class TLClassifier:
def __init__(self, is_site):
#TODO load classifier
assert not is_site
weights_file = r'light_classification/models/squeezenet_weights.h5' #Replace with real world classifier
image_shape = (224, 224, 3)
self.states = (TrafficLight.RED, TrafficLight.YELLOW,
TrafficLight.GREEN, TrafficLight.UNKNOWN)
print('Loading model..')
self.model = SqueezeNet(len(self.states), *image_shape)
self.model.load_weights(weights_file, by_name=True)
self.model._make_predict_function()
print('Loaded weights: %s' % weights_file)
def get_classification(self, image):
"""Determines the color of the traffic light in the image
Args:
image (cv::Mat): image containing the traffic light
Returns:
int: ID of traffic light color (specified in styx_msgs/TrafficLight)
"""
mini_batch = cv2.resize(
image, (224, 224),
cv2.INTER_AREA).astype('float')[np.newaxis, ..., ::-1] / 255.
light = self.states[np.argmax(self.model.predict(mini_batch))]
return light
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--checkpoint-path', required=True)
parser.add_argument('--image', nargs='+', required=True)
parser.add_argument('--num-classes', type=int, required=True)
args = parser.parse_args()
model = SqueezeNet(weights=None, classes=args.num_classes)
model.load_weights(args.checkpoint_path)
xs = []
for path in args.image:
img = image.load_img(path, target_size=(SIZE, SIZE))
x = image.img_to_array(img)
xs.append(x)
xs = np.array(xs)
xs = preprocess_input(xs)
probs = model.predict(xs)
print('')
for i, path in enumerate(args.image):
print('%s' % path)
print(' Prediction: %s' % np.argmax(probs[i]))
def predict(img_local_path):
model = SqueezeNet(weights='imagenet')
img = image.load_img(img_local_path, target_size=(227, 227))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
preds = model.predict(x)
result = decode_predictions(preds)
return result
def build_model(self):
self.model = SqueezeNet(include_top=True,
weights=None,
classes=3,
input_shape=(self.IMAGE_SIZE, self.IMAGE_SIZE,
3))
self.model.summary()
opt = Adam()
self.model.compile(loss="binary_crossentropy",
optimizer=opt,
metrics=["accuracy"])
return self.model
def create_discriminator():
"""
Makes the discriminator network
:return: The discriminator network
"""
# We are using the SqueezeNet because it is fairly resource-light and GANs can already be hard to train
disc_net = SqueezeNet(input_width=input_size[0], classes=2)
# Compile with loss
disc_net.compile(loss="binary_crossentropy",
optimizer=create_adam(),
metrics=["accuracy"])
return disc_net
def __init__(self, is_site):
#TODO load classifier
assert not is_site
weights_file = r'light_classification/models/squeezenet_weights.h5' #Replace with real world classifier
image_shape = (224, 224, 3)
self.states = (TrafficLight.RED, TrafficLight.YELLOW,
TrafficLight.GREEN, TrafficLight.UNKNOWN)
print('Loading model..')
self.model = SqueezeNet(len(self.states), *image_shape)
self.model.load_weights(weights_file, by_name=True)
self.model._make_predict_function()
print('Loaded weights: %s' % weights_file)
def get_pretrained_model(model_name,
input_shape,
include_top=False,
weights="imagenet",
pooling=None):
if model_name == "vgg16":
return VGG16(include_top=include_top,
weights=weights,
input_shape=input_shape,
pooling=pooling)
if model_name == "vgg19":
return VGG19(include_top=include_top,
weights=weights,
input_shape=input_shape,
pooling=pooling)
if model_name == "resnet50":
