os.environ["CUDA_VISIBLE_DEVICES"] = args.device
learning_rate = args.lr
train_data = cifargen(batchsize)
valid_data = cifargen(batchsize, train=False)
h = train_data.x.shape[1]
w = train_data.x.shape[2]
if train_data.x.ndim > 3:
c = train_data.x.shape[3]
else:
c = 1
archi = Classifier(brickname='lenet', dropouts=[0.5, 0.5], fcdropouts=[0.5])
clf = CLF(archi, height=h, width=w, colors=c, learning_rate=learning_rate, optimizer='SGD')
print(clf)
trainlvals = []
trainaccvals = []
validlvals = []
validaccvals = []
accplot = []
for e in range(epochs):
print('EPOCH: ' + str(e + 1) + str('/') + str(epochs))
valid_data = cifargen(batchsize, train=False)
print('Y shape: ', train_data.y.shape)
h = train_data.x.shape[1]
w = train_data.x.shape[2]
if train_data.x.ndim > 3:
c = train_data.x.shape[3]
else:
c = 1
epochs = 1
archi = Classifier(brickname='reference')
clf = CLF(archi, height=h, width=w, colors=c, learning_rate=0.001)
print(clf)
trainlvals = []
trainaccvals = []
validlvals = []
validaccvals = []
for e in range(epochs):
print('EPOCH: ' + str(e + 1) + str('/') + str(epochs))
print('TRAIN')
w = 125
c = 3
archi = Classifier(brickname='reference',
filters=[32, 64, 128],
kernels=[4, 5, 6],
strides=[1, 1, 1],
dropouts=[0.1, 0.2, 0.25],
fc=[1024, 1024],
fcdropouts=[0.5, 0.5],
conv_activations=['relu', 'relu', 'relu'],
fc_activations=['relu', 'relu'],
end_activation='softmax',
output_channels=3)
clf = CLF(archi, height=h, width=w, colors=c, n_classes=3, learning_rate=lr, optimizer=opt)
print(clf)
trainlvals = []
trainaccvals = []
validlvals = []
validaccvals = []
accuracyplot = []
for e in range(epochs):
print('EPOCH: ' + str(e + 1) + str('/') + str(epochs))
dirname = os.path.join(logfolder, 'dependency' + str(n))
# first of all, evaluate reference alone
if n == 0:
archidep = Classifier(brickname='dependency' + str(n),
dropouts=[0.5, 0.5],
fcdropouts=[0.5])
archiref = Classifier(brickname='reference',
dropouts=[0.5, 0.5],
fcdropouts=[0.5])
clf = CLF(archiref,
height=h,
width=w,
colors=c,
learning_rate=0.0001,
model_path=os.path.join(refnetwork_folder, 'model.ckpt'))
print('SCORE REFERENCE')
validprogressbar = tqdm(valid_data, total=valid_data.steps)
accvals = []
for x, y in validprogressbar:
_, accval = clf.validate(x, y)
accvals.append(accval)
metrics = [('acc', numpy.mean(accvals))]
desc = monitor(metrics, 4)
refarchi = Classifier(brickname='reference',
filters=[32, 64, 128],
kernels=[4, 5, 6],
strides=[1, 1, 1],
dropouts=[0.1, 0.2, 0.25],
fc=[1024, 1024],
fcdropouts=[0.5, 0.5],
conv_activations=['relu', 'relu', 'relu'],
fc_activations=['relu', 'relu'],
end_activation='softmax',
output_channels=3)
clf = CLF(refarchi,
height=h,
width=w,
colors=c,
n_classes=3,
learning_rate=0.001,
model_path=ref_path,
optimizer="SGD")
outputdir = os.path.join(experimentfolder, 'slides_prediction')
if not os.path.isdir(outputdir):
os.makedirs(outputdir)
# predict slides
predict_slides(clf, slidedir, outputdir)
clf.close()
c = 1
archi = VGG(brickname='kerasCNN',
filters=[32, 64],
kernels=[3, 3],
strides=[1, 1],
dropouts=[0.25, 0.25],
thicknesses=[2, 2, 2],
fc=[512],
fcdropouts=[0.5],
conv_activations=['relu', 'relu'],
fc_activations=['relu'],
end_activation='softmax',
output_channels=10)
clf = CLF(archi, height=h, width=w, colors=c, learning_rate=0.01, optimizer='SGD', model_path=modelpath)
print(clf)
trainlvals = []
trainaccvals = []
validlvals = []
validaccvals = []
accplot = []
for e in range(epochs):
print('EPOCH: ' + str(e + 1) + str('/') + str(epochs))
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = args.device
# load dataset (CIFAR10 here)
# ---------------------------
xtrain, ytrain, xtest, ytest = cifar10.load_data()
ytrain = to_categorical(ytrain, 10)
ytest = to_categorical(ytest, 10)
# load trained classifier (reference)
# -----------------------------------
# first, build architecture
refarchi = Classifier(brickname='lenet', dropouts=[0.5, 0.5], fcdropouts=[0.5])
# then, create the model
refclf = CLF(refarchi, model_path=basenet)
# predict with classifier
predref = refclf.predict(xtest)
# load trained classifier (dependency)
# ------------------------------------
# first, build architecture
deparchi = Classifier(brickname='explorer',
dropouts=[0.5, 0.5],
fcdropouts=[0.5])
# then, create the model
depclf = CLF(deparchi, model_path=depnet)
# predict with classifier
preddep = depclf.predict(xtest)
predrefclass = numpy.argmax(predref, axis=1)