原版本：import gin.tf.external_configurables会报错： AttributeError: module 'tensorflow._api.v1.keras.optimizers' has no attribute 'Ftrl' 我删去了gin.tf.external_configurables 的两行话之后就没问题了。调整版本在此。 安装方法：

git clone https://github.com/Li-Shu14/gin-config.git
cd gin-config
python -m setup install

Gin provides a lightweight configuration framework for Python, based on dependency injection. Functions or classes can be decorated with @gin.configurable, allowing default parameter values to be supplied from a config file (or passed via the command line) using a simple but powerful syntax. This removes the need to define and maintain configuration objects (e.g. protos), or write boilerplate parameter plumbing and factory code, while often dramatically expanding a project's flexibility and configurability.

Gin is particularly well suited for machine learning experiments (e.g. using TensorFlow), which tend to have many parameters, often nested in complex ways.

Authors: Dan Holtmann-Rice, Sergio Guadarrama, Nathan Silberman Contributors: Oscar Ramirez, Marek Fiser

This is not an official Google product.

[TOC]

This section provides a high-level overview of Gin's main features, ordered roughly from "basic" to "advanced". More details on these and other features can be found in the user guide.

Install Gin with pip:

pip install gin-config

Install Gin from source:

git clone https://github.com/google/gin-config
cd gin-config
python -m setup.py install

Import Gin (without TensorFlow functionality):

import gin

Import additional TensorFlow-specific functionality via the TF module:

import gin.tf

At its most basic, Gin can be seen as a way of providing or changing default values for function or constructor parameters. To make a function's parameters "configurable", Gin provides the gin.configurable decorator:

@gin.configurable
def dnn(inputs,
        num_outputs,
        layer_sizes=(512, 512),
        activation_fn=tf.nn.relu):
  ...

This decorator registers the dnn function with Gin, and automatically makes all of its parameters configurable. To set ("bind") a value for the layer_sizes parameter above within a ".gin" configuration file:

# Inside "config.gin"
dnn.layer_sizes = (1024, 512, 128)

Bindings have syntax function_name.parameter_name = value. All Python literal values are supported as value (numbers, strings, lists, tuples, dicts). Once the config file has been parsed by Gin, any future calls to dnn will use the Gin-specified value for layer_sizes (unless a value is explicitly provided by the caller).

Classes can also be marked as configurable, in which case the configuration applies to constructor parameters:

@gin.configurable
class DNN(object):
  # Constructor parameters become configurable.
  def __init__(self,
               num_outputs,
               layer_sizes=(512, 512),
               activation_fn=tf.nn.relu):
    ...

  def __call__(inputs):
    ...

Within a config file, the class name is used when binding values to constructor parameters:

# Inside "config.gin"
DNN.layer_sizes = (1024, 512, 128)

Finally, after defining or importing all configurable classes or functions, parse your config file to bind your configurations (to also permit multiple config files and command line overrides, see gin.parse_config_files_and_bindings):

gin.parse_config_file('config.gin')

Note that no other changes are required to the Python code, beyond adding the gin.configurable decorator and a call to one of Gin's parsing functions.

In addition to accepting Python literal values, Gin also supports passing other Gin-configurable functions or classes. In the example above, we might want to change the activation_fn parameter. If we have registered, say tf.nn.tanh with Gin (see registering external functions), we can pass it to activation_fn by referring to it as @tanh (or @tf.nn.tanh):

# Inside "config.gin"
dnn.activation_fn = @tf.nn.tanh

Gin refers to @name constructs as configurable references. Configurable references work for classes as well:

def train_fn(..., optimizer_cls, learning_rate):
  optimizer = optimizer_cls(learning_rate)
  ...

Then, within a config file:

# Inside "config.gin"
train_fn.optimizer_cls = @tf.train.GradientDesecentOptimizer
...

Sometimes it is necessary to pass the result of calling a specific function or class constructor. Gin supports "evaluating" configurable references via the @name() syntax. For example, say we wanted to use the class form of DNN from above (which implements __call__ to "behave" like a function) in the following Python code:

def build_model(inputs, network_fn, ...):
  logits = network_fn(inputs)
  ...

We could pass an instance of the DNN class to the network_fn parameter:

# Inside "config.gin"
build_model.network_fn = @DNN()

To use evaluated references, all of the referenced function or class's parameters must be provided via Gin. The call to the function or constructor takes place just before the call to the function to which the result is passed, In the above example, this would be just before build_model is called.

The result is not cached, so a new DNN instance will be constructed for each call to build_model.

What happens if we want to configure the same function in different ways? For instance, imagine we're building a GAN, where we might have a "generator" network and a "discriminator" network. We'd like to use the dnn function above to construct both, but with different parameters:

def build_model(inputs, generator_network_fn, discriminator_network_fn, ...):
  ...

