Skip to content

P
Paddle

机器未来 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle

体验新版 GitCode，发现更多精彩内容 >>

提交 6ab6c356 编写于 作者: W wangkuiyi 提交者： GitHub
浏览文件
操作

Merge pull request #1322 from wangkuiyi/design_doc_new_api 

Update API design doc according to discussions in issue #1315
上级 a30f6aa9 6915cb29
隐藏空白更改
内联 并排
Showing with 155 addition and 170 deletion
doc/design/api.md
浏览文件 @ 6ab6c356
......@@ -2,140 +2,148 @@
## Ingredients
As the first step of our design, we list important concepts in deep
learning and try to figure their relationship, as shown below:
As our design principle is starting from the essence: how could we
allow users to express and solve their problems at neural networks.
Some essential concepts that our API have to provide include:
```
Model = {topology, parameters}
1. A *topology* is an expression of *layers*.
Evaluator = {Model*, activations}
- forward
- test(cost, ...)
1. A layer could be any kind of computation, including *cost*.
GradientMachine = {Evaluator*, gradients}
- backward
1. Some layers have parameters, some don't. Most costs don't have
parameters.
Optimizer = {GradientMachine*}
- train(cost, ...)
- update
- checkpoint
```
1. In some topologies, layers share parameters. For
example,
[the network for training a ranking model](https://github.com/PaddlePaddle/Paddle/issues/1311#issuecomment-279121850).
where the pair of curly braces `{` and `}` indicate *composition*, `*`
indicates a *reference*, and `-` marks a "class method".
1. At programming time, users specify topologies and possible sharing
of parameters. PaddlePaddle can figure out and create parameters
required (and possibly shared) by one or more topologies.
### Model
## Starting from Examples
We used to think that parameters are part of the topology (or layers).
But that is not true because multiple layers could share the same
parameter matrix. An example is a network that compares two text
segments in a semantic space:
As a summarization
of
[our disucssion](https://github.com/PaddlePaddle/Paddle/issues/1315),
let us present two examples here:
```
semantic
text A -> projection ---\
layer A \
cosine
similarity -> output
layer
semantic /
text B -> projection ---/
layer B
```
In this network, the two semantic projection layers (A and B) share
the same parameter matrix.
### Example 1. Sharing Parameters between Layers
For more information about our API that specifies topology and
parameter sharing, please refer to [TODO: API].
We use
the
[3-branch ranking](https://github.com/PaddlePaddle/Paddle/issues/1311#issuecomment-279121850) model
in this example. For your convenience, I copy-a-paste the model's
topology as follows:
```
A -> f -\
Q -> f --> cost
B -> f -/
```
### Evaluator
Supposed that we have a trained ranking model, we should be able to
use it in our search engine. The search engine's Web server is a
concurrent program so to serve many HTTP requests simultaneously. It
doesn't make sense for each of these threads to have its own copy of the model because that would duplicate topologies and parameters.
However, each thread should be able to record layer outputs, i.e.,
activations, computed from an input, derived from the request. With
*Evaluator* that saves activations, we can write the over-simplified
server program as:
The following program trains the topology including the cost, and then
use the sub-network in the trained topology in inference:
```python
m = paddle.model.load("trained.model")
http.handle("/",
lambda req:
e = paddle.evaluator.create(m)
e.forward(req)
e.activation(layer="output")) # returns activations of layer "output"
def f(in):
e = paddle.layer.embedding(in, parameter_name="embedding")
o = paddle.layer.softmax(e, parameter_name="semantic")
return o
# Create 3 topologies (subnets), they share parameters because all
# correspoinding layers have the same parameter names.
fA = f(paddle.layer.data(input_name="A"))
fB = f(paddle.layer.data(input_name="B"))
fQ = f(paddle.layer.data(input_name="Q"))
topology = paddle.layer.less_than(
paddle.layer.cross_entropy(fA, fQ),
paddle.layer.corss_entropy(fB, fQ))
# Derive parameters required in topology and create them in model.
parameters = paddle.parameters.create(topology)
# Estimate parameters used in topology from data.
paddle.train(topology, parameters, reader=read_ranking_model_data)
# Inference using fA (or fB or fC, as they share their parameters).
[testA, testB, testQ] = read_ranking_model_data()
print "The sematic-vector of testA: ", paddle.infer(fA, parameters, testA)
```
### GradientMachine
Similar to the evaluation, the training needs to compute gradients so
to update model parameters. Because an [optimizer](#optimizer) might
run multiple simultaneous threads to update the same model, gradients
should be separated from the model. Because gradients are only used
in training, but not serving, they should be separate from Evaluator.
Hence the `GradientMachine`.
### Optimizer
None of Model, Evaluator, nor GradientMachine implements the training
