" \u003ca target=\"_blank\" href=\"https://www.tensorflow.org/official_models/nlp/customize_encoder\"\u003e\u003cimg src=\"https://www.tensorflow.org/images/tf_logo_32px.png\" /\u003eView on TensorFlow.org\u003c/a\u003e\n",
" \u003c/td\u003e\n",
" \u003ctd\u003e\n",
" \u003ca target=\"_blank\" href=\"https://colab.research.google.com/github/tensorflow/models/blob/master/official/colab/nlp/customize_encoder.ipynb\"\u003e\u003cimg src=\"https://www.tensorflow.org/images/colab_logo_32px.png\" /\u003eRun in Google Colab\u003c/a\u003e\n",
" \u003c/td\u003e\n",
" \u003ctd\u003e\n",
" \u003ca target=\"_blank\" href=\"https://github.com/tensorflow/models/blob/master/official/colab/nlp/customize_encoder.ipynb\"\u003e\u003cimg src=\"https://www.tensorflow.org/images/GitHub-Mark-32px.png\" /\u003eView source on GitHub\u003c/a\u003e\n",
"The [TensorFlow Models NLP library](https://github.com/tensorflow/models/tree/master/official/nlp/modeling) is a collection of tools for building and training modern high performance natural language models.\n",
"\n",
"The [TransformEncoder](https://github.com/tensorflow/models/blob/master/official/nlp/modeling/networks/encoder_scaffold.py) is the core of this library, and lots of new network architectures are proposed to improve the encoder. In this Colab notebook, we will learn how to customize the encoder to employ new network architectures."
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "YYxdyoWgsl8t"
},
"source": [
"## Install and import"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "fEJSFutUsn_h"
},
"source": [
"### Install the TensorFlow Model Garden pip package\n",
"\n",
"* `tf-models-nightly` is the nightly Model Garden package created daily automatically.\n",
"* `pip` will install all models and dependencies automatically."
"Before learning how to customize the encoder, let's firstly create a canonical BERT enoder and use it to instantiate a `BertClassifier` for classification task."
"`canonical_classifier_model` can be trained using the training data. For details about how to train the model, please see the colab [fine_tuning_bert.ipynb](https://github.com/tensorflow/models/blob/master/official/colab/fine_tuning_bert.ipynb). We skip the code that trains the model here.\n",
"\n",
"After training, we can apply the model to do prediction.\n"
"One BERT encoder consists of an embedding network and multiple transformer blocks, and each transformer block contains an attention layer and a feedforward layer."
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "rmwQfhj6fmKz"
},
"source": [
"We provide easy ways to customize each of those components via (1)\n",
"[EncoderScaffold](https://github.com/tensorflow/models/blob/master/official/nlp/modeling/networks/encoder_scaffold.py) and (2) [TransformerScaffold](https://github.com/tensorflow/models/blob/master/official/nlp/modeling/layers/transformer_scaffold.py)."
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "xsMgEVHAui11"
},
"source": [
"### Use EncoderScaffold\n",
"\n",
"`EncoderScaffold` allows users to provide a custom embedding subnetwork\n",
" (which will replace the standard embedding logic) and/or a custom hidden layer class (which will replace the `Transformer` instantiation in the encoder)."
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "-JBabpa2AOz8"
},
"source": [
"#### Without Customization\n",
"\n",
"Without any customization, `EncoderScaffold` behaves the same the canonical `TransformerEncoder`.\n",
"\n",
"As shown in the following example, `EncoderScaffold` can load `TransformerEncoder`'s weights and output the same values:"
"Next, we show how to use a customized embedding network.\n",
"\n",
"We firstly build an embedding network that will replace the default network. This one will have 2 inputs (`mask` and `word_ids`) instead of 3, and won't use positional embeddings."
"User can also override the [hidden_cls](https://github.com/tensorflow/models/blob/master/official/nlp/modeling/networks/encoder_scaffold.py#L103) argument in `EncoderScaffold`'s constructor to employ a customized Transformer layer.\n",
"\n",
"See [ReZeroTransformer](https://github.com/tensorflow/models/blob/master/official/nlp/modeling/layers/rezero_transformer.py) for how to implement a customized Transformer layer.\n",
"\n",
"Following is an example of using `ReZeroTransformer`:\n"
"# Assert that the variable `rezero_alpha` from ReZeroTransformer exists.\n",
"assert 'rezero_alpha' in ''.join([x.name for x in classifier_model.trainable_weights])"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "6PMHFdvnxvR0"
},
"source": [
"### Use [TransformerScaffold](https://github.com/tensorflow/models/blob/master/official/nlp/modeling/layers/transformer_scaffold.py)\n",
"\n",
"The above method of customizing `Transformer` requires rewriting the whole `Transformer` layer, while sometimes you may only want to customize either attention layer or feedforward block. In this case, [TransformerScaffold](https://github.com/tensorflow/models/blob/master/official/nlp/modeling/layers/transformer_scaffold.py) can be used.\n",
"\n"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "D6FejlgwyAy_"
},
"source": [
"#### Customize Attention Layer\n",
"\n",
"User can also override the [attention_cls](https://github.com/tensorflow/models/blob/master/official/nlp/modeling/layers/transformer_scaffold.py#L45) argument in `TransformerScaffold`'s constructor to employ a customized Attention layer.\n",
"\n",
"See [TalkingHeadsAttention](https://github.com/tensorflow/models/blob/master/official/nlp/modeling/layers/talking_heads_attention.py) for how to implement a customized `Attention` layer.\n",
"\n",
"Following is an example of using [TalkingHeadsAttention](https://github.com/tensorflow/models/blob/master/official/nlp/modeling/layers/talking_heads_attention.py):"
"# Assert that the variable `pre_softmax_weight` from TalkingHeadsAttention exists.\n",
"assert 'pre_softmax_weight' in ''.join([x.name for x in classifier_model.trainable_weights])"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "kuEJcTyByVvI"
},
"source": [
"#### Customize Feedforward Layer\n",
"\n",
"Similiarly, one could also customize the feedforward layer.\n",
"\n",
"See [GatedFeedforward](https://github.com/tensorflow/models/blob/master/official/nlp/modeling/layers/gated_feedforward.py) for how to implement a customized feedforward layer.\n",
"\n",
"Following is an example of using [GatedFeedforward](https://github.com/tensorflow/models/blob/master/official/nlp/modeling/layers/gated_feedforward.py)."
"# Assert that the variable `gate` from GatedFeedforward exists.\n",
"assert 'gate' in ''.join([x.name for x in classifier_model.trainable_weights])"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "a_8NWUhkzeAq"
},
"source": [
"### Build a new Encoder using building blocks from KerasBERT.\n",
"\n",
"Finally, you could also build a new encoder using building blocks in the modeling library.\n",
"\n",
"See [AlbertTransformerEncoder](https://github.com/tensorflow/models/blob/master/official/nlp/modeling/networks/albert_transformer_encoder.py) as an example:\n"