return ResNet50(include_top=include_top,
weights=weights,
input_shape=input_shape,
pooling=pooling)
if model_name == "inceptionV3":
return InceptionV3(include_top=include_top,
weights=weights,
input_shape=input_shape,
pooling=pooling)
if model_name == "xception":
return Xception(include_top=include_top,
weights=weights,
input_shape=input_shape,
pooling=pooling)
if model_name == "squeezenet":
return SqueezeNet(include_top=include_top,
weights=weights,
input_shape=input_shape)
def PSPNet(classnum=81, input_shape=(473, 473), sq=False):
inp = Input(input_shape + (3,), name="psp_input")
if sq:
res = SqueezeNet(input_tensor=inp, include_top=False, input_shape=(800,800) + (3,))
x = Lambda(Interp, arguments={'shape': (60, 60)})(res.output)
else:
x = ResNet(inp, layers=50)
x = Lambda(Interp, arguments={'shape': (60, 60)})(x)
psp = build_psp(x, input_shape)
x = Conv2D(512, (3, 3), strides=(1, 1), padding="same", name="conv5_4", use_bias=False)(psp)
x = BN(name="conv5_4_bn")(x)
embedFeatures = Activation('relu')(x)
x = Dropout(0.1)(embedFeatures)
classFeatures = Conv2D(classnum, (1, 1), strides=(1, 1), name="lastConv")(x)
x = Lambda(Interp, arguments={'shape': (input_shape[0], input_shape[1])})(classFeatures)
classes = Activation('sigmoid')(x)
model = Model(inputs=inp, outputs=classes)
if not sq:
model.load_weights('/mnt/course/weights/pretrained.h5', by_name=True)
# for layer in model.layers[:-8]:
# layer.trainable = False
sgd = SGD(lr=0.01, momentum=0.9, nesterov=True)
model.compile(optimizer=sgd,
loss='mse',
metrics=['accuracy'])
return model
def get_squeezenet(input_shape, n_classes, **params):
"""
"""
optimizer = '' if 'optimizer' not in params else params['optimizer']
lr = 0.01 if 'lr' not in params else params['lr']
loss = '' if 'loss' not in params else params['loss']
weights = 'imagenet' if 'weights' not in params else params['weights']
final_activation = 'sigmoid'
snet = SqueezeNet(input_shape=input_shape,
classes=n_classes,
include_top=False,
weights=weights)
x = snet.outputs[0]
x = GlobalAveragePooling2D()(x)
out = Activation(final_activation, name='tag_vector')(x)
model = Model(inputs=snet.inputs, outputs=out)
model.name = "SqueezeNet"
if optimizer == 'adadelta':
opt = Adadelta(lr=lr)
elif optimizer == 'adam':
opt = Adam(lr=lr)
elif optimizer == 'sgd':
opt = SGD(lr=lr, momentum=0.9, decay=0.00001, nesterov=True)
else:
opt = None
if opt is not None:
model.compile(loss=loss, optimizer=opt, metrics=[precision, recall])
return model
def create_model(input_shape, n_out):
input_tensor = Input(shape=(SIZE, SIZE, 3))
#bn = BatchNormalization()(input_tensor)
#conv = Conv2D(3,(3,3),padding='same',activation='relu')(bn)
base_model = SqueezeNet(include_top=False,
weights='imagenet',
input_shape=(SIZE, SIZE, 3),
input_tensor=input_tensor)
enc_out, short_cuts = encoder(base_model)
x0 = GlobalAveragePooling2D()(squeeze_excite_block(enc_out))
x1 = GlobalAveragePooling2D()(squeeze_excite_block(short_cuts[0]))
x2 = GlobalAveragePooling2D()(squeeze_excite_block(short_cuts[1]))
x3 = GlobalAveragePooling2D()(squeeze_excite_block(short_cuts[2]))
#x4 = GlobalAveragePooling2D()(squeeze_excite_block(short_cuts[3]))
x = Concatenate()([x0, x1, x2, x3])
x = BatchNormalization()(x)
x = Dropout(0.5)(x)
x = Dense(256, activation='relu')(x)
#x = BatchNormalization()(x)
#x = Dropout(0.5)(x)
#x = Dense(256, activation='relu')(x)
x = BatchNormalization()(x)
x = Dropout(0.5)(x)
