23.8. The d2l API Document
Open the notebook in Colab
Open the notebook in Colab
Open the notebook in Colab
Open the notebook in Colab
Open the notebook in SageMaker Studio Lab

This section displays classes and functions (sorted alphabetically) in the d2l package, showing where they are defined in the book so you can find more detailed implementations and explanations. See also the source code on the GitHub repository.

23.8.1. Classes

classd2l.torch.AdditiveAttention(num_hiddens, dropout, **kwargs)[source]

Bases: Module

Additive attention.

Defined in Section 11.3.2.2

forward(queries, keys, values, valid_lens)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

classd2l.torch.AddNorm(norm_shape, dropout)[source]

Bases: Module

The residual connection followed by layer normalization.

Defined in Section 11.7.2

forward(X, Y)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

classd2l.torch.AttentionDecoder[source]

Bases: Decoder

The base attention-based decoder interface.

Defined in Section 11.4

propertyattention_weights
classd2l.torch.Classifier(plot_train_per_epoch=2, plot_valid_per_epoch=1)[source]

Bases: Module

The base class of classification models.

Defined in Section 4.3

accuracy(Y_hat, Y, averaged=True)[source]

Compute the number of correct predictions.

Defined in Section 4.3

layer_summary(X_shape)[source]

Defined in Section 7.6

loss(Y_hat, Y, averaged=True)[source]

Defined in Section 4.5

validation_step(batch)[source]
classd2l.torch.DataModule(root='../data', num_workers=4)[source]

Bases: HyperParameters

The base class of data.

Defined in Section 3.2.2

get_dataloader(train)[source]
get_tensorloader(tensors, train, indices=slice(0, None, None))[source]

Defined in Section 3.3

train_dataloader()[source]
val_dataloader()[source]
classd2l.torch.Decoder[source]

Bases: Module

The base decoder interface for the encoder–decoder architecture.

Defined in Section 10.6

forward(X, state)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

init_state(enc_all_outputs, *args)[source]
classd2l.torch.DotProductAttention(dropout)[source]

Bases: Module

Scaled dot product attention.

Defined in Section 11.3.2.2

forward(queries, keys, values, valid_lens=None)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

classd2l.torch.Encoder[source]

Bases: Module

The base encoder interface for the encoder–decoder architecture.

Defined in Section 10.6

forward(X, *args)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

classd2l.torch.EncoderDecoder(encoder, decoder)[source]

Bases: Classifier

The base class for the encoder–decoder architecture.

Defined in Section 10.6

forward(enc_X, dec_X, *args)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

predict_step(batch, device, num_steps, save_attention_weights=False)[source]

Defined in Section 10.7.6

classd2l.torch.FashionMNIST(batch_size=64, resize=(28, 28))[source]

Bases: DataModule

The Fashion-MNIST dataset.

Defined in Section 4.2

get_dataloader(train)[source]

Defined in Section 4.2

text_labels(indices)[source]

Return text labels.

Defined in Section 4.2

visualize(batch, nrows=1, ncols=8, labels=[])[source]

Defined in Section 4.2

classd2l.torch.GRU(num_inputs, num_hiddens, num_layers, dropout=0)[source]

Bases: RNN

The multilayer GRU model.

Defined in Section 10.3

classd2l.torch.HyperParameters[source]

Bases: object

The base class of hyperparameters.

save_hyperparameters(ignore=[])[source]

Save function arguments into class attributes.

Defined in Section 23.7

classd2l.torch.LeNet(lr=0.1, num_classes=10)[source]

Bases: Classifier

The LeNet-5 model.

Defined in Section 7.6

classd2l.torch.LinearRegression(lr)[source]

Bases: Module

The linear regression model implemented with high-level APIs.

Defined in Section 3.5

configure_optimizers()[source]

Defined in Section 3.5

forward(X)[source]

Defined in Section 3.5

get_w_b()[source]

Defined in Section 3.5

loss(y_hat, y)[source]

Defined in Section 3.5

classd2l.torch.LinearRegressionScratch(num_inputs, lr, sigma=0.01)[source]

Bases: Module

The linear regression model implemented from scratch.

Defined in Section 3.4

configure_optimizers()[source]

Defined in Section 3.4

forward(X)[source]

Defined in Section 3.4

loss(y_hat, y)[source]

Defined in Section 3.4

classd2l.torch.Module(plot_train_per_epoch=2, plot_valid_per_epoch=1)[source]

Bases: Module, HyperParameters

The base class of models.

