pipeline

delete _internal_config_dict_converted

convert init_subclass to wrapper

trainer_configmixin

sha256 and pretrained. fix wrapper

download from obs

remove test code

image_classification_trainer

append sys

add dict

mindspore dataset

fix bug findFromPreConfig

fix path bug

fix path bug

pathlib

class check

add document

checkfiles

autopipeline

documentation

doc

init

docs/en/source/pipeline/overview.md

@@ -0,0 +1,196 @@
+# Pipeline
+
+## Introduction
+
+To enable fast downloading of pre-trained models and perform model training and inference with minimal code, tinyms provides pipeline API. With the pipeline, you can simply use a few lines of code to download the model from the cloud and perform inference locally. Besides, you can also use just a few lines of code to train and fine-tune the model.
+
+## Examples
+
+### Load a pretrained model
+
+``` python
+>>> from tinyms.pipeline import AutoModelPileline
+>>> model = AutoModelPileline.from_pretrained("model_cache_path", "model_repo")
+>>> pred = model(your_input)
+```
+
+### Load a trainer
+
+```python
+>>> from tinyms.pipeline import AutoTrainerPipeline
+>>> trainer = AutoTrainerPipeline.from_pretrained("trainer_cache_path", "trainer_repo")
+>>> trainer.init_model(model)
+>>> trainer.train()
+```
+
+You can also pass some arguments such as epoch to the `train` method. The passed arguments will override the default value.
+
+After training, you can save your model by invoking `save_pretrained` method.
+
+```python
+>>> model.save_pretrained("path_to_save")
+```
+
+This will save config and checkpoint to `path_to_save`.
+
+
+
+
+## For Develeper
+
+## details for config
+
+### structure of folder
+
+A the folder structure using `from_pretrained` to load is displayed below. 
+
+
+    model_cache_path
+    ├── demo_model
+        ├── model
+        │   ├── config.json
+        │   ├── filelist.txt
+        │   └── weight.ckpt
+        ├── model_code
+        │   ├── filelist.txt
+        │   ├── networks_test.py
+        └── trainer
+            ├── config.json
+            ├── filelist.txt
+            └── train_config
+                └── config.json
+
+`model_cache_path` is the path to save the downloaded repo.
+
+`demo_model` is the name of downloaded repo.
+
+`model` is the folder to save config and checkpoint.
+
+`model_code` is the folder to save the extra-code. If no extra-code, there is no this folder.
+
+`trainer` is the folder to save config about training the model.
+
+The config folder can be generated by invoking `save_pretrained`.
+
+### structure of config.json
+
+Blocks below is an example of the `config.json` in `trainer`.
+
+``` json
+{
+  "__module__": "tinyms.pipeline.image_classification_trainer.Trainer",
+  "__version__": "0.3.2",
+  "build_config": null,
+  "eval_config": null,
+  "fit_config": null,
+  "loss": {
+    "loss": "SoftmaxCrossEntropyWithLogits",
+    "params": {
+      "sparse": true
+    }
+  },
+  "metrics": [
+    "accuracy"
+  ],
+  "optim": {
+    "optimizer": "Momentum",
+    "params": {
+      "learning_rate": 0.1,
+      "momentum": 0.9
+    }
+  },
+  "predict_config": null,
+  "train_config": {
+    "__module__": "tinyms.pipeline.image_classification_trainer.TrainConfig",
+    "__subfolder__": null
+  }
+}
+```
+
+This is an example of `config.json`. `save_pretrained` will record the `__module__` needed and the tinyms version. Other keys such as `loss` is used to instantiates module `tinyms.pipeline.image_classification_trainer.Trainer`. If a dictionary has key `__subfolder__`, the details of this arguments will be saved into a subfoleder with name of this dictionary.
+
+
+## Define new model
+
+To define new model class, you need define a new class inherited from `ConfigMixin`.
+
+```python
+>>> from tinyms.pipeline import ConfigMixin, save_config, Ignore, SubFolder
+>>> class Model(ConfigMixin):
+...     @save_config
+...     def __init__(