output = Dense(n_out, activation='sigmoid')(x)
model = Model(input_tensor, output)
# transfer imagenet weights
#res_img = ResNet34(include_top=False, weights='imagenet', input_shape=(SIZE, SIZE, 3))
#offset = 2
#for i, l in enumerate(base_model.layers[offset+1:]):
# l.set_weights(res_img.layers[i + 1].get_weights())
return model
def get_squeezenet_model(layers):
base_model = SqueezeNet(content_layers)
all_n = []
all_n = get_ckpt_weights('../squeezenet_weights/squeezenet.ckpt')
all_weights = []
for w in base_model.model.weights:
all_weights.append(all_n[weight_to_weight[w.name]])
base_model.set_weights(all_weights)
base_model.trainable = False
dream_model = base_model
return dream_model
def get_squeezenet21(input_shape, n_classes, **params):
"""
"""
optimizer = '' if 'optimizer' not in params else params['optimizer']
lr = 0.01 if 'lr' not in params else params['lr']
loss = '' if 'loss' not in params else params['loss']
weights = 'imagenet' if 'weights' not in params else params['weights']
snet = SqueezeNet(input_shape=input_shape,
classes=n_classes,
include_top=False,
weights=weights)
x = snet.outputs[0]
x = Flatten()(x)
x = Dense(96, activation='relu', name='d1')(x)
x = Dropout(0.5)(x)
x = Dense(n_classes, name='d2')(x)
out = Activation('sigmoid', name='tags')(x)
model = Model(inputs=snet.inputs, outputs=out)
model.name = "SqueezeNet21"
if optimizer == 'adadelta':
opt = Adadelta(lr=lr)
elif optimizer == 'adam':
opt = Adam(lr=lr)
elif optimizer == 'sgd':
opt = SGD(lr=lr, momentum=0.9, decay=0.00001, nesterov=True)
else:
opt = None
if opt is not None:
model.compile(loss=loss, optimizer=opt, metrics=[precision, recall])
return model
def create_squeezenet(nclass, train=True):
input = Input(shape=(HEIGHT, WIDTH, 3), name="img_input")
base_model = SqueezeNet(include_top=False,
weights=None,
input_tensor=input)
x = base_model.output
if train:
base_model.load_weights(
"models/squeezenet_weights_tf_dim_ordering_tf_kernels_notop.h5")
x = Dropout(0.2, name='drop9')(x)
x = Convolution2D(nclass, (1, 1), padding='valid', name='conv10')(x)
x = Activation('relu', name='relu_conv10')(x)
x = GlobalAveragePooling2D()(x)
x = Activation('softmax', name='loss')(x)
model = Model(input, x, name='squeezenet')
return model
def __init__(self, session):
self.session = session
self.add_placeholders()
self.squeezenet = SqueezeNet(save_path='squeezenet/squeezenet.ckpt',
sess=session)
self.y_out = self.add_prediction_op()
self.loss = self.add_loss_op(self.y_out)
self.training_op = self.add_optimization_op(self.loss)
self.correct_prediction, self.accuracy = self.add_accuracy_op()
def main():
test_patterns = [
('VGGNetBN', VGGNetBN(17), 224), ('VGGNetBNHalf', VGGNetBN(17,
32), 224),
('VGGNetBNQuater', VGGNetBN(17, 16), 224),
('GoogLeNetBN', GoogLeNetBN(17), 224),
('GoogLeNetBNHalf', GoogLeNetBN(17, 16), 224),
('GoogLeNetBNQuater', GoogLeNetBN(17, 8), 224),
('ResNet50', ResNet50(17), 224), ('ResNet50Half', ResNet50(17,
32), 224),
('ResNet50Quater', ResNet50(17, 16), 224),
('SqueezeNet', SqueezeNet(17), 224),
('SqueezeNetHalf', SqueezeNet(17, 8), 224),
('MobileNet', MobileNet(17), 224),
('MobileNetHalf', MobileNet(17, 16), 224),
('MobileNetQuater', MobileNet(17, 8), 224),
('InceptionV4', InceptionV4(dim_out=17), 299),
('InceptionV4S',
InceptionV4(dim_out=17,
base_filter_num=6,
ablocks=2,
bblocks=1,
cblocks=1), 299),
('InceptionResNetV2', InceptionResNetV2(dim_out=17), 299),
('InceptionResNetV2S',
InceptionResNetV2(dim_out=17,
