def save_model_bin(self, binFilePath, modelPath=None):
from fxpmath import Fxp
if modelPath != None:
self.model = load_model(modelPath, compile="False")
with open(binFilePath, 'wb') as binFile:
largest_inaccuracy = 0
for layer in self.model.layers:
g = layer.get_config()
h = layer.get_weights()
print(g)
#binFile.write(json.dumps(g).encode(encoding="ascii",errors="unknown char"))
# embedding = 1 * 68 * 8
# simple rnn = 8 * 8, 8 * 8, 8 * 1
# drop out: none
# dense: 8 * 1, 1 * 1
# activation: sigmoid
for i in h:
i = np.array(i)
for index, x in np.ndenumerate(i):
print(x)
h_fxp = Fxp(x, signed=True, n_word=16, n_frac=8)
difference = abs(h_fxp.get_val() - x)
if difference > largest_inaccuracy:
largest_inaccuracy = difference
print(h_fxp.bin())
binFile.write(h_fxp.bin().encode(
encoding="ascii", errors="unknown char"))
print("largest difference")
print(str(largest_inaccuracy))
def param_convert(x, a):
y = Fxp(x,
signed=True,
n_word=a + 2,
n_frac=a,
overflow='saturate',
rounding='around')
y = y.get_val()
if (a == 100):
y = torch.from_numpy(x)
else:
y = torch.from_numpy(y)
y = y.type(torch.FloatTensor)
return y
files = [f for f in listdir('./') if ".csv" in f]
word_bits = 16
frac_bits = 11
error = []
for f_ in files:
with open(f_, 'r') as handle:
data = np.genfromtxt(handle, delimiter=',')
fxp_val = np.zeros_like(data)
fxp_bin = np.ndarray(data.shape, dtype='U16')
for line in range(len(data)):
if "bias" in f_:
fxp_sample = Fxp(data[line], True, word_bits, frac_bits)
fxp_val[line] = fxp_sample.get_val()
fxp_bin[line] = fxp_sample.bin()
error.append(
np.nan_to_num((data[line] - fxp_val[line]) / data[line]))
else:
for column in range(len(data[line])):
fxp_sample = Fxp(data[line][column], True, word_bits,
frac_bits)
fxp_val[line][column] = fxp_sample.get_val()
fxp_bin[line][column] = fxp_sample.bin()
error.append(
np.nan_to_num(
(data[line][column] - fxp_val[line][column]) /
data[line][column]))
with open('report.txt', 'a') as r:
r.write(f_ + ' min: ' + str(np.min(fxp_val)) + ' - max: ' +
from fxpmath import Fxp
import numpy as np
x = np.array([[1 / 3, 1 / 3]])
y = np.array([[1 / 3], [1 / 3]])
print("X original = ", x)
print("Y original = ", y)
print("-----------")
x_fxp = Fxp(x, signed=False, n_word=20, n_frac=15)
y_fxp = Fxp(y, signed=False, n_word=20, n_frac=15)
print("X fxp =", x_fxp.get_val())
print("Y fxp =", y_fxp.get_val())
print("-----------")
print(x_fxp.info(verbose=3))
print("-----------")
print("Dot product without scaling = ", x_fxp.get_val().dot(y_fxp.get_val()))
print("Dot prod scaled = ",
Fxp(x_fxp.get_val().dot(y_fxp.get_val()), n_word=20, n_frac=15))
print("Dot prod without specifications = ",
Fxp(x_fxp.get_val().dot(y_fxp.get_val())))
dot_fpx = Fxp(None, signed=True, n_word=20, n_frac=15)
dot_fpx.equal(x_fxp().dot(y_fxp()))