def overfeat(path=example1):
# input = iml.ImageLoader.getOutputNpArray(path, crop=True)
layerContainer = [
#3, 224, 224
conv.ConvLayer(padding=1,
filtershape=(96, 3, 7, 7),
stride_length=2,
pool=pool.PoolLayer(pool_size=(3, 3), stride_length=3),
ishape=(3, 224, 224)),
#96, 36, 36
conv.ConvLayer(padding=1,
filtershape=(256, 96, 5, 5),
stride_length=1,
pool=pool.PoolLayer(pad=1,
pool_size=(2, 2),
stride_length=2),
ishape=(96, 36, 36)),
#256, 18, 18
conv.ConvLayer(padding=1,
filtershape=(512, 256, 3, 3),
stride_length=1,
ishape=(256, 18, 18)),
#512, 18, 18
conv.ConvLayer(padding=1,
filtershape=(512, 512, 3, 3),
stride_length=1,
ishape=(256, 18, 18)),
#512, 18, 18
conv.ConvLayer(padding=1,
filtershape=(512, 512, 3, 3),
stride_length=1,
ishape=(256, 18, 18)),
#512, 18, 18
conv.ConvLayer(padding=1,
filtershape=(512, 512, 3, 3),
stride_length=1,
ishape=(256, 18, 18)),
#512, 18, 18
conv.ConvLayer(padding=1,
filtershape=(512, 512, 3, 3),
stride_length=1,
ishape=(256, 18, 18)),
]
# model.learn()
overfeat = model.Model(learning_rate=None,
dataset=None,
layerContainer=layerContainer)
output = overfeat.compute(input)
# overfeat.saveLayers(["OVconv1", "OVconv2", "OVconv3", "OVconv4", "OVconv5", "OVconv6", "OVconv7", "OVclassifier"])
overfeat.learn()
def test_learn():
layerContainer = [
#3, 150, 150
conv.ConvLayer(optimizer=adam.AdamConv(),
load_path="ff2c1",
filtershape=(32, 3, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2), stride_length=2),
ishape=(3, 150, 150)),
#32, 74, 74
conv.ConvLayer(optimizer=adam.AdamConv(),
load_path="ff2c2",
filtershape=(64, 32, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2), stride_length=2),
ishape=(32, 74, 74)),
#64, 36, 36
conv.ConvLayer(optimizer=adam.AdamConv(),
load_path="ff2c3",
filtershape=(128, 64, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2), stride_length=2),
ishape=(64, 36, 36)),
#128, 17, 17
fcnetwork.FCLayer(optimizer=adam.AdamFC(),
load_path="ff2fcn1",
arch=[36992, 512, 128],
activation_func="relu",
is_classifier=False),
softmax.SoftmaxLayer(optimizer=adam.AdamFC(),
load_path="ff2softm",
arch=[128, 5])
]
# signal.signal(signal.SIGINT, signal_handler)
## here learning rate is useless
model_FAndF = model.Model(learning_rate=0.001,
dataset=None,
layerContainer=layerContainer)
pic = iml.ImageLoader.getOutputNpArray(example1,
crop=True,
crop_size=(0, 0, 150, 150))
y = model_FAndF.compute(pic)
print(y)
def simple_adam_optimizer_test(self):
train, test_data = dataloader.load_some_flowers(5,
0,
crop_size=(0, 0, 150,
150))
x = list(train)[0][0]
c1 = conv.ConvLayer(optimizer=adam.AdamConv(),
filtershape=(32, 3, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2),
stride_length=2),
ishape=(3, 150, 150))
initFilters, initBiases = c1.getFiltersAndBiases()
out = c1.compute(x, learn=True)
false_delta = numpy.random.randn(*(32, 74, 74))
c1.learn(false_delta)
c1.modify_weights(learning_rate=0.1, batch_size=1)
