import h5py
import datetime
# Import and preprocess data
if __name__ == "__main__":
h5f = h5py.File("./data/infarction-healthy.h5", "r")
h5f_test = h5py.File("./data/infarction-healthy-test.h5", "r")
inData = h5f["inData"]
inLabelsOH = h5f["inLabels"]
inData_test = h5f_test["inData"]
inLabelsOH_test = h5f_test["inLabels"]
# Neural net (two-channel)
sess = tf.InteractiveSession()
model = getCNN(2) # 2 classes: healthy, infarcted
# Train the model, leaving out the kfold not being used
model.fit(inData, inLabelsOH, batch_size=100, n_epoch=20, show_metric=True)
dt = str(datetime.datetime.now().replace(second=0,
microsecond=0).isoformat("_"))
model.save("./models/" + dt + "_3d-2channel-fakedata_infarction.tflearn")
# Get sensitivity and specificity
illTest = []
healthTest = []
inLabels_test = inLabelsOH_test[:, 1]
for index, item in enumerate(inLabels_test):
if item == 1:
illTest.append(inData_test[index])
if item == 0:
i = int(sys.argv[1]) # i is current kfold
k = 5 # k folds
inData = np.load("./data/shufData.npy")
inLabels = np.load("./data/shufLab.npy")
inLabelsOH = np.eye(2)[inLabels.astype(int)] # One hot encode
# k fold the data
kfoldData = np.array_split(inData, k)
kfoldLabels = np.array_split(inLabels, k)
kfoldLabelsOH = np.array_split(inLabelsOH, k)
# Neural net (two-channel)
sess = tf.InteractiveSession()
model = getCNN(2, finetune=True)
# Train the model, leaving out the kfold not being used
dummyData = np.reshape(
np.concatenate(kfoldData[:i] + kfoldData[i + 1:], axis=0),
[-1, 34, 34, 34, 2])
dummyLabels = np.reshape(
np.concatenate(kfoldLabelsOH[:i] + kfoldLabelsOH[i + 1:], axis=0),
[-1, 2])
model.load("./models/placeholder")
model.fit(dummyData,
dummyLabels,
batch_size=5,
n_epoch=150,
show_metric=True) # In practice learning stops ~150 epochs.
dt = str(datetime.datetime.now().replace(second=0,
import h5py
import datetime
# Import and preprocess data
if __name__ == "__main__":
h5f = h5py.File("./data/mixed-healthy.h5", "r")
h5f_test = h5py.File("./data/mixed-health-test.h5", "r")
inData = h5f["inData"]
inLabelsOH = h5f["inLabels"]
inData_test = h5f_test["inData"]
inLabelsOH_test = h5f_test["inLabels"]
# Neural net (two-channel)
sess = tf.InteractiveSession()
model = getCNN(5) # 2 classes: healthy, mixed
# Train the model, leaving out the kfold not being used
model.fit(inData, inLabelsOH, batch_size=100, n_epoch=20, show_metric=True)
dt = str(datetime.datetime.now().replace(second=0, microsecond=0).isoformat("_"))
model.save("./models/"+dt+"_3d-2channel-fakedata_mixed.tflearn")
# Get sensitivity and specificity
illTest = []
healthTest = []
inLabels_test = inLabelsOH_test[:,1]
for index, item in enumerate(inLabels_test):
if item == 1:
illTest.append(inData_test[index])
if item == 0:
healthTest.append(inData_test[index])
# Initialise Horovod
if args.dist:
print("Initialising Horovod")
hvd.init()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(hvd.local_rank())
K.set_session(tf.Session(config=config))
print("Hvd current rank:", str(hvd.local_rank()))
print("Seed:", str(args.SEED))
print("Current kfold:", str(args.i), "of", str(args.k - 1))
# Neural net (two-channel)
model = getCNN(2) # 2 classes: healthy, ischaemia
if args.dist:
opt = keras.optimizers.Adam(lr=0.001 * hvd.size())
opt = hvd.DistributedOptimizer(opt)
else:
opt = keras.optimizers.Adam(lr=0.001)
model.compile(optimizer=opt,
loss='categorical_crossentropy',
metrics=['accuracy'])
# callbacks
cb = []
if args.dist:
cb.append(hvd.callbacks.BroadcastGlobalVariablesCallback(0))
