def print_prediction(config):
cache_m = CacheManager()
if not os.path.isfile(cache_m.fileLocation('test_pred.pik')):
return None
#Load predictions
(expected, Y_pred, nclasses) = cache_m.load('test_pred.pik')
y_pred = np.argmax(Y_pred, axis=1)
#Output metrics
if nclasses > 2:
f1 = metrics.f1_score(expected, y_pred, average='weighted')
else:
f1 = metrics.f1_score(expected, y_pred, pos_label=1)
print("F1 score: {0:.2f}".format(f1))
m_conf = PrintConfusionMatrix(y_pred, expected, nclasses, config, "TILs")
#ROC AUC
#Get positive scores (binary only)
if nclasses == 2:
scores = Y_pred.transpose()[1]
fpr, tpr, thresholds = metrics.roc_curve(expected, scores, pos_label=1)
print("AUC: {0:f}".format(metrics.roc_auc_score(expected, scores)))
print("Accuracy: {0:.3f}".format(m_conf[nclasses + 2][nclasses]))
if config.verbose > 1:
print("False positive rates: {0}".format(fpr))
print("True positive rates: {0}".format(tpr))
print("Thresholds: {0}".format(thresholds))
class KNet(GenericModel):
"""
Implements abstract methods from GenericModel.
Model is the same as in: https://keras.io/examples/mnist_cnn/
"""
def __init__(self, config, ds, name=None):
super().__init__(config, ds, name=name)
if name is None:
self.name = "KerasNet"
self._modelCache = "{0}-model.h5".format(self.name)
self._weightsCache = "{0}-weights.h5".format(self.name)
self._mgpu_weightsCache = "{0}-mgpu-weights.h5".format(self.name)
self.cache_m = CacheManager()
self.cache_m.registerFile(
os.path.join(config.model_path, self._modelCache),
self._modelCache)
self.cache_m.registerFile(
os.path.join(config.weights_path, self._weightsCache),
self._weightsCache)
self.cache_m.registerFile(
os.path.join(config.weights_path, self._mgpu_weightsCache),
self._mgpu_weightsCache)
self.single = None
self.parallel = None
def get_model_cache(self):
"""
Returns path to model cache
"""
return self.cache_m.fileLocation(self._modelCache)
def get_weights_cache(self):
"""
Returns path to model cache
"""
return self.cache_m.fileLocation(self._weightsCache)
def get_mgpu_weights_cache(self):
"""
Returns path to model cache
"""
return self.cache_m.fileLocation(self._mgpu_weightsCache)
def build(self, **kwargs):
model, parallel_model = self._build(**kwargs)
self.single = model
self.parallel = parallel_model
return (model, parallel_model)
def _build(self, **kwargs):
"""
@param pre_trained <boolean>: returned model should be pre-trained or not
@param data_size <int>: size of the training dataset
"""
width, height, channels = self._check_input_shape()
if 'data_size' in kwargs:
self.data_size = kwargs['data_size']
if 'training' in kwargs:
training = kwargs['training']
else:
training = True
if 'feature' in kwargs:
feature = kwargs['feature']
else:
feature = False
if backend.image_data_format() == 'channels_first':
input_shape = (channels, height, width)
else:
input_shape = (height, width, channels)
self.cache_m = CacheManager()
model = self._build_architecture(input_shape, training, feature)
#Check if previous training and LR is saved, if so, use it
lr_cache = "{0}_learning_rate.txt".format(self.name)
self.cache_m.registerFile(os.path.join(self._config.cache, lr_cache),
lr_cache)
l_rate = 0.0005
if os.path.isfile(self.cache_m.fileLocation(
lr_cache)) and not self._config.new_net:
l_rate = float(self.cache_m.read(lr_cache))
if self._config.info:
print("Found previous learning rate: {0}".format(l_rate))
#opt = optimizers.SGD(lr=l_rate, decay=1.5e-4, momentum=0.9, nesterov=True)
#opt = optimizers.Adam(lr = l_rate)
opt = optimizers.Adadelta()
#Return parallel model if multiple GPUs are available
parallel_model = None
if self._config.gpu_count > 1:
with tf.device('/cpu:0'):
model.compile(loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy'])
parallel_model = multi_gpu_model(model,
gpus=self._config.gpu_count)
parallel_model.compile(
loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy'],
#options=p_opt,
#run_metadata=p_mtd
)
else:
model.compile(
loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy'],
#options=p_opt,
#run_metadata=p_mtd
)
return (model, parallel_model)
def _build_architecture(self, input_shape, training=None, feature=False):
model = Sequential()
model.add(
Convolution2D(32,
kernel_size=(3, 3),
activation='relu',
input_shape=input_shape))
model.add(Convolution2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(self._ds.nclasses, activation='softmax'))
return model
class Inception(GenericModel):
"""
Implements abstract methods from GenericModel.