finalFilters, finalBiases = c1.getFiltersAndBiases()
# print(f"init f {initFilters}")
# print(f"final f {finalFilters}")
self.assertFalse((initFilters == finalFilters).all())
self.assertFalse((initBiases == finalBiases).all())
def test_bias_gradient1(self):
train, test_data = dataloader.load_some_flowers(5,
0,
crop_size=(0, 0, 150,
150))
x = list(train)[0][0]
c1 = conv.ConvLayer(filtershape=(32, 3, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2),
stride_length=2),
ishape=(3, 150, 150))
out = c1.compute(x, learn=True)
self.assertEqual(out.shape, (32, 74, 74))
false_delta = numpy.ones((32, 74, 74))
c1.learn(false_delta)
nabla_b = c1.getNablaB()
self.assertTrue(nabla_b.shape, (32, ))
def flowerAndFun(path=example1):
input = iml.ImageLoader.getOutputNpArray(path,
crop=True,
crop_size=(0, 0, 150, 150))
layerContainer = [
#3, 150, 150
conv.ConvLayer(optimizer=adam.AdamConv(),
filtershape=(32, 3, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2), stride_length=2),
ishape=(3, 150, 150)),
#32, 74, 74
conv.ConvLayer(optimizer=adam.AdamConv(),
filtershape=(64, 32, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2), stride_length=2),
ishape=(32, 74, 74)),
#64, 36, 36
conv.ConvLayer(optimizer=adam.AdamConv(),
filtershape=(128, 64, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2), stride_length=2),
ishape=(64, 36, 36)),
#128, 17, 17
fcnetwork.FCLayer(optimizer=adam.AdamFC(), arch=[36992, 512, 128, 5])
]
learning_rate = 0.0001
model_FAndF = model.Model(learning_rate=learning_rate,
dataset=None,
layerContainer=layerContainer)
# output = model_FAndF.compute(input, learn=True)
# model_FAndF.soft_learn()
model_FAndF.test_learn(epoch=50)
def test_max_pool3d_1(self):
maxPoolLayer = pool.PoolLayer()
input = numpy.arange(start=1, stop=49, step=1).reshape(3, 4, 4)
expected_res = numpy.array([[[6, 8], [14, 16]], [[22, 24], [30, 32]],
[[38, 40], [46, 48]]])
output = maxPoolLayer.compute(input)
self.assertEqual(output.shape, (3, 2, 2))
# print(f"output = {output}")
cmp = (output == expected_res)
self.assertEqual(cmp.all(), True)
def test_derivative_with_depth3(self):
input = numpy.array([[[1, 2, 11, 0.3], [0, 4, 4, 36], [0, 12, 27, 34],
[62, 12, 11, 10]],
[[11, 58, -1, 2], [1, 36, 12, 8], [27, 0, 64, 1],
[1, -1, 4, 3]],
[[1, 3, 24, 5], [12, 2, 0, 1], [-1, 2, 3, 11],
[12, 27, 18, 2]]])
p = pool.PoolLayer()
p.compute(input)
expected_res = numpy.array([[[0, 0, 0, 0], [0, 1, 0, 1], [0, 0, 0, 1],
[1, 0, 0, 0]],
[[0, 1, 0, 0], [0, 0, 1, 0], [1, 0, 1, 0],
[0, 0, 0, 0]],
[[0, 0, 1, 0], [1, 0, 0, 0], [0, 0, 0, 0],
[0, 1, 1, 0]]])
res = p.get_derivative()
print(res)
cmp = (res == expected_res)
self.assertEqual(cmp.all(), True)
def test_get_next_delta2(self):
# with depth and filter
# input = (3, 5, 5)
# filter = (4, 3, 2, 2)
pathdir = "./tensorfiles"
filename1 = "conv_next_delta_test1"
input = numpy.array([[[0.1, 2, 0.11, 0.3, 1], [0, 0.4, 0.4, 0.36, 1],
[0, 0.12, 0.27, 0.34, -3],
