Indexing and slicing tensors in TensorFlow - Model Metrics & Evaluation

Metrics & Evaluation - Indexing and slicing tensors

Which metric matters for this concept and WHY

When working with tensors, the main goal is to correctly access and manipulate data parts. The key metric here is correctness of indexing and slicing. This means the selected slices or elements must match the intended data positions exactly. Errors in indexing lead to wrong data being used, which can cause model mistakes or training failures.

Confusion matrix or equivalent visualization

While confusion matrices apply to classification, here we use a simple tensor example to visualize indexing correctness:

Original tensor (3x3):
[[10, 20, 30],
 [40, 50, 60],
 [70, 80, 90]]

Indexing slice: tensor[1:3, 0:2]
Expected output:
[[40, 50],
 [70, 80]]

If output matches expected, indexing is correct.

Precision vs Recall (or equivalent tradeoff) with concrete examples

In indexing and slicing, the tradeoff is between selecting too much data (over-selection) and selecting too little data (under-selection).

• Over-selection: Picking extra rows or columns by mistake. This wastes computation and may confuse the model.
• Under-selection: Missing important data parts. This leads to incomplete training or wrong predictions.

Example: If you want the first two rows but slice tensor[:3], you get an extra row (over-selection). If you slice tensor[:1], you get only one row (under-selection).

What "good" vs "bad" metric values look like for this use case

Good indexing means the output tensor shape and values exactly match the intended slice.

• Good: Output shape and values match expected slice. For example, slicing a (3,3) tensor with [1:3, 0:2] returns shape (2,2) with correct values.
• Bad: Output shape is wrong (too big or too small), or values do not match expected positions. This indicates incorrect indexing.

Metrics pitfalls (accuracy paradox, data leakage, overfitting indicators)

Common pitfalls when indexing and slicing tensors include:

• Off-by-one errors: Forgetting that Python slicing excludes the end index leads to wrong slices.
• Mixing up axes: Confusing rows and columns causes wrong data extraction.
• Negative indices misuse: Negative indices count from the end, which can cause unexpected slices if misunderstood.
• Shape mismatch: Using slices that produce unexpected shapes can break model input requirements.

Self-check: Your model has 98% accuracy but 12% recall on fraud. Is it good?

This question is about model evaluation, but relates to indexing because wrong slicing of data can cause such issues.

Answer: No, it is not good. The low recall means the model misses most fraud cases. This could happen if the fraud data was sliced incorrectly during training or evaluation, causing the model to learn poorly. Correct indexing and slicing ensure the model sees the right data and metrics reflect true performance.

Key Result

Correct indexing and slicing ensure the model uses the right data parts, preventing errors and improving training quality.

Practice

1. What does indexing a tensor in TensorFlow do?

easy

A. Selects a single element by its position

B. Changes the shape of the tensor

C. Adds new elements to the tensor

D. Deletes elements from the tensor

Solution

1. Step 1: Understand indexing

   Indexing means picking one element from a tensor by its position, like choosing one item from a list.

2. Step 2: Compare with other options

   Changing shape, adding, or deleting elements are different operations, not indexing.

3. Final Answer:

   Selects a single element by its position -> Option A

4. Quick Check:

   Indexing = single element pick [OK]

Hint: Indexing picks one element, slicing picks many [OK]

Common Mistakes:

• Thinking indexing changes tensor shape
• Confusing indexing with adding elements
• Assuming indexing deletes elements

2. Which of the following is the correct syntax to slice a 1D tensor t from index 2 to 5 (exclusive) in TensorFlow?

easy

A. t[2:5]

B. t.slice(2, 5)

C. t[2, 5]

D. t.slice(2:5)

Solution

1. Step 1: Recall slicing syntax

   TensorFlow uses Python-style slicing: t[start:stop] to get elements from start up to but not including stop.

2. Step 2: Check each option

   t[2:5] uses correct Python slice syntax. t.slice(2, 5) and D use incorrect method calls or syntax. t[2, 5] uses comma which is invalid for 1D slicing.

3. Final Answer:

   t[2:5] -> Option A

4. Quick Check:

   Slice syntax = t[start:stop] [OK]

Hint: Use square brackets with colon for slicing [OK]

Common Mistakes:

• Using commas instead of colons in slices
• Trying to call slice as a method incorrectly
• Confusing slice stop index as inclusive

3. Given the tensor t = tf.constant([[1, 2, 3], [4, 5, 6], [7, 8, 9]]), what is the output of t[1:, :2].numpy()?

medium

A. [[5 6]
[8 9]]

B. [[1 2]
[4 5]]

C. [[4 5 6]
[7 8 9]]

D. [[4 5]
[7 8]]

Solution

1. Step 1: Understand slicing t[1:, :2]

   1: means rows from index 1 to end (rows 1 and 2). :2 means columns from start to index 2 (columns 0 and 1).

2. Step 2: Extract the sliced elements

   Rows 1 and 2 are [[4,5,6], [7,8,9]]. Taking first two columns gives [[4,5], [7,8]].

3. Final Answer:

   [[4 5] [7 8]] -> Option D

4. Quick Check:

   Rows 1+ and cols 0-1 = [[4 5],[7 8]] [OK]

Hint: Remember slice stop is exclusive, so :2 means columns 0 and 1 [OK]

Common Mistakes:

• Including column index 2 mistakenly
• Starting rows from 0 instead of 1
• Confusing rows and columns order

4. What is wrong with this TensorFlow slicing code?

t = tf.constant([10, 20, 30, 40, 50])
slice = t[1:6]

medium

A. Index 6 is out of range, causing an error

B. Slicing with stop index beyond length is allowed, no error

C. Syntax error due to missing colon

D. TensorFlow does not support slicing

Solution

1. Step 1: Check slicing behavior with stop index

   In Python and TensorFlow, slicing stop index can be beyond tensor length without error; it stops at the end.

2. Step 2: Analyze given code

   Tensor t has length 5, slicing 1:6 extracts elements from index 1 to end safely.

3. Final Answer:

   Slicing with stop index beyond length is allowed, no error -> Option B

4. Quick Check:

   Slice stop > length is safe [OK]

Hint: Slice stop can exceed length without error [OK]

Common Mistakes:

• Expecting IndexError for slice stop beyond length
• Confusing slicing with indexing single element
• Thinking slicing syntax is invalid

5. You have a 3D tensor t = tf.constant([[[1,2],[3,4]], [[5,6],[7,8]], [[9,10],[11,12]]]). How do you extract the second element from each 2D matrix (i.e., elements 2, 4, 6, 8, 10, 12) using indexing and slicing?

hard

A. t[1, :, :]

B. t[:, 1, :]

C. t[:, :, 1]

D. t[:, 1]

Solution

1. Step 1: Understand tensor shape and indexing

   The tensor shape is (3, 2, 2): 3 matrices, each 2x2. We want the second element in the last dimension (index 1).

2. Step 2: Apply slicing to get second element in last dimension

   Using t[:, :, 1] selects all matrices (:), all rows (:), and the second element (index 1) in the last dimension.

3. Final Answer:

   t[:, :, 1] -> Option C

4. Quick Check:

   Last dim index 1 selects second elements [OK]

Hint: Use colon for all dims except last, index last dim 1 [OK]

Common Mistakes:

• Mixing row and column indices
• Using incomplete slicing like t[:, 1]
• Selecting wrong dimension index