Huawei_Technology
/
mindspore-mindinsight
Not watched
1
0
Fork 0
Code
Releases 11 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

Add platform announcement of mindconverter.

tags/v1.1.0
liuchongming 5 years ago
parent
ecd1e3d8cd
commit
3b13b49e14
2 changed files with 33 additions and 34 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +15
    -16
      mindinsight/mindconverter/README.md
  2. +18
    -18
      mindinsight/mindconverter/README_CN.md

+ 15
- 16
mindinsight/mindconverter/README.md View File

@@ -79,12 +79,11 @@ optional arguments:

### PyTorch Model Scripts Migration

**MindConverter provides two modes for PyTorch：**
#### MindConverter Provides Two Modes for PyTorch：

1. **Abstract Syntax Tree (AST) based conversion**: Use the argument `--in_file` will enable the AST mode.
2. **Computational Graph based conversion**: Use `--model_file` and `--shape` arguments will enable the Graph mode.


> The AST mode will be enabled, if both `--in_file` and `--model_file` are specified.

For the Graph mode, `--shape` is mandatory.
@@ -104,7 +103,6 @@ Please note that your original PyTorch project is included in the module search

> AST mode is not supported for TensorFlow, only computational graph based mode is available.


## Scenario

MindConverter provides two modes for different migration demands.
@@ -122,7 +120,7 @@ Some typical image classification networks such as ResNet and VGG have been test
> 2. The Dropout operator will be lost after conversion because the inference mode is used to load the PyTorch or TensorFlow model. Manually re-implement is necessary.
> 3. The Graph-based mode will be continuously developed and optimized with further updates.

Supported models list:
Supported models list (Models in below table have been tested based on PyTorch 1.14.0 and TensorFlow 1.15.0, X86 Ubuntu released version):

| Supported Model | PyTorch Script | TensorFlow Script |
| :----: | :----: | :----: |
@@ -236,7 +234,6 @@ class Classifier(nn.Cell):

#### TensorFlow Model Scripts Conversion


To use TensorFlow model script migration, you need to export TensorFlow model to Pb format first, and obtain the model input node and output node name. You can refer to the following methods to export and obtain the node name:

```python
@@ -267,7 +264,8 @@ print(f"Input node name: {INPUT_NODE}, output node name: {OUTPUT_NODE}")
After the above code is executed, the model will be saved to `/home/user/xxx/frozen_model.pb`. `INPUT_NODE` can be passed into `--input_nodes`, and `OUTPUT_NODE` is the corresponding `--output_nodes`.

Suppose the input node name is `input_1:0`, output node name is `predictions/Softmax:0`, the input shape of model is `1,224,224,3`, the following command can be used to generate the script:
```shell script

```bash
mindconverter --model_file /home/user/xxx/frozen_model.pb --shape 1,224,224,3 \
--input_node input_1:0 \
--output_node predictions/Softmax:0 \
@@ -277,12 +275,9 @@ mindconverter --model_file /home/user/xxx/frozen_model.pb --shape 1,224,224,3 \

After executed MindSpore script, and report file can be found in corresponding directory.


The format of conversion report generated by script generation scheme based on graph structure is the same as that of AST scheme. However, since the graph based scheme is a generative method, the original pytorch script is not referenced in the conversion process. Therefore, the code line and column numbers involved in the generated conversion report refer to the generated script.


In addition, for operators that are not converted successfully, the input and output shape of tensor of the node will be identified in the code_ shape`, `output_ For example, please refer to [PyTorch Model Scripts Conversion](#manual_modify).

In addition, for operators that are not converted successfully, the input and output shape of tensor of the node will be identified in the code by `input_shape` and `output_shape`. For example, please refer to [PyTorch Model Scripts Conversion](#manual_modify).

## Caution

@@ -296,17 +291,18 @@ In addition, for operators that are not converted successfully, the input and ou

Classes and functions that can't be converted:

* The use of shape, ndim and dtype member of torch.Tensor.
* torch.nn.AdaptiveXXXPoolXd and torch.nn.functional.adaptive_XXX_poolXd()
* torch.nn.functional.Dropout
* torch.unsqueeze() and torch.Tensor.unsqueeze()
* torch.chunk() and torch.Tensor.chunk()
1. The use of `.shape`, `.ndim` and `.dtype` member of `torch.Tensor`.
2. `torch.nn.AdaptiveXXXPoolXd` and `torch.nn.functional.adaptive_XXX_poolXd()`.
3. `torch.nn.functional.Dropout`.
4. `torch.unsqueeze()` and `torch.Tensor.unsqueeze()`.
5. `torch.chunk()` and `torch.Tensor.chunk()`.

### Situation2

Subclassing from the subclasses of nn.Module

e.g. (code snip from torchvision.models.mobilenet)