Defined in Section 3.2

apply_init(inputs, init=None)[source]

Defined in Section 6.4

configure_optimizers()[source]

Defined in Section 4.3

forward(X)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

loss(y_hat, y)[source]
plot(key, value, train)[source]

Plot a point in animation.

training_step(batch)[source]
validation_step(batch)[source]
classd2l.torch.MTFraEng(batch_size, num_steps=9, num_train=512, num_val=128)[source]

Bases: DataModule

The English-French dataset.

Defined in Section 10.5

build(src_sentences, tgt_sentences)[source]

Defined in Section 10.5.3

get_dataloader(train)[source]

Defined in Section 10.5.3

classd2l.torch.MultiHeadAttention(num_hiddens, num_heads, dropout, bias=False, **kwargs)[source]

Bases: Module

Multi-head attention.

Defined in Section 11.5

forward(queries, keys, values, valid_lens)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

transpose_output(X)[source]

Reverse the operation of transpose_qkv.

Defined in Section 11.5

transpose_qkv(X)[source]

Transposition for parallel computation of multiple attention heads.

Defined in Section 11.5

classd2l.torch.PositionalEncoding(num_hiddens, dropout, max_len=1000)[source]

Bases: Module

Positional encoding.

Defined in Section 11.6

forward(X)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

classd2l.torch.PositionWiseFFN(ffn_num_hiddens, ffn_num_outputs)[source]

Bases: Module

The positionwise feed-forward network.

Defined in Section 11.7

forward(X)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

classd2l.torch.ProgressBoard(xlabel=None, ylabel=None, xlim=None, ylim=None, xscale='linear', yscale='linear', ls=['-', '--', '-.', ':'], colors=['C0', 'C1', 'C2', 'C3'], fig=None, axes=None, figsize=(3.5, 2.5), display=True)[source]

Bases: HyperParameters

The board that plots data points in animation.

Defined in Section 3.2

draw(x, y, label, every_n=1)[source]

Defined in Section 23.7

classd2l.torch.Residual(num_channels, use_1x1conv=False, strides=1)[source]

Bases: Module

The Residual block of ResNet models.

Defined in Section 8.6

forward(X)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

classd2l.torch.ResNeXtBlock(num_channels, groups, bot_mul, use_1x1conv=False, strides=1)[source]

Bases: Module

The ResNeXt block.

Defined in Section 8.6.2

forward(X)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

classd2l.torch.RNN(num_inputs, num_hiddens)[source]

Bases: Module

The RNN model implemented with high-level APIs.

Defined in Section 9.6

forward(inputs, H=None)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

classd2l.torch.RNNLM(rnn, vocab_size, lr=0.01)[source]

Bases: RNNLMScratch

The RNN-based language model implemented with high-level APIs.

Defined in Section 9.6

init_params()[source]
output_layer(hiddens)[source]

Defined in Section 9.5

classd2l.torch.RNNLMScratch(rnn, vocab_size, lr=0.01)[source]

Bases: Classifier

The RNN-based language model implemented from scratch.

Defined in Section 9.5

forward(X, state=None)[source]

Defined in Section 9.5

init_params()[source]
one_hot(X)[source]

Defined in Section 9.5

output_layer(rnn_outputs)[source]

Defined in Section 9.5

predict(prefix, num_preds, vocab, device=None)[source]

Defined in Section 9.5

training_step(batch)[source]
validation_step(batch)[source]
classd2l.torch.RNNScratch(num_inputs, num_hiddens, sigma=0.01)[source]

Bases: Module

The RNN model implemented from scratch.

Defined in Section 9.5

forward(inputs, state=None)[source]

Defined in Section 9.5

classd2l.torch.Seq2Seq(encoder, decoder, tgt_pad, lr)[source]

Bases: EncoderDecoder

The RNN encoder–decoder for sequence to sequence learning.

Defined in Section 10.7.3

configure_optimizers()[source]

Defined in Section 4.3

validation_step(batch)[source]
classd2l.torch.Seq2SeqEncoder(vocab_size, embed_size, num_hiddens, num_layers, dropout=0)[source]

Bases: Encoder

The RNN encoder for sequence-to-sequence learning.

Defined in Section 10.7

forward(X, *args)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

classd2l.torch.SGD(params, lr)[source]

Bases: HyperParameters

Minibatch stochastic gradient descent.

Defined in Section 3.4

step()[source]
zero_grad()[source]
classd2l.torch.SoftmaxRegression(num_outputs, lr)[source]

Bases: Classifier

The softmax regression model.