loop, hence Optimizer. We can define a concurrent optimizer that runs
multiple simultaneous threads to train a model -- just let each
thread has its own GradientMachine object.
Most models should be able to be trained using the
`paddle.optimizer.SGD` by calling its `train` method. Many
customizations to the SGD algorithm happens with the update equation,
e.g., momentum and the Adam SGD algorithm. We make `train` calls
`update` to do an update, so that we can derive a `paddle.optimizer.Adam`
from `paddle.optimizer.SGD` by overrides only the `update` method.
### Example 2. Sharing Parameters between "Models"
We use [GAN](https://github.com/PaddlePaddle/book/tree/develop/gan) in
this example. In the following example program, `d0` and `d1`
correspond to the two networks in the following figure:
## Programming Interface
A fictive example of PaddlePaddle program looks like the following:
<img src="https://github.com/wangyang59/book/raw/00036f4b0da5225041a6824587c1a01cf20159b1/gan/image/gan_ig.png" width=400 />
```python
import paddle
def G(in):
# over-simplified example as G has only one layers:
return paddle.layer.fc(in, parameter_name="G")
def D(in);
# again, over-simplified:
return paddle.layer.fc(in, parameter_name="D")
# Construct the first topology, which contains both D and G.
# By learning this topology, we update parameters of G.
d0 = paddle.layer.should_be_false(D(G(paddle.layer.data())))
# Construct a second topology d1, which contains only D. By
# training this topology, we update parameters of D. Note
# that d1 share parameters with d0.
d1 = paddle.layer.should_be_true(D(paddle.layer.data()))
# Create parameters from a list of multiple topologies (models) for
# the chance to share parameters between these topologies.
parameters = paddle.parameters.create([d0, d1])
# Iterative training of GAN.
for ...:
train(d0, parameters, reader=read_from_rng, immutable_parameters={"D"})
train(d1, parameters, reader=read_from_realistic_images)
# Use d1 for inference:
print "D thinks a batch of images are realistic ", infer(d1, parameters, read_mnist_images)
```
def read(args):
f = open_file(args["filename"])
mb = read_a_minibatch(f)
end_pass = eof(f)
if end_pass:
f = open_file(args["filename"]) # rewind for reading again
yield mb, end_pass
input = paddle.layer.data(...)
intermediate = paddle.layers.fc(input)
output = paddle.layer.softmax(intermediate)
### Summarization
model = paddle.model.create(output)
paddle.train(model, data_provider=read)
```
Above two programs reveal some important design concerns:
This shows some important part of a program:
1. Users describe a topology as an expression of layers. Every layer
has a *parameter name*. If the users don't specify it explicitly, it's automatically generated as a unique name. By
specifying the parameter name, users can specify the sharing of
parameters between layers and even between topologies.
1. Define how to read (and augment) data by defining a function, in
this example, `read`, that `yields` a minibatch and a boolean flag
`eof_of_pass`.
1. `paddle.parameters.create` figures out parameters required by one
or more topologies from parameter names of layers. It creates these
parameters and returns a `ParameterSet` object, which is in essence
a map from *parameter names* to *parameters*.
1. Define the topology, `input`, `intermediate`, and `output` in this
example.
1. At training and inference time, `paddle.train` and `paddle.infer`
requires both a topology and the parameter set that holds the parameters of that topology. There are some reasons:
1. Create parameters from the topology thus forms the model by calling
`paddel.model.create`.
1. This prevents users from forgetting to call
`paddle.parameters.create`.
1. `paddle.train` needs to know which parameter set to update.
1. Users could load another (pre-trained) parameter set and use it
with a topology in `train.infer`.
1. Train the model by calling `paddle.train`.
1. By specifying the `immutable_parameters` parameter of
`paddle.train`, we can forbid the update of these parameters.
### Reader
## Reader
Not all programming frameworks allow users to define I/O functions.
An example is Google MapReduce, which can only read from text,
......@@ -145,91 +153,67 @@ readers and writers by deriving from base classes `Reader` and
decide to provide the flexibility to users to define their readers.
#### A Synthetic Data Reader
There are some open questions here:
Sometimes we want to test a topology and/or a training algorithm using
synthetic data. We can do this by defining the reader a synthesizer:
1. **Should a reader return a Python dictionary?**
```python
def read(args):
x = sample_from_uniform(0.0, 1.0)
y = sample_from_gauss(2 * x, sigma)
yield {x, y}, False # no end-of-file so no end-of-pass
```
1. **How to map multiple outputs from a reader to multiple data layers?**
#### A Reader for Online Learning
1. **How to easily compose some existing readers to read more data and
feed a topology with more data layers?**
Readers can also read an infinite data stream, e.g., a log stream from
