---
tags: NTU-IoT-Cow, tutorial
---
# FSL Training Code
[](https://github.com/WesleyCh3n/FSL)
[](https://hackmd.io/Y-MdlpaZRBKtj-nBfWyyJw)
## Overview
In `train_*.py`, it mainly contain the following parts:
- Dataset: `tf.data.Dataset`
- Model: `tf.keras.models.Model`
- Model Structure
- Training loop
- Loss
- Optirmizer: `tf.keras.optimizers`
- Callbacks
- Tensorboard: `tf.keras.callbacks.TensorBoard`
- Model checkpoint: `tf.keras.callbacks.ModelCheckpoint`
- Early stopping: `tf.keras.callbacks.EarlyStopping`
---
## Dataset
### `./model/input_fn.py`
- **def dataset_pipeline(is_training: bool, batch=True, \*\*params)**
```python=
"""
Read and preprocess images then return dataset format and its length.
Data directory which should contain sub-directory for each class. Such as:
'''
.
├── data_dir/
│ ├── Class_1/
│ ├── Class_2/
...
│ └── Class_N/
'''
Args:
is_training (bool): if dataset is going to be trained or not
batch (bool): return batched dataset or not
params: keyword arguments (parameters dictionary)
Return:
ds (tf.data.Dataset): tensorflow dataset format
ds_count (int): dataset length
"""
```
- **def dataset_pipeline_balance_label(is_training: bool, batch=True, \*\*params)**
```python=
"""
Read and preprocess images then return dataset format and its length. Also,
Balance each class number per batch.
Data directory which should contain sub-directory for each class. Such as:
'''
.
├── data_dir/
│ ├── Class_1/
│ ├── Class_2/
...
│ └── Class_N/
'''
Args:
is_training (bool): if dataset is going to be trained or not
batch (bool): return batched dataset or not
params: keyword arguments (parameters dictionary)
Return:
dataset (tf.data.Dataset): tensorflow dataset format
total_num (int): dataset length
"""
```
## Model
### `./model/softmax_model_fn.py`
- **class CustomModel(tf.keras.Model)**
```python=
"""
Inherited from `tf.keras.Model`.
Custom training step, test step, metrics.
self.compiled_loss is SparseCategoricalCrossentropy.
metrics include {train loss, val loss, train acc, val acc}
"""
```
- **def model_fn(is_training=True, \*\*params)**
```python=
"""
Create base model with MobileNetV2 + Dense layer (n class).
Wrap up with CustomModel process.
Args:
is_training (bool): if it is going to be trained or not
params: keyword arguments (parameters dictionary)
"""
```
### `./model/triplet_model_fn.py`
- **class CustomModel(tf.keras.Model)**
```python=
"""
Inherited from `tf.keras.Model`.
Custom training step, metrics.
self.compiled_loss is triplet loss.
metrics include {loss,
hardest positive distance (HPD),
hardest negative distance (HND)}
"""
```
- **def model_fn(is_training=True, \*\*params)**
```python=
"""
Create feature extractor model with MobileNetV2 + Dense layer (128).
Wrap up with CustomModel process.
Args:
is_training (bool): if it is going to be trained or not
params: keyword arguments (parameters dictionary)
"""
```
- **def transfer_model_fn(is_training=True, \*\*params)**
```python=
"""
Load MobileNetV2 + Dense layer (n class) weight, replace
Dense layer (n class) with Dense layer (128).
Wrap up with CustomModel process. Then train with triplet loss.
Args:
is_training (bool): if it is going to be trained or not
params: keyword arguments (parameters dictionary)
"""
```
- **def fine_tune_model_fn(is_training=True, \*\*params)**
```python=
"""
Load MobileNetV2 + Dense layer (128) weight, freeze weight to
block_16_depthwise_relu layer, only train last 2 layer
Wrap up with CustomModel process. Then train with triplet loss.
Args:
is_training (bool): if it is going to be trained or not
params: keyword arguments (parameters dictionary)
"""
```
### `./model/triplet_loss.py`
- **def batch_hard_triplet_loss(labels, embeddings, margin, squared=False)**
```python=
"""Build the triplet loss over a batch of embeddings.
For each anchor, we get the hardest positive and hardest negative to form a triplet.
Args:
labels: labels of the batch, of size (batch_size,)
embeddings: tensor of shape (batch_size, embed_dim)
margin: margin for triplet loss
squared: Boolean. If true, output is the pairwise squared euclidean distance matrix.
If false, output is the pairwise euclidean distance matrix.
Returns:
triplet_loss: scalar tensor containing the triplet loss
"""
```
## Callbacks
- `tf.keras.callbacks.TensorBoard`: Record metrics
- `tf.keras.callbacks.ModelCheckpoint`: Save model checkpoint
- `tf.keras.callbacks.EarlyStopping`: Early stop training
Sign in with Wallet
Connect another wallet