NetFrame is a flexible and scalable deep-learning framework to build segmentation model on large scale pathological images.

---

Environment

python2.7

pip install -r requirements.txt

Configuration

NetFame uses a configuration file of yaml to set up all the parameters.

Argument	Comment	Example
random_mirror	Randomly mirrors the image	True
color_jitter	Randomly distorts color	True
random_blur	Randomly blurs the image	True
random_scale	Randomly scales the images between 0.75 to 1.25 times of the original size.	True
num_classes	Number of classes	2
learning_rate	Initial learning rate	0.003
lr_decay_step	The number of steps in a learning rate decay cycle	15000
lr_decay_ratio	Learning rate decay ratio	0.5
batch_size	Batch size	32
resnet_layer	Number of layers of the ResNet backbone	50
input_size	Input size	320
patch_size	Patch size	320
l2_loss_lambda	Factor for weight decay loss	!!float 1e-5
restore_iters	Restored model corresponding to the iterations	0
log_label	Logging the metrics of the label	1
save_step	The number of steps model is saved in	2000
max_epoch	The maxmum training epoch	10
gpu	Specified the list of gpus to use	[0,1]

Build a New Model

Training

1. Create a configuration file setting all the parameters used for training;

2. Implemente the data and model module under the project directory by inheriting the default parent class;

3. Run the training program, eg:

   python main.py \
       --config_file ./config/stomach/config_stomach_v0.yaml \
       --version v0_0 \
       --mode train \
   

Inference

1. Change the value of argument mode from train to test;

2. Specify the number of iterations of the model to be tested;

3. Run the test program, eg:

   python main.py \
       --config_file ./config/stomach/config_stomach_v0.yaml \
       --version v0_0 \
       --mode test \
       --restore_iters 50000