```python
from torch import nn

@@ -324,4 +320,7 @@ class ConvBNReLU(nn.Sequential):

Q1. `terminate called after throwing an instance of 'std::system_error', what(): Resource temporarily unavailable, Aborted (core dumped)`:
> Answer: This problem is caused by TensorFlow. First step of conversion process is loading TensorFlow model into memory using TensorFlow module, and TensorFlow starts to apply for needed resource. When required resource is unavailable, such as exceeding max process number of Linux system limit, etc., TensorFlow will raise an error from its C/C++ layer. For more detail, please refer to TensorFlow official repository. There are some known issue for reference only:
[TF ISSUE 14885](https://github.com/tensorflow/tensorflow/issues/14885), [TF ISSUE 37449](https://github.com/tensorflow/tensorflow/issues/37449)
[TF ISSUE 14885](https://github.com/tensorflow/tensorflow/issues/14885), [TF ISSUE 37449](https://github.com/tensorflow/tensorflow/issues/37449)

Q2. Can MindConverter run on ARM platform?
> Answer: MindConverter usability on X86 Ubuntu machine has been verified, yet, on ARM has not.

+ 18
- 18
mindinsight/mindconverter/README_CN.md View File

@@ -79,7 +79,7 @@ optional arguments:

### PyTorch模型脚本迁移

**MindConverter提供两种PyTorch模型脚本迁移方案：**
#### MindConverter提供两种PyTorch模型脚本迁移方案：

1. **基于抽象语法树(Abstract syntax tree, AST)的脚本转换**：指定`--in_file`的值，将使用基于AST的脚本转换方案；
2. **基于图结构的脚本生成**：指定`--model_file`与`--shape`将使用基于图结构的脚本生成方案。
@@ -92,8 +92,7 @@ optional arguments:

另外，当使用基于图结构的脚本生成方案时，请确保原PyTorch项目已在Python包搜索路径中，可通过CLI进入Python交互式命令行，通过import的方式判断是否已满足；若未加入，可通过`--project_path`命令手动将项目路径传入，以确保MindConverter可引用到原PyTorch脚本。

> 假设用户项目目录为`/home/user/project/model_training`，用户可通过如下命令手动项目添加至包搜索路径中：`export PYTHONPATH=/home/user/project/model_training:$PYTHONPATH`

> 假设用户项目目录为`/home/user/project/model_training`，用户可通过如下命令手动项目添加至包搜索路径中：`export PYTHONPATH=/home/user/project/model_training:$PYTHONPATH`；
> 此处MindConverter需要引用原PyTorch脚本，是因为PyTorch模型反向序列化过程中会引用原脚本。

### TensorFlow模型脚本迁移
@@ -102,7 +101,6 @@ optional arguments:

> AST方案不支持TensorFlow模型脚本迁移，TensorFlow脚本迁移仅支持基于图结构的方案。


## 使用场景

MindConverter提供两种技术方案，以应对不同脚本迁移场景：
@@ -120,9 +118,9 @@ MindConverter提供两种技术方案，以应对不同脚本迁移场景：
> 2. 基于图结构的脚本生成方案，由于要加载PyTorch、TensorFlow模型，会导致转换后网络中Dropout算子丢失，需要用户手动补齐；
> 3. 基于图结构的脚本生成方案持续优化中。

支持网络列表:
支持的模型列表（如下模型已基于x86 Ubuntu发行版，PyTorch 1.4.0以及TensorFlow 1.15.0测试通过）:

| 支持网络 | PyTorch脚本 | TensorFlow脚本 |
| 模型 | PyTorch脚本 | TensorFlow脚本 |
| :----: | :----: | :----: |
| ResNet18 | [脚本链接](https://github.com/pytorch/vision/blob/v0.5.0/torchvision/models/resnet.py) | 暂未测试 |
| ResNet34 | [脚本链接](https://github.com/pytorch/vision/blob/v0.5.0/torchvision/models/resnet.py) | 暂未测试 |
@@ -235,12 +233,12 @@ class Classifier(nn.Cell):
#### TensorFlow模型脚本生成示例

使用TensorFlow模型脚本迁移，需要先将TensorFlow模型导出为pb格式，并且获取模型输入节点、输出节点名称，可参考如下方法进行导出、获取节点名称：

```python
import tensorflow as tf
from tensorflow.python.framework import graph_io
from tensorflow.keras.applications.inception_v3 import InceptionV3


def freeze_graph(graph, session, output):
saved_path = "/home/user/xxx"
with graph.as_default():
@@ -257,13 +255,13 @@ INPUT_NODE = base_model.inputs[0].op.name # Get input node name of TensorFlow.
OUTPUT_NODE = base_model.outputs[0].op.name # Get output node name of TensorFlow.
freeze_graph(session.graph, session, [out.op.name for out in base_model.outputs])
print(f"Input node name: {INPUT_NODE}, output node name: {OUTPUT_NODE}")