[0.62, 0.12, 0.11, 10, 1],
[0, 0.56, 0.11, 0.44, 0.23]],
[[0.1, 2, 0.11, 0.3, 1], [0, 0.4, 0.4, 0.36, 1],
[0, 0.12, 0.27, 0.34, -3],
[0.62, 0.12, 0.11, 10, 1],
[0, 0.56, 0.11, 0.23, 0.44]],
[[0.1, 2, 0.11, 0.3, 1], [0, 0.4, 0.4, 0.36, 1],
[0, 0.12, 0.27, 0.34, -3],
[0.62, 0.12, 0.11, 10, 1],
[0, 0.56, 0.11, 0.23, 0.44]]])
# print(f"input shape = {input.shape}")
if not os.path.exists(os.path.join(pathdir, filename1 + ".bs1.npy")):
f = numpy.array([[[-0.13, 0.15], [-0.51, 0.62]],
[[-0.13, 0.15], [-0.51, 0.62]],
[[-0.13, 0.15], [-0.51, 0.62]]])
containerfilter = numpy.ndarray((0, 3, 2, 2))
containerfilter = numpy.insert(containerfilter, 0, f, axis=0)
containerfilter = numpy.insert(containerfilter, 0, f, axis=0)
containerfilter = numpy.insert(containerfilter, 0, f, axis=0)
containerfilter = numpy.insert(containerfilter, 0, f, axis=0)
#print(f"container filter shape = {containerfilter.shape}")
biases = numpy.zeros((4, ))
tm = TensorFileManager("./tensorfiles")
# print(f"shape = {f.shape}")
tm.save(filename1 + ".bs1", biases)
tm.save(filename1 + ".ws1", containerfilter)
l1 = conv.ConvLayer(load_path=filename1,
filtershape=(4, 3, 2, 2),
pool=pool.PoolLayer(),
activation_function="sigmoid")
res1 = l1.compute(input)
# print(f"result conv shape {res1.shape}")
ws = numpy.array([[
0.61, 0.82, 0.96, -1, 0.9, 0.71, 0.3, 0.276, 0.11, 0.12, 0.17, 0.5,
0.1, 0.2, 0.11, 0.6
],
[
0.02, -0.5, 0.23, 0.17, 0.9, 0.1, 0.4, 0.9, 0.2,
0.12, 0.11, 0.3, 0.1, 0.2, 0.7, 0.8
]])
prev_delta = numpy.array([0.25, -0.15]) #delta from the link
delta = numpy.dot(prev_delta, ws)
# print(f"delta = {delta}")
l1.learn(delta)
dx = l1.getLastDelta()
self.assertEqual(dx.shape, (3, 5, 5))
# print(dx)
self.assertEqual(numpy.isclose(dx[0][2][2], 2.837872562200001e-10),
True)
def test_learn_with_depth_and_multiple_filter(self):
# input 3, 5, 5
# filter 3, 3, 2, 2 => output 3, 4, 4
input = numpy.array([[[0.1, 2, 0.11, 0.3, 1], [0, 0.4, 0.4, 0.36, 1],
[0, 0.12, 0.27, 0.34, -3],
[0.62, 0.12, 0.11, 10, 1], [0, 56, 11, 23, 44]],
[[0.11, 0.58, -1, 2, 0.35],
[0.1, 0.36, 0.12, 0.8, 0.27],
[0.27, 0, 0.64, 1, 0.12], [1, -1, 0.4, 3, 11],
[0, 0.56, 0.11, 0.23, 0.44]],
[[1, 3, 0.24, 5, -1], [0.12, 2, 0, 1, 11],
[-0.1, 0.2, 0.3, 0.11, 0.22],
[12, 0.27, 0.18, 0.2, 0.34],
[0, 0.56, 0.11, 0.23, 0.44]]])
pathdir = "./tensorfiles"
filename = "conv_test_depth_multiple_filter"
fullypath = pathdir + "/" + filename
if not os.path.exists(os.path.join(pathdir, filename + ".bs1.npy")):
tm = TensorFileManager("./tensorfiles")
containerfilter = numpy.ndarray((0, 3, 2, 2))
f3 = numpy.array([[[0.3, 0.3], [0.3, 0.3]], [[0.3, 0.3],
[0.3, 0.3]],
[[0.3, 0.3], [0.3, 0.3]]])
containerfilter = numpy.insert(containerfilter, 0, f3, 0)
f2 = numpy.array([[[0.2, 0.2], [0.2, 0.2]], [[0.2, 0.2],
[0.2, 0.2]],
[[0.2, 0.2], [0.2, 0.2]]])
containerfilter = numpy.insert(containerfilter, 0, f2, 0)
f1 = numpy.array([[[0.1, 0.1], [0.1, 0.1]], [[0.1, 0.1],
[0.1, 0.1]],
