def test_rescale_01(self):
X = np.random.rand(10, 3, 3, 3)
Xr = emlib.rescale_01(X, perChannel=False)
self.assertTrue(np.max(Xr) <= 1.0)
self.assertTrue(np.min(Xr) >= 0.0)
Xr = emlib.rescale_01(X, perChannel=True)
self.assertTrue(np.max(Xr) <= 1.0)
self.assertTrue(np.min(Xr) >= 0.0)
X = np.random.rand(10, 3, 100, 100)
X[:, 0, ...] *= 1000.0
Xr = emlib.rescale_01(X, perChannel=False)
self.assertTrue(np.max(Xr[:, 1, ...]) <= .1)
Xr = emlib.rescale_01(X, perChannel=True)
self.assertTrue(np.max(Xr[:, 1, ...]) > .8)
def deploy_model(X, weightsFile,
log=None,
slices=[],
modelName='ciresan_n3',
evalPct=1.0,
outFile=None):
""" Applies a previously trained CNN to new data.
Xtrain : Tensor of features with dimensions as specified in (1)
trainSlices : A list of slice indices to evalute
(or [] to use all the data)
log : a logging object (for reporting status)
outFile : File name where class probabilities should be stored
evalPct : Fraction of volume to evalute; \in [0,1]
"""
# Setup output file/dirctory
if not outFile:
if log: log.warning('No output file specified - are you sure this is what you want?')
elif not os.path.exists(os.path.dirname(outFile)):
os.makedirs(os.path.dirname(outFile))
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# preprocess data
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
if slices:
X = X[slices,:,:]
# rescale features to live in [0, 1]
X = emlib.rescale_01(X, perChannel=True)
if log: log.info('X volume dimensions: %s' % str(X.shape))
if log: log.info('X values min/max: %g, %g' % (np.min(X), np.max(X)))
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# initialize CNN
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
if log: log.info('initializing CNN...')
model = getattr(emm, modelName)()
model.compile(optimizer='sgd', # not used, but required by keras
loss='categorical_crossentropy')
model.load_weights(weightsFile)
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# Do it
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
if log: log.info('evaluating volume...')
Prob = _evaluate(model, X, log=log, evalPct=evalPct)
if log: log.info('Complete!')
if outFile:
np.save(outFile, Prob)
scipy.io.savemat(outFile.replace('.npy', '.mat'), {'P' : Prob})
if log: log.info('Probabilites stored in file %s' % outFile)
return Prob
def train_model(Xtrain, Ytrain,
Xvalid, Yvalid,
trainSlices=[],
validSlices=[],
omitLabels=[],
modelName='ciresan_n3',
learnRate0=0.01,
weightDecay=1e-6,
momentum=0.9,
maxMbPerEpoch=sys.maxint,
nEpochs=30,
log=None,
outDir=None):
"""Trains a CNN using Keras.
Some of the key parameters include:
Xtrain, Ytrain : Tensors of features and per-pixel class labels with
dimensions as specified in (1),(2)
Xvalid, Yalid : Tensors of features and per-pixel class labels with
dimensions as specified in (1),(2). Presumed to
be held-out data (i.e. disjoint from X/Ytrain)
trainSlices : A list of slice indices to include in training
(or [] to use all the data)
validSlices : A list of slice indices to include in validation
(or [] to use all the data)
omitLabels : A list of class labels whose corresponding pixel data
should be omitted from train and test. If [], uses
all data.
maxMbPerEpoch : The maximum number of minibatches to run in each
epoch (default is to process entire data volume
each epoch).
log : a logging object (for reporting status)
outDir : if not None, a directory where model weights
will be stored (highly recommended)
"""
if not outDir:
if log: log.warning('No output directory specified - are you sure this is what you want?')
