Dataset from tensors in TensorFlow - ML Experiment: Train & Evaluate

Experiment - Dataset from tensors

Problem:You want to create a TensorFlow dataset from tensors and use it to train a simple model. Currently, the dataset is created but the model training is slow and inefficient.

Current Metrics:Training loss after 5 epochs: 0.85, Training accuracy: 65%, Validation loss: 0.90, Validation accuracy: 60%

Issue:The dataset is created from tensors but lacks batching and shuffling, causing slow training and poor model generalization.

Your Task

Improve the dataset pipeline by adding batching and shuffling to increase training speed and validation accuracy to above 70%.

You must use TensorFlow's tf.data API to create the dataset.

Do not change the model architecture.

Keep the number of epochs to 5.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

TensorFlow

import tensorflow as tf
import numpy as np

# Create sample data tensors
features = tf.constant(np.random.rand(1000, 10), dtype=tf.float32)
labels = tf.constant(np.random.randint(0, 2, size=(1000, 1)), dtype=tf.int32)

# Create dataset from tensors
raw_dataset = tf.data.Dataset.from_tensor_slices((features, labels))

# Shuffle and batch the dataset
batch_size = 32
shuffled_batched_dataset = raw_dataset.shuffle(buffer_size=1000).batch(batch_size)

# Define a simple model
model = tf.keras.Sequential([
    tf.keras.layers.Dense(16, activation='relu', input_shape=(10,)),
    tf.keras.layers.Dense(1, activation='sigmoid')
])

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

# Train the model
history = model.fit(shuffled_batched_dataset, epochs=5, validation_data=shuffled_batched_dataset.take(100 // batch_size))

Added dataset.shuffle(buffer_size=1000) to shuffle the data before each epoch.

Added dataset.batch(batch_size=32) to process data in batches.

Kept the model architecture and epochs unchanged.

Results Interpretation

Before: Training accuracy 65%, Validation accuracy 60%, Loss around 0.85-0.90

After: Training accuracy 78%, Validation accuracy 72%, Loss reduced to 0.55-0.60

Shuffling and batching datasets improve training efficiency and model generalization by mixing data order and processing multiple samples at once.

Bonus Experiment

Try adding dataset.prefetch(tf.data.AUTOTUNE) to the pipeline to further improve training speed.

💡 Hint

Prefetching overlaps data preparation and model execution, reducing idle time.

Practice

1. What does tf.data.Dataset.from_tensor_slices() do in TensorFlow?

easy

A. It merges multiple datasets into one.

B. It converts a dataset back into tensors.

C. It creates a dataset by slicing the input tensors row-wise.

D. It shuffles the dataset randomly.

Solution

1. Step 1: Understand the function purpose

   tf.data.Dataset.from_tensor_slices() takes tensors and creates a dataset by slicing them row-wise, so each element is one slice.

2. Step 2: Compare with other options

   Options B, C, and D describe different dataset operations, not the slicing creation step.

3. Final Answer:

   It creates a dataset by slicing the input tensors row-wise. -> Option C

4. Quick Check:

   Dataset from tensor slices = row-wise slicing [OK]

Hint: Remember: from_tensor_slices splits tensors row-wise [OK]

Common Mistakes:

• Confusing from_tensor_slices with shuffling
• Thinking it merges datasets
• Assuming it converts datasets back to tensors

2. Which of the following is the correct syntax to create a dataset from a tensor data_tensor using TensorFlow?

easy

A. dataset = tf.data.Dataset.from_tensor_slices(data_tensor)

B. dataset = tf.data.Dataset.create_from_tensor(data_tensor)

C. dataset = tf.data.Dataset.tensor_slices(data_tensor)

D. dataset = tf.data.from_tensor_slices(data_tensor)

Solution

1. Step 1: Recall the correct method name

   The correct TensorFlow method to create a dataset from tensor slices is tf.data.Dataset.from_tensor_slices().

2. Step 2: Check syntax correctness

   dataset = tf.data.Dataset.from_tensor_slices(data_tensor) matches the exact syntax. Options A, B, and D use incorrect method names or missing parts.

