow 2.0

NOTE: Use scipy 1.2.0, as scipy.misc.toimage is deprecated in later versions.

import tensorflow as tf
import numpy as np
import pandas as pd
import os
import scipy.misc

def get_file_list(root_path):
    """
    # Returns:
        file_list: _list_, list of full paths to all files found
    """
    file_list = []
    for root, dirs, files in os.walk(root_path):
        for name in files:
            file_list.append(os.path.join(root, name))
    return file_list

def gen_rand_img_labels(n_rand_imgs, path_img, path_label):
    for i in range(n_rand_imgs):
        img_rand = np.random.randint(0, 256, size=img_dim)
        PIL.Image.fromarray(img_rand, cmin=0, cmax=255).save(os.path.join(path_img, 'img{}.png'.format(i)))

        label_rand = np.random.randint(0, n_classes, size=(img_dim[0], img_dim[1]))
        print('label_rand.shape: ', label_rand.shape)
        PIL.Image.fromarray(label_rand, cmin=0, cmax=255).save(os.path.join(path_label, 'img{}.png'.format(i)))

if __name__ == "__main__":
    img_dim = (100, 200, 3)  # height, width, channels
    batch_size = 1
    nr_epochs = 1

    n_classes = 5
    n_rand_imgs = 10
    savepath_img = r''
    savepath_label = r''

    # --- generate random images and random labels and save them to disk
    gen_rand_img_labels(n_rand_imgs, savepath_img, savepath_label)


    # --- build Data Generator
    train_df = pd.DataFrame(columns=['path', 'label'])

    list_img_names = get_file_list(savepath_img)

    for fname in list_img_names:
        fname_pure = os.path.split(fname)[1]

        # read in png label file as numpy array
        y = scipy.misc.imread(os.path.join(savepath_label, fname_pure))
        y = tf.keras.utils.to_categorical(y, n_classes)
        print('shape y: {}'.format(y.shape))
        train_df.loc[len(train_df)] = [fname, y]

    datagen = tf.keras.preprocessing.image.ImageDataGenerator(rescale=1/255.0, validation_split=0.25)
    train_generator = datagen.flow_from_dataframe(
        dataframe=train_df,
        x_col="path",
        y_col="label",
        subset="training",
        batch_size=batch_size,
        class_mode="raw",
        target_size=(img_dim[0], img_dim[1]))


    valid_generator = datagen.flow_from_dataframe(
        dataframe=train_df,
        x_col="path",
        y_col="label",
        subset="validation",
        batch_size=batch_size,
        class_mode="raw",
        target_size=(img_dim[0], img_dim[1]))

    # --- create the model and train it
    input_ = tf.keras.Input(shape=img_dim)
    x = tf.keras.layers.Conv2D(n_classes, (3, 3), activation='relu', padding='same', name='conv1')(input_)
    model = tf.keras.Model(inputs=input_, outputs=[x])
    model.summary()

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

    # Train model
    STEP_SIZE_TRAIN = train_generator.n // train_generator.batch_size
    STEP_SIZE_VALID = valid_generator.n // valid_generator.batch_size
    model.fit_generator(generator=train_generator,
                        steps_per_epoch=STEP_SIZE_TRAIN,
                        validation_data=valid_generator,
                        validation_steps=STEP_SIZE_VALID,
                        epochs=nr_epochs)