def main():
# define array
receive_bit = np.zeros(const.BITN, dtype=int)
transmit_bit = np.zeros(const.BITN, dtype=int)
receive_signal = np.zeros((const.SYMBOLN + const.GI), dtype=complex)
transmit_signal = np.zeros((const.SYMBOLN + const.GI), dtype=complex)
try:
f = open(const.FILENAME, 'w')
except IOError as err:
print("File error", str(err))
finally:
print(
"[%s] LOOPN=%d, symbol number=%d, SNR from %d~%d dB, channel=%s" %
(time.asctime(time.localtime(time.time())), const.LOOPN,
const.SYMBOLN, const.SNR_START, const.SNR_STOP, const.CHANNEL))
f.write(
"[ %s ] LOOPN=%d, symbol number=%d, SNR from %d~%d dB, channel=%s"
% (time.asctime(time.localtime(time.time())), const.LOOPN,
const.SYMBOLN, const.SNR_START, const.SNR_STOP, const.CHANNEL))
CNR = 0.0
for Eb_N0 in range(const.SNR_START, const.SNR_STOP + 1):
CNR = float(Eb_N0) + 3.0
for loop in range(const.LOOPN):
# transmission
transmitter.ofdm_transmitter(transmit_bit, transmit_signal)
if const.CHANNEL == "AWGN":
channel.AWGN(transmit_signal, receive_signal, CNR)
elif const.CHANNEL == "RAYLEIGH":
channel.Rayleigh(transmit_signal, receive_signal, CNR)
receiver.ofdm_receiver(receive_signal, receive_bit)
ber.ber(loop, transmit_bit, receive_bit, f, CNR)
f.close()
def __init__(self, ref_data, output_dim):
input_dim = ref_data.shape[1]
ref_data_sh = theano.shared(numpy.array(ref_data, dtype=numpy.float32), name='ref_data')
rng = RandomStreams()
ae_bricks = []
ae_input = ref_data_sh
ae_costs = []
for i, (idim, odim) in enumerate(zip([input_dim] + ae_dims[:-1], ae_dims)):
ae_mlp = MLP(activations=[ae_activations[i]],
dims=[idim, odim],
name='enc%i'%i)
enc = ae_mlp.apply(ae_input)
enc_n = ae_mlp.apply(ae_input + rng.normal(size=ae_input.shape, std=ae_f_noise_std))
ae_mlp_dec = MLP(activations=[ae_activations[i]],
dims=[odim, idim],
name='dec%i'%i)
dec = ae_mlp_dec.apply(enc_n)
cost = tensor.sqrt(((ae_input - dec) ** 2).sum(axis=1)).mean() + \
ae_l1_pen * abs(enc).sum(axis=1).mean()
ae_costs.append(cost)
ae_input = enc
ae_bricks = ae_bricks + [ae_mlp, ae_mlp_dec]
self.ae_costs = ae_costs
ref_data_enc = ae_input
# Construct the model
j = tensor.lvector('j')
r = ref_data_enc[j, :]
x = tensor.fmatrix('x')
y = tensor.ivector('y')
# input_dim must be nr
mlp = MLP(activations=activation_functions,
dims=[ae_dims[-1]] + hidden_dims + [n_inter], name='inter_gen')
mlp2 = MLP(activations=activation_functions_2 + [None],
dims=[n_inter] + hidden_dims_2 + [output_dim],
name='end_mlp')
inter_weights = mlp.apply(r)
if inter_bias == None:
ibias = Bias(n_inter)
ibias.biases_init = Constant(0)
ibias.initialize()
inter = ibias.apply(tensor.dot(x, inter_weights))
else:
inter = tensor.dot(x, inter_weights) - inter_bias
inter = inter_act_fun.apply(inter)
final = mlp2.apply(inter)
cost = Softmax().categorical_cross_entropy(y, final)
confidence = Softmax().apply(final)
pred = final.argmax(axis=1)
# error_rate = tensor.neq(y, pred).mean()
ber = balanced_error_rate.ber(y, pred)
# Initialize parameters
for brick in ae_bricks + [mlp, mlp2]:
brick.weights_init = IsotropicGaussian(0.01)
brick.biases_init = Constant(0.001)
brick.initialize()
# apply regularization
cg = ComputationGraph([cost, ber])
if r_dropout != 0:
# - dropout on input vector r : r_dropout
cg = apply_dropout(cg, [r], r_dropout)
if x_dropout != 0:
cg = apply_dropout(cg, [x], x_dropout)
if s_dropout != 0:
# - dropout on intermediate layers of first mlp : s_dropout
s_dropout_vars = list(set(VariableFilter(bricks=[Tanh], name='output')
(ComputationGraph([inter_weights])))
- set([inter_weights]))
cg = apply_dropout(cg, s_dropout_vars, s_dropout)
if i_dropout != 0:
# - dropout on input to second mlp : i_dropout
cg = apply_dropout(cg, [inter], i_dropout)
if a_dropout != 0:
# - dropout on hidden layers of second mlp : a_dropout
a_dropout_vars = list(set(VariableFilter(bricks=[Tanh], name='output')
(ComputationGraph([final])))
- set([inter_weights]) - set(s_dropout_vars))
cg = apply_dropout(cg, a_dropout_vars, a_dropout)
if r_noise_std != 0:
cg = apply_noise(cg, [r], r_noise_std)
if w_noise_std != 0:
# - apply noise on weight variables
weight_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, weight_vars, w_noise_std)
[cost_reg, ber_reg] = cg.outputs
if s_l1pen != 0:
s_weights = VariableFilter(bricks=mlp.linear_transformations, roles=[WEIGHT])(cg)
cost_reg = cost_reg + s_l1pen * sum(abs(w).sum() for w in s_weights)
if i_l1pen != 0:
cost_reg = cost_reg + i_l1pen * abs(inter).sum()
if a_l1pen != 0:
a_weights = VariableFilter(bricks=mlp2.linear_transformations, roles=[WEIGHT])(cg)
cost_reg = cost_reg + a_l1pen * sum(abs(w).sum() for w in a_weights)
self.cost = cost
self.cost_reg = cost_reg
self.ber = ber
self.ber_reg = ber_reg
self.pred = pred
self.confidence = confidence
def __init__(self, ref_data, output_dim):
input_dim = ref_data.shape[1]
ref_data_sh = theano.shared(numpy.array(ref_data, dtype=numpy.float32),
name='ref_data')
rng = RandomStreams()
ae_bricks = []
ae_input = ref_data_sh
ae_costs = []
for i, (idim,
odim) in enumerate(zip([input_dim] + ae_dims[:-1], ae_dims)):
ae_mlp = MLP(activations=[ae_activations[i]],
dims=[idim, odim],
name='enc%i' % i)
enc = ae_mlp.apply(ae_input)
enc_n = ae_mlp.apply(
ae_input + rng.normal(size=ae_input.shape, std=ae_f_noise_std))
ae_mlp_dec = MLP(activations=[ae_activations[i]],
dims=[odim, idim],
name='dec%i' % i)
dec = ae_mlp_dec.apply(enc_n)
cost = tensor.sqrt(((ae_input - dec) ** 2).sum(axis=1)).mean() + \
ae_l1_pen * abs(enc).sum(axis=1).mean()
ae_costs.append(cost)
ae_input = enc
ae_bricks = ae_bricks + [ae_mlp, ae_mlp_dec]
self.ae_costs = ae_costs
ref_data_enc = ae_input
# Construct the model
j = tensor.lvector('j')
r = ref_data_enc[j, :]
x = tensor.fmatrix('x')
y = tensor.ivector('y')
# input_dim must be nr
mlp = MLP(activations=activation_functions,
dims=[ae_dims[-1]] + hidden_dims + [n_inter],
name='inter_gen')
