DyNet

2.1.2 verified Mon Apr 27 auth: no python

DyNet is the Dynamic Neural Network Toolkit, a C++ library with Python bindings for training neural networks, particularly focused on natural language processing. Current version 2.1.2 supports implicitly cast parameters as expressions, Python 3.8+, and advanced slicing. Release cadence is irregular.

pip install dynet

Common errors

error ImportError: No module named dynet ↓

cause DyNet not installed or Python environment not correct.

fix

Run pip install dynet and ensure you are using the correct Python interpreter.

error TypeError: 'Parameter' object is not callable ↓

cause Using `p()` instead of `p.expr()` in DyNet <2.1, or using `p.expr()` in DyNet >=2.1 where parameters are already expressions.

fix

In DyNet >=2.1 just use p (not p() or p.expr()). In older versions use p.expr().

error ValueError: The model has already been trained, or the computation graph has already been built. ↓

cause Attempting to reuse a computation graph without calling `dy.renew_cg()`.

fix

Call dy.renew_cg() before each training iteration or when building a new graph.

Warnings

gotcha In DyNet 2.1, parameters are implicitly cast to expressions. Do not call `dy.parameter(p)` or `p.expr()`; just use `p` directly. ↓

fix Use `p` instead of `p.expr()` or `dy.parameter(p)`.

breaking DyNet 2.0 removed dependency on Boost. Model files saved with v1.x are incompatible. You must re-train or convert models. ↓

fix Retrain models with v2.0+ or use a conversion script (if available).

gotcha Always call `dy.renew_cg()` before creating new computation graph nodes in a loop, otherwise memory usage grows unbounded. ↓

fix Call `dy.renew_cg()` at the start of each training iteration.

Imports

dynet
wrong
```
from dynet import *
```
correct
```
import dynet as dy
```
Wildcard imports pollute namespace; convention is to import as dy.
Model
wrong
```
import dynet.Model
```
correct
```
from dynet import Model
```
dynet.Model is not a submodule; directly import the class.

Quickstart

Minimal example: create a model, parameters, and perform a simple optimization loop.

import dynet as dy

model = dy.Model()
trainer = dy.SimpleSGDTrainer(model)
pW = model.add_parameters((2, 2))
pb = model.add_parameters(2)

x = dy.vecInput(2)
y = dy.scalarInput(0)

for epoch in range(5):
    loss = dy.pickneglogsoftmax(pW.expr() * x, y)
    loss.backward()
    trainer.update()
    print(f"Epoch {epoch}: loss = {loss.value()}")