def create_stimuli(root, ksize, stride, bitwidth_data_in, bitwidth_data_out,
bitwidth_weights, channel_in, channel_out, width, height):
a_rand = random_fixed_array((1, channel_in, height, width),
bitwidth_data_in)
a_in = v_to_fixedint(a_rand)
np.savetxt(join(root, "gen", f"input_{ksize}_{stride}_{channel_in}.csv"),
flatten(a_in),
delimiter=", ",
fmt="%3d")
a_weights_rand = random_fixed_array(
(channel_out, channel_in, ksize, ksize), bitwidth_weights)
a_bias_rand = random_fixed_array((channel_out, ), bitwidth_weights)
# weights and bias to txt
weights_to_files(a_weights_rand, a_bias_rand,
f"conv_{ksize}_{stride}_{channel_in}", join(root, "gen"))
# assign the outputs
conv_out = v_to_fixedint(
conv(a_rand, a_weights_rand, a_bias_rand, (ksize, stride),
bitwidth_data_out.as_tuple))
filename = join(root, "gen", f"output_{ksize}_{stride}_{channel_in}.csv")
with open(filename, "w") as outfile:
np.savetxt(outfile, flatten(conv_out), delimiter=", ", fmt="%3d")
def create_stimuli(root, ksize, stride, bitwidth, shape):
a_rand = random_fixed_array(shape, bitwidth)
a_in = v_to_fixedint(a_rand)
np.savetxt(join(root, "src", "input_%d_%d.csv" % (ksize, stride)),
flatten(a_in),
delimiter=", ",
fmt="%3d")
# assign the outputs
filename = join(root, "src", "output_%d_%d.csv" % (ksize, stride))
max_out = v_to_fixedint(max_pool(a_rand, ksize, stride))
with open(filename, "w") as outfile:
np.savetxt(outfile, flatten(max_out), delimiter=", ", fmt="%3d")
def create_stimuli(root, ksize, stride, total_bits, frac_bits, channel,
width, height):
int_bits = total_bits - frac_bits
a_rand = np.random.randint(2 ** total_bits,
size=(1, channel, height, width))
np.savetxt(join(root, "src", "input_%d_%d.csv" % (ksize, stride)),
flatten(a_rand), delimiter=", ", fmt="%3d")
# assign the outputs
filename = join(root, "src", "output_%d_%d.csv" % (ksize, stride))
max_out = max_pool(a_rand, ksize, stride, (int_bits, frac_bits))
with open(filename, "w") as outfile:
np.savetxt(outfile, flatten(max_out), delimiter=", ", fmt="%3d")
def create_arrays(root, w, h, ch):
id_ = "one" if ch == 1 else "multiple"
a_rand = random_fixed_array((1, ch, h, w), 8, 0)
a_in = v_float2fixedint(a_rand, 8, 0)
np.savetxt(join(root, "src", "input_%s.csv" % id_),
flatten(a_in),
delimiter=", ",
fmt="%3d")
a_out = v_float2fixedint(zero_pad(a_rand), 8, 0)
np.savetxt(join(root, "src", "output_%s.csv" % id_),
flatten(a_out),
delimiter=", ",
fmt="%3d")
def create_stimuli(root, model_name):
model = onnx.load(join(root, model_name))
shape = cnn_onnx.parse_param.get_input_shape(model)
in_ = np.random.randint(256, size=shape)
out_ = cnn_onnx.inference.numpy_inference(model, in_)
np.savetxt(join(root, "input.csv"),
flatten(in_),
delimiter=", ",
fmt="%3d")
np.savetxt(join(root, "output.csv"), out_, delimiter=", ", fmt="%3d")
def create_stimuli(root, w, h, ch, total_bits, frac_bits):
int_bits = total_bits - frac_bits
a_rand = np.random.randint(256, size=(1, ch, h, w), dtype=np.uint8)
np.savetxt(join(root, "src", "input.csv"),
flatten(a_rand),
delimiter=", ",
fmt="%3d")
a_out = avg_pool(a_rand, (int_bits, frac_bits))
np.savetxt(join(root, "src", "output.csv"),
a_out,
delimiter=", ",
fmt="%3d")
def create_stimuli(root, model_name):
model = onnx.load(join(root, model_name))
shape = cnn_onnx.parse_param.get_input_shape(model)
a_rand = random_fixed_array(shape, Bitwidth(8, 8, 0), signed=False)