Defined in Section 4.5

forward(X)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

classd2l.torch.SyntheticRegressionData(w, b, noise=0.01, num_train=1000, num_val=1000, batch_size=32)[source]

Bases: DataModule

Synthetic data for linear regression.

Defined in Section 3.3

get_dataloader(train)[source]

Defined in Section 3.3

classd2l.torch.TimeMachine(batch_size, num_steps, num_train=10000, num_val=5000)[source]

Bases: DataModule

The Time Machine dataset.

Defined in Section 9.2

build(raw_text, vocab=None)[source]

Defined in Section 9.2

get_dataloader(train)[source]

Defined in Section 9.3.3

classd2l.torch.Trainer(max_epochs, num_gpus=0, gradient_clip_val=0)[source]

Bases: HyperParameters

The base class for training models with data.

Defined in Section 3.2.2

clip_gradients(grad_clip_val, model)[source]

Defined in Section 9.5

fit(model, data)[source]
fit_epoch()[source]

Defined in Section 3.4

prepare_batch(batch)[source]

Defined in Section 6.7

prepare_data(data)[source]
prepare_model(model)[source]

Defined in Section 6.7

classd2l.torch.TransformerEncoder(vocab_size, num_hiddens, ffn_num_hiddens, num_heads, num_blks, dropout, use_bias=False)[source]

Bases: Encoder

The Transformer encoder.

Defined in Section 11.7.4

forward(X, valid_lens)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

classd2l.torch.TransformerEncoderBlock(num_hiddens, ffn_num_hiddens, num_heads, dropout, use_bias=False)[source]

Bases: Module

The Transformer encoder block.

Defined in Section 11.7.2

forward(X, valid_lens)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

classd2l.torch.Vocab(tokens=[], min_freq=0, reserved_tokens=[])[source]

Bases: object

Vocabulary for text.

to_tokens(indices)[source]
propertyunk

23.8.2. Functions

d2l.torch.add_to_class(Class)[source]

Register functions as methods in created class.

Defined in Section 3.2

d2l.torch.bleu(pred_seq, label_seq, k)[source]

Compute the BLEU.

Defined in Section 10.7.6

d2l.torch.check_len(a, n)[source]

Check the length of a list.

Defined in Section 9.5

d2l.torch.check_shape(a, shape)[source]

Check the shape of a tensor.

Defined in Section 9.5

d2l.torch.corr2d(X, K)[source]

Compute 2D cross-correlation.

Defined in Section 7.2

d2l.torch.cpu()[source]

Get the CPU device.

Defined in Section 6.7

d2l.torch.gpu(i=0)[source]

Get a GPU device.

Defined in Section 6.7

d2l.torch.init_cnn(module)[source]

Initialize weights for CNNs.

Defined in Section 7.6

d2l.torch.init_seq2seq(module)[source]

Initialize weights for sequence-to-sequence learning.

Defined in Section 10.7

d2l.torch.masked_softmax(X, valid_lens)[source]

Perform softmax operation by masking elements on the last axis.

Defined in Section 11.3

d2l.torch.num_gpus()[source]

Get the number of available GPUs.

Defined in Section 6.7

d2l.torch.plot(X, Y=None, xlabel=None, ylabel=None, legend=[], xlim=None, ylim=None, xscale='linear', yscale='linear', fmts=('-', 'm--', 'g-.', 'r:'), figsize=(3.5, 2.5), axes=None)[source]

Plot data points.

Defined in Section 2.4

d2l.torch.set_axes(axes, xlabel, ylabel, xlim, ylim, xscale, yscale, legend)[source]

Set the axes for matplotlib.

Defined in Section 2.4

d2l.torch.set_figsize(figsize=(3.5, 2.5))[source]

Set the figure size for matplotlib.

Defined in Section 2.4

d2l.torch.show_heatmaps(matrices, xlabel, ylabel, titles=None, figsize=(2.5, 2.5), cmap='Reds')[source]

Show heatmaps of matrices.

Defined in Section 11.1

d2l.torch.show_list_len_pair_hist(legend, xlabel, ylabel, xlist, ylist)[source]

Plot the histogram for list length pairs.

Defined in Section 10.5

d2l.torch.try_all_gpus()[source]

Return all available GPUs, or [cpu(),] if no GPU exists.

Defined in Section 6.7

d2l.torch.try_gpu(i=0)[source]

Return gpu(i) if exists, otherwise return cpu().

Defined in Section 6.7

d2l.torch.use_svg_display()[source]

Use the svg format to display a plot in Jupyter.

Defined in Section 2.4