@ -5,9 +5,11 @@ In a large scale machine learning setup where the size of the training data is h
Polyak and Juditsky (1992) showed that the test performance of simple average of parameters obtained by Stochastic Gradient Descent (SGD) is as good as that of parameter values that are obtained by training the model over and over again, over the training dataset.
Hence, to accelerate the speed of Stochastic Gradient Descent, Averaged Stochastic Gradient Descent (ASGD) was proposed in Polyak and Juditsky (1992). For ASGD, the running average of parameters obtained by SGD, is used as the estimator for <imgsrc="./images/theta_star.gif"/><br/> . The averaging is done as follows:
Hence, to accelerate the speed of Stochastic Gradient Descent, Averaged Stochastic Gradient Descent (ASGD) was proposed in Polyak and Juditsky (1992). For ASGD, the running average of parameters obtained by SGD, is used as the estimator for <imgsrc="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/theta_star.gif"/><br/> . The averaging is done as follows:
@ -14,11 +14,29 @@ In programming languages, a block is a pair of curly braces that includes local
Blocks work with control flow structures like `if`, `else`, and `for`, which have equivalents in deep learning:
| programming languages | PaddlePaddle |
|-----------------------|-----------------------|
| for, while loop | RNN, WhileOp |
| if, if-else, switch | IfElseOp, SwitchOp |
| sequential execution | a sequence of layers |
<table>
<thead>
<tr>
<th>programming languages</th>
<th>PaddlePaddle</th>
</tr>
</thead>
<tbody>
<tr>
<td>for, while loop </td>
<td>RNN, WhileOp </td>
</tr>
<tr>
<td>if, if-else, switch </td>
<td>IfElseOp, SwitchOp </td>
</tr>
<tr>
<td>sequential execution </td>
<td>a sequence of layers </td>
</tr>
</tbody>
</table>
A key difference is that a C++ program describes a one pass computation, whereas a deep learning program describes both the forward and backward passes.
@ -26,12 +44,33 @@ A key difference is that a C++ program describes a one pass computation, whereas
The existence of the backward pass makes the execution of a block of PaddlePaddle different from traditional programs:
We'd like to have Python bindings to operators in package `paddle.operator`, and Python compositions of operators in package `paddle.layer`. So we have the following concepts in above illustrative example:
Like other deep learning systems, PaddlePaddle supports training models from sequence data. Also, like other systems, PaddlePaddle represent a mini-batch of sequences as a Tensor. What is different is that PaddlePaddle doesn't require all sequences in a mini-batch to be of the same length. Thus no need for padding zeros.
PaddlePaddle achieves this flexibility by passing through a new data type, *LoD Tensor*, which is a Tensor attached with segmentation index known as *LoD*, between operators. The LoD index doesn't only segment a tensor, but also recursively segments sub-sequences. This document presents the design of LoD and LoDTensor.
@ -10,10 +10,27 @@ PaddlePaddle uses proto message to describe compile time program because :
The computation `Program` consists of nested `Blocks`. Each `Block` will consist of data(i.e. `Variable`) and `Operations`. The concept to represent them is in the table below.
@ -116,7 +146,7 @@ An explanation of the above program:
- `fluid.k8s` is a package that provides access to Kubernetes API.
- `fluid.k8s.get_worker_addrs` returns the list of IP and ports of all pods of the current job except for the current one (the master pod).
- `fluid.tensor_array` creates a [tensor array](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor_array.h). `fluid.parallel_for` creates a `ParallelFor` intrinsic, which, when executed,
- `fluid.tensor_array` creates a [tensor array](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor_array.h). `fluid.parallel_for` creates a `ParallelFor` intrinsic, which, when executed,
1. creates `len(L)` scopes, each for the concurrent running of the sub-block (block 1 in this case), and initializes a variable named "index" in the scope to an integer value in the range `[0, len(L)-1]`, and
2. creates `len(L)` threads by calling into the `ThreadPool` singleton, each thread
@ -40,11 +40,11 @@ computation is only specified in Python code which sits outside of PaddlePaddle,
Similar to how a compiler uses an intermediate representation (IR) so that the programmer does not need to manually optimize their code for most of the cases, we can have an intermediate representation in PaddlePaddle as well. The compiler optimizes the IR as follows:
PaddlePaddle can support model parallelism by converting the IR so that the user no longer needs to manually perform the computation and operations in the Python component:
The IR for PaddlePaddle after refactoring is called a `Block`, it specifies the computation dependency graph and the variables used in the computation.