base_filter_num=8,
ablocks=1,
bblocks=2,
cblocks=1), 299),
('FaceClassifier100x100V', FaceClassifier100x100V(17), 100),
('FaceClassifier100x100V2', FaceClassifier100x100V2(17), 100)
]
for model_name, model, test_size in test_patterns:
oltp_cpu, batch_gpu = check_speed(model, test_images[test_size])
print('{}\t{:.02f}\t{:.02f}'.format(model_name, oltp_cpu * 1000,
batch_gpu * 1000))
def train(img_local_path, label_path, model_object_key):
model = SqueezeNet(weights='imagenet')
img = image.load_img(img_local_path, target_size=(227, 227))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
label_file = open(label_path)
y = np.array([label_file.read()])
label_file.close()
model.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
history = model.fit(x, y)
model.summary()
model.save_weights(tmp_path + model_object_key)
return history.history
def main():
pl_train, pl_labels = get_dataset('./Pan_Licence/')
pl_labels = to_categorical(pl_labels, num_classes=36)
x_train, x_val, y_train, y_val = train_test_split(pl_train,
pl_labels,
test_size=0.2,
random_state=2064)
tb = TensorBoard(log_dir='./logs/Squeezenet', write_graph=True)
model = SqueezeNet()
model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
print(model.summary())
history = model.fit(x_train,
y_train,
batch_size=32,
epochs=5,
validation_split=0.1,
shuffle=True,
callbacks=[tb])
## Save Model
json_model = model.to_json()
with open('model_squeezenet.json', 'w') as f:
f.write(json_model)
model.save_weights('model_squeezenet.h5')
print('Model Saved')
print('Evaluating Model')
predict = model.evaluate(x=x_val, y=y_val, batch_size=1)
print('Score', predict[1] * 100.00)
print('Loss', predict[0])
def init():
global model
model_root = Model.get_model_path(
'kd_teach_the_student') #, _workspace=ws)
# model_root = model_root.strip('model.json')
print(model_root)
# load json and create model
weight_file = os.path.join(model_root, "squeezenet_weights.hdf5")
model = SqueezeNet(weight_decay=0.0,
image_size=299,
trainable=False,
weight_file=weight_file)
# model.load_weights(os.path.join(model_root, "squeezenet_weights.hdf5"))
model = Model_keras(model.input, model.outputs)
def get_model(learning_rate=1e-3, num_classes=200):
model = SqueezeNet(num_classes=num_classes)
# set the first layer not trainable
model.features[0].weight.requires_grad = False
# all conv layers except the first and the last
all_conv_weights = [
(n, p) for n, p in model.named_parameters()
if 'weight' in n and not 'bn' in n and not 'features.1.' in n
]
weights_to_be_quantized = [
p for n, p in all_conv_weights
if not ('classifier' in n or 'features.0.' in n)
]
# the last layer
weights = [model.classifier[1].weight]
biases = [model.classifier[1].bias]
# parameters of batch_norm layers
bn_weights = [
p for n, p in model.named_parameters()
if ('bn' in n or 'features.1.' in n) and 'weight' in n
]
bn_biases = [
p for n, p in model.named_parameters()
if ('bn' in n or 'features.1.' in n) and 'bias' in n
]
params = [{
'params': weights,
'weight_decay': 1e-4
}, {
'params': weights_to_be_quantized
}, {
'params': biases
}, {
'params': bn_weights
}, {
'params': bn_biases
}]
optimizer = optim.Adam(params, lr=learning_rate)
# loss function
criterion = nn.CrossEntropyLoss().cuda()
# move the model to gpu
model = model.cuda()
return model, criterion, optimizer
def train_chesspiece_model():
"""Trains the chesspiece model based on SqueezeNet-v1.1."""
