Skip to content

Latest commit

 

History

History
199 lines (150 loc) · 8.42 KB

tf1_hub_module.md

File metadata and controls

199 lines (150 loc) · 8.42 KB
Raw

TF1 Hub format

At its launch in 2018, TensorFlow Hub offered a single type of asset: TF1 Hub format for import into TensorFlow 1 programs.

This page explains how to use TF1 Hub format in TF1 (or the TF1 compatibility mode of TF2) with the hub.Module class and associated APIs. (The typical use is to build a tf.Graph, possibly inside a TF1 Estimator, by combining one or more models in TF1 Hub format with tf.compat.layers or tf.layers).

Users of TensorFlow 2 (outside TF1 compatibility mode) must use the new API with hub.load() or hub.KerasLayer. The new API loads the new TF2 SavedModel asset type, but also has limited support for loading TF1 Hub format into TF2.

Using a model in TF1 Hub format

Instantiating a model in TF1 Hub format

A model in TF1 Hub format is imported into a TensorFlow program by creating a hub.Module object from a string with its URL or filesystem path, such as:

m = hub.Module("path/to/a/module_dir")

This adds the module's variables to the current TensorFlow graph. Running their initializers will read their pre-trained values from disk. Likewise, tables and other state is added to the graph.

Caching Modules

When creating a module from a URL, the module content is downloaded and cached in the local system temporary directory. The location where modules are cached can be overridden using TFHUB_CACHE_DIR environment variable. For details, see Caching.

Applying a Module

Once instantiated, a module m can be called zero or more times like a Python function from tensor inputs to tensor outputs:

y = m(x)

Each such call adds operations to the current TensorFlow graph to compute y from x. If this involves variables with trained weights, these are shared between all applications.

Modules can define multiple named signatures in order to allow being applied in more than one way (similar to how Python objects have methods). A module's documentation should describe the available signatures. The call above applies the signature named "default". Any signature can be selected by passing its name to the optional signature= argument.

If a signature has multiple inputs, they must be passed as a dict, with the keys defined by the signature. Likewise, if a signature has multiple outputs, these can be retrieved as a dict by passing as_dict=True, under the keys defined by the signature (the key "default" is for the single output returned if as_dict=False). So the most general form of applying a Module looks like:

outputs = m(dict(apples=x1, oranges=x2), signature="fruit_to_pet", as_dict=True)
y1 = outputs["cats"]
y2 = outputs["dogs"]

A caller must supply all inputs defined by a signature, but there is no requirement to use all of a module's outputs. TensorFlow will run only those parts of the module that end up as dependencies of a target in tf.Session.run(). Indeed, module publishers may choose to provide various outputs for advanced uses (like activations of intermediate layers) along with the main outputs. Module consumers should handle additional outputs gracefully.

Trying out alternative modules

Whenever there are multiple modules for the same task, TensorFlow Hub encourages to equip them with compatible signatures (interfaces) such that trying different ones is as easy as varying the module handle as a string-valued hyperparameter.

To this end, we maintain a collection of recommended Common Signatures for popular tasks.

Creating a New Module

Compatibility note

The TF1 Hub format is geared towards TensorFlow 1. It is only partially supported by TF Hub in TensorFlow 2. Please do consider publishing in the new TF2 SavedModel format instead.

The TF1 Hub format is similar to the SavedModel format of TensorFlow 1 on a syntactic level (same file names and protocol messages) but semantically different to allow for module reuse, composition and re-training (e.g., different storage of resource initializers, different tagging conventions for metagraphs). The easiest way to tell them apart on disk is the presence or absence of the tfhub_module.pb file.

General approach

To define a new module, a publisher calls hub.create_module_spec() with a function module_fn. This function constructs a graph representing the module's internal structure, using tf.placeholder() for inputs to be supplied by the caller. Then it defines signatures by calling hub.add_signature(name, inputs, outputs) one or more times.

For example:

def module_fn():
  inputs = tf.placeholder(dtype=tf.float32, shape=[None, 50])
  layer1 = tf.layers.dense(inputs, 200)
  layer2 = tf.layers.dense(layer1, 100)
  outputs = dict(default=layer2, hidden_activations=layer1)
  # Add default signature.
  hub.add_signature(inputs=inputs, outputs=outputs)

...
spec = hub.create_module_spec(module_fn)

The result of hub.create_module_spec() can be used, instead of a path, to instantiate a module object within a particular TensorFlow graph. In such case, there is no checkpoint, and the module instance will use the variable initializers instead.

Any module instance can be serialized to disk via its export(path, session) method. Exporting a module serializes its definition together with the current state of its variables in session into the passed path. This can be used when exporting a module for the first time, as well as when exporting a fine tuned module.

For compatibility with TensorFlow Estimators, hub.LatestModuleExporter exports modules from the latest checkpoint, just like tf.estimator.LatestExporter exports the entire model from the latest checkpoint.

Module publishers should implement a common signature when possible, so that consumers can easily exchange modules and find the best one for their problem.

Real example

Take a look at our text embedding module exporter for a real-world example of how to create a module from a common text embedding format.

Fine-Tuning

Training the variables of an imported module together with those of the model around it is called fine-tuning. Fine-tuning can result in better quality, but adds new complications. We advise consumers to look into fine-tuning only after exploring simpler quality tweaks, and only if the module publisher recommends it.

For Consumers

To enable fine-tuning, instantiate the module with hub.Module(..., trainable=True) to make its variables trainable and import TensorFlow's REGULARIZATION_LOSSES. If the module has multiple graph variants, make sure to pick the one appropriate for training. Usually, that's the one with tags {"train"}.

Choose a training regime that does not ruin the pre-trained weights, for example, a lower learning rate than for training from scratch.

For Publishers

To make fine-tuning easier for consumers, please be mindful of the following:

  • Fine-tuning needs regularization. Your module is exported with the REGULARIZATION_LOSSES collection, which is what puts your choice of tf.layers.dense(..., kernel_regularizer=...) etc. into what the consumer gets from tf.losses.get_regularization_losses(). Prefer this way of defining L1/L2 regularization losses.

  • In the publisher model, avoid defining L1/L2 regularization via the l1_ and l2_regularization_strength parameters of tf.train.FtrlOptimizer, tf.train.ProximalGradientDescentOptimizer, and other proximal optimizers. These are not exported alongside the module, and setting regularization strengths globally may not be appropriate for the consumer. Except for L1 regularization in wide (i.e. sparse linear) or wide & deep models, it should be possible to use individual regularization losses instead.

  • If you use dropout, batch normalization, or similar training techniques, set their hyperparameters to values that make sense across many expected uses. The dropout rate may have to be adjusted to the target problem's propensity to overfitting. In batch normalization, the momentum (a.k.a. decay coefficient) should be small enough to enable fine-tuning with small datasets and/or large batches. For advanced consumers, consider adding a signature that exposes control over critcial hyperparameters.