def test_Distribute():
#initilalize master node
CNN = [{
"l_type": "conv",
"kernel": "W1",
"hparams": {
"stride": 1,
"pad": 0
}
}, {
"l_type": "max",
"hparams": {
"stride": 1,
"f": 2
}
}]
nodes = [{"ip": "localhost", "port": 9998}]
edge = {"ip": "localhost", "port": 9000}
image = np.array([1, 2])
master_node = Master(CNN, nodes, edge, image)
np.random.seed(1)
X1 = np.random.randn(1, 3, 3, 1)
out = vecConv(X1[0, :, :, :], kernels["W1"], {"stride": 1, "pad": 0})
out2 = master_node.thread_Compute(X1, CNN[0])
np.testing.assert_array_equal(out2, out)
def test_conv():
np.random.seed(1)
A_prev = np.random.randn(1, 3, 3, 3)
hparameters = {"pad": 0, "stride": 1}
w = np.ones((2, 2, 3, 1))
b = np.zeros(((1, 1, 1, 1)))
c_out = conv_forward(A_prev, w, b, hparameters)
print(c_out.shape)
print(c_out[0, :, :, :])
v_out = vecConv(A_prev[0, :, :, :], w, hparameters)
print(v_out.shape)
print(v_out)
np.testing.assert_array_equal(v_out, c_out[0, :, :, :])
def layerResult(self, layer, X, pos):
"""
layer-dictonary l_type,kernel,hparams
X-numpy array(n_H_prev, n_W_prev, n_C_prev)
pooling-
X-numpy array(n_H_prev, n_W_prev, n_C_prev)
hparameters-"f" and "stride"
Pooling(X,hparameters,mode="max")
conv
vecConv(X,kernel,hparameters):
X- numpy arrya shape (n_H_prev, n_W_prev, n_C_prev)
kernel-numpy array of shape (f, f, n_C_prev, n_C)
hparameters-- python dictinory containing stride and pad
"""
if (layer["l_type"] == "conv"):
w = kernels[layer["kernel"]]
hparam = layer["hparams"]
return vecConv(X, w[:, :, :, pos[0]:pos[1]], hparam)
else:
hparam = layer["hparams"]
mode = layer["l_type"]
#batch size of 1
return Pooling(X[:, :, pos[0]:pos[1]], hparam, mode)
data_variable={"data":X,"pos":a,"layer":layer}
a=(start,end)
layer={"l_type":"conv","kernel":"W1","hparams":{"stride":1,"pad":0}}
pooling layer testing
layer2={"l_type":"max","hparams":{"stride":1,"f":2}}
"""
c, addr = s.accept()
#receive data from client
tic = time.process_time()
data_variable = receive_array(data, payload_size, c)
print('Connect with', addr, data_variable["data"].shape)
#imgout=y.conv_forward(data_variable["data"], w.W1[:,:,:,data_variable["pos"]:], w.b1[:,:,:,data_variable["pos"]:],data_variable["hpara"])
#out={"data":imgout}
X = data_variable["data"]
hparam = data_variable["layer"]["hparams"]
mode = data_variable["layer"]["l_type"]
if (mode == "conv"):
pos = data_variable["pos"]
w = kernels[data_variable["layer"]["kernel"]]
out = vecConv(X, w[:, :, :, pos[0]:pos[1]], hparam)
else:
out = Pooling(X, hparam, mode)
dout = {"data": out}
send(c, dout)
toc = time.process_time()
print("Computation time for conv part2 = " + str(1000 * (toc - tic)) +
"ms")
#send data to client
#c.send(bytes("Welcome to server",'utf-8'))
c.close()
print('Connect with', addr)
X = data_variable["data"]
index = data_variable["index"]
#FOG node computation goes here
for index, layer in enumerate(CNN[index:], index):
mode = layer["l_type"]
hparam = layer["hparams"]
#check offload
offload = False
if (offload):
#offload to the cloud
#X=
break
else:
if (mode == 'conv'):
w = kernels["kernel"]
X = vecConv(X, w, hparam)
X += kernels["bias"]
X = ActivationFunc(X, layer["act"])
else:
X = Pooling(X, hparam, mode)
#X - output send to roof node
dout = {"data": X}
send(c, dout)
toc = time.process_time()
print("Computation time FOG server = " + str(1000 * (toc - tic)) + "ms")
c.close()
from vec import vecConv
import numpy as np
import time
channel = [128, 64, 32, 16, 8, 4, 2, 1]
np.random.seed(1)
for i in channel:
x = np.random.randn(64, 64, i)
k = np.random.randn(9, 9, i, 256)
tic = time.process_time()
vecConv(x, k, {"pad": 0, "stride": 1})
toc = time.process_time()
print("i = " + str(i) + " Computation time for conv= " +
str(1000 * (toc - tic)) + "ms")