[[0.1, 0.1], [0.1, 0.1]]])
containerfilter = numpy.insert(containerfilter, 0, f1, 0)
biases = numpy.zeros((3, ))
tm.save("conv_test_depth_multiple_filter.bs1", biases)
tm.save("conv_test_depth_multiple_filter.ws1", containerfilter)
l1 = conv.ConvLayer(load_path=filename,
filtershape=(3, 3, 2, 2),
pool=pool.PoolLayer(),
activation_function="relu")
output = l1.compute(input)
filename = "fcn_test_depth_multiple_filter"
if not os.path.exists(os.path.join(pathdir, filename + ".bs1.npy")):
tm = TensorFileManager("./tensorfiles")
ws = numpy.array([[
0.1, 0.3, 0.5, 0.12, 0.9, 0.12, 0.9, 0.10, 0.1, 0.11, 0.12,
0.13
],
[
0.34, 0.3, 0.64, 0.12, 1, 0.12, 0.1, 0.1,
0.12, 0.13, 0.15, 0.11
]])
biases = numpy.zeros((2, ))
tm.save("fcn_test_depth_multiple_filter.bs1", biases)
tm.save("fcn_test_depth_multiple_filter.ws1", ws)
l2 = fcnetwork.FCLayer(arch=[12, 2],
load_path="fcn_test_depth_multiple_filter")
expected_res = numpy.array([1, 0])
l2.compute(input=output, learn=True)
l2.learn(expected_res)
delta = l2.getLastDelta()
l1.learn(delta)
nabla_w = l1.getNablaW()
self.assertEqual(numpy.isclose(nabla_w[0][0][0][0], 4.1609570961e-09),
True)
self.assertEqual(numpy.isclose(nabla_w[1][1][0][0], 1.8135273233e-09),
True)
input = numpy.array([(0.51, 0.9, 0.88, 0.84, 0.05),
(0.4, 0.62, 0.22, 0.59, 0.1),
(0.11, 0.2, 0.74, 0.33, 0.14),
(0.47, 0.01, 0.85, 0.7, 0.09),
(0.76, 0.19, 0.72, 0.17, 0.57)])
filter = numpy.array([[-0.13, 0.15], [-0.51, 0.62]])
biases = numpy.zeros((1, ))
if not os.path.exists(os.path.join(pathdir, filename1 + ".bs1.npy")):
tm = TensorFileManager("./tensorfiles")
tm.save("convtest.bs1", biases)
tm.save("convtest.ws1", filter)
l1 = conv.ConvLayer(load_path=filename1, pool=pool.PoolLayer())
res1 = l1.compute(input)
print(res1)
filename2 = "networktest"
############# no need to build a fcn actually
ws = numpy.array([[0.61, 0.82, 0.96, -1], [0.02, -0.5, 0.23, 0.17]])
# biases = numpy.zeros((2,))
# if not os.path.exists(os.path.join(pathdir, filename2 + ".bs1.npy")):
# tm = TensorFileManager("./tensorfiles")
# tm.save("networktest.bs1", biases)
# tm.save("networktest.ws1", ws)
def flowerAndFun2(path=example1):
def signal_handler(sig, frame):
# saveModel.saveLayers(["ff2c1", "ff2c2", "ff2c3", "ff2fcn1", "ff2softm"])
pic = iml.ImageLoader.getOutputNpArray(example1,
crop=True,
crop_size=(0, 0, 150, 150))
y = saveModel.compute(pic)
saveModel.saveLayers([
"ff2c1", "d1", "ff2c2", "d2", "ff2c3", "d3", "ff2fcn1", "d4",
"ff2softm"
])
print(y)
sys.exit(0)
input = iml.ImageLoader.getOutputNpArray(path,
crop=True,
crop_size=(0, 0, 150, 150))