base_model = SqueezeNet(input_shape=(227, 227, 3), include_top=False,
weights='imagenet')
# First train only the top layers
for layer in base_model.layers:
layer.trainable = False
model = build_model(base_model)
train_generator, validation_generator = data_generators(
preprocess_input, (227, 227), 64)
callbacks = model_callbacks(5, "./models/SqueezeNet1p1_pre.h5", 0.1, 10)
history = train_model(model, 20, train_generator, validation_generator,
callbacks, use_weights=False, workers=5)
plot_model_history(history, "./models/SqueezeNet1p1_pre_acc.png",
"./models/SqueezeNet1p1_pre_loss.png")
evaluate_model(model, validation_generator)
# Also train fire 7-9
for layer in model.layers[:41]:
layer.trainable = False
for layer in model.layers[41:]:
layer.trainable = True
model.compile(optimizer='Adam', loss='categorical_crossentropy',
metrics=['accuracy'])
callbacks = model_callbacks(20, "./models/SqueezeNet1p1.h5", 0.2, 8)
history = train_model(model, 100, train_generator, validation_generator,
callbacks, use_weights=False, workers=5)
plot_model_history(history, "./models/SqueezeNet1p1_acc.png",
"./models/SqueezeNet1p1_loss.png")
evaluate_model(model, validation_generator)
model.save("./models/SqueezeNet1p1_last.h5")
def get_model():
model = SqueezeNet()
# create different parameter groups
weights = [
p for n, p in model.named_parameters()
if 'weight' in n and not 'bn' in n and not 'features.1.' in n
]
biases = [model.classifier[1].bias]
bn_weights = [
p for n, p in model.named_parameters()
if ('bn' in n or 'features.1.' in n) and 'weight' in n
]
bn_biases = [
p for n, p in model.named_parameters()
if ('bn' in n or 'features.1.' in n) and 'bias' in n
]
# initialize batch norm params
for p in bn_weights:
constant(p, 1.0)
for p in bn_biases:
constant(p, 0.0)
params = [{
'params': weights,
'weight_decay': 1e-3
}, {
'params': biases
}, {
'params': bn_weights
}, {
'params': bn_biases
}]
optimizer = optim.SGD(params, lr=4e-2, momentum=0.95, nesterov=True)
# loss function
criterion = nn.CrossEntropyLoss().cuda()
# move the model to gpu
model = model.cuda()
return model, criterion, optimizer
def get_model():
model = SqueezeNet()
# create different parameter groups
weights = [
p for n, p in model.named_parameters()
if 'weight' in n and 'bn' not in n and 'features.1.' not in n
]
biases = [model.classifier[1].bias]
bn_weights = [
p for n, p in model.named_parameters()
if ('bn' in n or 'features.1.' in n) and 'weight' in n
]
bn_biases = [
p for n, p in model.named_parameters()
if ('bn' in n or 'features.1.' in n) and 'bias' in n
]
for p in bn_weights:
constant(p, 1.0)
for p in bn_biases:
constant(p, 0.0)
params = [{
'params': weights,
'weight_decay': 3e-4
}, {
'params': biases
}, {
'params': bn_weights
}, {
'params': bn_biases
}]
optimizer = optim.SGD(params, lr=4e-2, momentum=0.95, nesterov=True)
loss = nn.CrossEntropyLoss().cuda()
model = model.cuda() # move the model to gpu
return model, loss, optimizer
model = DenseNet121()
elif args.model == 'densenet121cov':
MODEL_NAME = 'DenseNet121Cov'
model = DenseNet121Cov()
elif args.model == 'densenet121covdropout':
MODEL_NAME = 'DenseNet121CovDropout'
model = DenseNet121CovDropout()
elif args.model == 'stndensenet121covdropout':
MODEL_NAME = 'STNDenseNet121CovDropout'
model = STNDenseNet121CovDropout()
elif args.model == 'inceptionresnetv2':
MODEL_NAME = 'InceptionResNetV2'
model = InceptionResNetV2()
elif args.model == 'squeezenet':
MODEL_NAME = 'SqueezeNet'
model = SqueezeNet()
elif args.model == 'attentionresnet56':
MODEL_NAME = 'AttentionResNet56'
model = AttentionResNet56(shape=(192, 192, 3),
n_classes=8,
n_channels=32)
elif args.model == 'lightcnn':
MODEL_NAME = 'LightCNN'
model = LightCNN()
elif args.model == 'lightcnncov':
MODEL_NAME = 'LightCNNCov'
model = LightCNNCov()
elif args.model == 'holonet':
MODEL_NAME = 'HoloNet'
model = HoloNet()
elif args.model == 'holonetcov':
cell_fw = tf.contrib.rnn.LSTMCell(
hps['num_hidden'],
initializer=tf.random_uniform_initializer(-0.1, 0.1, seed=123),
state_is_tuple=True)
(emb_encoder_outputs,
fw_state) = tf.contrib.rnn.static_rnn(cell_fw,
emb_encoder_inputs,
dtype=tf.float32,
sequence_length=question_lens)
return (emb_encoder_outputs, fw_state)
#load CNN
sess = tf.InteractiveSession()
CNN = SqueezeNet(save_path=PATH_TO_SAVED_NET, sess=sess)
#load question answer data
val_data = load_data()
print("=========== Data loaded ===========")
#process loaded data
# words = [set(sample['question'].split(" ")) for sample in val_data]
# set_words = set()
# for s in words:
# set_words = set_words.union(s)
# hps['vsize'] = set_words.__len__()
words_to_idx = wordsToIdx()
idx_to_words = idxToWords()
import numpy as np
from keras.applications.imagenet_utils import preprocess_input, decode_predictions
from keras.preprocessing import image
from squeezenet import SqueezeNet
#import matplotlib.pyplot as plt
#import matplotlib.image as mpimg
model = SqueezeNet()
img = image.load_img('pexels-photo-280207.jpeg', target_size=(227, 227))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
preds = model.predict(x)
all_results = decode_predictions(preds)
for results in all_results:
for result in results:
print('Probability %0.2f%% => [%s]' % (100*result[2], result[1]))
#result_text= 'Probability %0.2f%% => [%s]' % (100*result[2], result[1])
#break
#plt.figure(num=1,figsize=(8, 6), dpi=80)
#plt.imshow(img)
#plt.text(130,90,result_text,horizontalalignment='center', verticalalignment='center',fontsize=16,color='black')
#plt.axis('off')
#plt.show()
training_data = gen(X_train, Y_train, nb=200)
validation_data = gen(X_test, Y_test, nb=50)