To handle this case, Gin provides "scopes", which provide a name for a specific set of bindings for a given function or class. In both bindings and references, the "scope name" precedes the function name, separated by a "/" (i.e., scope_name/function_name):

# Inside "config.gin"
build_model.generator_network_fn = @generator/dnn
build_model.discriminator_network_fn = @discriminator/dnn

generator/dnn.layer_sizes = (128, 256)
generator/dnn.num_outputs = 784

discriminator/dnn.layer_sizes = (512, 256)
discriminator/dnn.num_outputs = 1

dnn.activation_fn = @tf.nn.tanh

In this example, the generator network has increasing layer widths and 784 outputs, while the discriminator network has decreasing layer widths and 1 output.

Any parameters set on the "root" (unscoped) function name are inherited by scoped variants (unless explicitly overridden), so in the above example both the generator and the discriminator use the tf.nn.tanh activation function.

The greatest degree of flexibility and configurability in a project is achieved by writing small modular functions and "wiring them up" hierarchically via (possibly scoped) references. For example, this code sketches a generic training setup that could be used with the tf.estimator.Estimator API:

@gin.configurable
def build_model_fn(network_fn, loss_fn, optimize_loss_fn):
  def model_fn(features, labels):
    logits = network_fn(features)
    loss = loss_fn(labels, logits)
    train_op = optimize_loss_fn(loss)
    ...
  return model_fn

@gin.configurable
def optimize_loss(loss, optimizer_cls, learning_rate):
  optimizer = optimizer_cls(learning_rate)
  return optimizer.minimize(loss)

@gin.configurable
def input_fn(file_pattern, batch_size, ...):
  ...

@gin.configurable
def run_training(train_input_fn, eval_input_fn, estimator, steps=1000):
  estimator.train(train_input_fn, steps=steps)
  estimator.evaluate(eval_input_fn)
  ...

In conjunction with suitable external configurables to register TensorFlow functions/classes (e.g., Estimator and various optimizers), this could be configured as follows:

# Inside "config.gin"
run_training.train_input_fn = @train/input_fn
run_training.eval_input_fn = @eval/input_fn

input_fn.batch_size = 64  # Shared by both train and eval...
train/input_fn.file_pattern = ...
eval/input_fn.file_pattern = ...


run_training.estimator = @tf.estimator.Estimator()
tf.estimator.Estimator.model_fn = @build_model_fn()

build_model_fn.network_fn = @dnn
dnn.layer_sizes = (1024, 512, 256)

build_model_fn.loss_fn = @tf.losses.sparse_softmax_cross_entropy

build_model_fn.optimize_loss_fn = @optimize_loss

optimize_loss.optimizer_cls = @tf.train.MomentumOptimizer
MomentumOptimizer.momentum = 0.9

optimize_loss.learning_rate = 0.01

Note that it is straightforward to switch between different network functions, optimizers, datasets, loss functions, etc. via different config files.

Additional features described in more detail in the user guide include:

• Automatic logging of all configured parameter values (the "operative config"), including TensorBoard integration.
• "Macros", to specify a value used in multiple places within a config, as well as Python-defined constants.
• Module imports and config file inclusion.
• Disambiguation of configurable names via modules.

At a high level, we recommend using the minimal feature set required to achieve your project's desired degree of configurability. Many projects may only require the features outlined in sections 2 or 3 above. Extreme configurability comes at some cost to understandability, and the tradeoff should be carefully evaluated for a given project.

Gin is still in alpha development and some corner-case behaviors may be changed in backwards-incompatible ways. We recommend the following best practices:

• Minimize use of evaluated configurable references (@name()), especially when combined with macros (where the fact that the value is not cached may be surprising to new users).
• Avoid nesting of scopes (i.e., scope1/scope2/function_name). While supported there is some ongoing debate around ordering and behavior.
• When passing an unscoped reference (@name) as a parameter of a scoped function (some_scope/fn.param), the unscoped reference gets called in the scope of the function it is passed to... but don't rely on this behavior.
• Wherever possible, prefer to use a function or class's name as its configurable name, instead of overriding it. In case of naming collisions, use module names (which are encouraged to be renamed to match common usage) for disambiguation.
• In fact, to aid readability for complex config files, we gently suggest always including module names to help make it easier to find corresponding definitions in Python code.
• When doing "full hierarchical configuration" (section 4 above), structure the code to minimize the number of "top-level" functions that are configured without themselves being passed as parameters. In other words, the configuration tree should have only one root.

In short, use Gin responsibly :)

A quick reference for syntax unique to Gin (which otherwise supports non-control-flow Python syntax, including literal values and line continuations). Note that where function and class names are used, these may include a dotted module name prefix (some.module.function_name).