convolution added to numpy backend

[ziofil]

User description

In order to compute the PNR conditional matrix we need the convolution function to work in the numpy backend. I added it.

---

PR Type

enhancement

---

Description

• Added a new convolution function to the numpy backend, enabling 2D convolution operations similar to TensorFlow's tf.nn.convolution.
• The convolution function uses scipy.signal.convolve2d for efficient computation and supports 'SAME' and 'VALID' padding options.
• Updated the poisson function in backend_manager.py to ensure consistent dtype usage with rate.

---

Changes walkthrough 📝

Relevant files

Enhancement

backend_manager.py Update dtype handling in `poisson` function                            mrmustard/math/backend_manager.py Updated poisson function to use rate.dtype for arange. +1/-2      backend_numpy.py Add 2D convolution function to numpy backend                          mrmustard/math/backend_numpy.py Added convolution function for 2D convolution using numpy. Utilized scipy.signal.convolve2d for efficient 2D convolution. Implemented padding logic for 'SAME' and 'VALID' options. +54/-0

---

💡 PR-Agent usage: Comment /help "your question" on any pull request to receive relevant information

[codiumai-pr-agent-pro]

PR-Agent was enabled for this repository. To continue using it, please link your git user with your CodiumAI identity here.

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🏅 Score: 85

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Performance Concern The convolution function performs the operation only for the first channel. This might lead to incorrect results for multi-channel inputs. Potential Bug The convolution function ignores the 'data_format' parameter, which might cause issues if the function is called with different data formats.

[codiumai-pr-agent-pro]

PR-Agent was enabled for this repository. To continue using it, please link your git user with your CodiumAI identity here.

PR Code Suggestions ✨

Explore these optional code suggestions:

Category	Suggestion	Score
Possible bug	Utilize all input channels in the convolution operation instead of only the first channel The current implementation only uses the first channel of the input array for convolution. Modify the code to utilize all input channels, which is typically expected in a convolution operation. mrmustard/math/backend_numpy.py [169-173] -array = np.pad( - array[:, :, :, 0], ((0, 0), (pad_h, pad_h), (pad_w, pad_w)), mode="constant" -) -... -array = array[:, :, :, 0] +if padding == "SAME": + array = np.pad( + array, ((0, 0), (pad_h, pad_h), (pad_w, pad_w), (0, 0)), mode="constant" + ) +# No need for the else case as we're not modifying the array for 'VALID' padding Suggestion importance[1-10]: 9 Why: The suggestion corrects a critical oversight by ensuring that all input channels are utilized in the convolution operation, which is a standard expectation. This change is crucial for the correctness and completeness of the convolution process.	9
Enhancement	Implement multi-channel convolution support to handle all output channels Consider using broadcasting to handle multiple output channels efficiently, instead of only processing the first channel. This can improve performance and functionality for multi-channel convolutions. mrmustard/math/backend_numpy.py [163-189] -# Reshape filter to 2D for convolution -filter_2d = filters[:, :, 0, 0] -... -# Perform convolution for each batch +# Perform convolution for each batch and channel for b in range(batch): - # Convolve using scipy's convolve2d which is more efficient than np.convolve for 2D - output[b, :, :, 0] = scipy_convolve2d( - array[b], - np.flip(np.flip(filter_2d, 0), 1), # Flip kernel for proper convolution - mode="valid", - ) + for c in range(out_channels): + output[b, :, :, c] = scipy_convolve2d( + array[b], + np.flip(np.flip(filters[:, :, 0, c], 0), 1), + mode="valid", + ) Suggestion importance[1-10]: 8 Why: The suggestion addresses a significant limitation in the current implementation by enabling multi-channel convolution, which is essential for handling all output channels efficiently. This enhancement improves both functionality and performance for multi-channel data.	8
Possible issue	Improve padding calculation to correctly handle both odd and even kernel sizes The padding calculation for 'SAME' mode assumes odd kernel sizes. For even kernel sizes, this might lead to asymmetric padding. Consider adjusting the padding calculation to handle both odd and even kernel sizes correctly. mrmustard/math/backend_numpy.py [166-171] # For SAME padding, calculate padding sizes if padding == "SAME": - pad_h = (kernel_h - 1) // 2 - pad_w = (kernel_w - 1) // 2 + pad_h = max(kernel_h - 1, 0) + pad_w = max(kernel_w - 1, 0) + pad_top = pad_h // 2 + pad_bottom = pad_h - pad_top + pad_left = pad_w // 2 + pad_right = pad_w - pad_left array = np.pad( - array[:, :, :, 0], ((0, 0), (pad_h, pad_h), (pad_w, pad_w)), mode="constant" + array[:, :, :, 0], ((0, 0), (pad_top, pad_bottom), (pad_left, pad_right)), mode="constant" ) Apply this suggestion Suggestion importance[1-10]: 7 Why: The suggestion corrects the padding calculation for 'SAME' mode to handle both odd and even kernel sizes, preventing potential asymmetric padding issues. This change enhances the robustness and correctness of the convolution operation.	7
Best practice	Add input validation to ensure correct tensor shapes and dimensions The function assumes the input array has exactly 4 dimensions. Consider adding input validation to ensure the input shape is correct and provide meaningful error messages if not. mrmustard/math/backend_numpy.py [159-160] +# Validate input shapes +if array.ndim != 4 or filters.ndim != 4: + raise ValueError(f"Expected 4D tensors, got array with {array.ndim}D and filters with {filters.ndim}D") + # Extract shapes batch, in_height, in_width, in_channels = array.shape -kernel_h, kernel_w, _, out_channels = filters.shape +kernel_h, kernel_w, filter_in_channels, out_channels = filters.shape +# Validate channel dimensions +if in_channels != filter_in_channels: + raise ValueError(f"Input channels ({in_channels}) must match filter input channels ({filter_in_channels})") + Apply this suggestion Suggestion importance[1-10]: 6 Why: Adding input validation ensures that the function receives tensors with the expected dimensions, providing meaningful error messages if not. This improves the reliability and user-friendliness of the function.	6

💡 Need additional feedback ? start a PR chat

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.80%. Comparing base (3a1d8bb) to head (b264136).

Additional details and impacted files

@@             Coverage Diff             @@
##           develop     #517      +/-   ##
===========================================
+ Coverage    89.77%   89.80%   +0.03%     
===========================================
  Files           92       92              
  Lines         6072     6072              
===========================================
+ Hits          5451     5453       +2     
+ Misses         621      619       -2     

Files with missing lines	Coverage Δ
mrmustard/math/backend_manager.py	98.16% <100.00%> (-0.01%)	⬇️
mrmustard/math/backend_numpy.py	100.00% <100.00%> (ø)

... and 1 file with indirect coverage changes

---

Continue to review full report in Codecov by Sentry.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 3a1d8bb...b264136. Read the comment docs.