Model is the same as in: https://github.com/keras-team/keras-applications/blob/master/keras_applications/inception_resnet_v2.py
Addapted to provide a Bayesian model
"""
def __init__(self, config, ds, name=None):
super().__init__(config, ds, name=name)
if name is None:
self.name = "Inception"
self._modelCache = "{0}-model.h5".format(self.name)
self._weightsCache = "{0}-weights.h5".format(self.name)
self._mgpu_weightsCache = "{0}-mgpu-weights.h5".format(self.name)
self.cache_m = CacheManager()
self.cache_m.registerFile(
os.path.join(config.model_path, self._modelCache),
self._modelCache)
self.cache_m.registerFile(
os.path.join(config.weights_path, self._weightsCache),
self._weightsCache)
self.cache_m.registerFile(
os.path.join(config.weights_path, self._mgpu_weightsCache),
self._mgpu_weightsCache)
self.single = None
self.parallel = None
def get_model_cache(self):
"""
Returns path to model cache
"""
return self.cache_m.fileLocation(self._modelCache)
def get_weights_cache(self):
"""
Returns path to model cache
"""
return self.cache_m.fileLocation(self._weightsCache)
def get_mgpu_weights_cache(self):
"""
Returns path to model cache
"""
return self.cache_m.fileLocation(self._mgpu_weightsCache)
def get_npweights_cache(self, add_ext=False):
"""
Returns path to model cache.
@param add_ext <boolean>: add numpy file extension to file name.
"""
if add_ext:
return "{}.npy".format(
self.cache_m.fileLocation(self._weightsCache).split('.')[0])
else:
return self.cache_m.fileLocation(self._weightsCache).split('.')[0]
def get_npmgpu_weights_cache(self, add_ext=False):
"""
Returns path to model cache
@param add_ext <boolean>: add numpy file extension to file name.
"""
if add_ext:
return "{}.npy".format(
self.cache_m.fileLocation(
self._mgpu_weightsCache).split('.')[0])
else:
return self.cache_m.fileLocation(
self._mgpu_weightsCache).split('.')[0]
def register_ensemble(self, m):
self._model_n = m
self._weightsCache = "{0}-EM{1}-weights.h5".format(self.name, m)
self._mgpu_weightsCache = "{0}-EM{1}-mgpu-weights.h5".format(
self.name, m)
self._modelCache = "{0}-EM{1}-model.h5".format(self.name, m)
self.cache_m.registerFile(
os.path.join(self._config.weights_path, self._weightsCache),
self._weightsCache)
self.cache_m.registerFile(
os.path.join(self._config.weights_path, self._mgpu_weightsCache),
self._mgpu_weightsCache)
self.cache_m.registerFile(
os.path.join(self._config.model_path, self._modelCache),
self._modelCache)
def return_model_n(self):
if hasattr(self, '_model_n'):
return self._model_n
else:
return -1
def build(self, **kwargs):
"""
@param pre_trained <boolean>: returned model should be pre-trained or not
@param data_size <int>: size of the training dataset
"""
model, parallel_model = self._build(**kwargs)
self.single = model
self.parallel = parallel_model
return (model, parallel_model)
def build_extractor(self, **kwargs):
"""
Builds a feature extractor.