mlp2 = MLP(activations=activation_functions_2 + [None],
dims=[n_inter] + hidden_dims_2 + [output_dim],
name='end_mlp')
inter_weights = mlp.apply(r)
if inter_bias == None:
ibias = Bias(n_inter)
ibias.biases_init = Constant(0)
ibias.initialize()
inter = ibias.apply(tensor.dot(x, inter_weights))
else:
inter = tensor.dot(x, inter_weights) - inter_bias
inter = inter_act_fun.apply(inter)
final = mlp2.apply(inter)
cost = Softmax().categorical_cross_entropy(y, final)
confidence = Softmax().apply(final)
pred = final.argmax(axis=1)
# error_rate = tensor.neq(y, pred).mean()
ber = balanced_error_rate.ber(y, pred)
# Initialize parameters
for brick in ae_bricks + [mlp, mlp2]:
brick.weights_init = IsotropicGaussian(0.01)
brick.biases_init = Constant(0.001)
brick.initialize()
# apply regularization
cg = ComputationGraph([cost, ber])
if r_dropout != 0:
# - dropout on input vector r : r_dropout
cg = apply_dropout(cg, [r], r_dropout)
if x_dropout != 0:
cg = apply_dropout(cg, [x], x_dropout)
if s_dropout != 0:
# - dropout on intermediate layers of first mlp : s_dropout
s_dropout_vars = list(
set(
VariableFilter(bricks=[Tanh], name='output')
(ComputationGraph([inter_weights]))) -
set([inter_weights]))
cg = apply_dropout(cg, s_dropout_vars, s_dropout)
if i_dropout != 0:
# - dropout on input to second mlp : i_dropout
cg = apply_dropout(cg, [inter], i_dropout)
if a_dropout != 0:
# - dropout on hidden layers of second mlp : a_dropout
a_dropout_vars = list(
set(
VariableFilter(bricks=[Tanh], name='output')
(ComputationGraph([final]))) - set([inter_weights]) -
set(s_dropout_vars))
cg = apply_dropout(cg, a_dropout_vars, a_dropout)
if r_noise_std != 0:
cg = apply_noise(cg, [r], r_noise_std)
if w_noise_std != 0:
# - apply noise on weight variables
weight_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, weight_vars, w_noise_std)
[cost_reg, ber_reg] = cg.outputs
if s_l1pen != 0:
s_weights = VariableFilter(bricks=mlp.linear_transformations,
roles=[WEIGHT])(cg)
cost_reg = cost_reg + s_l1pen * sum(
abs(w).sum() for w in s_weights)
if i_l1pen != 0:
cost_reg = cost_reg + i_l1pen * abs(inter).sum()
if a_l1pen != 0:
a_weights = VariableFilter(bricks=mlp2.linear_transformations,
roles=[WEIGHT])(cg)
cost_reg = cost_reg + a_l1pen * sum(
abs(w).sum() for w in a_weights)
self.cost = cost
self.cost_reg = cost_reg
self.ber = ber
self.ber_reg = ber_reg
self.pred = pred
self.confidence = confidence
def __init__(self, ref_data, output_dim):
ref_data_sh = theano.shared(numpy.array(ref_data, dtype=numpy.float32),
name='ref_data')
# Construct the model
j = tensor.lvector('j')
x = tensor.fmatrix('x')
y = tensor.ivector('y')
last_outputs = []
s_dropout_vars = []
r_dropout_vars = []
i_dropout_vars = []
for i in range(nparts):
fs = numpy.random.binomial(1,
part_r_proba,
size=(ref_data.shape[1], ))
input_dim = int(fs.sum())
fs_sh = theano.shared(fs)
r = ref_data_sh[j, :][:, fs_sh.nonzero()[0]]
mlp = MLP(activations=activation_functions,
dims=[input_dim] + hidden_dims + [n_inter],
