def generate_c(self, link_idx, inp_shape):
name = self.cname + str(link_idx)
text = []
# BinaryLinear bl
l = self.bl
lName = l.name
lname=name+'_'+lName
for p in l.params():
pname=p.name
if pname == 'W':
text += [bu.np_to_uint8C(bu.binarize_real(p.data.T), lname+'_'+pname, 'col_major', pad='1')]
num_classes = p.data.shape[0]
fc_size = p.data.shape[1]
elif pname == 'b':
text += [bu.np_to_floatC(p.data, lname+'_'+pname, 'row_major')]
text = "\n".join(text)
m = 1
n = fc_size
k = num_classes
ftext = "void {name}(uint8_t* input, uint8_t* output){{\n"
ftext += " blinear_layer(input, {name}_bl_W, output, {name}_bl_b, {m}, {n}, {k}); \n}}\n\n"
ftext = ftext.format(name=name, m=m, n=n, k=k)
text += ftext
return text
def generate_c(self, link_idx, inp_shape):
name = self.cname + str(link_idx)
text = []
# BinaryLinear bl
l = self.bl
lName = l.name
lname=name+'_'+lName
for p in l.params():
pname=p.name
if pname == 'W':
text += [bu.np_to_uint8C(bu.binarize_real(p.data.T), lname+'_'+pname, 'col_major', pad='1')]
num_classes = p.data.shape[0]
fc_size = p.data.shape[1]
elif pname == 'b':
text += [bu.np_to_floatC(p.data, lname+'_'+pname, 'row_major')]
# BatchNormalization bn
l = self.bn
lName = l.name
lname=name+'_'+lName
for p in l.params():
pname=p.name
if pname == 'gamma':
text += [bu.np_to_floatC(p.data, lname+'_'+pname, 'row_major')]
elif pname == 'beta':
text += [bu.np_to_floatC(p.data, lname+'_'+pname, 'row_major')]
for p in l._persistent:
pname=p
persistent = l.__dict__[p]
if pname == 'avg_mean':
text += [bu.np_to_floatC(persistent, lname+'_mean', 'row_major')]
elif pname == 'avg_var':
text += [bu.np_to_floatC(np.sqrt(persistent, dtype=persistent.dtype), lname+'_std', 'row_major')]
text = "\n".join(text)+'\n'
m = 1
n = fc_size
k = num_classes
ftext = "void {name}(uint8_t* input, uint8_t* output){{\n"
ftext += " blinear_layer(input, {name}_bl_W, output, {name}_bl_b, {name}_bn_gamma, {name}_bn_beta, {name}_bn_mean, {name}_bn_std, {m}, {n}, {k}); \n}}\n\n"
ftext = ftext.format(name=name, m=m, n=n, k=k)
text += ftext
return text
def generate_c(self, link_idx, inp_shape):
#if not hasattr(self,'inp_shape'):
# raise Exception("no input shape found")
# return ""
w, h = inp_shape[2:4]
name = self.cname + str(link_idx)
text = []
m = 1
sw, sh = self.bconv.stride
pw, ph = self.bconv.pad
pl_w, pl_h = 1, 1
pl_sw, pl_sh = 1, 1
pl_pw, pl_ph = 0, 0
# Bconv
l = self.bconv
lname = name + '_' + l.name
for p in l.params():
pname = p.name
if pname == 'W':
num_f, n, kw, kh = p.data.shape
bin_data = bu.binarize_real(p.data).reshape(
p.data.shape[0] * p.data.shape[1], -1)
text += [
bu.np_to_uint8C(bin_data,
lname + '_' + pname,
'row_major',
pad='1')
]
elif pname == 'b':
text += [
bu.np_to_floatC(p.data, lname + '_' + pname, 'row_major')
]
# BatchNormalization bn
l = self.bn
lName = l.name
lname = name + '_' + lName
for p in l.params():
pname = p.name
if pname == 'gamma':
text += [
bu.np_to_floatC(p.data, lname + '_' + pname, 'row_major')
]
elif pname == 'beta':
text += [
bu.np_to_floatC(p.data, lname + '_' + pname, 'row_major')
]
for p in l._persistent:
pname = p
persistent = l.__dict__[p]
if pname == 'avg_mean':
text += [
bu.np_to_floatC(persistent, lname + '_mean', 'row_major')
]
elif pname == 'avg_var':
text += [
bu.np_to_floatC(
np.sqrt(persistent, dtype=persistent.dtype),
lname + '_std', 'row_major')
]
text = "\n".join(text) + '\n'
ftext = "void {name}(uint8_t* input, uint8_t* output){{\n"
ftext += " bconv_layer(input, {name}_bconv_W, output, {name}_bconv_b, {name}_bn_gamma, {name}_bn_beta, {name}_bn_mean, {name}_bn_std, {m}, {num_f}, {w}, {h}, {n}, {kw}, {kh}, {sw}, {sh}, {pw}, {ph}, {pl_w}, {pl_h}, {pl_sw}, {pl_sh}, {pl_pw}, {pl_ph});\n}}\n\n"
ftext = ftext.format(name=name,
m=m,
n=n,
w=w,
h=h,
num_f=num_f,
kw=kw,
kh=kh,
sw=sw,
sh=sh,
pw=pw,
ph=ph,
pl_w=pl_w,
pl_h=pl_h,
pl_sw=pl_sw,
pl_sh=pl_sh,
pl_pw=pl_pw,
pl_ph=pl_ph)
text += ftext
return text
