def runTensorflow(self, images):
with tf.Session() as sess:
with tf.device("/cpu:0"):
# Tensorflow does not know how to release /GPU:0 resources
# without process termination
# Issue #1727 in the tensorflow/tensorflow git repository
# Letting it just use CPU for this forward run instead
img = tf.placeholder(tf.float32,
shape=[1, 224, 224, 3],
name="images")
model = vgg.VGG()
model.buildGraph(img, network_version="VGG19")
images = images.reshape((1, 224, 224, 3))
sess.run(tf.global_variables_initializer())
# print self.model.layers.keys()
self.output = sess.run(
[model.layers[_] for _ in self.testLayers],
feed_dict={img: images})
sess.close()
return
def runGraph(self, im):
with self.test_session() as sess, tf.device("/gpu:0"):
img = tf.placeholder(tf.float32, im_data.shape, name="images")
net = vgg.VGG("vgg16test_1")
net.buildGraph(img,
train=False,
weightsPath=s.DEF_FRCNN_WEIGHTS_PATH,
biasesPath=s.DEF_FRCNN_BIASES_PATH,
network_version="VGG16CONV")
sess.run(tf.global_variables_initializer())
output = sess.run(net.layers['relu4_1'],
feed_dict={img: im_data})
sess.close()
# Does output come in list form if only one output is produced? [probably]
# Blob name is conv4_1, not relu4_1; relu is done in-place by caffe
return output
def runGraph(self):
with self.test_session() as sess, tf.device("/gpu:0"):
try:
conv4_1_in = self.reference_activations['conv4_1']
except KeyError:
print(
"Warning: conv4_1 not found in reference_activations. Something \
wrong with .npz file")
net = vgg.VGG(namespace="vgg16test_2")
net.buildGraph(tf.placeholder(dtype=tf.float32),
train=False,
weightsPath=s.DEF_FRCNN_WEIGHTS_PATH,
biasesPath=s.DEF_FRCNN_BIASES_PATH,
network_version="VGG16CONV")
conv4_1 = net.layers['conv4_1']
sess.run(tf.global_variables_initializer())
output = sess.run(net.layers['relu5_3'],
feed_dict={conv4_1: conv4_1_in})
sess.close()
return output
if shuffle:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# Creating PT data samplers and loaders:
train_sampler = SubsetRandomSampler(train_indices)
valid_sampler = SubsetRandomSampler(val_indices)
trainloader = dataset.CSILoader(data, opt,sampler=train_sampler)
testloader = dataset.CSILoader(data, opt,sampler=valid_sampler)
print('==> Building model..')
# net = VGG('VGG19')
net = vgg.VGG('VGG11')
# net = ResNet18()
# net = GoogLeNet()
# net = DenseNet121()
# net = ResNeXt29_2x64d()
# net = MobileNet()
# net = DPN92()
# net = ShuffleNetG2()
# net = SENet18()
# result_folder = './results/'
# if not os.path.exists(result_folder):
# os.makedirs(result_folder)
#
# logname = result_folder + net.__class__.__name__ + '_' + opt.sess + '_' + str(opt.seed) + '.csv'
# dataset_size = len(data)
# indices = list(range(dataset_size))
# split = 0.8
# split = int(np.floor(split * dataset_size))
# print(split)
# if shuffle:
# np.random.seed(opt.seed)
# np.random.shuffle(indices)
#train_indices, val_indices = indices[split:], indices[:split]
# train_indices, val_indices = indices[:split], indices[split:]
trainloader = dataset.CSILoader(data_train, opt, shuffle=True)
testloader = dataset.CSILoader(data_test, opt, shuffle=True)
print('==> Building model..')
net = vgg.VGG('VGG11', linear_in=2048)
# net = ResNet18()
#net = LeNet.LeNet(in_channel=3, linear_in=9216)
# net = DenseNet121()
# net = ResNeXt29_2x64d()
# net = MobileNet()
# net = DPN92()
# net = ShuffleNetG2()
# net = SENet18()
# result_folder = './results/'
# if not os.path.exists(result_folder):
# os.makedirs(result_folder)
#
# logname = result_folder + net.__class__.__name__ + '_' + opt.sess + '_' + str(opt.seed) + '.csv'
def testSave(self):
# Layers to be tested
testLayers = ['conv1_1', 'conv2_1', 'conv3_3', 'conv4_3', 'fc6', 'fc7']
img = tf.placeholder(tf.float32, [1, 224, 224, 3], name="images")
model = vgg.VGG("frcnn")
model.buildGraph(img,
network_version="VGG16",
weightsPath=s.DEF_FRCNN_WEIGHTS_PATH,
biasesPath=s.DEF_FRCNN_BIASES_PATH)
testLayerVariables = []
for layer in testLayers:
testLayerVariables.append("frcnn/" + layer + "/Weights:0")
testLayerVariables.append("frcnn/" + layer + "/Bias:0")
# Find the weights and biases for several layers in a model
weightsFn = None
biasFn = None
try:
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
output = sess.run(testLayerVariables)
randNum = None
while True:
# Try different file names until you find one that isn't taken
randNum = random.random()
weightsFn = str(randNum) + "testWeights"
biasFn = str(randNum) + "testBias"
if os.path.isfile("models/" +
weightsFn) or os.path.isfile("models/" +
biasFn):
continue
else:
model.save(weightsFn, biasFn)
break
# Load the weights/biases into a new model
model2 = vgg.VGG()
model2.buildGraph(img,
train=False,
weightsPath="models/" + weightsFn + ".npz",
biasesPath="models/" + biasFn + ".npz")
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
output2 = sess.run(testLayerVariables)
self.assertEqual(len(output2),
2 * len(testLayers),
msg="Incorrect number of output layers")
# Check to make sure that the values are the same
for i, var in enumerate(output2):
np.testing.assert_equal(
output[i],
output2[i],
err_msg="Output number %i was not equal" % i)
finally:
if weightsFn is not None and os.path.isfile("models/" + weightsFn +
".npz"):
os.remove("models/" + weightsFn + ".npz")
if biasFn is not None and os.path.isfile("models/" + biasFn +
".npz"):
os.remove("models/" + biasFn + ".npz")
best_acc = 0
use_cuda = torch.cuda.is_available()
opt = parser.parse_args()
torch.manual_seed(opt.seed)
# Creating data indices for training and validation splits:
data_train = dataset.CSISet(data_x_train, data_y_train)
data_test = dataset.CSISet(data_x_test, data_y_test)
trainloader = dataset.CSILoader(data_train, opt, shuffle=True)
testloader = dataset.CSILoader(data_test, opt, shuffle=True)
print('==> Building model..')
# net = VGG('VGG19')
net = vgg.VGG('VGG11', in_channels=16, num_classes=9, linear_in=1536)
# net = ResNet18()
# net = GoogLeNet()
# net = DenseNet121()
# net = ResNeXt29_2x64d()
# net = MobileNet()
# net = DPN92()
# net = ShuffleNetG2()
# net = SENet18()
# result_folder = './results/'
# if not os.path.exists(result_folder):
# os.makedirs(result_folder)
#
# logname = result_folder + net.__class__.__name__ + '_' + opt.sess + '_' + str(opt.seed) + '.csv'