@ -60,7 +60,7 @@ For a detailed explanation, refer to this document -
The revamped distributed training architecture can address the above discussed limitations. Below is the illustration of how it does so:
@ -171,7 +171,7 @@ In the future, a more general placement algorithm should be implemented, which m
The local training architecture will be the same as the distributed training architecture, the difference is that everything runs locally, and there is just one PaddlePaddle runtime:
The above diagram shows an RNN unrolled into a full network.
@ -22,7 +22,7 @@ There are several important concepts here:
There could be local variables defined in each step-net. PaddlePaddle runtime realizes these variables in *step-scopes* which are created for each step.
@ -93,7 +93,7 @@ For example, we could have a 2-level RNN, where the top level corresponds to par
The following figure illustrates feeding in text into the lower level, one sentence at a step, and the feeding in step outputs to the top level. The final top level output is about the whole text.
Batch normalization is a frequently-used method in deep network training. It adjusts the mean and variance of a layer's output, and make the data distribution easier for next layer's training.
Batch normalization is a frequently-used method in deep network training. It adjusts the mean and variance of a layer's output, and make the data distribution easier for next layer's training.
The principle of batch normalization can be summarized into a simple function:
@ -66,7 +66,7 @@ As most C++ operators do, `batch_norm_op` is defined by inputs, outputs, attribu
The following graph showes the training computational process of `batch_norm_op`:
@ -119,15 +119,15 @@ for pass_id in range(PASS_NUM):
if pass_id % 100 == 0:
net.infer(test_image) # run inferencing model
# ...
```
```
`is_infer` is an attribute. Once an operator is created, its attributes can not be changed. It suggests us that we shall maintain two `batch_norm_op` in the model, one's `is_infer` is `True`(we call it `infer_batch_norm_op`) and the other one's is `False`(we call it `train_batch_norm_op`). They share all parameters and variables, but be placed in two different branches. That is to say, if a network contains a `batch_norm_op`, it will fork into two branches, one go through `train_batch_norm_op` and the other one go through `infer_batch_norm_op`:
Just like what is shown in the above graph, the net forks before `batch_norm_op` and will never merge again. All the operators after `batch_norm_op` will duplicate.
Just like what is shown in the above graph, the net forks before `batch_norm_op` and will never merge again. All the operators after `batch_norm_op` will duplicate.
When the net runs in training mode, the end of the left branch will be set as the running target, so the dependency tracking process will ignore right branch automatically. When the net runs in inferencing mode, the process is reversed.
@ -6,23 +6,23 @@ A central problem in machine learning is how to design an algorithm that will pe
### Parameter Norm Penalties
Most common regularization approaches in deep learning are based on limiting the capacity of the models by adding a parameter norm penalty to the objective function `J`. This is given as follows:
The parameter `alpha` is a hyperparameter that weights the relative contribution of the norm penalty term, `omega`, relative to the standard objective function `J`.
The most commonly used norm penalties are the L2 norm penalty and the L1 norm penalty. These are given as follows:
A much more detailed mathematical background of regularization can be found [here](http://www.deeplearningbook.org/contents/regularization.html).
## Regularization Survey
A detailed survey of regularization in various deep learning frameworks can be found [here](https://github.com/PaddlePaddle/Paddle/wiki/Regularization-Survey).
A detailed survey of regularization in various deep learning frameworks can be found [here](https://github.com/PaddlePaddle/Paddle/wiki/Regularization-Survey).
## Proposal for Regularization in PaddlePaddle
@ -32,41 +32,35 @@ In the new design, we propose to create new operations for regularization. For n
- L2_regularization_op
- L1_regularization_op
These ops can be like any other ops with their own CPU/GPU implementations either using Eigen or separate CPU and GPU kernels. As the initial implementation, we can implement their kernels using Eigen following the abstraction pattern implemented for [Activation Ops](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/accuracy_op.h). This abstraction pattern can make it very easy to implement new regularization schemes other than L1 and L2 norm penalties.
These ops can be like any other ops with their own CPU/GPU implementations either using Eigen or separate CPU and GPU kernels. As the initial implementation, we can implement their kernels using Eigen following the abstraction pattern implemented for [Activation Ops](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/accuracy_op.h). This abstraction pattern can make it very easy to implement new regularization schemes other than L1 and L2 norm penalties.
