Interformer_Pytorch

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

Interformer_Pytorch

(★ 17)

No description available.

Interformer_Pytorch Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# InterFormer PyTorch Port

This repository is a PyTorch reading-and-reimplementation project based on the paper:

- Paper: `InterFormer: Effective Heterogeneous Interaction Learning for Click-Through Rate Prediction`
- arXiv: `https://arxiv.org/abs/2411.09852`

The repository follows the same engineering layout as `OneTrans_Pytorch`, but replaces the backbone with an InterFormer-style architecture that combines:

- non-sequence interaction learning
- sequence modeling with personalized FFN
- cross-modal summarization between the two branches

## What This Repository Contains

There are two layers in this codebase:

1. The backbone port
   This is the PyTorch implementation of the InterFormer building blocks.
2. The training wrapper
   This adds dataset loading, feature tensorization, metrics, mixed precision, checkpointing, and CLI training support.

This repo is therefore a runnable local training scaffold around an InterFormer-style backbone, not just a shape demo.

## Implementation Notes

The model design follows the paper-level architecture:

- non-sequence feature preprocessing into per-field tokens
- sequence preprocessing through a MaskNet-like block
- Interaction Arch for behavior-aware non-sequence learning
- Sequence Arch for context-aware sequence modeling
- Cross Arch for low-dimensional information exchange

The default Interaction Arch follows the paper's reported benchmark setting:

- `DHEN`
- with `DOT` and `DCN` as the ensembled interaction modules

Because the runnable demo script reuses the local TAAC tensorization pipeline from `OneTrans_Pytorch`, the data interface is adapted to the available numeric tensors:

- non-sequence features are treated as numeric fields and tokenized field-wise
- sequence features are treated as per-step numeric channels and unified by `SequenceMaskNet`

This keeps the project runnable while preserving the main architecture of the paper.

## Repository Structure

### `main_pytorch.py`

This file is the backbone reference implementation. It contains:

- `NumericFieldTokenizer`
- `SequenceMaskNet`
- `LinearCompressedEmbedding`
- `SelfGating`
- `DotProductInteraction`
- `DCNInteraction`
- `DHENInteraction`
- `PersonalizedFFN`
- `RotaryMultiHeadAttention`
- `InterFormerBlock`
- `InterFormerEncoder`

This is the closest file to the core paper architecture.

### `models/`

Task-level model wrappers live here.

Current file:

- `models/taac_interformer.py`

This wraps the backbone into a trainable classifier:

- tokenizes non-sequence features into field tokens
- applies MaskNet-style sequence preprocessing
- stacks InterFormer blocks
- summarizes the final non-sequence and sequence states
- applies a classification head

### `utils/`

Reusable non-model logic lives here.

- `utils/common.py`
  General helpers such as seed setup and split generation.
- `utils/metrics.py`
  Accuracy and AUC computation.
- `utils/taac_data.py`
  Dataset loading, schema handling, feature conversion, and tensor construction.

### `scripts/`

Runnable entrypoints live here.

Current file:

- `scripts/run_taac2026_sample.py`

This script is the main training entrypoint. It handles:

- dataset download or local parquet reading
- feature tensorization
- model construction
- AMP setup
- training and validation loops
- checkpoint save and resume

## Backbone Architecture

At a high level, the model flow is:

1. Map non-sequence features into field tokens.
2. Map sequence channels into dense sequence tokens through `SequenceMaskNet`.
3. Compute an initial non-sequence summary and prepend it as CLS tokens.
4. Repeat `L` InterFormer blocks:
   - summarize non-sequence tokens
   - summarize sequence tokens through CLS, PMA, and recent tokens
   - update non-sequence tokens through the Interaction Arch
   - update sequence tokens through PFFN and rotary attention
5. Recompute final non-sequence and sequence summaries.
6. Flatten and concatenate both summaries for downstream prediction.

## Input Shapes

- non-sequential input:
  `[batch_size, non_seq_dim]`
- sequential input:
  `[batch_size, seq_len, seq_feature_dim]`

## Recommended Entry Points

### Backbone sanity check

```bash
python main_pytorch.py
```

### Training run

```bash
python scripts/run_taac2026_sample.py --epochs 5 --batch-size 32
```

### Training with the default paper-style Interaction Arch

```bash
python scripts/run_taac2026_sample.py --epochs 5 --batch-size 32 --interaction-backbone dhen
```

### Resume training

```bash
python scripts/run_taac2026_sample.py --epochs 10 --batch-size 32 --resume best_model_20260416_120000.pt --save-checkpoint
```

## What This README Tries To Clarify

This repository is not an official release from the paper authors. It is:

- a local PyTorch reimplementation of the published InterFormer architecture
- adapted to the engineering conventions used in `OneTrans_Pytorch`
- wrapped in a runnable training scaffold over the same local TAAC demo pipeline

Interformer_Pytorch

Interformer_Pytorch Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Interformer_Pytorch

Interformer_Pytorch Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README