# Horovod: average metrics among workers at the end of every epoch.
# Note: This callback must be in the list before the ReduceLROnPlateau,
inDataAbs = np.fabs(inData)
inDataMax = np.amax(inData)
normalisedData = inDataAbs / inDataMax
return normalisedData
if __name__ == "__main__":
# inData are heartcubes with same shape as those used in the CNN
# So far this has only been implemented for healthy/ill pairs. It should be easy enough to generalise it to n classes though.
# Generalisation to n classes would require grouping of healthy and ill diagnoses as in cnnAll.py, or a rethink of the current loss function.
ppt = 20
h5f = h5py.File("./data/twoThousand.h5", "r")
inData = h5f["inData"][ppt]
inData = inData[np.newaxis, ...]
model = getCNN(2)
model.load(
"./data/placeholderModel") # The model that we want to test goes here
inLabel = model.predict(inData)[:, 1]
maskWidth = 8 # Might be more representative to have this as even.
lossCube = np.zeros(inData.shape[1:4])
for j in np.arange(inData.shape[1] - maskWidth + 1):
for k in np.arange(inData.shape[2] - maskWidth + 1):
for l in np.arange(inData.shape[3] - maskWidth + 1):
loss = getLoss(inData, inLabel, maskWidth, j, k, l)
lossCube[j + maskWidth / 2, k + maskWidth / 2,
l + maskWidth / 2] = loss
print(j + maskWidth / 2, k + maskWidth / 2, l + maskWidth / 2,
import h5py
import datetime
# Import and preprocess data
if __name__ == "__main__":
h5f = h5py.File("./data/artefact-healthy.h5", "r")
h5f_test = h5py.File("./data/artefact-healthy-test.h5", "r")
inData = h5f["inData"]
inLabelsOH = h5f["inLabels"]
inData_test = h5f_test["inData"]
inLabelsOH_test = h5f_test["inLabels"]
# Neural net (two-channel)
sess = tf.InteractiveSession()
model = getCNN(2) # 2 classes: healthy, artefact
# Train the model, leaving out the kfold not being used
model.fit(inData, inLabelsOH, batch_size=100, n_epoch=20, show_metric=True)
dt = str(datetime.datetime.now().replace(second=0,
microsecond=0).isoformat("_"))
model.save("./models/" + dt + "_3d-2channel-fakedata_artefact.tflearn")
# Get sensitivity and specificity
illTest = []
healthTest = []
inLabels_test = inLabelsOH_test[:, 1]
for index, item in enumerate(inLabels_test):
if item == 1:
illTest.append(inData_test[index])
if item == 0:
return e_x / e_x.sum()
# Import and preprocess data
if __name__ == "__main__":
h5f = h5py.File("./data/all.h5", "r")
h5f_test = h5py.File("./data/allTest.h5", "r")
inData = h5f["inData"]
inLabelsOH = h5f["inLabels"]
inData_test = h5f_test["inData"]
inLabelsOH_test = h5f_test["inLabels"]
# Neural net (two-channel)
sess = tf.InteractiveSession()
model = getCNN(
5) # 5 classes: healthy, ischaemic, infarcted, mixed, artefacted
# Train the model
model.fit(
inData,
inLabelsOH,
batch_size=100,
n_epoch=30,
show_metric=True,
validation_set=0.1) # Need validation so I can see when learning stops
dt = str(datetime.datetime.now().replace(second=0,
microsecond=0).isoformat("_"))
model.save("./models/" + dt + "_3d-2channel-fakedata-all.tflearn")
# Retrieve indices of the different cubes and get the accuracies of each type
inLabels_test = np.argwhere(np.array(inLabelsOH_test) == 1)[:, 1]