# Unzip to two lists.
# images, classes = zip(*images_classes)
# [images, classes], [x, y] = mnist.load_data()
# images = images[0:500]
# classes = classes[0:500]
# images = np.array([cv2.resize(cv2.cvtColor(im, cv2.COLOR_GRAY2RGB), (227, 227)) for im in images])
# images = np.array(images)
# print images.shape
# classes = to_categorical(classes, nb_classes=nr_classes)
print('Loading model..')
model = SqueezeNet(nb_classes, input_shape=input_shape)
adam = Adam(lr=0.040)
sgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss="categorical_crossentropy",
optimizer='adam',
metrics=['accuracy'])
if os.path.isfile(weights_file):
print('Loading weights: %s' % weights_file)
model.load_weights(weights_file, by_name=True)
print('Fitting model')
# model.fit(images, classes, batch_size=batch_size, nb_epoch=nb_epoch, verbose=1, validation_split=0.2, initial_epoch=0)
model.fit_generator(training_data,
samples_per_epoch=samples_per_epoch,
validation_data=validation_data,
nb_val_samples=nb_val_samples,
def train_factory(MODEL_NAME):
config = tf.ConfigProto()
config.gpu_options.allocator_type = 'BFC'
config.gpu_options.allow_growth = True
set_session(tf.Session(config=config))
# model = CCR(input_shape=(img_width,img_height,1),classes=charset_size)
# model = LeNet.build(width=img_width, height=img_height, depth=1, classes=charset_size)
# model = ResNet.build_model(SHAPE=(img_width,img_height,1), classes=charset_size)
# vgg net 5
# MODEL_PATH='trained_model/vggnet5.hdf5'
# model=VGGNet5.vgg(input_shape=(img_width,img_height,1),classes=charset_size)
model=None
if(MODEL_NAME=='inception_resnet_v2'):
model=InceptionResNetV2.inception_resnet_v2(input_shape=(img_width,img_height,3),classes=charset_size,weights='./trained_model/inception_resnet_v2/inception_resnet_v2.12-0.8244.hdf5')
elif(MODEL_NAME=='xception'):
# xeception
model=Xception.Xception((img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='mobilenet_v2'):
#mobilenet v2
model=MobileNetv2.MobileNet_v2((img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='inception_v3'):
#mobilenet v2
model=Inception_v3.inception((img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='vgg16'):
model=VGGNet.vgg(input_shape=(img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='vgg19'):
model=VGG19.VGG19(input_shape=(img_width,img_height,3),classes=charset_size,weights='weights/vgg19_weights_tf_dim_ordering_tf_kernels.h5')
elif(MODEL_NAME=='resnet50'):
model=ResNet50.resnet(input_shape=(img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='inception_v4'):
model=inception_v4.inception_v4(input_shape=(img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='resnet34'):
model=ResNet34.ResNet34(input_shape=(img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='densenet121'):
model=DenseNet.DenseNet(input_shape=(img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='densenet161'):
model=DenseNet.DenseNet(input_shape=(img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='shufflenet_v2'):
model=ShuffleNetV2.ShuffleNetV2(input_shape=(img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='resnet_attention_56'):
model=Resnet_Attention_56.Resnet_Attention_56(input_shape=(img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='squeezenet'):
model=SqueezeNet.SqueezeNet(input_shape=(img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='seresnet50'):
model=SEResNet50.SEResNet50(input_shape=(img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='se_resnext'):
model=SEResNext.SEResNext(input_shape=(img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='nasnet'):
model=NASNet.NASNetLarge(input_shape=(img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='custom'):
model=Custom_Network.Custom_Network(input_shape=(img_width,img_height,3),classes=charset_size)
elif(MODEL_NAME=='resnet18'):
model=ResnetBuilder.build_resnet_18(input_shape=(img_width,img_height,3),num_outputs=charset_size)
print(model.summary())
train(model,MODEL_NAME)
def model_fn(features, labels, mode, params):
assert params['n_classes'] > 0
logits = SqueezeNet(features,
classes=params['n_classes'],
training=(mode == tf.estimator.ModeKeys.TRAIN))
onehot_labels = tf.one_hot(labels, params['n_classes'], axis=1)
