classifier_trains 1.2.1

A PyTorch-based deep learning classifier training framework.

Classifier-trains

Install as a package

pip3 install classifier-trains

Examples of Pipeline Configuration

For example of pipeline configuration, please see pipeline_config_only_train.yml, pipeline_config_only_eval.yml, full_pipeline_config.yml.

# Run training or evaluation
python3 -m classifier_trains run --config <path to yml config> --output_dir <output_dir>

# Run training or evaluation with profiling, which will generate a profile report
python3 -m classifier_trains profile --config <path to yml config> --output_dir <output_dir>

# Compute mean and std of dataset
python3 -m classifier_trains compute-mean-and-std --dir-path <path to dataset>

# Get output mapping of dataset
python3 -m classifier_trains get-output-mapping --dir-path <path to dataset>

Expected Folder Structure for Dataset

Dataset Directory
├── train
│   ├── <class1>
│   │   ├── <image1>
│   │   └── ...
│   └── <class2>
│       ├── <image1>
│       └── ...
├── val
│   ├── <class1>
│   │   ├── <image1>
│   │   └── ...
│   └── <class2>
│       ├── <image1>
│       └── ...
└── eval
    ├── <class1>
    │   ├── <image1>
    │   └── ...
    └── <class2>
        ├── <image1>
        └── ...

Pipeline Parameter Table

Pipeline Parameters

Parameter	Description	Type	Default	Choices
enable_training	Enable training	bool	False	True, False
enable_evaluation	Enable evaluation	bool	False	True, False

Model Parameters

Parameter	Description	Type	Default	Choices
model	Model architecture to use	str	/	resnet18, resnet34, resnet50, resnet152, vgg11, vgg11_bn, vgg13, vgg13_bn, vgg16, vgg16_bn, vgg19, vgg19_bn, squeezenet1_0, squeezenet1_1, densenet121, densenet161, densenet169, densenet201, inception_v3
num_classes	Number of classes	int	/	Any positive integer
weights	Pretrained weights	str	DEFAULT	Check PyTorch Documentation
checkpoint_path	Path to checkpoint	Optional[str]	None	Self defined path

Dataloader Parameters

Parameter	Description	Type	Default	Choices
batch_size	Batch size	int	/	Any positive integer
num_workers	Number of workers	int	0	Any non negative integer

Resize Parameters

Parameter	Description	Type	Default	Choices
width	Width of resized image	int	/	Any positive integer
height	Height of resized image	int	/	Any positive integer
interpolation	Interpolation method	str	bicubic	nearest, nearest_exact, bilinear, bicubic
padding	Padding	Optional[str]	None	top_left, top_right, bottom_left, bottom_right, center
maintain_aspect_ratio	Maintain aspect ratio	bool	False	True, False

Spatial Transform Parameters

Parameter	Description	Type	Default	Choices
hflip_prob	Horizontal flip probability	float	/	Any float between 0 and 1
vflip_prob	Vertical flip probability	float	/	Any float between 0 and 1
max_rotate_in_degree	Maximum rotation in degree	float	0.0	Any non negative float
allow_center_crop	Allow center crop	bool	False	True, False
allow_random_crop	Allow random crop	bool	False	True, False

Color Transform Parameters

Parameter	Description	Type	Default	Choices
allow_gray_scale	Allow gray scale	bool	False	True, False
allow_random_color	Allow random color	bool	False	True, False

Preprocessing Parameters

Parameter	Description	Type	Default	Choices
mean	Mean of dataset	List[float]	/	Any list of floats
std	Standard deviation of dataset	List[float]	/	Any list of floats

Optimizer Parameters

Parameter	Description	Type	Default	Choices
name	Optimizer to use	str	/	sgd, adam, rmsprop, adamw
lr	Learning rate	float	/	Any positive float
weight_decay	Weight decay	float	0.0	Any non negative float
momentum	Momentum	Optional[float]	None	Any non negative float
alpha	Alpha	Optional[float]	None	Any non negative float
betas	Betas	Optional[Tuple[float, float]]	None	Any tuple of non negative floats

Scheduler Parameters

Parameter	Description	Type	Default	Choices
name	Scheduler to use	str	/	step, consine
lr_min	Minimum learning rate	Optional[float]	None	Any non negative float
step_size	Step size	Optional[int]	None	Any positive integer
gamma	Gamma	Optional[float]	None	Any non negative float

Training Parameters

Parameter	Description	Type	Default	Choices
name	Name of the experiment	str	/	Any string
num_epochs	Number of epochs	int	/	Any positive integer
trainset_dir	Path to training dataset	str	/	Any string
valset_dir	Path to validation dataset	str	/	Any string
testset_dir	Path to testing dataset	Optional[str]	None	Any string
device	Device to use	str	cuda	cpu, cuda, etc.
max_num_hrs	Maximum number of hours to run	Optional[float]	None	Any non negative float
criterion	What criterion to use for best model	str	loss	loss, accuracy
validate_every	Validate every n epochs	int	1	Any positive integer
save_every	Save model every n epochs	int	3	Any positive integer
patience	Patience for early stopping	int	5	Any positive integer
random_seed	Random seed	int	42	Any positive integer
precision	Precision	int	64	16, 32, 64
export_last_as_onnx	Export last model as ONNX	bool	False	True, False
export_best_as_onnx	Export best model as ONNX	bool	False	True, False

Evaluation Parameters

Parameter	Description	Type	Default	Choices
name	Name of the experiment	str	/	Any string
evalset_dir	Path to evaluation dataset	str	/	Any string
device	Device to use	str	cuda	cpu, cuda, etc.
precision	Precision	int	64	16, 32, 64
random_seed	Random seed	int	42	Any positive integer