An implementation of Equivariant e2 convolutional kernals into a convolutional self attention network, applied to radio astronomy data. This branch include the development of the Fisher Analysis of the network.
This work extends and builds on previous work:
- FR Classification using e2cnn, whose code is here.
- Attention for FR Classificaion, whose code is here.
This work makes extensive use of the
e2cnnpackage: e2cnn.
import os
import torch
import e2cnn
import pandas as pd
import numpy as np
import configparser as ConfigParser
import utils
To load a network import the network files, read in a config file, and load the network with the config file parameters:
from networks import AGRadGalNet, VanillaLeNet, testNet, DNSteerableLeNet, DNSteerableAGRadGalNet
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
config = ConfigParser.ConfigParser(allow_no_value=True)
config_path = "configs/CONFIG_NAME"
config.read(config_path)
net = locals()[config['model']['base']](**config['model']).to(device)
Alternatively, if you know which architecture you want to use, only import that and load in the config parameters (i.e. without using locals()):
from networks import DNSteerableAGRadGalNet
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
config = ConfigParser.ConfigParser(allow_no_value=True)
config_path = "configs/CONFIG_NAME"
config.read(config_path)
net = DNSteerableAGRadGalNet(**config['model']).to(device)
To load the best performing model (according to early stopping) from a given model training session:
path_supliment = config['data']['augment']+'/'
model = utils.utils.load_model(config, load_model='best', device=device, path_supliment=path_supliment)
Follow the example of evaluation.ipynb for model evaluation.
To define your own model simply create a new config which follows the examples provided in configs.
To train a model use the train.py using your config file:
python3.8 train.py --config YOUR_CONFIG_NAME.cfg >& logs/YOUR_CONFIG_NAME.log
To copy weights over of all layers from two networks, and freeze apart from the last layer use:
utils.fisher.WeightTransfer(referenceModel, newModel)
To Initialise the Process of Calculating the FIM and Related Metrics (for Emperical Fisher use approximation=emperical)
Fishers, Rank, FR = utils.fisher.CalcFIM(model, train_loader, n_iterations, approximation=None)