elif not os.path.exists(outDir):
os.makedirs(outDir)
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# preprocess data
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# discard unneeded slices (if any)
if trainSlices:
Xtrain = Xtrain[trainSlices,:,:,:]
Ytrain = Ytrain[trainSlices,:,:]
if validSlices:
Xvalid = Xvalid[validSlices,:,:,:]
Yvalid = Yvalid[validSlices,:,:]
# rescale features to live in [0 1]
# XXX: technically, should probably use scale factors from
# train volume on validation data...
Xtrain = emlib.rescale_01(Xtrain, perChannel=True)
Xvalid = emlib.rescale_01(Xvalid, perChannel=True)
# Remap class labels to consecutive natural numbers.
# Note that any pixels that should be omitted from the
# analysis are mapped to -1 by this function.
Ytrain = emlib.number_classes(Ytrain, omitLabels)
Yvalid = emlib.number_classes(Yvalid, omitLabels)
if log:
log.info('training volume dimensions: %s' % str(Xtrain.shape))
log.info('training values min/max: %g, %g' % (np.min(Xtrain), np.max(Xtrain)))
log.info('training class labels: %s' % str(np.unique(Ytrain)))
for yi in np.unique(Ytrain):
cnt = np.sum(Ytrain == yi)
log.info(' class %d has %d instances' % (yi, cnt))
log.info('')
log.info('validation volume dimensions: %s' % str(Xvalid.shape))
log.info('validation values min/max: %g, %g' % (np.min(Xvalid), np.max(Xvalid)))
log.info('validation class labels: %s' % str(np.unique(Yvalid)))
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# create and configure CNN
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
if log: log.info('creating CNN')
model = getattr(emm, modelName)()
sgd = SGD(lr=learnRate0, decay=weightDecay, momentum=momentum, nesterov=True)
model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# Do training
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
for epoch in range(nEpochs):
if log: log.info('starting training epoch %d (of %d)' % (epoch, nEpochs));
acc, loss = _train_one_epoch(model, Xtrain, Ytrain,
log=log,
omitLabels=[-1,],
nBatches=maxMbPerEpoch)
if outDir:
# save a snapshot of current model weights
weightFile = os.path.join(outDir, "weights_epoch_%03d.h5" % epoch)
if os.path.exists(weightFile):
os.remove(weightFile)
model.save_weights(weightFile)
# also save accuracies (for diagnostic purposes)
accFile = os.path.join(outDir, 'acc_epoch_%03d.npy' % epoch)
np.save(accFile, acc)
# Evaluate performance on validation data.
if log: log.info('epoch %d complete. validating...' % epoch)
Prob, acc = _evaluate(model, Xvalid, Yvalid, omitLabels=[-1,], log=log)
if log: log.info('accuracy on validation data: %0.2f%%' % acc)
if outDir:
estFile = os.path.join(outDir, "validation_epoch_%03d.npy" % epoch)
np.save(estFile, Prob)
if log: log.info('Finished!')
return model
def deploy_model(X,
weightsFile,
log=None,
slices=[],
modelName='ciresan_n3',
evalPct=1.0,
outFile=None):
""" Applies a previously trained CNN to new data.
Xtrain : Tensor of features with dimensions as specified in (1)
trainSlices : A list of slice indices to evalute
(or [] to use all the data)
log : a logging object (for reporting status)
outFile : File name where class probabilities should be stored
evalPct : Fraction of volume to evalute; \in [0,1]
"""
# Setup output file/dirctory
if not outFile:
if log:
log.warning(
'No output file specified - are you sure this is what you want?'
)
elif not os.path.exists(os.path.dirname(outFile)):
os.makedirs(os.path.dirname(outFile))
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# preprocess data
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
if slices:
X = X[slices, :, :]
# rescale features to live in [0, 1]
X = emlib.rescale_01(X, perChannel=True)
if log: log.info('X volume dimensions: %s' % str(X.shape))
if log: log.info('X values min/max: %g, %g' % (np.min(X), np.max(X)))
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# initialize CNN
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
if log: log.info('initializing CNN...')
model = getattr(emm, modelName)()
model.compile(
optimizer='sgd', # not used, but required by keras
loss='categorical_crossentropy')
model.load_weights(weightsFile)
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# Do it
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
if log: log.info('evaluating volume...')
Prob = _evaluate(model, X, log=log, evalPct=evalPct)
if log: log.info('Complete!')
if outFile:
np.save(outFile, Prob)
scipy.io.savemat(outFile.replace('.npy', '.mat'), {'P': Prob})
if log: log.info('Probabilites stored in file %s' % outFile)
return Prob