# layerContainer = [
# #3, 150, 150
# conv.ConvLayer(optimizer=adam.AdamConv(), filtershape=(32, 3, 3, 3), stride_length=1, pool=pool.PoolLayer(pool_size=(2, 2), stride_length=2), ishape=(3, 150, 150)),
# dropout.DropoutLayer(p=0.2, ishape=(32, 74, 74)),
# #32, 74, 74
# conv.ConvLayer(optimizer=adam.AdamConv(), filtershape=(64, 32, 3, 3), stride_length=1, pool=pool.PoolLayer(pool_size=(2, 2), stride_length=2), ishape=(32, 74, 74)),
# dropout.DropoutLayer(p=0.2, ishape=(64, 36, 36)),
# #64, 36, 36
# conv.ConvLayer(optimizer=adam.AdamConv(), filtershape=(128, 64, 3, 3), stride_length=1, pool=pool.PoolLayer(pool_size=(2, 2), stride_length=2), ishape=(64, 36, 36)),
# dropout.DropoutLayer(p=0.2, ishape=(128, 17, 17)),
# #128, 17, 17
# fcnetwork.FCLayer(optimizer=adam.AdamFC(), arch=[36992, 512, 128], activation_func="relu", is_classifier=False),
# dropout.DropoutLayer(p=0.2, ishape=(128,)),
# softmax.SoftmaxLayer(optimizer=adam.AdamFC(), arch=[128, 5])
# ]
# # load net
layerContainer = [
#3, 150, 150
conv.ConvLayer(optimizer=adam.AdamConv(),
load_path="ff2c1",
filtershape=(32, 3, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2), stride_length=2),
ishape=(3, 150, 150)),
dropout.DropoutLayer(p=0.2, ishape=(32, 74, 74)),
#32, 74, 74
conv.ConvLayer(optimizer=adam.AdamConv(),
load_path="ff2c2",
filtershape=(64, 32, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2), stride_length=2),
ishape=(32, 74, 74)),
dropout.DropoutLayer(p=0.2, ishape=(64, 36, 36)),
#64, 36, 36
conv.ConvLayer(optimizer=adam.AdamConv(),
load_path="ff2c3",
filtershape=(128, 64, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2), stride_length=2),
ishape=(64, 36, 36)),
dropout.DropoutLayer(p=0.2, ishape=(128, 17, 17)),
#128, 17, 17
fcnetwork.FCLayer(optimizer=adam.AdamFC(),
load_path="ff2fcn1",
arch=[36992, 512, 128],
activation_func="relu",
is_classifier=False),
dropout.DropoutLayer(p=0.2, ishape=(128, )),
softmax.SoftmaxLayer(optimizer=adam.AdamFC(),
load_path="ff2softm",
arch=[128, 5])
]
signal.signal(signal.SIGINT, signal_handler)
## here learning rate is useless
model_FAndF = model.Model(learning_rate=0.001,
dataset=None,
layerContainer=layerContainer)
saveModel = model_FAndF
# saveModel.saveLayers(["ff2c1", "ff2c2", "ff2c3", "ff2fcn1", "ff2softm"])
model_FAndF.test_learn(epoch=50)
def flowerAndFunModel(path=example1):
#it crops by 224x224 by default
print(f"Model load this picture as input: {path}")
input = iml.ImageLoader.getOutputNpArray(path,
crop=True,
crop_size=(0, 0, 150, 150))
dir = "./tensorfiles"
if not os.path.exists(dir + "/" + "ff2c1" + ".ws1.npy"):
layerContainer = [
#3, 150, 150
conv.ConvLayer(optimizer=adam.AdamConv(),
filtershape=(32, 3, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2),
stride_length=2),
ishape=(3, 150, 150)),
#32, 74, 74
conv.ConvLayer(optimizer=adam.AdamConv(),
filtershape=(64, 32, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2),
stride_length=2),
ishape=(32, 74, 74)),
#64, 36, 36
conv.ConvLayer(optimizer=adam.AdamConv(),
filtershape=(128, 64, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2),
stride_length=2),
ishape=(64, 36, 36)),
#128, 17, 17
fcnetwork.FCLayer(optimizer=adam.AdamFC(),
arch=[36992, 512, 128],
activation_func="relu",
is_classifier=False),