```

上述代码执行完毕，模型将会保存至`/home/user/xxx/frozen_model.pb`。其中，`INPUT_NODE`为输入节点名称，`OUTPUT_NODE`为输出节点名称。

假设输入节点名称为`input_1:0`、输出节点名称为`predictions/Softmax:0`，模型输入样本尺寸为`1,224,224,3`，则可使用如下命令进行脚本生成：
```shell script

```bash
mindconverter --model_file /home/user/xxx/frozen_model.pb --shape 1,224,224,3 \
--input_node input_1:0 \
--output_node predictions/Softmax:0 \
@@ -273,13 +271,10 @@ mindconverter --model_file /home/user/xxx/frozen_model.pb --shape 1,224,224,3 \

执行该命令，MindSpore代码文件、转换报告生成至相应目录。


基于图结构的脚本生成方案产生的转换报告格式与AST方案相同。然而，由于基于图结构方案属于生成式方法，转换过程中未参考原TensorFlow脚本，因此生成的转换报告中涉及的代码行、列号均指生成后脚本。


另外，对于未成功转换的算子，在代码中会相应的标识该节点输入、输出Tensor的shape（以`input_shape`、`output_shape`标识），便于用户手动修改，示例见[PyTorch模型脚本生成示例](#manual_modify)。


## 注意事项

1. PyTorch、TensorFlow、TF2ONNX(1.7.1)不作为MindInsight明确声明的依赖库。若想使用基于图结构的脚本生成工具，需要用户手动安装与生成PyTorch模型版本一致的PyTorch库（MindConverter推荐使用PyTorch 1.4.0或PyTorch 1.6.0进行脚本生成），或TensorFlow；
@@ -291,17 +286,19 @@ mindconverter --model_file /home/user/xxx/frozen_model.pb --shape 1,224,224,3 \
### 场景1

部分类和方法目前无法转换：
* 使用```torch.Tensor```的```shape```，```ndim```和```dtype```成员
* ```torch.nn.AdaptiveXXXPoolXd```和```torch.nn.functional.adaptive_XXX_poolXd()```
* ```torch.nn.functional.Dropout```
* ```torch.unsqueeze()```和```torch.Tensor.unsqueeze()```
* ```torch.chunk()```和```torch.Tensor.chunk()```

1. 使用`torch.Tensor`的`shape`，`ndim`和`dtype`成员；
2. `torch.nn.AdaptiveXXXPoolXd`和`torch.nn.functional.adaptive_XXX_poolXd()`；
3. `torch.nn.functional.Dropout`；
4. `torch.unsqueeze()`和`torch.Tensor.unsqueeze()`；
5. `torch.chunk()`和`torch.Tensor.chunk()`；

### 场景2

继承的父类是nn.Module的子类。

例如：(如下代码片段摘自torchvision.models.mobilenet)

```python
from torch import nn

@@ -319,4 +316,7 @@ class ConvBNReLU(nn.Sequential):

Q1. `terminate called after throwing an instance of 'std::system_error', what(): Resource temporarily unavailable, Aborted (core dumped)`:

> 答: 该问题由TensorFlow导致。脚本转换时，需要通过TensorFlow库加载TensorFlow的模型文件，此时TensorFlow会申请相关资源进行初始化，若申请资源失败（可能由于系统进程数超过Linux最大进程数限制），TensorFlow C/C++层会出现Core Dumped问题。详细信息请参考TensorFlow官方ISSUE，如下ISSUE仅供参考：[TF ISSUE 14885](https://github.com/tensorflow/tensorflow/issues/14885), [TF ISSUE 37449](https://github.com/tensorflow/tensorflow/issues/37449)
> 答: 该问题由TensorFlow导致。脚本转换时，需要通过TensorFlow库加载TensorFlow的模型文件，此时TensorFlow会申请相关资源进行初始化，若申请资源失败（可能由于系统进程数超过Linux最大进程数限制），TensorFlow C/C++层会出现Core Dumped问题。详细信息请参考TensorFlow官方ISSUE，如下ISSUE仅供参考：[TF ISSUE 14885](https://github.com/tensorflow/tensorflow/issues/14885), [TF ISSUE 37449](https://github.com/tensorflow/tensorflow/issues/37449)

Q2. MindConverter是否可以在ARM平台运行？
> 答：MindConverter已基于X86 Ubuntu发行版进行测试，ARM平台未进行测试。

Write Preview
Loading…
Cancel
Save