The idea of building ops for regularization is in sync with the refactored Paddle philosophy of using operators to represent any computation unit. The way these ops will be added to the computation graph, will be decided by the [layer functions](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/python_api.md#layer-function) in Python API.
The idea of building ops for regularization is in sync with the refactored Paddle philosophy of using operators to represent any computation unit. The way these ops will be added to the computation graph, will be decided by the [layer functions](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/python_api.md#layer-function) in Python API.
### Computation Graph
Below is an example of a really simple feed forward neural network.
Using the low level ops, `L2_regularization_op` and `L1_regularization_op`, any user can add regularization to their computation graphs. However, this will require a lot of lines of code and we should design Python APIs that support regularization. An example of such an API can be seen in [Keras](https://keras.io/regularizers/). As per the PaddlePaddle [Python API design](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/python_api.md), the layer functions are responsible for creating operators, operator parameters and variables. Since regularization is a property of parameters, it makes sense to create these in the layer functions.
Using the low level ops, `L2_regularization_op` and `L1_regularization_op`, any user can add regularization to their computation graphs. However, this will require a lot of lines of code and we should design Python APIs that support regularization. An example of such an API can be seen in [Keras](https://keras.io/regularizers/). As per the PaddlePaddle [Python API design](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/python_api.md), the layer functions are responsible for creating operators, operator parameters and variables. Since regularization is a property of parameters, it makes sense to create these in the layer functions.
#### Creation of Regularization ops
There are two possibilities for creating the regularization ops:
1. We create these ops immediately while building the computation graph.
2. We add these ops in a lazy manner, just before the backward, similar to the way the optimization ops are added.
1. We create these ops immediately while building the computation graph.
2. We add these ops in a lazy manner, just before the backward, similar to the way the optimization ops are added.
The proposal is to add these ops in a lazy manner just before the backward pass.
The proposal is to add these ops in a lazy manner just before the backward pass.
#### Storage of Regularization attributes
Since we want to create the regularization ops in a lazy manner, the regularization attributes (type of regularization and weight of regularization penalty) can be stored as attributes of the [`Parameter`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/framework.py#L421) class. This is because regularization is a property of the parameters and storing regularization properties with Parameters also allows for shared parameters.
Since we want to create the regularization ops in a lazy manner, the regularization attributes (type of regularization and weight of regularization penalty) can be stored as attributes of the [`Parameter`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/framework.py#L421) class. This is because regularization is a property of the parameters and storing regularization properties with Parameters also allows for shared parameters.
#### High-level API
In PaddlePaddle Python API, users will primarily rely on [layer functions](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/python_api.md#layer-function) to create neural network layers. Hence, we also need to provide regularization functionality in layer functions. The design of these APIs can be postponed for later right now. A good reference for these APIs can be found in [Keras](https://keras.io/regularizers/) and also by looking at Tensorflow in [`tf.contrib.layers`](https://www.tensorflow.org/api_guides/python/contrib.layers).
@ -10,11 +10,37 @@ Fluid is the answer. Fluid is similar to PyTorch and TensorFlow Eager Execution
Deep learning infrastructure is one of the fastest evolving technologies. Within four years, there have already been three generations of technologies invented.
| Existed since | model as sequence of layers | model as graph of operators | No model |
From the above table, we see that the deep learning technology is evolving towards getting rid of the concept of a model. To understand the reasons behind this direction, a comparison of the *programming paradigms* or the ways to program deep learning applications using these systems, would be helpful. The following section goes over these.
The word *graph* is interchangeable with *block* in this document. A graph consists of computation steps and local variables similar to a C++/Java program block, or a pair of parentheses(`{` and `}`).
We are planning to build Deep Speech 2 (DS2) \[[1](#references)\], a powerful Automatic Speech Recognition (ASR) engine, on PaddlePaddle. For the first-stage plan, we have the following short-term goals:
@ -68,11 +68,33 @@ We roughly break down the project into 14 tasks:
@ -102,37 +124,82 @@ We don't have to persist on this 2-3-7-1-1-1 depth \[[2](#references)\]. Similar
Key ingredients about the layers:
- **Data Layers**:
- **Data Layers**:
- Frame sequences data of audio **spectrogram** (with FFT).
- Token sequences data of **transcription** text (labels).
- Token sequences data of **transcription** text (labels).