softmax.SoftmaxLayer(optimizer=adam.AdamFC(), arch=[128, 5])
]
ffM = model.Model(learning_rate=None,
dataset=None,
layerContainer=layerContainer)
ffM.saveLayers(["ff2c1", "ff2c2", "ff2c3", "ff2fcn1", "ff2softm"])
layerContainer = [
conv.ConvLayer(optimizer=adam.AdamConv(),
load_path="ff2c1",
filtershape=(32, 3, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2), stride_length=2),
ishape=(3, 150, 150)),
#32, 74, 74
conv.ConvLayer(optimizer=adam.AdamConv(),
load_path="ff2c2",
filtershape=(64, 32, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2), stride_length=2),
ishape=(32, 74, 74)),
#64, 36, 36
conv.ConvLayer(optimizer=adam.AdamConv(),
load_path="ff2c3",
filtershape=(128, 64, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(2, 2), stride_length=2),
ishape=(64, 36, 36)),
#128, 17, 17
fcnetwork.FCLayer(optimizer=adam.AdamFC(),
load_path="ff2fcn1",
arch=[36992, 512, 128],
activation_func="relu",
is_classifier=False),
softmax.SoftmaxLayer(optimizer=adam.AdamFC(),
load_path="ff2softm",
arch=[128, 5])
]
ffM = model.Model(learning_rate=None,
dataset=None,
layerContainer=layerContainer)
try:
output = ffM.compute(input)
except:
print("error occured in zf5 model")
return "error"
return return_response(output)
# print(zf5model())
# overfeat()
# a = numpy.array(
# [[1, 2, 3, 4, 5, 6],
# [1, 2, 3, 4, 5, 6],
# [1, 2, 3, 4, 5, 6],
# [1, 2, 3, 4, 5, 6],
# [1, 2, 3, 4, 5, 6],
# [1, 2, 3, 4, 5, 6]])
def zf5model(path=example1):
# from this architecture:
# https://www.researchgate.net/figure/Architecture-of-ZF-model-An-3-channels-image-with-224224-is-as-the-input-It-is_fig5_318577329
# image = iml.ImageLoader.getOutputNpArray(rp_dataset + "daisy/" + "5547758_eea9edfd54_n.jpg", gray=True)
# image = iml.ImageLoader.getCropedImage(rp_dataset + "daisy/" + "5547758_eea9edfd54_n.jpg")#.getOutputNpArray(rp_dataset + "daisy/" + "5547758_eea9edfd54_n.jpg")
#it crops by 224x224 by default
print(f"Model load this picture as input: {path}")
input = iml.ImageLoader.getOutputNpArray(path,
crop=True,
crop_size=(0, 0, 224, 224))
dir = "./tensorfiles"
if not os.path.exists(dir + "/" + "conv1" + ".ws1.npy"):
layerContainer = [
#3, 224, 224
conv.ConvLayer(padding=1,
filtershape=(96, 3, 7, 7),
stride_length=2,
pool=pool.PoolLayer(pad=1,
pool_size=(3, 3),
stride_length=2),
ishape=(3, 224, 224)),
#96, 55, 55
conv.ConvLayer(filtershape=(256, 96, 5, 5),
stride_length=2,
pool=pool.PoolLayer(pad=1,
pool_size=(3, 3),
stride_length=2),
ishape=(96, 55, 55)),
#256, 26, 26
conv.ConvLayer(padding=1,
filtershape=(384, 256, 3, 3),
stride_length=1,
ishape=(256, 26, 26)),
#384, 13, 13
conv.ConvLayer(padding=1,
filtershape=(384, 384, 3, 3),
stride_length=1,
ishape=(384, 13, 13)),
#384, 13, 13
conv.ConvLayer(padding=1,
filtershape=(256, 384, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(3, 3),
stride_length=2),
ishape=(384, 13, 13)),
#do use he initialization here
fcnetwork.FCLayer(arch=[9216, 4096, 4096, 5])
]