3. Final Answer:

   dataset = tf.data.Dataset.from_tensor_slices(data_tensor) -> Option A

4. Quick Check:

   Correct method name and syntax = dataset = tf.data.Dataset.from_tensor_slices(data_tensor) [OK]

Hint: Use exact method: Dataset.from_tensor_slices() [OK]

Common Mistakes:

• Using wrong method names
• Missing Dataset class before method
• Confusing with other dataset creation functions

3. What will be the output of the following code?

import tensorflow as tf
x = tf.constant([[1, 2], [3, 4], [5, 6]])
dataset = tf.data.Dataset.from_tensor_slices(x)
for element in dataset:
    print(element.numpy())

medium

A. [[1 2]
[3 4]
[5 6]]

B. [[1], [2], [3], [4], [5], [6]]

C. [1, 2, 3, 4, 5, 6]

D. [1 2]
[3 4]
[5 6]

Solution

1. Step 1: Understand from_tensor_slices behavior

   The method slices the tensor row-wise, so each element is a 1D tensor representing one row.

2. Step 2: Analyze the loop output

   Each iteration prints one row as a numpy array, so output lines are [1 2], then [3 4], then [5 6].

3. Final Answer:

   [1 2] [3 4] [5 6] -> Option D

4. Quick Check:

   Row-wise slices printed line by line = [1 2] [3 4] [5 6] [OK]

Hint: from_tensor_slices outputs row slices printed separately [OK]

Common Mistakes:

• Expecting full tensor printed at once
• Confusing row slices with flattened output
• Assuming column-wise slicing

4. Identify the error in this code snippet:

import tensorflow as tf
x = tf.constant([1, 2, 3])
dataset = tf.data.Dataset.from_tensor_slices(x)
for element in dataset:
    print(element.numpy())
print(dataset.batch(2))

medium

A. Calling batch() after iteration does not return a new dataset.

B. print(dataset.batch(2)) prints a dataset object, not batches.

C. from_tensor_slices() requires a list, not a tensor.

D. The loop should use dataset.batch(2) instead of dataset.

Solution

1. Step 1: Understand batch() output

   The batch() method returns a new dataset object that groups elements, but printing it directly shows the object info, not the batch contents.

2. Step 2: Check what print(dataset.batch(2)) does

   It prints a dataset representation, not the actual batched data. To see batches, you must iterate over it.

3. Final Answer:

   print(dataset.batch(2)) prints a dataset object, not batches. -> Option B

4. Quick Check:

   Printing dataset.batch() shows object info, not data [OK]

Hint: Iterate to see batches; print shows object info only [OK]

Common Mistakes:

• Expecting print to show batch data
• Thinking batch modifies original dataset in place
• Confusing tensor and list input types

5. You have two tensors:

features = tf.constant([[1, 2], [3, 4], [5, 6]])
labels = tf.constant([0, 1, 0])

You want to create a dataset that pairs each feature row with its label for training. Which code correctly creates this dataset?

hard

A. dataset = tf.data.Dataset.from_tensor_slices((features, labels))

B. dataset = tf.data.Dataset.from_tensor_slices(features).zip(labels)

C. dataset = tf.data.Dataset.from_tensor_slices(features + labels)

D. dataset = tf.data.Dataset.from_tensor_slices(features).batch(labels)

Solution

1. Step 1: Understand pairing tensors in dataset

   To pair features and labels, pass a tuple of tensors to from_tensor_slices(). This creates dataset elements as (feature_row, label) pairs.

2. Step 2: Evaluate each option

   dataset = tf.data.Dataset.from_tensor_slices((features, labels)) correctly uses a tuple. dataset = tf.data.Dataset.from_tensor_slices(features).zip(labels) tries to zip a tensor, which is invalid. dataset = tf.data.Dataset.from_tensor_slices(features + labels) adds tensors incorrectly. dataset = tf.data.Dataset.from_tensor_slices(features).batch(labels) misuses batch() with labels.

3. Final Answer:

   dataset = tf.data.Dataset.from_tensor_slices((features, labels)) -> Option A

4. Quick Check:

   Tuple input pairs tensors row-wise = dataset = tf.data.Dataset.from_tensor_slices((features, labels)) [OK]

Hint: Use tuple inside from_tensor_slices to pair tensors [OK]

Common Mistakes:

• Trying to zip a tensor directly
• Adding tensors instead of pairing
• Using batch() incorrectly with labels