Weights should be loaded by caller!
Key word arguments:
preload_w: return model with weights already loaded? True -> Yes
parallel: return parallel model (overrides gpu_count avaliation)? True -> Yes
"""
#Weight loading for the feature extraction is done latter by requesting party
kwargs['preload_w'] = False
if 'parallel' in kwargs and not kwargs['parallel']:
s, p = self._build(**kwargs)
return (s, None)
else:
return self._build(**kwargs)
def build_ensemble(self, **kwargs):
"""
Builds an ensemble of M Inception models.
Weights are loaded here because of the way ensembles should be built.
Default build: avareges the output of the corresponding softmaxes
@param npfile <boolean>: loads weights from numpy files
"""
if 'npfile' in kwargs:
npfile = kwargs['npfile']
else:
npfile = False
s_models = []
p_models = []
for m in range(self._config.emodels):
self.register_ensemble(m)
single, parallel = self._build(**kwargs)
if not parallel is None:
if npfile and hasattr(model, 'get_npmgpu_weights_cache'):
parallel.set_weights(
np.load(model.get_npmgpu_weights_cache(),
allow_pickle=True))
if self._config.info:
print("[Inception] loaded ensemble weights: {}".format(
model.get_npmgpu_weights_cache()))
elif os.path.isfile(model.get_mgpu_weights_cache()):
parallel.load_weights(model.get_mgpu_weights_cache(),
by_name=True)
if self._config.info:
print("[Inception] loaded ensemble weights: {}".format(
model.get_mgpu_weights_cache()))
else:
parallel = None
if npfile and hasattr(model, 'get_npweights_cache'):
single.set_weights(
np.load(model.get_weights_cache(), allow_pickle=True))
if self._config.info:
print("[Inception] loaded ensemble weights: {}".format(
model.get_npweights_cache()))
elif os.path.isfile(model.get_weights_cache()):
single.load_weights(model.get_weights_cache(), by_name=True)
else:
if self._config.info:
print(
"[Inception] Could not load ensemble weights (model {})"
.format(m))
single = None
s_models.append(single)
p_models.append(parallel)
s_inputs = [inp for s in s_models for inp in s.inputs]
s_outputs = [out for s in s_models for out in s.outputs]
p_models = list(filter(lambda x: not x is None, p_models))
if len(p_models) > 0:
p_inputs = [inp for p in p_models for inp in p.inputs]
p_outputs = [out for p in p_models for out in p.outputs]
else:
p_inputs = None
p_outputs = None
#Build the ensemble output from individual models
s_model, p_model = None, None
##Single GPU model
x = Average()(s_outputs)
s_model = Model(inputs=s_inputs, outputs=x)
##Parallel model
if not p_inputs is None:
x = Average()(p_outputs)
p_model = Model(inputs=p_inputs, outputs=x)
return s_model, p_model
def _build(self, **kwargs):
width, height, channels = self._check_input_shape()
if 'data_size' in kwargs:
self.data_size = kwargs['data_size']
if 'training' in kwargs:
training = kwargs['training']
else:
training = True
if 'feature' in kwargs:
feature = kwargs['feature']
else:
feature = False
if 'preload_w' in kwargs:
preload = kwargs['preload_w']
else:
preload = True
if 'allocated_gpus' in kwargs and not kwargs['allocated_gpus'] is None:
allocated_gpus = kwargs['allocated_gpus']
else:
allocated_gpus = self._config.gpu_count
if backend.image_data_format() == 'channels_first':
input_shape = (channels, height, width)
else:
input_shape = (height, width, channels)
self.cache_m = CacheManager()
model = self._build_architecture(input_shape, training, feature,
preload)
#Check if previous training and LR is saved, if so, use it
lr_cache = "{0}_learning_rate.txt".format(self.name)
