How to read, save and display images, encoded in csv format - csv

Question

I've got some images for training and testing a tensorflow model encoded in csv format. Is there a way to extract those images and / or save them in jpg like format?
Part of the file can be seen above as a opened in excel as a screenshot. If you prefer text to hyperlinks, here is a part of it in a form of a text:
label pixel1 pixel2 ...
6 149 149 ...
5 126 128 ...
10 85 88 ...
0 203 205 ...
There are 785 columns and 7173 rows in total. I have no idea how to deal with that.

Answer 1

You can do it like this
# first i create a dummy dataset to work on
data = make_classification(10000, n_features=784, random_state=1234)
df = pd.DataFrame(data[0], columns=[str(f'col_{i}') for i in range(784)])
df['label'] = data[1]
# Now we create a img_vector and labels array from the dataframe
img_vector = df[[str(f'col_{i}') for i in range(784)]].values
labels = df['label'].values
# splitting the data
# Now we creating the dataset
def get_img(inputs, labels):
# here you have 784 pixels which usually represent a 28*28 image with 1 channel
# hence I reshape it that way
img = tf.reshape(inputs, (28,28,1))
# you can also add some augmentation
img = tf.image.flip_left_right(img)
img = tf.image.flip_up_down(img)
return img, labels
# We pass the img_vector and labels to the make the dataset
train_dataset = tf.data.Dataset.from_tensor_slices((train_mat, train_label))
# Map the dataset to get images form it.
train_dataset = train_dataset.map(get_img).batch(16)
# same for valid dataset
valid_dataset = tf.data.Dataset.from_tensor_slices((valid_mat, valid_label))
valid_dataset = valid_dataset.map(get_img).batch(16)
# A sanity check
import matplotlib.pyplot as plt
sample = None
for i in train_dataset:
sample = i
break
plt.imshow(sample[0][0])
# Creating a model
model = tf.keras.Sequential([
tf.keras.layers.Conv2D(3,3, input_shape=(28,28,1)),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(1, activation='sigmoid')
])
# Compile the model
model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['acc'])
# Finally train the model
model.fit(train_dataset,
epochs=10,
validation_data=valid_dataset)
Also, if you ever take a dataset from Kaggle you will usually find a sample notebook for that dataset in the code section.

Answer 2

You can read any row, plot it and save it as image like this:
import numpy as np
import pandas as pd
# read csv file
df = pd.read_csv("data.csv")
# read pixels
images = np.array(df.iloc[:,1:])
labels = np.array(df.iloc[:,0])
# select random number between 0 and 7172
index = 2
# reshape 784 rows to 28 height x 28 width
sample_image = images[index,:].reshape(28,28)
# import plt for displaying image
from matplotlib import pyplot as plt
# plot image
plt.imshow(sample_image)
plt.axis('off')
# plot it's label
print(labels[index])
# save image
plt.savefig("./image{}_label{}".format(index,labels[index]))