a search engine and collected by Kafka:
```python
def read(args):
log_stream = kafka.open_channel(args["kafka channel name"])
yeild log_stream.read(), False # no end-of-pass in online learning
```
## Training
### Topology
By default, layers don't have names. But if we want to refer to a
layer later some time, for example, when we do serving using the model
and wants activations/outputs of a layer, we should give it a name.
The recommended way to training a model is to call `paddle.train`,
which simply calls `paddle.trainer.Default`, a global variable of
type `paddle.trainer.SGD`. Equivalently, we can do
```python
input = paddle.layer.data(...)
intermediate = paddle.layer.fc(input, name="inter", ...)
output = paddle.layer.softmax(intermediate, name="output", ...)
m = paddle.model.create(output)
e = paddle.evaluator.create(model)
e.forward(read_an_input()) # compute activations of all layers.
print e.activations(layer="inter") # retrieve the activations of layer "inter"
print e.activations(layer="output") # retrieve the activations of layer "output"
opt = paddle.trainer.SGD(..., paddle.updater.Adam(...))
opt.train(topology, parameters, reader=read, ...)
```
#### Sharing Parameters
### Updater
In [above section](#model) we shows a network whose two layers share
the same parameter matrix. To specify such cases, we give "parameter
names" to layers. If some layers have the same paraemter names,
`paddle.model.create` creates a single parameter matrix for these
layers:
Please be aware that a trainer can accept an updater as its data
member, where an updater is a class derived from
`paddle.trainer.Updater`. This is to make it easier to customize
trainers, as discussed
[here](https://github.com/PaddlePaddle/Paddle/issues/1319).
```python
text1 = paddle.layer.data(...)
sematic1 = paddle.layer.fc(text1, ..., parameter_name="sematic_projection")
text2 = paddle.layer.data(...)
sematic2 = paddle.layer.fc(text2, ..., parameter_name="sematic_projection")
out = paddle.layer.cosine(semantic1, semantic2)
```
### Event Handler
We can also share parameter matrices between layers in different
models. To do this, we need an additional parameter that refers to a
model:
`paddle.train` and `paddle.trainer.XXX.train` take an optional
parameter `event_handler`, which should be either `None` or a function
that handle some events:
```python
model1_input = paddle.layer.data(...)
model1_output = paddle.layer.softmax(model1_input, ...,
parameter_name="a_parameter_matrix")
model1 = paddle.model.create(model1_output)
# Another model
model2_semantic = paddle.layer.fc(text2, ...,
parameter_name="a_parameter_matrix",
parameter_model=model1)
```
1. BeginTraining
1. EndTraining
1. BeginIteration
1. EndIteration
1. BeginPass
1. EndPass
### Training
where EndPass is sent if and only if the reader yields
`end_pass=True`.
The recommended way to training a model is to call `paddle.train`,
which simply calls `paddle.optimizer.Default`, a global variable of
type `paddle.optimizer.SGD`. Equivalently, we can do
An example as follows:
```python
opt = paddle.optimizer.SGD(...)
opt.train(model, reader=read, ...)
def event_handler(event):
if ininstance(event, paddle.event.EndIteration):
print paddle.test(...)
paddle.train(topology, parameters, reader, event_handler)
```
#### Distributed Training
If we are writing a PaddlePaddle program in and for iPython/Jypyter,
we can use metaplotlib in the event handler to plot a curve of
cost/error versus iterations, as shown
[here](https://blog.dominodatalab.com/interactive-dashboards-in-jupyter/).
### Distributed Training
If users want to do distributed training on a cluster, s/he should
call `paddle.dist_train` and provides access tokens to the cluster as
......@@ -240,8 +224,9 @@ access a Kubernetes cluster, s/he should be able to call
```python
paddle.dist_train(model,
trainer=paddle.trainer.SGD(...,
paddle.updater.Adam(...)),
reader=read,
optimizer=paddle.optimizer.SGDOptimizer(...),
k8s_user="yi",
k8s_token="kube_cluster_tls.pem",
k8s_job="hello",
......@@ -251,7 +236,7 @@ paddle.dist_train(model,
The pseudo code if `paddle.dist_train` is as follows:
```python
def dist_train():
def dist_train(topology, parameters, trainer, reader, ...):
if os.getenv("KUBERNETES_SERVICE_HOST") == None:
image_name = k8s_user + '/' + k8s_job
docker_build(image_name)
......@@ -264,13 +249,13 @@ def dist_train():
elif rank < 15:
parameter_server()
else:
optimizer.train(model, reader=read)
trainer.train(model, reader=read)
```
Please be aware that if a process is running on the Kubernetes
cluster, it will have some environment variables pre-defined.
If `dist_train` doesn't see these environment variables, it knowns
If `dist_train` doesn't see these environment variables, it knows
that it's running on users' personal computer, and it should work as a
*launcher*. Otherwise, it knows that it's running on the cluster and
need to figure out its role as either the master, or a trainer, or a
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或