self.cache_m.registerFile(os.path.join(self._config.cache, lr_cache),
lr_cache)
l_rate = 0.00005
if os.path.isfile(self.cache_m.fileLocation(
lr_cache)) and not self._config.new_net:
l_rate = float(self.cache_m.read(lr_cache))
if self._config.info:
print("Found previous learning rate: {0}".format(l_rate))
#opt = optimizers.SGD(lr=l_rate, decay=1.5e-4, momentum=0.9, nesterov=True)
opt = optimizers.Adam(lr=l_rate)
#opt = optimizers.Adadelta(lr=l_rate)
#Return parallel model if multiple GPUs are available
parallel_model = None
if allocated_gpus > 1:
with tf.device('/cpu:0'):
model.compile(loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy'])
parallel_model = multi_gpu_model(model, gpus=allocated_gpus)
parallel_model.compile(
loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy'],
#options=p_opt,
#run_metadata=p_mtd
)
else:
model.compile(
loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy'],
#options=p_opt,
#run_metadata=p_mtd
)
return (model, parallel_model)
def _build_architecture(self,
input_shape,
training=None,
feature=False,
preload=True):
from . import inception_resnet_v2
kwargs = {
'training': training,
'feature': feature,
'custom_top': False,
'preload': preload,
'batch_n': True if self._config.gpu_count <= 1 else False
}
inp = Input(shape=input_shape)
inception_body = inception_resnet_v2.InceptionResNetV2(
include_top=False,
weights='imagenet',
input_tensor=inp,
input_shape=input_shape,
pooling='avg',
classes=self._ds.nclasses,
**kwargs)
return inception_body
class VGG16(GenericModel):
"""
Implements abstract methods from GenericModel.
Producess a VGG16 model as implemented by Keras, with convolutional layers
FC layers are substituted by Conv2D, as defined in:
https://github.com/ALSM-PhD/quip_classification/blob/master/NNFramework_TF/sa_networks/vgg.py
"""
def __init__(self, config, ds, name=None):
super().__init__(config, ds, name=name)
if name is None:
self.name = "VGG16_A1"
self._modelCache = "{0}-model.h5".format(self.name)
self._weightsCache = "{0}-weights.h5".format(self.name)
self._mgpu_weightsCache = "{0}-mgpu-weights.h5".format(self.name)
self.cache_m = CacheManager()
self.cache_m.registerFile(
os.path.join(config.model_path, self._modelCache),
self._modelCache)
self.cache_m.registerFile(
os.path.join(config.weights_path, self._weightsCache),
self._weightsCache)
self.cache_m.registerFile(
os.path.join(config.weights_path, self._mgpu_weightsCache),
self._mgpu_weightsCache)
def get_model_cache(self):
"""
Returns path to model cache
"""
return self.cache_m.fileLocation(self._modelCache)
def get_weights_cache(self):
"""
Returns path to model cache
"""
return self.cache_m.fileLocation(self._weightsCache)
def get_mgpu_weights_cache(self):
"""
Returns path to model cache
"""
return self.cache_m.fileLocation(self._mgpu_weightsCache)
def build(self, **kwargs):
"""
Returns a VGG 16 model instance, final fully-connected layers are substituted by Conv2Ds
@param pre_trained <boolean>: returned model should be pre-trained or not
"""
width, height, channels = self._check_input_shape()
if backend.image_data_format() == 'channels_first':
input_shape = (channels, height, width)
else:
input_shape = (height, width, channels)
if 'data_size' in kwargs:
self.data_size = kwargs['data_size']
self.cache_m = CacheManager()
model = self._build_architecture(input_shape)
#Check if previous training and LR is saved, if so, use it
lr_cache = "{0}_learning_rate.txt".format(self.name)
self.cache_m.registerFile(os.path.join(self._config.cache, lr_cache),
