MindSpore Deep Probabilistic Programming (MDP) is a programming library for Bayesian deep learning. MDP is cooperatively developed with [ZhuSuan](https://zhusuan.readthedocs.io/en/latest/), which provides deep learning style primitives and algorithms for building probabilistic models and applying Bayesian inference.
MindSpore Deep Probabilistic Programming (MDP) is a programming library for Bayesian deep learning. MDP is cooperatively developed with [ZhuSuan](https://zhusuan.readthedocs.io/en/latest/), which provides deep learning style primitives and algorithms for building probabilistic models and applying Bayesian inference.
The objective of MDP is to integrate deep learning with Bayesian learning. On the one hand, similar to other Deep Probabilistic Programming Languages (DPPL) (e.g., TFP, Pyro), for the professional Bayesian learning researchers, MDP provides probability sampling, inference algorithms, and model building libraries; On the other hand, MDP provides high-level APIs for DNN researchers that are unfamiliar with Bayesian models, making it possible to take advantage of Bayesian models without the need of changing their DNN programming logics.
The objective of MDP is to integrate deep learning with Bayesian learning. On the one hand, similar to other Deep Probabilistic Programming Languages (DPPL) (e.g., TFP, Pyro), for the professional Bayesian learning researchers, MDP provides probability sampling, inference algorithms, and model building libraries; On the other hand, MDP provides high-level APIs for DNN researchers that are unfamiliar with Bayesian models, making it possible to take advantage of Bayesian models without the need of changing their DNN programming logics.
...
@@ -21,6 +21,7 @@ The objective of MDP is to integrate deep learning with Bayesian learning. On th
...
@@ -21,6 +21,7 @@ The objective of MDP is to integrate deep learning with Bayesian learning. On th
**Layer 1-2: Probabilistic inference algorithms**
**Layer 1-2: Probabilistic inference algorithms**
- SVI([mindspore.nn.probability.infer.variational](https://gitee.com/mindspore/mindspore/tree/master/mindspore/nn/probability/infer/variational)): A unified interface for stochastic variational inference.
- SVI([mindspore.nn.probability.infer.variational](https://gitee.com/mindspore/mindspore/tree/master/mindspore/nn/probability/infer/variational)): A unified interface for stochastic variational inference.
- MC: Algorithms for approximating integrals via sampling.
- MC: Algorithms for approximating integrals via sampling.
...
@@ -47,6 +48,7 @@ MDP requires MindSpore version 0.7.0-beta or later. MDP is actively evolving. In
...
@@ -47,6 +48,7 @@ MDP requires MindSpore version 0.7.0-beta or later. MDP is actively evolving. In
```
```
import mindspore.nn as nn
import mindspore.nn as nn
from mindspore.nn.probability import bnn_layers
from mindspore.nn.probability import bnn_layers
import mindspore.ops.operations as P
class BNNLeNet5(nn.Cell):
class BNNLeNet5(nn.Cell):
"""
"""
...
@@ -92,33 +94,24 @@ class BNNLeNet5(nn.Cell):
...
@@ -92,33 +94,24 @@ class BNNLeNet5(nn.Cell):
The way to construct Bayesian Neural Network by bnn_layers is the same as DNN. It's worth noting that bnn_layers and traditional layers of DNN can be combined with each other.
The way to construct Bayesian Neural Network by bnn_layers is the same as DNN. It's worth noting that bnn_layers and traditional layers of DNN can be combined with each other.
3. Define the Loss Function and Optimizer
3. Define the Loss Function and Optimizer
- Defining the Loss Function
The loss function `SoftmaxCrossEntropyWithLogits` and the optimizer `AdamWeightDecay` are used in the example. Call the loss function and optimizer in the `__main__` function.
The loss function `SoftmaxCrossEntropyWithLogits` is used in the example.
```
form mindspore.nn.loss import SoftmaxCrossEntropyWithLogits
```
Call the defined loss function in the `__main__` function.
The process of Bayesian network training is basically the same as that of DNN, the only differance is that WithLossCell is replaced with WithBNNLossCell suitable for BNN.
The process of Bayesian network training is basically the same as that of DNN, the only differance is that WithLossCell is replaced with WithBNNLossCell suitable for BNN.
Based on the two parameters `backbone` and `loss_fn` in WithLossCell, WithBNNLossCell adds two parameters of `dnn_factor` and `bnn_factor`. Those two parameters are used to trade off backbone's loss and kl loss to prevent kl loss from being too large to cover backbone's loss.
Based on the two parameters `backbone` and `loss_fn` in WithLossCell, WithBNNLossCell adds two parameters of `dnn_factor` and `bnn_factor`. Those two parameters are used to trade off backbone's loss and kl loss to prevent kl loss from being too large to cover backbone's loss.
The method `transform_to_bnn_model` can transform both convolutional layer and full connection layer of DNN model to BNN model. Its code is as follows:
The method `transform_to_bnn_model` can transform both convolutional layer and full connection layer of DNN model to BNN model. Its code is as follows:
```
```
...
@@ -349,15 +386,16 @@ Arg `get_dense_args` specifies which arguments to be gotten from full connection
...
@@ -349,15 +386,16 @@ Arg `get_dense_args` specifies which arguments to be gotten from full connection
The method `transform_to_bnn_layer` can transform a specific type of layers (nn.Dense or nn.Conv2d) in DNN model to corresponding BNN layer. Its code is as follows:
The method `transform_to_bnn_layer` can transform a specific type of layers (nn.Dense or nn.Conv2d) in DNN model to corresponding BNN layer. Its code is as follows:
The uncertainty estimation toolbox is based on MindSpore Deep Probabilistic Programming (MDP), and it is suitable for mainstream deep learning models, such as regression, classification, target detection and so on. In the inference stage, with the uncertainy estimation toolbox, developers only need to pass in the trained model and training dataset, specify the task and the samples to be estimated, then can obtain the aleatoric uncertainty and epistemic uncertainty. Based the uncertainty information, developers can understand the model and the dataset better.
The uncertainty estimation toolbox is based on MindSpore Deep Probabilistic Programming (MDP), and it is suitable for mainstream deep learning models, such as regression, classification, target detection and so on. In the inference stage, with the uncertainy estimation toolbox, developers only need to pass in the trained model and training dataset, specify the task and the samples to be estimated, then can obtain the aleatoric uncertainty and epistemic uncertainty. Based the uncertainty information, developers can understand the model and the dataset better.
-**Classification Task**
In classification task, for example, the model is lenet model, and the training dataset is mnist dataset. For evaluating the uncertainty of test examples, the use of the toolbox is as follows:
In classification task, for example, the model is lenet model. The MNIST dateset is used in the example. Data processing is consistent with [Implementing an Image Classification Application](https://www.mindspore.cn/tutorial/en/master/quick_start/quick_start.html) in Tutorial. For evaluating the uncertainty of test examples, the use of the toolbox is as follows:
```
```
evaluation = UncertaintyEvaluation(model=lenet,
from mindspore.nn.probability.toolbox.uncertainty_evaluation import UncertaintyEvaluation
train_dataset=mnist,
from mindspore.train.serialization import load_checkpoint, load_param_into_net
In regression task, for example, the model is MLP model, the training dataset is boston_housing. For evaluating the uncertainty of test examples, the use of the toolbox is as follows: