Add tensorflow adapted ASL

[Mooster1]

Add tf_losses.py file which initializes edited version of AsymmetricLoss.forward() function for use with tensorflow.

[sonarqubecloud]

Quality Gate failed

Failed conditions
B Maintainability Rating on New Code (required ≥ A)

See analysis details on SonarQube Cloud

Catch issues before they fail your Quality Gate with our IDE extension SonarQube for IDE

[mrT23]

/describe

[qodo-code-review]

User description

Add tf_losses.py file which initializes edited version of AsymmetricLoss.forward() function for use with tensorflow.

---

PR Type

Enhancement

---

Description

• Add TensorFlow adaptation of Asymmetric Loss function

• Implements ASL with configurable gamma parameters

• Supports multi-label classification with asymmetric emphasis

• Returns compiled loss function for TensorFlow models

---

Diagram Walkthrough

flowchart LR
  A["ASL Wrapper Function"] -- "gamma_neg, gamma_pos" --> B["ASL_func"]
  B -- "logits, targets" --> C["Probability Calculation"]
  C --> D["Cross Entropy Loss"]
  D --> E["Asymmetric Focusing"]
  E --> F["Weighted Loss Output"]

Loading

File Walkthrough

Relevant files

Enhancement

tf_losses.py TensorFlow Asymmetric Loss implementation                                src/loss_functions/tf_losses.py New file implementing TensorFlow adaptation of Asymmetric Loss Wrapper function AsymmetricLoss accepts gamma_neg and gamma_pos parameters Inner ASL_func computes asymmetric cross-entropy loss with focusing mechanism Calculates probabilities, applies basic CE, and applies asymmetric weighting +42/-0

---

[mrT23]

/agentic_review

[qodo-code-review]

Code Review by Qodo

🐞 Bugs (3) 📘 Rule violations (2) 📎 Requirement gaps (0)

1. tf.math.log() can hit -inf 📘 Rule violation ⛯ Reliability

Description

The loss computes tf.math.log(xs_pos) and tf.math.log(xs_neg) without clipping/validating that
x is in (0,1), which can produce -inf/NaN and break training. This is missing boundary/edge-case
handling and input validation.

Code

src/loss_functions/tf_losses.py[R23-30]

+    # Calculating Probabilities
+    xs_pos = x
+    xs_neg = 1 - x
+
+    # Basic CE calculation
+    los_pos = y * tf.math.log(xs_pos)
+    los_neg = (1 - y) * tf.math.log(xs_neg)
+    loss = los_pos + los_neg

Evidence

PR Compliance ID 3 and 6 require explicit handling/validation of boundary values and inputs; the new
code performs logarithms on values derived directly from x and 1 - x with no clipping or
validation, so x at 0/1 (or outside [0,1]) will cause invalid logs.

Rule 3: Generic: Robust Error Handling and Edge Case Management
Rule 6: Generic: Security-First Input Validation and Data Handling
src/loss_functions/tf_losses.py[23-30]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
`tf.math.log()` is applied to `x` and `1 - x` without ensuring values are strictly within (0,1), which can produce `-inf/NaN` for boundary/out-of-range inputs.

## Issue Context
This function is intended to be used as a Keras loss. Model outputs may be logits (unbounded) or probabilities that can reach exactly 0/1 depending on activations and numerics.

## Fix Focus Areas
- src/loss_functions/tf_losses.py[23-30]

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

---

2. Logits used as probs 🐞 Bug ✓ Correctness

Description

ASL_func’s docstring says x is logits, but the implementation uses x directly as a probability
(xs_pos=x, xs_neg=1-x) and omits the sigmoid step used by the reference ASL, producing an incorrect
loss when model outputs are logits.

Code

src/loss_functions/tf_losses.py[R23-26]

+    # Calculating Probabilities
+    xs_pos = x
+    xs_neg = 1 - x
+

Evidence

The TF version directly treats x as probability, while the repo’s reference ASL explicitly applies
sigmoid to logits before computing xs_pos/xs_neg; training code also documents that logits are
passed and sigmoid is applied inside the loss.

src/loss_functions/tf_losses.py[15-26]
src/loss_functions/losses.py[15-27]
train.py[110-113]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
`src/loss_functions/tf_losses.py` documents `x` as logits but computes probabilities as `xs_pos = x` / `xs_neg = 1 - x`, which breaks ASL semantics when the model outputs logits.

## Issue Context
The repo’s reference ASL (`src/loss_functions/losses.py`) applies `sigmoid` to logits inside the loss, and training code passes logits with an explicit comment that sigmoid happens in the loss.

## Fix Focus Areas
- src/loss_functions/tf_losses.py[15-26]
- src/loss_functions/tf_losses.py[23-31]
- src/loss_functions/losses.py[23-27]
- train.py[110-113]

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

---

3. ASL_func uses y, x 📘 Rule violation ✓ Correctness

Description

The new loss function uses single-letter parameters (y, x) and a nonstandard order (targets
first, logits second), making it easy to misuse and reducing readability. This violates the
meaningful naming/self-documenting requirement.