name='inter_gen_%d' % i)
mlp2 = MLP(activations=activation_functions_2 + [None],
dims=[n_inter] + hidden_dims_2 + [output_dim],
name='end_mlp_%d' % i)
inter_weights = mlp.apply(r)
ibias = Bias(n_inter, name='inter_bias_%d' % i)
inter = ibias.apply(tensor.dot(x, inter_weights))
inter = inter_act_fun.apply(inter)
out = mlp2.apply(inter)
last_outputs.append(out)
r_dropout_vars.append(r)
s_dropout_vars = s_dropout_vars + (VariableFilter(
bricks=[Tanh], name='output')(ComputationGraph([inter_weights
])))
i_dropout_vars.append(inter)
# Initialize parameters
for brick in [mlp, mlp2, ibias]:
brick.weights_init = IsotropicGaussian(0.01)
brick.biases_init = Constant(0.001)
brick.initialize()
final = tensor.concatenate([o[:, :, None] for o in last_outputs],
axis=2).mean(axis=2)
cost = Softmax().categorical_cross_entropy(y, final)
confidence = Softmax().apply(final)
pred = final.argmax(axis=1)
# error_rate = tensor.neq(y, pred).mean()
ber = balanced_error_rate.ber(y, pred)
# apply regularization
cg = ComputationGraph([cost, ber])
if r_noise_std != 0:
cg = apply_noise(cg, r_dropout_vars, r_noise_std)
if w_noise_std != 0:
# - apply noise on weight variables
weight_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, weight_vars, w_noise_std)
if s_dropout != 0:
cg = apply_dropout(cg, s_dropout_vars, s_dropout)
if x_dropout != 0:
cg = apply_dropout(cg, [x], x_dropout)
if r_dropout != 0:
cg = apply_dropout(cg, r_dropout_vars, r_dropout)
if i_dropout != 0:
cg = apply_dropout(cg, i_dropout_vars, i_dropout)
[cost_reg, ber_reg] = cg.outputs
self.cost = cost
self.cost_reg = cost_reg
self.ber = ber
self.ber_reg = ber_reg
self.pred = pred
self.confidence = confidence
def __init__(self, ref_data, output_dim):
ref_data_sh = theano.shared(numpy.array(ref_data, dtype=numpy.float32), name='ref_data')
# Construct the model
j = tensor.lvector('j')
x = tensor.fmatrix('x')
y = tensor.ivector('y')
last_outputs = []
s_dropout_vars = []
r_dropout_vars = []
i_dropout_vars = []
for i in range(nparts):
fs = numpy.random.binomial(1, part_r_proba, size=(ref_data.shape[1],))
input_dim = int(fs.sum())
fs_sh = theano.shared(fs)
r = ref_data_sh[j, :][:, fs_sh.nonzero()[0]]
mlp = MLP(activations=activation_functions,
dims=[input_dim] + hidden_dims + [n_inter], name='inter_gen_%d'%i)
mlp2 = MLP(activations=activation_functions_2 + [None],
dims=[n_inter] + hidden_dims_2 + [output_dim],
name='end_mlp_%d'%i)
inter_weights = mlp.apply(r)
ibias = Bias(n_inter, name='inter_bias_%d'%i)
inter = ibias.apply(tensor.dot(x, inter_weights))
inter = inter_act_fun.apply(inter)
out = mlp2.apply(inter)
last_outputs.append(out)
r_dropout_vars.append(r)
s_dropout_vars = s_dropout_vars + (
VariableFilter(bricks=[Tanh], name='output')
(ComputationGraph([inter_weights]))
)
i_dropout_vars.append(inter)
# Initialize parameters
for brick in [mlp, mlp2, ibias]:
brick.weights_init = IsotropicGaussian(0.01)
brick.biases_init = Constant(0.001)