a_in = v_to_fixedint(a_rand)
a_out = v_to_fixedint(cnn_onnx.inference.numpy_inference(model, a_rand))
np.savetxt(join(root, "input.csv"),
flatten(a_in),
delimiter=", ",
fmt="%3d")
np.savetxt(join(root, "output.csv"), a_out, delimiter=", ", fmt="%3d")
def create_stimuli(root, shape, bitwidth):
a_rand = random_fixed_array(shape, bitwidth)
a_in = v_to_fixedint(a_rand)
np.savetxt(join(root, "src", "input.csv"),
flatten(a_in),
delimiter=", ",
fmt="%3d")
a_out = v_to_fixedint(avg_pool(a_rand))
np.savetxt(join(root, "src", "output.csv"),
a_out,
delimiter=", ",
fmt="%3d")
def create_stimuli(root, ksize, stride, total_bits_data, frac_bits_data_in,
frac_bits_data_out, total_bits_weight, frac_bits_weight,
channel_in, channel_out, width, height):
int_bits_data_in = total_bits_data - frac_bits_data_in
int_bits_data_out = total_bits_data - frac_bits_data_out
a_rand = np.random.randint(256,
size=(1, channel_in, height, width),
dtype=np.uint8)
np.savetxt(join(root, "gen", f"input_{ksize}_{stride}_{channel_in}.csv"),
flatten(a_rand),
delimiter=", ",
fmt="%3d")
int_bits_weight = total_bits_weight - frac_bits_weight
scale = 2**frac_bits_weight
a_weights_rand = np.random.randint(-2**7,
2**7 - 1,
size=(channel_out, channel_in, ksize,
ksize),
dtype=np.int8)
a_bias_rand = np.random.randint(-2**7,
2**7 - 1,
size=(channel_out, ),
dtype=np.int8)
# weights and bias to txt
weights_to_files(a_weights_rand / scale, a_bias_rand / scale,
(int_bits_weight, frac_bits_weight),
f"conv_{ksize}_{stride}_{channel_in}", join(root, "gen"))
# assign the outputs
conv_out = conv(a_rand, a_weights_rand, a_bias_rand, (ksize, stride),
(int_bits_data_in, frac_bits_data_in, int_bits_data_out,
frac_bits_data_out, int_bits_weight, frac_bits_weight))
filename = join(root, "gen", f"output_{ksize}_{stride}_{channel_in}.csv")
with open(filename, "w") as outfile:
np.savetxt(outfile, flatten(conv_out), delimiter=", ", fmt="%3d")
def create_stimuli(root, model_name):
model = onnx.load(join(root, model_name))
shape = cnn_onnx.parse_param.get_input_shape(model)
a_rand = random_fixed_array(shape, Bitwidth(8, 8, 0), signed=False)
a_in = v_to_fixedint(a_rand)
a_out = v_to_fixedint(cnn_onnx.inference.numpy_inference(model, a_rand))
# ONNX runtime prediction, TODO: doesn't work right now
# https://github.com/microsoft/onnxruntime/issues/2964
# sess = rt.InferenceSession(join(root, model_name))
# input_name = sess.get_inputs()[0].name
# pred_onnx = sess.run(None, {input_name: in_.astype(np.float32)})[0]
# print(pred_onnx)
np.savetxt(join(root, "input.csv"), flatten(a_in),
delimiter=", ", fmt="%3d")
np.savetxt(join(root, "output.csv"), a_out,
delimiter=", ", fmt="%3d")
def create_stimuli(root, model_name):
model = onnx.load(join(root, model_name))
shape = cnn_onnx.parse_param.get_input_shape(model)
in_ = np.random.randint(256, size=shape, dtype=np.uint8)
out_ = cnn_onnx.inference.numpy_inference(model, in_)
# ONNX runtime prediction, TODO: doesn't work right now
# https://github.com/microsoft/onnxruntime/issues/2964
# sess = rt.InferenceSession(join(root, model_name))
# input_name = sess.get_inputs()[0].name
# pred_onnx = sess.run(None, {input_name: in_.astype(np.float32)})[0]
# print(pred_onnx)
np.savetxt(join(root, "input.csv"),
flatten(in_),
delimiter=", ",
fmt="%3d")
np.savetxt(join(root, "output.csv"), out_, delimiter=", ", fmt="%3d")