- These two type of sequences do not have the same lengthes, thus a CTC-loss layer is required.
- **2D Convolution Layers**:
- **2D Convolution Layers**:
- Not only temporal convolution, but also **frequency convolution**. Like a 2D image convolution, but with a variable dimension (i.e. temporal dimension).
- With striding for only the first convlution layer.
- No pooling for all convolution layers.
- **Uni-directional RNNs**
- **Uni-directional RNNs**
- Uni-directional + row convolution: for low-latency inference.
- Bi-direcitional + without row convolution: if we don't care about the inference latency.
- **Row convolution**:
- For looking only a few steps ahead into the feature, instead of looking into a whole sequence in bi-directional RNNs.
- Not nessesary if with bi-direcitional RNNs.
- Not nessesary if with bi-direcitional RNNs.
- "**Row**" means convolutions are done within each frequency dimension (row), and no convolution kernels shared across.
- **Batch Normalization Layers**:
- Added to all above layers (except for data and loss layer).
- Sequence-wise normalization for RNNs: BatchNorm only performed on input-state projection and not state-state projection, for efficiency consideration.
Required Components | PaddlePaddle Support | Need to Develop
- The **Beam Search Decoder** for DS2 CTC-trained network follows the similar approach in \[[3](#references)\] as shown in Figure 2, with two important modifications for the ambiguous parts:
- 1) in the iterative computation of probabilities, the assignment operation is changed to accumulation for one prefix may comes from different paths;
- The **Beam Search Decoder** for DS2 CTC-trained network follows the similar approach in \[[3](#references)\] as shown in Figure 2, with two important modifications for the ambiguous parts:
- 1) in the iterative computation of probabilities, the assignment operation is changed to accumulation for one prefix may comes from different paths;
- 2) the if condition ```if l^+ not in A_prev then``` after probabilities' computation is deprecated for it is hard to understand and seems unnecessary.
- An **external scorer** would be passed into the decoder to evaluate a candidate prefix during decoding whenever a white space appended in English decoding and any character appended in Mandarin decoding.
- Such external scorer consists of language model, word count or any other custom scorers.
- The **language model** is built from Task 5, with parameters should be carefully tuned to achieve minimum WER/CER (c.f. Task 7)
- This decoder needs to perform with **high efficiency** for the convenience of parameters tuning and speech recognition in reality.
- This decoder needs to perform with **high efficiency** for the convenience of parameters tuning and speech recognition in reality.
GAN (General Adversarial Net [https://arxiv.org/abs/1406.2661]) is an important model for unsupervised learning and widely used in many areas.
GAN (General Adversarial Net [https://arxiv.org/abs/1406.2661]) is an important model for unsupervised learning and widely used in many areas.
It applies several important concepts in machine learning system design, including building and running subgraphs, dependency tracing, different optimizers in one executor and so forth.
In our GAN design, we wrap it as a user-friendly easily customized python API to design different models. We take the conditional DC-GAN (Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks [https://arxiv.org/abs/1511.06434]) as an example due to its good performance on image generation.
Figure 1. The overall running logic of GAN. The black solid arrows indicate the forward pass; the green dashed arrows indicate the backward pass of generator training; the red dashed arrows indicate the backward pass of the discriminator training. The BP pass of the green (red) arrow should only update the parameters in the green (red) boxes. The diamonds indicate the data providers. d\_loss and g\_loss marked in red and green are the two targets we would like to run.
</p>
The operators, layers and functions required/optional to build a GAN demo is summarized in https://github.com/PaddlePaddle/Paddle/issues/4563.
Figure 2. Photo borrowed from the original DC-GAN paper.
</p>
## The Conditional-GAN might be a class.
## The Conditional-GAN might be a class.
This design we adopt the popular open source design in https://github.com/carpedm20/DCGAN-tensorflow and https://github.com/rajathkmp/DCGAN. It contains following data structure:
- DCGAN(object): which contains everything required to build a GAN model. It provides following member functions methods as API:
@ -29,7 +29,7 @@ This design we adopt the popular open source design in https://github.com/carped
Returns a generated image.
- discriminator(image):
Given an image, decide if it is from a real source or a fake one.
Given an image, decide if it is from a real source or a fake one.
Returns a 0/1 binary label.
- build_model(self):
@ -47,7 +47,7 @@ To be more detailed, we introduce our design of DCGAN as following:
Yancey1989
7 years ago