brick.initialize()
final = tensor.concatenate([o[:, :, None] for o in last_outputs], axis=2).mean(axis=2)
cost = Softmax().categorical_cross_entropy(y, final)
confidence = Softmax().apply(final)
pred = final.argmax(axis=1)
# error_rate = tensor.neq(y, pred).mean()
ber = balanced_error_rate.ber(y, pred)
# apply regularization
cg = ComputationGraph([cost, ber])
if r_noise_std != 0:
cg = apply_noise(cg, r_dropout_vars, r_noise_std)
if w_noise_std != 0:
# - apply noise on weight variables
weight_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, weight_vars, w_noise_std)
if s_dropout != 0:
cg = apply_dropout(cg, s_dropout_vars, s_dropout)
if x_dropout != 0:
cg = apply_dropout(cg, [x], x_dropout)
if r_dropout != 0:
cg = apply_dropout(cg, r_dropout_vars, r_dropout)
if i_dropout != 0:
cg = apply_dropout(cg, i_dropout_vars, i_dropout)
[cost_reg, ber_reg] = cg.outputs
self.cost = cost
self.cost_reg = cost_reg
self.ber = ber
self.ber_reg = ber_reg
self.pred = pred
self.confidence = confidence
def __init__(self, ref_data, output_dim):
if pca_dims is not None:
covmat = numpy.dot(ref_data.T, ref_data)
ev, evec = numpy.linalg.eig(covmat)
best_i = ev.argsort()[-pca_dims:]
best_evecs = evec[:, best_i]
best_evecs = best_evecs / numpy.sqrt(
(best_evecs**2).sum(axis=0)) #normalize
ref_data = numpy.dot(ref_data, best_evecs)
input_dim = ref_data.shape[1]
ref_data_sh = theano.shared(numpy.array(ref_data, dtype=numpy.float32),
name='ref_data')
# Construct the model
j = tensor.lvector('j')
r = ref_data_sh[j, :]
x = tensor.fmatrix('x')
y = tensor.ivector('y')
# input_dim must be nr
mlp = MLP(activations=activation_functions,
dims=[input_dim] + hidden_dims + [n_inter],
name='inter_gen')
mlp2 = MLP(activations=activation_functions_2 + [None],
dims=[n_inter] + hidden_dims_2 + [output_dim],
name='end_mlp')
inter_weights = mlp.apply(r)
if inter_bias == None:
ibias = Bias(n_inter)
ibias.biases_init = Constant(0)
ibias.initialize()
inter = ibias.apply(tensor.dot(x, inter_weights))
else:
inter = tensor.dot(x, inter_weights) - inter_bias
inter = inter_act_fun.apply(inter)
final = mlp2.apply(inter)
cost = Softmax().categorical_cross_entropy(y, final)
confidence = Softmax().apply(final)
pred = final.argmax(axis=1)
# error_rate = tensor.neq(y, pred).mean()
ber = balanced_error_rate.ber(y, pred)
# Initialize parameters
for brick in [mlp, mlp2]:
brick.weights_init = IsotropicGaussian(0.01)
brick.biases_init = Constant(0.001)
brick.initialize()
# apply regularization
cg = ComputationGraph([cost, ber])
if r_dropout != 0:
# - dropout on input vector r : r_dropout
cg = apply_dropout(cg, [r], r_dropout)
if x_dropout != 0:
cg = apply_dropout(cg, [x], x_dropout)
if s_dropout != 0:
# - dropout on intermediate layers of first mlp : s_dropout
s_dropout_vars = list(
set(
VariableFilter(bricks=[Tanh], name='output')
(ComputationGraph([inter_weights]))) -
set([inter_weights]))
cg = apply_dropout(cg, s_dropout_vars, s_dropout)
