contextual-conv 0.6.4

PyTorch convolutional layers with global context conditioning

ContextualConv

ContextualConv – PyTorch convolutional layers with global context conditioning: per‑channel bias, scale, or modulated FiLM-style scaling.

---

🚀 Quick start

from contextual_conv import ContextualConv2d
import torch

# FiLM‑style (scale + bias)
conv = ContextualConv2d(
    in_channels=16,
    out_channels=32,
    kernel_size=3,
    padding=1,
    context_dim=10,   # size of global vector c
    h_dim=64,         # optional MLP hidden dim
    use_scale=True,   # γ(c)
    use_bias=True,    # β(c)
    scale_mode="film" # or "scale"
)

x = torch.randn(8, 16, 32, 32)  # feature map
c = torch.randn(8, 10)          # context vector

out = conv(x, c)  # shape: (8, 32, 32, 32)

Modes at a glance

use_scale	use_bias	scale_mode	Behaviour
False	True	–	Contextual bias only
True	False	"scale"	Scale only: out * γ
True	True	"film"	FiLM: out * (1 + γ) + β
True	True	"scale"	Scale + shift: out * γ + β
False	False	–	Plain convolution (no modulation)

If context_dim is provided, at least one of use_scale or use_bias must be True.

---

🔧 Key features

• ⚙️ Drop‑in replacement for nn.Conv1d / nn.Conv2d
  → Same arguments + optional context options.
• 🧠 Global vector conditioning via learnable γ(c) and/or β(c)
• 🔀 Modulation modes:
  • scale_mode="film": out * (1 + γ)
  • scale_mode="scale": out * γ
• 🪶 Lightweight – one small MLP (or single Linear) per layer
• 🧑‍🔬 FiLM ready – reproduce Feature‑wise Linear Modulation with two lines
• 🧩 Modular – combine with any architecture, works on CPU / GPU
• 📤 Infer context vectors from unmodulated outputs with .infer_context()
• ✅ Unit‑tested and documented

---

📦 Installation

pip install contextual-conv  # version 0.6.3 on PyPI

Or install from source:

git clone https://github.com/abbassix/ContextualConv.git
cd ContextualConv
pip install -e .[dev]

---

📐 Context vector details

• Shape: (B, context_dim)
  (one global descriptor per sample – class label embedding, latent code, etc.)
• Processed by a ContextProcessor:
  • Linear(context_dim, out_dim) (bias‑only / scale‑only)
  • Small MLP if h_dim is set.
• Output dims:
  • out_channels → bias or scale
  • 2 × out_channels → FiLM (scale + bias)

---

🔎 Context inference

You can extract the context vector inferred from the output using:

context = conv.infer_context(x)

To also get the unmodulated output from the convolution layer:

context, raw_out = conv.infer_context(x, return_raw_output=True)

This is useful when you need both the input’s context and its original unmodulated features.

---

🧪 Running tests

Run the full test suite with coverage:

pytest --cov=contextual_conv --cov-report=term-missing

---

📘 Documentation

Full API reference & tutorials: https://contextualconv.readthedocs.io

---

🤝 Contributing

Bug reports, feature requests, and PRs are welcome! See CONTRIBUTING.md.

---

📄 License

GNU GPLv3 – see LICENSE file for details.