lr_cache)
l_rate = 0.0005
if os.path.isfile(self.cache_m.fileLocation(
lr_cache)) and not self._config.new_net:
l_rate = float(self.cache_m.read(lr_cache))
if self._config.info:
print("Found previous learning rate: {0}".format(l_rate))
sgd = optimizers.SGD(lr=l_rate,
decay=1.5e-4,
momentum=0.9,
nesterov=True)
#adam = optimizers.Adam(lr = l_rate)
#Return parallel model if multiple GPUs are available
parallel_model = None
if self._config.gpu_count > 1:
with tf.device('/cpu:0'):
model.compile(loss='categorical_crossentropy',
optimizer=sgd,
metrics=['accuracy'])
parallel_model = multi_gpu_model(model,
gpus=self._config.gpu_count)
parallel_model.compile(
loss='categorical_crossentropy',
optimizer=sgd,
metrics=['accuracy'],
#options=p_opt,
#run_metadata=p_mtd
)
else:
model.compile(
loss='categorical_crossentropy',
optimizer=sgd,
metrics=['accuracy'],
#options=p_opt,
#run_metadata=p_mtd
)
self.single = model
self.parallel = parallel_model
return (model, parallel_model)
def _build_architecture(self, input_shape):
original_vgg16 = vgg16.VGG16(
weights=self.cache_m.fileLocation('vgg16_weights_notop.h5'),
include_top=False,
input_shape=input_shape)
#Freeze initial layers, except for the last 3:
#for layer in original_vgg16.layers[:-2]:
# layer.trainable = False
model = Sequential()
model.add(original_vgg16)
model.add(
Convolution2D(4096, (7, 7),
strides=1,
padding='valid',
kernel_initializer='he_normal'))
model.add(Activation('relu'))
model.add(Dropout(0.75))
model.add(
Convolution2D(4096, (1, 1),
strides=1,
padding='valid',
kernel_initializer='he_normal'))
model.add(Activation('relu'))
model.add(Dropout(0.75))
model.add(
Convolution2D(self._ds.nclasses, (1, 1),
strides=1,
padding='valid',
kernel_initializer='he_normal'))
model.add(Flatten())
model.add(Dense(self._ds.nclasses))
model.add(Activation('softmax'))
return model
def run_test(self, model, x_test=None, y_test=None, load_full=True):
"""
This should be executed after a model has been trained
"""
cache_m = CacheManager()
split = None
if os.path.isfile(cache_m.fileLocation('split_ratio.pik')):
split = cache_m.load('split_ratio.pik')
else:
print(
"[Predictor] A previously trained model and dataset should exist. No previously defined spliting found."
)
return Exitcodes.RUNTIME_ERROR
#Priority is for given data as parameters. If None is given, try to load metadata as configured
if x_test is None or y_test is None:
if self._config.testdir is None:
#Load sampled data if required by command line
if self._config.sample < 1.0:
_, _, (x_test, y_test) = self._ds.split_metadata(
split=split,
data=self._ds.sample_metadata(self._config.sample))
else:
_, _, (x_test, y_test) = self._ds.split_metadata(split)
else:
x_test, y_test = self._ds._run_dir(self._config.testdir)
if self._config.verbose > 0:
unique, count = np.unique(y_test, return_counts=True)
l_count = dict(zip(unique, count))
if len(unique) > 2:
print("Test items:")
print("\n".join([
"label {0}: {1} items".format(key, l_count[key])
for key in unique
]))
else:
if not 1 in l_count:
l_count[1] = 0
print(
"Test labels: {0} are 0; {1} are 1;\n - {2:.2f} are positives"
.format(l_count[0], l_count[1],
(l_count[1] / (l_count[0] + l_count[1]))))
print("Test set: {} items".format(len(y_test)))
X, Y = self._ds.load_data(data=(x_test, y_test), keepImg=self._keep)
if self._config.verbose > 1:
print("Y original ({1}):\n{0}".format(Y, Y.shape))
Y = to_categorical(Y, self._ds.nclasses)
# session setup
sess = K.get_session()
ses_config = tf.ConfigProto(