if i_dropout != 0:
# - dropout on input to second mlp : i_dropout
cg = apply_dropout(cg, [inter], i_dropout)
if a_dropout != 0:
# - dropout on hidden layers of second mlp : a_dropout
a_dropout_vars = list(
set(
VariableFilter(bricks=[Tanh], name='output')
(ComputationGraph([final]))) - set([inter_weights]) -
set(s_dropout_vars))
cg = apply_dropout(cg, a_dropout_vars, a_dropout)
if r_noise_std != 0:
cg = apply_noise(cg, [r], r_noise_std)
if w_noise_std != 0:
# - apply noise on weight variables
weight_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, weight_vars, w_noise_std)
[cost_reg, ber_reg] = cg.outputs
if s_l1pen != 0:
s_weights = VariableFilter(bricks=mlp.linear_transformations,
roles=[WEIGHT])(cg)
cost_reg = cost_reg + s_l1pen * sum(
abs(w).sum() for w in s_weights)
if i_l1pen != 0:
cost_reg = cost_reg + i_l1pen * abs(inter).sum()
if a_l1pen != 0:
a_weights = VariableFilter(bricks=mlp2.linear_transformations,
roles=[WEIGHT])(cg)
cost_reg = cost_reg + a_l1pen * sum(
abs(w).sum() for w in a_weights)
self.cost = cost
self.cost_reg = cost_reg
self.ber = ber
self.ber_reg = ber_reg
self.pred = pred
self.confidence = confidence
def main():
rrcos = rrcosfilter(NBAUDS * UPSAMPLE, ROLL_OFF, 1. / BAUD_RATE,
SAMPLE_RATE)[1]
rrcos = rrcos / np.sqrt(UPSAMPLE)
rrcos_fixed = arrayFixedInt(COEF_NBITS, COEF_FBITS, rrcos)
prbs_r = prbs(SEED_R)
tx_r = tx(rrcos_fixed, UPSAMPLE, COEF_NBITS, COEF_FBITS, TX_NBITS,
TX_FBITS)
rx_r = rx(rrcos_fixed, UPSAMPLE, COEF_NBITS, COEF_FBITS, TX_NBITS,
TX_FBITS)
ber_r = ber(SEQ_LEN)
rrcos_float = [i.fValue for i in rrcos_fixed]
prbs_r_v = []
tx_r_v = []
rx_r_v = []
rx_full_v = []
prbs_r.reset()
tx_r.reset()
rx_r.reset()
ber_r.reset()
phase = DX_SWITCH_SEL
prbs_r_s = prbs_r.prbs_out
tx_r_s = tx_r.tx_out
rx_r_s = rx_r.rx_out
enable_prbs = 0
enable_tx = 1
enable_rx = 1
enable_ber = 0
counter = 0
for i in range(NCLK):
prbs_r_s = prbs_r.prbs_out
tx_r_s = tx_r.tx_out
rx_r_s = rx_r.rx_out
rx_full_out = rx_r.rx_full_out
prbs_r_v.append(prbs_r_s)
tx_r_v.append(tx_r_s.fValue)
rx_r_v.append(rx_r_s)
rx_full_v.append(rx_full_out.fValue)
prbs_r.run(enable_prbs)
ber_r.run(prbs_r_s, rx_r_s, enable_ber)
rx_r.run(tx_r_s, phase, enable_rx)
tx_r.run(prbs_r_s, enable_tx)
if counter == 0:
enable_prbs = 1
enable_ber = 1
else:
enable_prbs = 0
enable_ber = 0
counter = (counter + 1) % 4
vector = zip(range(NCLK), prbs_r_v, tx_r_v, rx_full_v, rx_r_v)
"""
for i in vector[0:20]:
print i
exit()
"""
plt.figure()
plt.grid()
plt.plot(tx_r_v[:200])
plt.figure()
plt.grid()
plt.plot(rx_full_v[:200])
rx_a = arrayFixedInt(8, 7, rx_full_v[:200])
rx_a = [i.fValue for i in rx_a]
plt.figure()
plt.grid()
plt.plot(rx_a)
"""
eyediagram(rx_full_v[12:], 4, 1, UPSAMPLE)
rrcos_float = [i.fValue for i in rrcos_fixed]
H,A,F = resp_freq(rrcos_float, 1./BAUD_RATE, 512)
plt.figure()
plt.grid()
plt.semilogx(F, 20*np.log(H))
plt.figure()
plt.grid()
plt.plot([i.fValue for i in rrcos_fixed])
"""
plt.show()
