Dive into Deep Learning (d2l)

1.0.3 verified Fri May 01 auth: no python

The d2l package (v1.0.3) provides utility functions and dataset loaders for the textbook 'Dive into Deep Learning' (d2l.ai). It includes common deep learning building blocks, data iterators, and plotting helpers to accompany the book. Release cadence is irregular, tied to book updates.

pip install d2l

Common errors

error ModuleNotFoundError: No module named 'd2l.utils' ↓

cause d2l.utils was removed in v1.0.0; utils are now top-level or in submodules like d2l.data.

fix

Replace from d2l.utils import something with from d2l import something (e.g., Animator, Accumulator).

error AttributeError: module 'd2l' has no attribute 'torch' ↓

cause d2l.torch submodule removed; use `from d2l import torch` or upgrade to d2l>=1.0.0 where torch is re-exported at top level.

fix

Run pip install --upgrade d2l and change import to from d2l import torch as d2l or import d2l.

error FileNotFoundError: [Errno 2] No such file or directory: '/content/d2l-zh.zip' ↓

cause The deprecated d2l-zh package (Chinese version) is often confused with d2l. The wrong package downloaded a zip file instead of a proper Python module.

fix

Uninstall d2l-zh: pip uninstall d2l-zh. Install the correct package: pip install d2l. Then use import d2l as described.

Warnings

deprecated d2l.torch (submodule) is deprecated; use `from d2l import torch` or `import d2l` directly. The old import path `d2l.torch.nn` etc. may break in future versions. ↓

fix Replace `from d2l.torch import ...` with `from d2l import torch as d2l` or simply `import d2l` and use `d2l.xxx`.

breaking In d2l 1.0.0, many utility functions were moved to top-level (e.g., Animator, Accumulator, Timer) and removed from submodules like d2l.utils. Scripts using old paths will raise ImportError. ↓

fix Update imports: use `from d2l import Animator, Accumulator` instead of `from d2l.utils import Animator`.

gotcha The d2l package downloads datasets on first use from a server that may be slow or unreachable in certain regions. If dataset download hangs or fails, check network connectivity or manually download from the book's website. ↓

fix Set environment variable `D2L_DATA_DIR` to a local directory with pre-downloaded datasets. For offline use, manually copy datasets to ~/.d2l/data/.

deprecated The `d2l.mxnet` and `d2l.tensorflow` backends are deprecated and no longer maintained. Only PyTorch (torch) backend is actively developed. ↓

fix Use the torch backend: `from d2l import torch` or `import d2l` (defaults to torch). Avoid importing d2l.mxnet or d2l.tensorflow.

Imports

d2l
```
import d2l
```
Commonly imported as a module; subpackages like d2l.torch, d2l.mxnet, d2l.tensorflow exist but torch is default
torch
wrong
```
from d2l.torch import *
```
correct
```
from d2l import torch as d2l_torch
```
d2l.torch is deprecated; use from d2l import torch and then call d2l.torch.xxx or simply import d2l and use d2l.xxx
data
wrong
```
from d2l.utils import data
```
correct
```
from d2l import data
```
d2l.utils was restructured; direct submodules are preferred

Animator

wrong

from d2l.plot import Animator

correct

from d2l import Animator

Animator is now top-level in d2l

Accumulator
```
from d2l import Accumulator
```
Top-level utility; no common mistake

Quickstart

Quickstart: Load Fashion-MNIST using d2l's data loader. The d2l package simplifies dataset access and common training loops.

import torch
from d2l import torch as d2l

# Load a small dataset (Fashion-MNIST)
train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size=256)

# Check batch
for X, y in train_iter:
    print(X.shape, y.shape)
    break