# Compute loss.
# See https://stats.stackexchange.com/questions/306862/cross-entropy-versus-mean-of-cross-entropy
loss = tf.reduce_mean(tf.losses.softmax_cross_entropy(onehot_labels=onehot_labels,
logits=logits))
# Compute predictions
predictions = tf.cast(tf.argmax(logits, 1), tf.int32)
labels = tf.cast(labels, tf.int64)
predictions = tf.Print(predictions, [predictions], message='predictions', summarize=10000)
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(mode, predictions={
"classes": tf.argmax(input=logits, axis=1),
"probabilities": tf.nn.softmax(logits, name="softmax_tensor")
})
# NOTE: Useful for debugging!
loss = tf.Print(loss, [loss, tf.argmax(tf.nn.softmax(logits), axis=1)],
message='[Loss|Logits]',
summarize=1 + params['batch_size'] * 3)
# Compute evaluation metrics.
accuracy = tf.metrics.accuracy(labels, predictions, name='accuracy_1')
metrics = {
"accuracy": accuracy,
"recall_at_5": tf.metrics.recall_at_k(labels, logits, 5),
"recall_at_1": tf.metrics.recall_at_k(labels, logits, 1)
}
tf.summary.scalar('accuracy', accuracy[1])
if mode == tf.estimator.ModeKeys.EVAL:
return tf.estimator.EstimatorSpec(mode,
loss=loss,
eval_metric_ops=metrics)
# Create training op.
assert mode == tf.estimator.ModeKeys.TRAIN
optimizer = None
if params['optimizer'] == 'poly':
# NOTE: Setting the learning rate to 0.04 gave `NanLossDuringTrainingError` as per the paper.
decay_steps = params['n_images'] * params['n_epochs'] // params['batch_size']
learning_rate = tf.train.polynomial_decay(learning_rate=params['lr'],
global_step=tf.train.get_global_step(),
decay_steps=decay_steps,
end_learning_rate=0.0005,
power=1.0,
cycle=False)
optimizer = tf.train.MomentumOptimizer(learning_rate=learning_rate,
momentum=0.9,
use_nesterov=False)
elif params['optimizer'] == 'rms':
optimizer = tf.train.RMSPropOptimizer(learning_rate=params['lr'],
momentum=0.9)
elif params['optimizer'] == 'sgd':
optimizer = tf.train.MomentumOptimizer(learning_rate=params['lr'],
momentum=0.9,
use_nesterov=True)
elif params['optimizer'] == 'adam':
optimizer = tf.train.AdamOptimizer(learning_rate=params['lr'])
else:
assert 'No optimizer defined in params!'
train_op = optimizer.minimize(loss, global_step=tf.train.get_global_step())
return tf.estimator.EstimatorSpec(mode, loss=loss, train_op=train_op)
testset = datasets.CIFAR10(root=data_location,
train=False,
download=True,
transform=test_transform)
testloader = torch.utils.data.DataLoader(testset,
shuffle=False,
batch_size=args.test_batch_size,
**loader_args)
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse',
'ship', 'truck')
#model = AlexNet(num_classes=10, small_input=True, use_ttq=args.ttq)
model = SqueezeNet(version=1.1,
num_classes=10,
small_input=True,
use_ttq=args.ttq)
#model = resnet18(use_ttq=args.ttq)
def fix_state(valid, new):
for k, v in valid.items():
parts = k.split('.')
