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 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 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]])