def main(argv):
global symbName
path = argv[0]
symbName = argv[1]
#optional parameter
folderPath = None
E = None
if len(argv) >= 3:
folderPath = argv[2]
if len(argv) >= 4:
E = argv[3]
try:
os.makedirs(savePath + symbName)
except:
if os.path.isfile(generateBettiSavePath()):
print("Betti data already exists for: " + generateBettiSavePath())
return
print('Obtaining adjacency matrix...')
#weights = loadWeights(path,['dense_1'])
weights = loadWeightsCustom(folderPath, E)
matrix = getWeightedAdjacencyMatrixNoBias(weights)
print('Transforming weight matrix into distance matrix...')
matrix = makeWeightAbsoluteDistance(matrix)
#np.save(savePath + symbName + "/" + 'originalMatrix.npy', matrix)
matrix = removeZeros(matrix)
print('Running Floyd-Warshall (Transform to Shortest Distance Matrix)...')
matrix = floyd_warshall_fastest(matrix)
#np.save(savePath + symbName + "/" + 'shortMatrix.npy', matrix)
print('Running VR Filtration...')
runVRFiltration(matrix)
return
def runPipeline(epoch):
print('Loading Weights...')
weights = wl.simpleLoader(nnSavePath(epoch))
print('Generating Adjacency Matrix...')
adjacencyMatrix = adj.getWeightedAdjacencyMatrixNoBias(weights)
print('Preprocessing')
processedMatrix = pp.standardVR(adjacencyMatrix)
filtration.VR(vrSavePath(epoch), processedMatrix)
print(f"Finished epoch: {epoch}")
def runFloydReplacement(config):
import NNGeneration.nnUtil as utils
import NNGeneration.NNDigits as digits
import weightLoader as wl
import generateAdjacencyMatrix as adj
import preprocessing as pp
from NNGeneration.DigitsPositive import base_model, get_dataset
#model = digits.emptyModel(config)
#xTrain,yTrain,xTest,yTest = utils.mnistTrainTest()
model = base_model()
xTrain, xTest, yTrain, yTest = get_dataset()
trainAccs = []
testAccs = []
for epoch in config["epochs"]:
weights = wl.simpleLoader(config["nnSaveFn"](epoch))
adjacencyMatrix = adj.getWeightedAdjacencyMatrixNoBias(weights.copy())
processedMatrix = pp.standardVR(adjacencyMatrix)
model = utils.replaceModelParams(model, epoch, processedMatrix, config,
weights)
train = utils.getAccuracy(model, xTrain, yTrain)
test = utils.getAccuracy(model, xTest, yTest)
trainAccs.append(train)
testAccs.append(test)
print(f"Finished epoch: {epoch}. Train: {train}, Test: {test}")
path = config["analysisSaveFn"] + "replacement/"
if not os.path.isdir(path):
os.mkdir(path)
with open(path + "acc.txt", "w+") as f:
for i in range(len(trainAccs)):
f.write(f"{i},\t{trainAccs[i]},\t{testAccs[i]}")
return
def main(argv):
global symbName
path = argv[0]
symbName = argv[1]
#optional parameter
folderPath = None
E = None
if len(argv) >= 3:
folderPath = argv[2]
if len(argv) >= 4:
E = argv[3]
print('Epoch', E)
print('Obtaining adjacency matrix...')
#weights = loadWeights(path,['dense_1'])
weights = loadWeightsCustom(folderPath, E)
matrix = getWeightedAdjacencyMatrixNoBias(weights)
print('Transforming weight matrix into distance matrix...')
matrix = adjacencyToDistance(matrix)
M1 = matrix
print('Running Floyd-Warshall (Transform to Shortest Distance Matrix)...')
matrix = floyd_warshall_fastest(matrix)
#np.save(savePath + symbName + "/" + 'shortMatrix.npy', matrix)
print(calculatePercentStatic(M1,matrix))
return
return ret
def symbName(e):
return "CIFAR-10_" + str(e)
def pathName(e):
return constant + "MODEL_Epoch" + str(e)
filepath = "NNGeneration/Saved Models/CIFAR-10-AML/model-{epoch:02d}.hdf5"
"NNGeneration/Saved Models/CIFAR-10-Variation2/MODEL_Epoch9_W7.npy"
W = loadweights.loadWeights(filepath.format(epoch=str(10)), ['dense_1'])
M1 = GAM.getWeightedAdjacencyMatrixNoBias(W)
M2 = makeAllGraphs.main(["", "0", "-w", "-nb", None, path + fn + str(e)])
print(np.allclose(M1, M2, atol=tol))
def check_symmetric(a, tol=1e-8):
return
assert False
epochs = 125
matrices = []
for e in range(epochs + 1):
m = makeAllGraphs.main(["", "0", "-w", "-nb", None, path + fn + str(e)])
matrices.append(m)