zf5 = model.Model(learning_rate=None,
dataset=None,
layerContainer=layerContainer)
# output = zf5.compute(input)
# print(f"first output = {output}")
zf5.saveLayers(
["conv1", "conv2", "conv3", "conv4", "conv5", "classifier"])
layerContainer = [
#3, 224, 224
conv.ConvLayer(load_path="conv1"),
#96, 55, 55
conv.ConvLayer(load_path="conv2"),
#256, 26, 26
conv.ConvLayer(load_path="conv3"),
#384, 13, 13
conv.ConvLayer(load_path="conv4"),
#384, 13, 13
conv.ConvLayer(load_path="conv5"),
#do use he initialization here
fcnetwork.FCLayer(arch=[9216, 4096, 4096, 5], load_path="classifier")
]
zf5 = model.Model(learning_rate=None,
dataset=None,
layerContainer=layerContainer)
# output = zf5.compute(input)
# res = zf5.learn()
# print(f"snd output = {res}")
try:
output = zf5.compute(input)
except:
print("error occured in zf5 model")
return "error"
# print(return_response(output))
return return_response(output)
def load(self, file):
with open(file, 'rb') as f:
self.__layers = pickle.load(f)
if __name__ == '__main__':
import numpy as np
import conv
import act
import pool
import dense
# build
net = Network()
net.add_layer(conv.ConvLayer(28, 28, 1, 3, 10))
net.add_layer(act.ReLULayer())
net.add_layer(pool.PoolLayer(f=2, stride=2))
net.add_layer(act.ReLULayer())
net.add_layer(conv.ConvLayer(13, 13, 10, 4, 16))
net.add_layer(act.ReLULayer())
net.add_layer(pool.PoolLayer(f=2, stride=2))
net.add_layer(act.ReLULayer())
net.add_layer(conv.ConvLayer(5, 5, 16, 2, 10))
net.add_layer(act.ReLULayer())
net.add_layer(pool.PoolLayer(f=4, stride=1))
net.add_layer(dense.DenseLayer(10, 10))
net.add_layer(act.SigmoidLayer())
# train
i = np.random.randn(1, 28, 28, 1)
o = np.array([0, 0, 1, 0, 0, 0, 0, 0, 0, 0]).reshape(10, 1)
net.train(i, o, 100, 0.1)
print(np.argmax(net.forward(i)) + 1)
# image = iml.ImageLoader.getOutputNpArray(rp_dataset + "daisy/" + "5547758_eea9edfd54_n.jpg", gray=True)
# image = iml.ImageLoader.getCropedImage(rp_dataset + "daisy/" + "5547758_eea9edfd54_n.jpg")#.getOutputNpArray(rp_dataset + "daisy/" + "5547758_eea9edfd54_n.jpg")
input = iml.ImageLoader.getOutputNpArray(image_path=rp_dataset + "daisy/" +
"5547758_eea9edfd54_n.jpg",
crop=True)
# from this architecture:
# https://www.researchgate.net/figure/Architecture-of-ZF-model-An-3-channels-image-with-224224-is-as-the-input-It-is_fig5_318577329
layerContainer = [
conv.ConvLayer(padding=1,
filtershape=(96, 3, 7, 7),
stride_length=2,
pool=pool.PoolLayer(pad=1,
pool_size=(3, 3),
stride_length=2)),
conv.ConvLayer(filtershape=(256, 96, 5, 5),
stride_length=2,
pool=pool.PoolLayer(pad=1,
pool_size=(3, 3),
stride_length=2)),
conv.ConvLayer(padding=1, filtershape=(384, 256, 3, 3), stride_length=1),
conv.ConvLayer(padding=1, filtershape=(384, 384, 3, 3), stride_length=1),
conv.ConvLayer(padding=1,
filtershape=(256, 384, 3, 3),
stride_length=1,
pool=pool.PoolLayer(pool_size=(3, 3), stride_length=2)),
fcnetwork.FCLayer(arch=[9216, 4096, 5])
]