if len(parts) > 1:
if parts[-1] == 'W_p':
if not k in new:
new[k] = torch.Tensor(1)
new[k][0] = valid['.'.join(parts[:-1] +
['weight'])].max() / 2
elif parts[-1] == 'W_n':
if not k in new:
import numpy as np
from keras.applications.imagenet_utils import preprocess_input, decode_predictions
from keras.preprocessing import image
from squeezenet import SqueezeNet
import matplotlib.pyplot as plt
# import matplotlib.image as mpimg
model = SqueezeNet()
img = image.load_img('2.jpg', target_size=(227, 227))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
preds = model.predict(x)
all_results = decode_predictions(preds)
for results in all_results:
for result in results:
print('Probability %0.2f%% => [%s]' % (100 * result[2], result[1]))
result_text = 'Probability %0.2f%% => [%s]' % (100 * result[2],
result[1])
break
plt.figure(num=1, figsize=(8, 6), dpi=80)
plt.imshow(img)
plt.text(130,
90,
result_text,
horizontalalignment='center',
verticalalignment='center',
fontsize=16,
color='black')
class DNNModel:
def __init__(self, image_path):
self.IMAGE_SIZE = 64
self.data = []
self.labels = []
self.model = self.build_model()
if image_path is not None:
self.image_path = image_path
else:
self.image_path = "/home/madi/deeplearning/raspberry-pi/datasets"
pass
def gen_training_image_set(self):
imagePaths = os.listdir(self.image_path)
# loop over the input images
for imagePath in imagePaths:
# load the image, pre-process it, and store it in the data list
imagePath = self.image_path + "/" + imagePath
print imagePath
image = cv2.imread(imagePath)
image = cv2.resize(image, (self.IMAGE_SIZE, self.IMAGE_SIZE))
image = img_to_array(image)
self.data.append(image)
# extract the class label from the image path and update the
# labels list]
if "left" in imagePath.split(os.path.sep)[-2]:
label = 1
elif "right" in imagePath.split(os.path.sep)[-2]:
label = 2
else:
label = 0
self.labels.append(label)
# scale the raw pixel intensities to the range [0, 1]
self.data = np.array(self.data, dtype="float") / 255.0
self.labels = np.array(self.labels)
def add_training_sample(self, data, label):
image = cv2.resize(data, (self.IMAGE_SIZE, self.IMAGE_SIZE))
image = img_to_array(image)
self.data.append(image)
self.labels.append(label)
def scale_and_norm_training_samples(self):
# scale the raw pixel intensities to the range [0, 1]
self.data = np.array(self.data, dtype="float") / 255.0
self.labels = np.array(self.labels)
def build_model(self):
self.model = SqueezeNet(include_top=True,
weights=None,
classes=3,
input_shape=(self.IMAGE_SIZE, self.IMAGE_SIZE,
3))
self.model.summary()
opt = Adam()
self.model.compile(loss="binary_crossentropy",
optimizer=opt,
metrics=["accuracy"])
return self.model
def train(self):
# split train and test set
(trainX, testX, trainY, testY) = train_test_split(self.data,
self.labels,
test_size=0.25,
random_state=42)
# convert the labels from integers to vectors
trainY = to_categorical(trainY, num_classes=3)
testY = to_categorical(testY, num_classes=3)
print trainX.shape
print trainY.shape
self.model.fit(trainX,
trainY,
batch_size=1,
epochs=50,
verbose=1,
validation_data=(testX, testY))
self.test()
pass
def predict(self, img_frame):
img_frame = cv2.resize(img_frame, (self.IMAGE_SIZE, self.IMAGE_SIZE))
img_frame = img_to_array(img_frame)
data = np.array([img_frame])
# scale the raw pixel intensities to the range [0, 1]
data = np.array(data, dtype="float") / 255.0
ret = self.model.predict(data)
if len(ret) > 0:
return ret[0]
pass
def save_model(self):
self.model.save("greenball_squeezenet_local.h5")
pass
def load_model(self, path):
self.model = load_model("greenball_squeezenet_local.h5")
pass
def test(self):
cnt = 0
for i in xrange(len(self.data)):
ret = self.model.predict(self.data[i])
pred = np.argmax(ret)
if pred == self.labels[i]:
cnt += 1
print "total correct number is %d" % cnt