how to resize cifar10 image from 32x32 to 227x227?

[xingxi...@gmail.com]

I have read the image from cifar-10-batches-python

import os

import numpy as np

from PIL import Image

from pylab import *

import matplotlib.pyplot as plt

from scipy.misc import imresize

# read data

data_dir = "F:\\dataSet\\cifar-10-batches-py"

testdata_dir="F:\\dataSet\\cifar-10-batches-py\\test_batch"

da=np.load(testdata_dir)

testdata=da['data']

testlabel=np.array(da['labels'])

train=np.empty((50000,3072))

label=np.empty((50000,))

for i in range(1,2):

    str='data_batch_'+np.str(i)

    path1=os.path.join(data_dir,str)

    data=np.load(path1)

    train[10000*(i-1):10000*i,:]=data['data']

    label[10000*(i-1):10000*i,]=data['labels']

def intlabel(label):

    for i in range(label.shape[0]):

        label[i,]=int(float(label[i,]))

    return label

def intdata(data):

    n=data.shape[0]

    for i in range(n):

        for j in range(3072):

            data[i,j]=int(float(data[i,j]))

    return data

label,train=intlabel(label),intdata(train)

train,label=np.array(train),np.array(label)

then I don't know how to resize the data.

[Klemen Grm]

from keras.datasets import cifar10
import cv2

(X_train, y_train), (X_test, y_test) = cifar10.load_data()

data_upscaled = np.zeros((50000, 3, 227, 227))
for i, img in enumerate(X_train):
    im = img.reshape((1, 2, 0))
    large_img = cv2.resize(im, dsize=(227, 227), interpolation=cv2.INTER_CUBIC)
    data_upscaled[i] = large_img.reshape((2, 0, 1))

[gopida...@gmail.com]

Hey I tried this in keras and I'm getting error like,

cannot reshape array of size 3072 into shape (1,2,0)

[f.ortega...@gmail.com]

Try this one (from Klemen Grm answer):

import numpy as np

from keras.datasets import cifar10
import cv2

(X_train, y_train), (X_test, y_test) = cifar10.load_data()

data_upscaled = np.zeros(50000, 3, 227, 227)

for i, img in enumerate(X_train):

    im = np.transpose(img, (1, 2, 0))

    large_img = cv2.resize(im, dsize=(227, 227), interpolation=cv2.INTER_CUBIC)

    data_upscaled[i] = np.transpose(large_img, (2, 0, 1))

Regards

[khaoula boutiche]

hey!

please can you explain to me why we transpose !!

[Lance Norskog]

The input data is stored as 3 single-color images, R, G, and B, or "channels-first" format. We store it as one image with 3 colors per pixel or "channels-last format".

Transpose() rearranges the indexes to achieve this. Here is a better explanation:

https://machinelearningmastery.com/a-gentle-introduction-to-channels-first-and-channels-last-image-formats-for-deep-learning/

--
You received this message because you are subscribed to the Google Groups "Keras-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to keras-users...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/keras-users/7c3672bf-b7b7-49a1-9f41-c631da8b0171n%40googlegroups.com.

--

Lance Norskog
lance....@gmail.com
Redwood City, CA

[François Chollet]

Just use the Resizing layer: https://keras.io/api/layers/preprocessing_layers/image_preprocessing/resizing/

To view this discussion on the web, visit https://groups.google.com/d/msgid/keras-users/CAP0ykh8nN2C8XVO59GidR8n%3DRiSu8AHGbmY4ao1tSosiBqGRWA%40mail.gmail.com.