Research map / Architectures and efficiency / RNNs and normalization
On-the-fly Operation Batching in Dynamic Computation Graphs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Automatically batch compatible graph operations with agenda-based scheduling instead of no automatic batching | DyNet BiLSTM · BiLSTM, CPU | sentences/second: 16.8 -> 156 | Helped |
| Automatically batch compatible graph operations with agenda-based scheduling instead of no automatic batching | DyNet BiLSTM · BiLSTM, GPU | sentences/second: 56.2 -> 367 | Helped |
| Automatically batch compatible graph operations with agenda-based scheduling instead of no automatic batching | DyNet BiLSTM with character LSTMs · BiLSTM w/ char, CPU | sentences/second: 15.7 -> 132 | Helped |
| Automatically batch compatible graph operations with agenda-based scheduling instead of no automatic batching | DyNet BiLSTM with character LSTMs · BiLSTM w/ char, GPU | sentences/second: 43.2 -> 275 | Helped |
| Automatically batch compatible graph operations with agenda-based scheduling instead of no automatic batching | DyNet TreeLSTM · TreeLSTM, CPU | sentences/second: 50.2 -> 357 | Helped |
| Automatically batch compatible graph operations with agenda-based scheduling instead of no automatic batching | DyNet TreeLSTM · TreeLSTM, GPU | sentences/second: 76.5 -> 661 | Helped |
| Automatically batch compatible graph operations with agenda-based scheduling instead of no automatic batching | DyNet transition-based parser · Transition-Parsing, CPU | sentences/second: 16.8 -> 61.2 | Helped |
| Automatically batch compatible graph operations with agenda-based scheduling instead of no automatic batching | DyNet transition-based parser · Transition-Parsing, GPU | sentences/second: 33.0 -> 90.1 | Helped |
| Use agenda-based scheduling instead of depth-based scheduling | DyNet TreeLSTM · TreeLSTM, GPU | sentences/second: 672 -> 661 | Hurt |
| Use agenda-based scheduling instead of depth-based scheduling | DyNet transition-based parser · Transition-Parsing, CPU | sentences/second: 61.0 -> 61.2 | No clear effect |
| Use agenda-based scheduling instead of depth-based scheduling | DyNet BiLSTM · BiLSTM, CPU | sentences/second: 139 -> 156 | Helped |
| Use agenda-based scheduling instead of depth-based scheduling | DyNet BiLSTM · BiLSTM, GPU | sentences/second: 337 -> 367 | Helped |
Closest papers by meaning
- ED-Batch: Efficient Automatic Batching of Dynamic Neural Networks via Learned Finite State Machines (ICML 2023 · similarity 0.91)
- Deep Learning with Dynamic Computation Graphs (ICLR 2017 · similarity 0.90)
- Static Automatic Batching In TensorFlow (ICML 2019 · similarity 0.89)
- PRES: Toward Scalable Memory-Based Dynamic Graph Neural Networks (ICLR 2024 · similarity 0.84)
- FlashRNN: I/O-Aware Optimization of Traditional RNNs on modern hardware (ICLR 2025 · similarity 0.83)
- Pfeife: Automatic Pipeline Parallelism for PyTorch (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.