Deep Learning Foundations With Code

OverviewA folder of image paths becomes model-ready only after metadata, decoding, resizing, batching, and visual validation are connected. The Data-to-Display Conveyor uses pandas for indexed metadata, tf.data for efficient input flow, and Keras preprocessing for consistent image tensors; use it whenever labels originate from a CSV or directory listing.

Quick ReferenceStage

API

Output

Metadata

pandas.read_csv()

DataFrame with paths and labels

Split

DataFrame.sample()

Training and validation rows

Decode

tf.io.read_file(), tf.image.decode_jpeg()

Image tensor

Resize

tf.image.resize()

Fixed-size tensor

Preprocess

keras.layers.Rescaling()

Floating-point values

Pipeline

tf.data.Dataset

Batched, prefetched data

Visual check

matplotlib.pyplot.imshow()

Image grid with labels

ParametersParameter

Type

Required

Description

csv_path

str

Yes

CSV containing filepath and label columns.

image_size

tuple[int, int]

No

Target height and width; default (224, 224).

batch_size

int

No

Images per batch; default 32.

validation_fraction

float

No

Fraction reserved for validation; default 0.2.

seed

int

No

Reproducible split and shuffle seed; default 42.

num_parallel_calls

int or tf.data.AUTOTUNE

No

Parallel mapping workers; default AUTOTUNE.

shuffle_buffer

int

No

Training shuffle buffer size; default 1000.

Code Exampleimport pandas as pd

import tensorflow as tf

import matplotlib.pyplot as plt

from tensorflow import keras

CSV columns: filepath,label

df = pd.read_csv("images.csv")

df["label_id"] = df["label"].astype("category").cat.codes

train_df = df.sample(frac=0.8, random_state=42)

valid_df = df.drop(train_df.index)

image_size = (224, 224)

batch_size = 32

num_classes = df["label_id"].nunique()

def decode_image(path, label):

image = tf.io.read_file(path)

image = tf.image.decode_jpeg(image, channels=3)

image = tf.image.resize(image, image_size)

image = tf.cast(image, tf.float32) / 255.0

return image, label

def make_dataset(frame, training=False):

paths = frame["filepath"].to_numpy()

labels = frame["label_id"].to_numpy()

ds = tf.data.Dataset.from_tensor_slices((paths, labels))

if training:

ds = ds.shuffle(1000, seed=42, reshuffle_each_iteration=True)

return (ds.map(decode_image, num_parallel_calls=tf.data.AUTOTUNE)

.batch(batch_size)

.prefetch(tf.data.AUTOTUNE))

train_ds = make_dataset(train_df, training=True)

valid_ds = make_dataset(valid_df)

Keras preprocessing can be inserted before a model.

preprocess = keras.Sequential([

keras.layers.RandomFlip("horizontal"),

keras.layers.RandomRotation(0.05),

])

images, labels = next(iter(train_ds))

augmented = preprocess(images, training=True)

plt.figure(figsize=(8, 8))

for i in range(min(9, len(images))):

plt.subplot(3, 3, i + 1)

plt.imshow(augmented[i])

plt.title(f"class={labels[i].numpy()}")

plt.axis("off")

plt.tight_layout()

plt.show()Response FormatThe pipeline returns batches with this structure:

{

"images": "float32 tensor with shape [batch_size, 224, 224, 3], values in [0, 1]",

"labels": "int32 tensor with shape [batch_size]",

"example": {

"image": "decoded RGB image resized to the configured dimensions",

"label": "integer class ID derived from the pandas category mapping"

}

}Notes & Best PracticesValidate paths before creating the dataset; missing files fail during iteration, not CSV loading.

Use stratified splitting when classes are imbalanced; a random split can omit rare labels from validation.

Match preprocessing to the model. Do not divide by 255 twice when using keras.layers.Rescaling(1./255).

Inspect unaugmented and augmented batches. Incorrect color channels, distorted aspect ratios, and label mismatches are easiest to detect before training.