device_count={
"CPU": self._config.cpu_count,
"GPU": self._config.gpu_count
},
intra_op_parallelism_threads=self._config.cpu_count
if self._config.gpu_count == 0 else self._config.gpu_count,
inter_op_parallelism_threads=self._config.cpu_count
if self._config.gpu_count == 0 else self._config.gpu_count,
log_device_placement=True if self._verbose > 1 else False)
sess.config = ses_config
K.set_session(sess)
#During test phase multi-gpu mode is not used (maybe done latter)
if self._ensemble:
#Weights should be loaded during ensemble build
if hasattr(model, 'build_ensemble'):
pred_model = model.build_ensemble(training=False, npfile=True)
else:
if self._config.info:
print(
'[Predictor] Model not prepared to build ensembles, implement or choose other model'
)
return None
elif load_full and os.path.isfile(model.get_model_cache()):
try:
pred_model = load_model(model.get_model_cache())
if self._config.info:
print("Model loaded from: {0}".format(
model.get_model_cache()))
except ValueError:
pred_model, _ = model.build(training=False, pre_load_w=False)
pred_model.load_weights(model.get_weights_cache())
if self._config.info:
print("Model weights loaded from: {0}".format(
model.get_weights_cache()))
elif os.path.isfile(model.get_weights_cache()):
pred_model, _ = model.build(training=False, pre_load_w=False)
pred_model.load_weights(model.get_weights_cache())
if self._config.info:
print("Model weights loaded from: {0}".format(
model.get_weights_cache()))
else:
if self._config.info:
print("No trained model or weights file found")
return None
bsize = self._config.batch_size
stp = round((len(X) / bsize) + 0.5)
image_generator = ImageDataGenerator(
samplewise_center=self._config.batch_norm,
samplewise_std_normalization=self._config.batch_norm)
if self._ensemble:
if not self._config.tdim is None:
fix_dim = self._config.tdim
else:
fix_dim = self._ds.get_dataset_dimensions()[0][
1:] #Only smallest image dimensions matter here
test_generator = SingleGenerator(
dps=(X, Y),
classes=self._ds.nclasses,
dim=fix_dim,
batch_size=self._config.batch_size,
image_generator=image_generator,
extra_aug=self._config.augment,
shuffle=False,
verbose=self._verbose,
input_n=self._config.emodels)
else:
test_generator = image_generator.flow(x=X,
y=Y,
batch_size=bsize,
shuffle=False)
if self._config.progressbar:
l = tqdm(desc="Making predictions...", total=stp)
Y_pred = np.zeros((len(X), self._ds.nclasses), dtype=np.float32)
for i in range(stp):
start_idx = i * bsize
example = test_generator.next()
Y_pred[start_idx:start_idx + bsize] = pred_model.predict_on_batch(
example[0])
if self._config.progressbar:
l.update(1)
elif self._config.info:
print("Batch prediction ({0}/{1})".format(i, stp))
if self._config.verbose > 1:
if not np.array_equal(Y[start_idx:start_idx + bsize],
example[1]):
print(
"Datasource label ({0}) and batch label ({1}) differ".
format(Y[start_idx:start_idx + bsize], example[1]))
del (X)
del (test_generator)
if self._config.progressbar:
l.close()
y_pred = np.argmax(Y_pred, axis=1)
expected = np.argmax(Y, axis=1)
if self._config.verbose > 0:
if self._config.verbose > 1:
np.set_printoptions(threshold=np.inf)
print("Predicted probs ({1}):\n{0}".format(
Y_pred, Y_pred.shape))
#print("Y ({1}):\n{0}".format(Y,Y.shape))
print("expected ({1}):\n{0}".format(expected, expected.shape))
print("Predicted ({1}):\n{0}".format(y_pred, y_pred.shape))
#Save predictions
cache_m.dump((expected, Y_pred, self._ds.nclasses), 'test_pred.pik')
#Output metrics
print_prediction(self._config)