+...        self, a: Ignore=1, b: SubFolder=2, c: Union[Ignore, int]=3, d: Union[SubFolder: int]=4):
+>>> model = Model()        
+>>> model.save_pretrained("config_path")
+```
+
+`ConfigMixin` is a base class for pipeline mixin. This class provides methods forsaving and loading model config. A class that inherits from this class can apply `@save_config` to `__init__` method to record the config of the class.
+
+If you wrap `__init__` with `@save_config`, the argument of Ignore type will not be saved into the config. The SubFolder type will be saved into a sub folder.
+
+Set `__prefix__` to change the name of the folder to save config and checkpoint.
+Set `__weight__` to change the name of the checkpoint file.
+
+
+tinyms also provide a function `wrap_config_mixin` to define a new model.
+
+This function will add ConfigMixin to the base class of the class and wrap `__init__` method with `@save_config`.
+
+```python
+>>> class Model:
+...     ...
+>>>
+>>> Model = wrap_config_mixin(Model)
+>>> model = Model()
+>>> model.save_pretrained("config_path")
+```
+
+### Extra-code
+
+If the defined model is not in tinyms, you need save the code containing the defination of the model class. The code should save to `model_code` folder in the repo folder. When calling `from_pretrained`, the path of `model_code` will append to `sys.path`.
+
+
+## Define a new trainer
+
+To define new trainer class, you need define a new class inherited from `TrainerConfigMixin`.
+
+`TrainerConfigMixin` is a base class for trainer pipeline mixin. This class provides methods for saving and loading model config. A class that is inherited from this class can apply `@save_config` to `__init__` method to record the config of the class.
+
+If you wrap `__init__` with `@save_config`, the argument of Ignore type will not be saved into the config. The SubFolder type will be saved into a sub folder.
+
+For a trainer, you may want to implement some methods like `train`, `eval`, `predict`. You can use `@set_from_config` to set the arguments from the config. The `FromConfig` type arguments having the following property.
+
+1. The arguments that are not in the config will be set to default value. 
+2. The arguments set  in running time will override the arguments in the config.
+
+Once you wrap a method with `@set_from_config`, you can use `BaseArgsFromConfig` to generate the arguments class. To use `BaseArgsFromConfig`, you should wrap the `__init__` method with `@copy_signature(Trainer.method)`. The arguments of `__init__` method should be `__init__(self, *args, **kwargs)`.
+
+You should define the arguments class in `__init__` method. The default name of the arguments class is `{method_name}_config`. You can change the name by passing the name to `@set_from_config(name)`.
+
+```python
+>>> class Trainer(TrainerConfigMixin):
+...     @save_config
+...     def __init__(self, train_args=None):
+...        self.train_args = train_args
+...
+...     @set_from_config
+...     def train(self, epoch: FromConfig):
+...        ...
+>>>
+>>> class TrainConfig(BaseArgsFromConfig):
+...     @copy_signature(Trainer.train)
+...     def __init__(self, *args, **kwargs):
+...         super().__init__(*args, **kwargs)
+>>>
+>>> train_config = TrainConfig(2)
+>>> trainer = Trainer(train_config=train_config)
+>>>
+>>> trainer.save_pretrained('model_config')
+>>> new_trainer = trainer.from_pretrained('model_config')
+>>>
+>>> new_trainer.train()
+>>>
+>>> new_trainer.train(4)
+
+```
+

tinyms/pipeline/__init__.py

@@ -0,0 +1,4 @@
+from .auto_pipeline import AutoModelPipeline, AutoTrainerPipeline
+from .configmixin import ConfigMixin, wrap_config_mixin, save_config
+from .trainer_configmixin import TrainerConfigMixin
+from .download import make_filelist

tinyms/pipeline/auto_pipeline.py

@@ -0,0 +1,5 @@
+from .configmixin import ConfigMixin
+from .trainer_configmixin import TrainerConfigMixin
+
+AutoModelPipeline = ConfigMixin
+AutoTrainerPipeline = TrainerConfigMixin