# tfpack

**Repository Path**: aidings/tfpack

## Basic Information

- **Project Name**: tfpack
- **Description**: tensorflow2.0训练框架
- **Primary Language**: Python
- **License**: Apache-2.0
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2020-06-25
- **Last Updated**: 2020-12-16

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# TF2.0训练框架

### [wiki]()

### 安装
`pip install git+https://gitlab.vmic.xyz/11110558/tfpack.git@版本号`<br>

- 依赖库: `tensorflow` > 2.0, `opencv-python`

### 脚本使用
- 1. DataFlow继承使用
```python
from tfpack import TFDataFlow
import numpy as np

class XXXDataFlow(TFDataFlow):
    def load(self, data_path):
        # load img_dirs and img_labs from file

        return np.array(img_dirs), np.array(img_labs)

    def parse(self, img_dir, img_lab):
        # load image and augumentation the image

        return image, label
```

- 2. Trainer继承使用
```python
from tfpack import Trainer

class XXXTrainer(Trainer):
    def loss(self, ypred, ytrue):
        # define your loss

        return loss

    def accuracy(self, ypred, ytrue):
        # define your accuracy

        return acc

```

- 3. 训练
```python
dataflow = XXXDataFlow(batch_size=32, parallel_size=8)

train_dataset = dataflow(train_list)
test_dataset = dataflow(test_list)


#define your model and optimizer
model = MobileNetV3()
optimizer = tf.keras.optimizer.SGD(lr=0.001)
model_dir = './XXX'

# build trainer
trainer = XXXTrainer(model=model, optimizer=optimizer)

# train your model
trainer.fit(train_dataset, test_dataset, model_dir=model_dir)

```

- 4. 查看训练loss和acc
```
tensorboard --logdir='./XXX' --bind_all
```

- 5. 小函数
    - 5.1 `utils.conv_name_dict(input)`: input:(str/list/dict), 其中str文件路径，每一行都是一个标签，返回为name_dict为 name:class_id
    - 5.2 `utils.one_hot(str_labs, name_dict)`: str_labs:(int/str/list)，返回为one_hot标签
    - 5.3 `utils.process_classify(line, name_dict)`: line:文本行，返回为(图像路径，图像标签)
    - 5.4 `TFDataFlow.imread(img_dir, size)`: img_dir:Tensor,返回向量图像
- 6. 查看模型输入输出
```bash
saved_model_cli show --dir [saved_model_path] --all
```