Performance penalty by using fewer taps¶
import os
import numpy as np
import pandas as pd
from tqdm.auto import tqdm
from functools import partial
import matplotlib.pyplot as plt
from commplax import util
from gdbp import gdbp_base as gb, data as gdat, plot as gplt, aux
WARNING:absl:No GPU/TPU found, falling back to CPU. (Set TF_CPP_MIN_LOG_LEVEL=0 and rerun for more info.)
ds_train, ds_test = gdat.load(1, 0, 4, 2)[0], gdat.load(2, 0, 4, 1)[0]
aux.dl_trained_params() # download pre-trained parameters
loc_trained_params = './trained_params/pretrained'
100% [..........................................................................] 1582928 / 1582928
# use Pandas.Dataframe to store the results
df_test_res = pd.DataFrame({c: pd.Series(dtype=t) for c, t in {'ntaps': 'int',
'dtaps': 'int',
'xi': 'float',
'Model': 'str',
'Q': 'float'}.items()})
def init_models(data, hparams, mode='train'):
model_conf = {'mode': mode,
'dtaps': hparams['dtaps'],
'ntaps': hparams['ntaps'],
'init_fn': gb.fdbp_init(data.a, steps=3, xi=hparams['xi'])}
model_init = partial(gb.model_init, data, model_conf)
fdbp = model_init([('fdbp_0',)] , name='FDBP')
edbp = model_init([('fdbp_0', r'DConv_\d')], name='EDBP')
gdbp = model_init([], name='GDBP')
return fdbp, edbp, gdbp
def sweep_hparams(hparams, df_test_res=df_test_res, use_pretrained_params=False, save_params=False):
for i, hp in enumerate(tqdm(hparams, desc='sweeping hparams')): # iterate hparams
models_train = init_models(ds_train, hp)
models_test = init_models(ds_test, hp, mode='test')
for j, (model_train, model_test) in enumerate(tqdm(zip(models_train, models_test),
total=len(models_train),
desc='iterating models',
leave=False)): # iterate models
params_file = os.path.join(loc_trained_params,
'snr_vs_dtaps_ntaps',
'params_%d_%d' % (i, j)) # params_{hparam}_{model}
if use_pretrained_params:
params = util.load_variable(params_file)
else:
params_queue = [None] * 3
for _, p, _ in gb.train(model_train, ds_train, n_iter=2000):
params_queue.append(p)
params = params_queue.pop(0)
if save_params:
util.save_variable(params, params_file)
test_Q = gb.test(model_test, params, ds_test)[0].QSq.total
# collect result
df_test_res = df_test_res.append({'dtaps': hp['dtaps'],
'ntaps': hp['ntaps'],
'xi': hp['xi'],
'Model': model_test.name,
'Q': test_Q},
ignore_index=True)
if i % 10 == 0:
util.clear_xla_cache() # clear compile cache periodically to save RAM
return df_test_res
# make sweeping grid
# caveats: taps must be odd number
NTAPS = np.arange(1, 41, 6)
DTAPS = np.arange(181, 261, 12)
OPTXI = pd.read_csv('../../source_data/fdbp_opt_xi_ntaps.csv').values[:, 0] # read optimal xi given dtaps
hparams = []
for ntaps, xi in zip(NTAPS, OPTXI):
for dtaps in DTAPS:
hparams.append({'ntaps': ntaps, 'dtaps': dtaps, 'xi': xi})
# it may take a while to finish
df_test_res = sweep_hparams(hparams, df_test_res, use_pretrained_params=True, save_params=False)
df_test_res
| ntaps | dtaps | xi | Model | Q | |
|---|---|---|---|---|---|
| 0 | 1 | 181 | 0.15 | FDBP | 7.239062 |
| 1 | 1 | 181 | 0.15 | EDBP | 7.244903 |
| 2 | 1 | 181 | 0.15 | GDBP | 8.223946 |
| 3 | 1 | 193 | 0.15 | FDBP | 7.611000 |
| 4 | 1 | 193 | 0.15 | EDBP | 7.618716 |
| ... | ... | ... | ... | ... | ... |
| 142 | 37 | 241 | 1.10 | EDBP | 8.596812 |
| 143 | 37 | 241 | 1.10 | GDBP | 8.741637 |
| 144 | 37 | 253 | 1.10 | FDBP | 8.590569 |
| 145 | 37 | 253 | 1.10 | EDBP | 8.670206 |
| 146 | 37 | 253 | 1.10 | GDBP | 8.746362 |
147 rows × 5 columns
fig, ax = plt.subplots(subplot_kw={"projection": "3d"}, figsize=(5, 5), dpi=200)
ax.view_init(19, -30)
curves = 20
cms = [plt.cm.Greys, plt.cm.Blues, plt.cm.Reds]
colors = plt.cm.RdPu_r(np.linspace(0., 0.5, 3))[::-1]
for cm, mod in zip(cms, ['FDBP', 'EDBP', 'GDBP']):
df = df_test_res.groupby('Model').get_group(mod).pivot('dtaps', 'ntaps', 'Q')
(X, Y), Z = np.meshgrid(df.columns.values, df.index.values), df.values
gplt.wireframe_cmap(ax, X, Y, Z, cmap=cm, offset=[-1.2, 0], label=mod)
ax.set_xlabel('N-filter length')
ax.set_ylabel('D-filter length')
ax.set_zlabel('Q-factor (dB)')
ax.invert_xaxis()
ax.legend(loc=(-0.18, 0.4))
<matplotlib.legend.Legend at 0x7fde1cb677f0>