Code

src/loss_functions/tf_losses.py[R15-21]

+  def ASL_func(y, x):
+    """"
+    Parameters
+    ----------
+    x: input logits (y hat)
+    y: targets (multi-label binarized vector)
+    """

Evidence

PR Compliance ID 2 requires identifiers to clearly express intent and discourages single-letter
names except for short-lived iterators; the function signature introduces y and x for long-lived
semantic inputs (targets and logits).

Rule 2: Generic: Meaningful Naming and Self-Documenting Code
src/loss_functions/tf_losses.py[15-21]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
The inner loss function signature uses ambiguous single-letter names (`y`, `x`) and an unusual argument order versus common TensorFlow/Keras conventions.

## Issue Context
Keras losses are typically called with `(y_true, y_pred)`; unclear naming/order increases the chance of wiring the loss incorrectly.

## Fix Focus Areas
- src/loss_functions/tf_losses.py[15-21]

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

---

4. TensorFlow dep unmanaged 🐞 Bug ⛯ Reliability

Description

The new module unconditionally imports TensorFlow, but the project’s requirements don’t include it;
any environment that imports this module without TensorFlow installed will fail immediately.

Code

src/loss_functions/tf_losses.py[1]

+import tensorflow as tf

Evidence

tf_losses.py requires TensorFlow at import time, while requirements.txt lists only torch-related
dependencies. This makes TensorFlow either an undeclared hard dependency or requires guarding for
optional use.

src/loss_functions/tf_losses.py[1-1]
requirements.txt[1-5]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
`tf_losses.py` imports TensorFlow unconditionally, but TensorFlow is not declared in dependencies.

## Issue Context
This repository appears torch-first (torch is in requirements), so TF may be intended as an optional capability.

## Fix Focus Areas
- src/loss_functions/tf_losses.py[1-1]
- requirements.txt[1-5]

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

---

5. Malformed docstrings 🐞 Bug ✓ Correctness

Description

The docstrings start with four quotes ("""") instead of triple quotes ("""), leaving a stray quote
at the beginning of the docstring and making the docstrings misleading/poorly formatted.

Code

src/loss_functions/tf_losses.py[R3-17]

+def AsymmetricLoss(gamma_neg=4.0, gamma_pos=1.0): # Wrapper for ASL function
+  """"
+  Tensorflow adaptation of "Official Pytorch Implementation of: 'Asymmetric Loss For Multi-Label Classification'(ICCV, 2021) paper" --> https://github.com/Alibaba-MIIL/ASL/blob/main/src/loss_functions/losses.py
+  Returns a loss function with asymmetric, specifiable emphases for false negatives & false positives. Output can be passed in as loss function for model.compile().
+  ----------
+  Parameters
+  ----------
+  gamma_neg: asymmetric emphasis on false negatives
+  gamma_pos: assymetric emphasis on false positives
+  """
+  
+  # Return ASL function with custom emphases
+  def ASL_func(y, x):
+    """"
+    Parameters

Evidence

Both the outer factory and inner loss function use """" as the opening delimiter rather than
standard triple quotes, which results in malformed docstring content.

src/loss_functions/tf_losses.py[3-12]
src/loss_functions/tf_losses.py[15-17]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
Docstrings are opened with four quotes instead of three, resulting in malformed docstring text.

## Issue Context
This impacts readability and tooling (help(), IDE doc popups, doc generation).

## Fix Focus Areas
- src/loss_functions/tf_losses.py[3-12]
- src/loss_functions/tf_losses.py[15-21]

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

---

ⓘ The new review experience is currently in Beta. Learn more

[qodo-code-review]

1. tf.math.log() can hit -inf 📘 Rule violation ⛯ Reliability

The loss computes tf.math.log(xs_pos) and tf.math.log(xs_neg) without clipping/validating that
x is in (0,1), which can produce -inf/NaN and break training. This is missing boundary/edge-case
handling and input validation.

Agent Prompt

## Issue description
`tf.math.log()` is applied to `x` and `1 - x` without ensuring values are strictly within (0,1), which can produce `-inf/NaN` for boundary/out-of-range inputs.

## Issue Context
This function is intended to be used as a Keras loss. Model outputs may be logits (unbounded) or probabilities that can reach exactly 0/1 depending on activations and numerics.

## Fix Focus Areas
- src/loss_functions/tf_losses.py[23-30]

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

[qodo-code-review]

2. Logits used as probs 🐞 Bug ✓ Correctness

ASL_func’s docstring says x is logits, but the implementation uses x directly as a probability
(xs_pos=x, xs_neg=1-x) and omits the sigmoid step used by the reference ASL, producing an incorrect
loss when model outputs are logits.

Agent Prompt

## Issue description
`src/loss_functions/tf_losses.py` documents `x` as logits but computes probabilities as `xs_pos = x` / `xs_neg = 1 - x`, which breaks ASL semantics when the model outputs logits.

## Issue Context
The repo’s reference ASL (`src/loss_functions/losses.py`) applies `sigmoid` to logits inside the loss, and training code passes logits with an explicit comment that sigmoid happens in the loss.

## Fix Focus Areas
- src/loss_functions/tf_losses.py[15-26]
- src/loss_functions/tf_losses.py[23-31]
- src/loss_functions/losses.py[23-27]
- train.py[110-113]

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools