def image_complexity(self, X, masks):
# 복잡도 기준
threshold = 25.0
# reshape
#masks = cp.reshape(masks, masks.shape[:-1])
R, G, B = X[:, :, :, 0:1], X[:, :, :, 1:2], X[:, :, :, 2:]
# compute rg = R - G
rg = cp.absolute(R - G)
# compute yb = 0.5 * (R + G) - B
yb = cp.absolute(0.5 * (R + G) - B)
_masks = cp.logical_not(masks)
# compute the mean and standard deviation of both `rg` and `yb` ()
# no masked_where on cupy
rb_masked = np.ma.masked_where(_masks, rg)
(rb_mean, rb_std) = (cp.mean(rb_masked, axis=(1, 2, 3)),
cp.std(rb_masked, axis=(1, 2, 3)))
yb_masked = np.ma.masked_where(_masks, yb)
(yb_mean, yb_std) = (cp.mean(yb_masked, axis=(1, 2, 3)),
cp.std(yb_masked, axis=(1, 2, 3)))
# combine the mean and standard deviations
std_root = cp.sqrt((rb_std**2) + (yb_std**2))
mean_root = cp.sqrt((rb_mean**2) + (yb_mean**2))
# derive the "colorfulness" metric and return it
complexity = std_root + (0.3 * mean_root)
# 수치 수정
print('image_complexity Done.')
return (complexity >= threshold).astype(cp.uint8)
def sortBatches2(ccb0):
# takes as input a matrix of nBatches by nBatches containing
# dissimilarities.
# outputs a matrix of sorted batches, and the sorting order, such that
# ccb1 = ccb0(isort, isort)
# put this matrix on the GPU
ccb0 = cp.asarray(ccb0, order='F')
# compute its svd on the GPU (this might also be fast enough on CPU)
u, s, v = svdecon(ccb0)
# HACK: consistency with MATLAB
u = u * cp.sign(u[0, 0])
v = v * cp.sign(u[0, 0])
# initialize the positions xs of the batch embeddings to be very small but proportional to
# the first PC
xs = .01 * u[:, 0] / cp.std(u[:, 0], ddof=1)
# 200 iterations of gradient descent should be enough
# TODO: move_to_config
niB = 200
# this learning rate should usually work fine, since it scales with the average gradient
# and ccb0 is z-scored
# TODO: move_to_config
eta = 1
for k in tqdm(range(niB), desc="Sorting %d batches" % ccb0.shape[0]):
# euclidian distances between 1D embedding positions
ds = (xs - xs[:, np.newaxis])**2
# the transformed distances go through this function
W = cp.log(1 + ds)
# the error is the difference between ccb0 and W
err = ccb0 - W
# ignore the mean value of ccb0
err = err - cp.mean(err, axis=0)
# backpropagate the gradients
err = err / (1 + ds)
err2 = err * (xs[:, np.newaxis] - xs)
D = cp.mean(
err2, axis=1) # one half of the gradients is along this direction
E = cp.mean(err2, axis=0) # the other half is along this direction
# we don't need to worry about the gradients for the diagonal because those are 0
# final gradients for the embedding variable
dx = -D + E.T
# take a gradient step
xs = xs - eta * dx
# sort the embedding positions xs
isort = cp.argsort(xs, axis=0)
# sort the matrix of dissimilarities
ccb1 = ccb0[isort, :][:, isort]
return ccb1, isort
def noiser(self, pitch_track, SNR):
self.clean = cp.empty((self.size))
self.clean[:] = self.data
RMS = cp.std(self.data[pitch_track > 0])
noise = cp.random.normal(0, RMS / (10**(SNR / 20)), self.size)
self.data += noise
def normalize_array(A):
mean, std = np.mean(A), np.std(A)
A_normed = (A - mean) / std
def restore_function(X):
return X * std + mean
return A_normed.astype(np.float32), restore_function
def mass2_gpu(ts, query):
"""
Compute the distance profile for the given query over the given time
series. This require cupy to be installed.
Parameters
----------
ts : array_like
The array to create a rolling window on.
query : array_like
The query.
Returns
-------
An array of distances.
Raises
------
ValueError
If ts is not a list or np.array.
If query is not a list or np.array.
If ts or query is not one dimensional.
"""
def moving_mean_std_gpu(a, w):
s = cp.concatenate([cp.array([0]), cp.cumsum(a)])
sSq = cp.concatenate([cp.array([0]), cp.cumsum(a**2)])
segSum = s[w:] - s[:-w]
segSumSq = sSq[w:] - sSq[:-w]
movmean = segSum / w
movstd = cp.sqrt(segSumSq / w - (segSum / w)**2)
return (movmean, movstd)
x = cp.asarray(ts)
y = cp.asarray(query)
n = x.size
m = y.size
meany = cp.mean(y)
sigmay = cp.std(y)
meanx, sigmax = moving_mean_std_gpu(x, m)
meanx = cp.concatenate([cp.ones(n - meanx.size), meanx])
sigmax = cp.concatenate([cp.zeros(n - sigmax.size), sigmax])
y = cp.concatenate((cp.flip(y, axis=0), cp.zeros(n - m)))
X = cp.fft.fft(x)
Y = cp.fft.fft(y)
Z = X * Y
z = cp.fft.ifft(Z)
dist = 2 * (m - (z[m - 1:n] - m * meanx[m - 1:n] * meany) /
(sigmax[m - 1:n] * sigmay))
dist = cp.sqrt(dist)
return cp.asnumpy(dist)
def fix(self):
if self.PITCH_HALF > 0:
nz_pitch = self.samp_values[self.samp_values > 0]
idx = self.samp_values < (cp.mean(nz_pitch) -
self.PITCH_HALF_SENS * cp.std(nz_pitch))
if self.PITCH_HALF == 1:
self.samp_values[idx] = 0
elif self.PITCH_HALF == 2:
self.samp_values[idx] = 2 * self.samp_values[idx]
if self.PITCH_DOUBLE > 0:
nz_pitch = self.samp_values[self.samp_values > 0]
idx = self.samp_values > (
cp.mean(nz_pitch) + self.PITCH_DOUBLE_SENS * cp.std(nz_pitch))
if self.PITCH_DOUBLE == 1:
self.samp_values[idx] = 0
elif self.PITCH_DOUBLE == 2:
self.samp_values[idx] = 0.5 * self.samp_values[idx]
def std_data(X):
mu = cp.mean(X, axis=1, keepdims=True)
sigma = cp.std(X, axis=1, keepdims=True)
X = (X - mu) / sigma
stdval_cache = {"mu": mu, "sigma": sigma}
return X, stdval_cache
def std(inp) -> 'Tensor':
_check_tensors(inp)
engine = _get_engine(inp)
return _create_tensor(
inp,
data=engine.std(inp.data),
func=wrapped_partial(std_backward, inp=inp)
)
def stablesoftmax(x):
"""Compute the softmax of vector x in a numerically stable way."""
assert x.ndim == 2
classes = x.shape[1]
x = x - np.mean(x, keepdims=True)
x = x / np.std(x, keepdims=True)
x = np.exp(x)
x = x / np.sum(x, -1, keepdims=True)
return x
def norm_percentile(signals, p1=5, p2=95, pcnt=True):
# n signals array are of dim (n,T)
out = cp.zeros(signals.shape)
for i in range(signals.shape[0]):
s = signals[i, :]
if pcnt == True:
[n, m] = cp.percentile(s, [p1, p2])
out[i, :] = (s - n) / (m - n)
else:
out[i, :] = (s - cp.mean(s)) / cp.std(s)
return out
def softmax_derivative(x):
assert x.ndim == 2
classes = x.shape[1]
x = x - np.mean(x, keepdims=True)
x = x / np.std(x, keepdims=True)
out = np.zeros((x.shape[0], classes, classes))
for i in range(classes):
for j in range(classes):
out[:, i,
j] = x[:, i] * (1 - x[:, i]) if (i
== j) else -x[:, i] * x[:, j]
return out
def evolve(organism, organism_loss, scale, size, learning_rate=0.001):
flat_organism = flatten_organism(organism)
loc = flat_organism.mean(axis=0, keepdims=True)
n_organisms = flat_organism.shape[0]
N = flat_organism - loc
A = (organism_loss - cp.mean(organism_loss)) / (1e-6 +
cp.std(organism_loss))
loc = loc - learning_rate / (n_organisms * scale) * cp.dot(N.T, A)
N = cp.random.normal(loc=0,
scale=scale,
size=(int(n_organisms / 2), flat_organism.shape[1]))
new_flat_organism = cp.concatenate((loc + N, loc - N), axis=0)
return reform_organism(new_flat_organism, size), loc, N
def normalize(data, mask=None, padding=0, return_space='cpu'):
""" Apply normalization on GPU
Applies normalisation (data - mean) / stdev
Args:
data (np/cp.array): Data to preprocess
mask (np.cp.array): 1D Channel mask for RFI flagging
padding (int): size of edge region to discard (e.g. coarse channel edges)
return_space ('cpu' or 'gpu'): Returns array in CPU or GPU space
Returns: d_gpu (cp.array): Normalized data
"""
# Copy over to GPU if required
d_gpu = cp.asarray(data.astype('float32', copy=False))
d_gpu_flagged = cp.asarray(data.astype('float32', copy=True))
paddingu = None if padding == 0 else -padding
# Need to correct stats
N_flagged = 0
N_tot = np.product(d_gpu[..., padding:paddingu].shape)
if mask is not None:
# Convert 1D-mask to match data dimensions
mask_gpu = cp.repeat(cp.asarray(mask.reshape((1, 1, len(mask)))),
d_gpu.shape[0],
axis=0)
cp.putmask(d_gpu_flagged, mask_gpu, 0)
N_flagged = mask_gpu[..., padding:paddingu].sum()
# Normalise
t0 = time.time()
# Compute stats based off flagged arrays
d_mean = cp.mean(d_gpu_flagged[..., padding:paddingu])
d_std = cp.std(d_gpu_flagged[..., padding:paddingu])
flag_correction = N_tot / (N_tot - N_flagged)
d_mean = d_mean * flag_correction
d_std = d_std * np.sqrt(flag_correction)
logger.debug(f"flag fraction correction factor: {flag_correction}")
# Apply to original data
d_gpu = (d_gpu - d_mean) / d_std
t1 = time.time()
logger.info(f"Normalisation time: {(t1-t0)*1e3:2.2f}ms")
if return_space == 'cpu':
return cp.asnumpy(d_gpu)
else:
return d_gpu
def load_mc(self, t_ij):
for j, (l, h) in enumerate(
zip(self.observables.lows, self.observables.highs)):
in_bounds = np.logical_and(t_ij[:, j] > l, t_ij[:, j] < h)
t_ij = t_ij[in_bounds]
self.t_ij = cp.asarray(t_ij)
self.w_i = cp.ones(t_ij.shape[0])
if self.bootstrap_binning is not None:
counts, _ = cp.histogramdd(cp.asarray(self.t_ij),
bins=self.bin_edges,
weights=cp.asarray(self.w_i))
self.counts = (cp.asarray(counts).flatten() / self.bin_vol /
cp.sum(cp.asarray(counts))).reshape(counts.shape)
self.sigma_j = cp.std(self.t_ij, axis=0)
self.h_ij = self._adapt_bandwidth()
for j, (l, h, refl) in enumerate(
zip(self.observables.lows, self.observables.highs,
self.reflect_axes)):
if not refl:
continue
if type(refl) == tuple:
low, high = refl
mask = self.t_ij[:, j] < low
t_ij_reflected_low = cp.copy(self.t_ij[mask, :])
h_ij_reflected_low = self.h_ij[mask, :]
w_i_reflected_low = self.w_i[mask, :]
t_ij_reflected_low[:, j] = 2 * l - t_ij_reflected_low[:, j]
mask = self.t_ij[:, j] > high
t_ij_reflected_high = cp.copy(self.t_ij[mask, :])
h_ij_reflected_high = self.h_ij[mask, :]
w_i_reflected_high = self.w_i[mask, :]
t_ij_reflected_high[:, j] = 2 * h - t_ij_reflected_high[:, j]
else:
t_ij_reflected_low = cp.copy(self.t_ij)
h_ij_reflected_low = self.h_ij
w_i_reflected_low = self.w_i
t_ij_reflected_low[:, j] = 2 * l - self.t_ij[:, j]
t_ij_reflected_high = cp.copy(self.t_ij)
h_ij_reflected_high = self.h_ij
w_i_reflected_high = self.w_i
t_ij_reflected_high[:, j] = 2 * h - self.t_ij[:, j]
self.t_ij = cp.concatenate(
[self.t_ij, t_ij_reflected_low, t_ij_reflected_high])
self.h_ij = cp.concatenate(
[self.h_ij, h_ij_reflected_low, h_ij_reflected_high])
self.w_i = cp.concatenate(
[self.w_i, w_i_reflected_low, w_i_reflected_high])
self.t_ij = cp.ascontiguousarray(self.t_ij)
self.h_ij = cp.ascontiguousarray(self.h_ij)
self.w_i = cp.ascontiguousarray(self.w_i)
def get_spike_amps(s, return_counts=False):
sig = cupy.asarray(s[:1024, :20000])
peaks = cusignal.peak_finding.peak_finding.argrelmin(sig, order=20, axis=1)
mean_std = cupy.mean(cupy.std(sig, axis=1))
significant_peaks = sig[peaks[0], peaks[1]] < (-10 * mean_std)
amps = np.median(cupy.asnumpy(sig[:, peaks[1][significant_peaks]] * -1),
axis=1)
if return_counts:
sig_peak_chans = peaks[0][significant_peaks]
chan_count = np.array(
[len(sig_peak_chans[sig_peak_chans == i]) for i in range(1024)])
return amps, chan_count
else:
return amps
def sk_flag(data,
metadata,
n_sigma_upper=5,
n_sigma_lower=5,
flag_upper=True,
flag_lower=True):
""" Apply spectral kurtosis flagging
Args:
data (np.array): Numpy array with shape (N_timestep, N_beam, N_channel)
metadata (dict): Metadata dictionary, should contain 'df' and 'dt'
(frequency and time resolution)
boxcar_mode (str): Boxcar mode to apply. mean/sum/gaussian.
n_sigma_upper (float): Number of stdev above SK estimate to flag (upper bound)
n_sigma_lower (float): Number of stdev below SK estmate to flag (lower bound)
flag_upper (bool): Flag channels with large SK (highly variable signals)
flag_lower (bool): Flag channels with small SK (very stable signals)
return_space ('cpu' or 'gpu'): Returns array in CPU or GPU space
Returns:
mask (np.array, bool): Array of True/False flags per channel
Notes:
sk_flag upper and lower stdev is computed on log2(sk), as the minimum
spectral kurtosis (for a CW signal) approaches 0.
"""
Fs = (1.0 / metadata['frequency_step'] / 2)
samps_per_sec = np.abs(Fs.to('s').value) # Nyq sample rate for channel
N_acc = int(metadata['time_step'].to('s').value / samps_per_sec)
sk = spectral_kurtosis(data, metadata)
#var_theoretical = 2.0 / np.sqrt(N_acc)
#std_theoretical = np.sqrt(var_theoretical)
log_sk = cp.log2(sk)
std_log = cp.std(log_sk)
mean_log = cp.mean(log_sk)
if flag_upper and flag_lower:
mask = log_sk < mean_log + (std_log * n_sigma_upper)
mask &= log_sk > mean_log - (std_log * n_sigma_lower)
elif flag_upper and not flag_lower:
mask = log_sk > mean_log + (std_log * n_sigma_upper)
elif flag_lower and not flag_upper:
mask = log_sk < mean_log - (std_log * n_sigma_lower)
else:
raise RuntimeError(
"No flags to process: need to flag upper and/or lower!")
return ~mask
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--gpu-id', '-g', default=0, type=int, help='GPU ID')
parser.add_argument('--n-options', default=1000, type=int)
parser.add_argument('--n-samples-per-thread', default=1000, type=int)
parser.add_argument('--n-threads-per-option', default=100000, type=int)
args = parser.parse_args()
cupy.cuda.Device(args.gpu_id).use()
def rand_range(m, M):
samples = cupy.random.rand(args.n_options)
return (m + (M - m) * samples).astype(numpy.float64)
print('initializing...')
stock_price = rand_range(5, 30)
option_strike = rand_range(1, 100)
option_years = rand_range(0.25, 10)
risk_free = 0.02
volatility = 0.3
@contextlib.contextmanager
def timer(message):
cupy.cuda.Stream.null.synchronize()
start = time.time()
yield
cupy.cuda.Stream.null.synchronize()
end = time.time()
print('%s:\t%f sec' % (message, end - start))
print('start computation')
print(' # of options: {}'.format(args.n_options))
print(' # of samples per option: {}'.format(args.n_samples_per_thread *
args.n_threads_per_option))
with timer('GPU (CuPy, Monte Carlo method)'):
call_mc = compute_option_prices(stock_price, option_strike,
option_years, risk_free, volatility,
args.n_threads_per_option,
args.n_samples_per_thread)
# Compute the error between the value of the exact solution
# and that of the Monte-Carlo simulation
call_bs, _ = black_scholes_kernel(stock_price, option_strike, option_years,
risk_free, volatility)
error = cupy.std(call_mc - call_bs)
print('Error: %f' % error)
return 0
def fit(self, X, y, Xt=None, yt=None):
if self.params['norm']:
Xmean, Xstd = cp.mean(X, axis=0, keepdims=True), cp.std(X, axis=0, keepdims=True)
X = (X - Xmean) / (Xstd + 1e-5)
self.Xmean, self.Xstd = Xmean, Xstd
if Xt is None:
test_size = self.params['validation_fraction']
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size = test_size)
train_dataset = MyDataSet(X_train, y_train, task=self.task)
valid_dataset = MyDataSet(X_test, y_test, task=self.task)
Ntr,Nva = X_train.shape[0], X_test.shape[0]
else:
if self.params['norm']:
Xt = (Xt - Xmean) / (Xstd + 1e-5)
train_dataset = MyDataSet(X, y, task=self.task)
valid_dataset = MyDataSet(Xt, yt, task=self.task)
Ntr,Nva = X.shape[0], Xt.shape[0]
batch_size = self.get_batch_size(N=Ntr, mb=1024)
train_dataloader = MyDataLoader(train_dataset, batch_size=batch_size,
shuffle=self.params['shuffle'], drop_last=True)
batch_size = self.get_batch_size(N=Nva, mb=1024)
valid_dataloader = MyDataLoader(valid_dataset, batch_size=batch_size,
shuffle=False, drop_last=False)
C = 1 if self.task == 'regression' else cp.unique(y).shape[0]
model = MLP(X.shape[1], self.params['hidden_layer_sizes'], C, **self.params)
if self.task == 'regression':
loss_func = torch.nn.functional.mse_loss
elif self.task == 'classification':
loss_func = torch.nn.functional.cross_entropy
else:
assert 0, "Unknown taks: "+self.task
learner = Learner(model, loss_func, **self.params)
learner.fit(train_dl=train_dataloader, valid_dl=valid_dataloader)
self.learner = learner
def zdist(Q, S, mode="fft", epsilon=1e-6):
"""
Rolling mean- and amplitude-adjusted Euclidean Distance
Arguments:
-------
Q: cupy.core.core.ndarray or numba.cuda.DeviceNDArray or cudf.Series or numpy.ndarray
the input query of length m to be aligned
S: cupy.core.core.ndarray or numba.cuda.DeviceNDArray or cudf.Series or numpy.ndarray
the input stream of length n>=m to be scanned
epsilon: float
non-negative number for regularizing zero stdev
mode: str
either "naive" or "fft"
Returns
-------
cupy.core.core.ndarray
the computed distance array of length n-m+1
"""
if not isinstance(Q, cp.core.core.ndarray):
Q = cp.asarray(Q)
if not isinstance(S, cp.core.core.ndarray):
S = cp.asarray(S)
assert (epsilon > 0)
assert (Q.dtype == S.dtype)
assert ((len(Q.shape) == len(S.shape) == 1 and Q.shape[0] <= S.shape[0]))
assert (cp.std(Q, ddof=0) > 0)
if mode == "fft":
Z = fft_zdist(Q, S, epsilon)
else:
stream = cuda.stream()
Z = cp.empty(len(S) - len(Q) + 1, dtype=Q.dtype)
zdist_kernel[80 * 32, 64, stream](znorm(Q, epsilon), S, Z, epsilon)
stream.synchronize()
return Z
def znorm(x, epsilon):
"""
Mean- and amplitude-adjustment of a given time series. Afterwards the time series
has vanishing mean, i.e. sum_i x[i] = 0 and unit standard devitation i.e.
sum_i x[i]*x[i] = n where n is the length of the sequence x
Arguments:
-------
x: cupy.core.core.ndarray
the input array of length n to be normalized
Returns
-------
cupy.core.core.ndarray
the mean-adjusted array of length n
"""
return (x-cp.mean(x))/max(cp.std(x, ddof=0), epsilon)
def collect_test_losses(num_folds):
# Run this after CV results are in. e.g:
# python -c "from deepmolecule.util import collect_test_losses; collect_test_losses(10)"
results = {}
for net_type in ['conv', 'morgan']:
results[net_type] = []
for expt_ix in range(num_folds):
fname = "Final_test_loss_{0}_{1}.pkl.save".format(
expt_ix, net_type)
try:
with open(fname) as f:
results[net_type].append(pickle.load(f))
except IOError:
print("Couldn't find file {0}".format(fname))
print("Results are:")
print(results)
print("Means:")
print({k: np.mean(v) for k, v in results.items()})
print("Std errors:")
print({k: np.std(v) / np.sqrt(len(v) - 1) for k, v in results.items()})
def run_gpu(X, eps, min_samples):
# Transfer inputs to GPU
X = cp.array(X)
# Begin computation
t0 = time.time()
mean = cp.mean(X, axis=0)
std = cp.std(X, axis=0)
cp.subtract(X, mean, out=X)
cp.divide(X, std, out=X)
print('Preprocessing:', time.time() - t0)
# Run DBSCAN
db = cuml.DBSCAN(eps=eps, min_samples=min_samples)
db = db.fit(X)
labels = db.labels_
# Transfer outputs to CPU
# labels = labels.to_pandas().to_numpy()
labels = cp.asnumpy(labels)
return labels
def filter_and_whiten(raw_traces, params, probe, whitening_matrix):
sample_rate = params.fs
high_pass_freq = params.fshigh
low_pass_freq = params.fslow
scaleproc = params.scaleproc
whitening_matrix_np = cp.asarray(whitening_matrix,
dtype=np.float32) / np.float(scaleproc)
filtered_data = gpufilter(buff=cp.asarray(raw_traces, dtype=np.float32),
chanMap=probe.chanMap,
fs=sample_rate,
fslow=low_pass_freq,
fshigh=high_pass_freq)
whitened_data = cp.dot(filtered_data, whitening_matrix_np)
array_means = cp.mean(whitened_data, axis=0)
array_stds = cp.std(whitened_data, axis=0)
whitened_array = (whitened_data - array_means) / array_stds
return whitened_array.get()
def to_periodogram(signal):
"""
Returns periodogram of signal for finding frequencies that have high energy.
:param signal: signal (time domain)
:type signal: cudf.Series
:return: CuPy array representing periodogram
:rtype: cupy.core.core.ndarray
"""
# convert cudf series to cupy array
signal_cp = cp.fromDlpack(signal.to_dlpack())
# standardize the signal
signal_cp_std = (signal_cp - cp.mean(signal_cp)) / cp.std(signal_cp)
# take fourier transform of signal
FFT_data = cp.fft.fft(signal_cp_std)
# create periodogram
prdg = (1 / len(signal)) * ((cp.absolute(FFT_data))**2)
return prdg
def estimate_stats(voltages, stats_calc_num_samples=10000):
"""
Estimate mean and standard deviation, truncating to at most `stats_calc_num_samples` samples
to reduce computation.
Parameters
----------
voltages : array
Array of voltages
stats_calc_num_samples : int, optional
Maximum number of samples for use in estimating noise statistics
Returns
-------
data_mean : float
Mean of voltages
data_sigma : float
Standard deviation of voltages
"""
calc_len = xp.amin(xp.array([stats_calc_num_samples, len(voltages)]))
data_sigma = xp.std(voltages[:calc_len])
data_mean = xp.mean(voltages[:calc_len])
return data_mean, data_sigma
h = self.bn_b(self.convb(h))
if self.proj:
x = self.bn_r(self.convr(x))
return F.relu(h + x)
train, test = cifar.get_cifar10()
train_iter = iterators.SerialIterator(train, 128)
test_iter = iterators.SerialIterator(test, 2000, repeat=False, shuffle=False)
gpu_id = 0
Itrain, Ttrain = concat_examples(train, gpu_id)
Train_mean = cp.mean(Itrain)
Train_std = cp.std(Itrain)
model = ResNet34().to_gpu(gpu_id)
optimizer = optimizers.MomentumSGD(lr=0.1, momentum=0.9)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.WeightDecay(0.0001))
max_epoch = 50000
plt.figure(figsize=(7, 5))
test_acc_array = []
train_acc_array = []
while train_iter.epoch < max_epoch:
def standardization(data):
mu = np.mean(data, axis=0)
sigma = np.std(data, axis=0)
return (data - mu) / sigma
return Array._new(np.prod(x._array, axis=axis, keepdims=keepdims))
def std(
x: Array,
/,
*,
axis: Optional[Union[int, Tuple[int, ...]]] = None,
correction: Union[int, float] = 0.0,
keepdims: bool = False,
) -> Array:
# Note: the keyword argument correction is different here
if x.dtype not in _floating_dtypes:
raise TypeError("Only floating-point dtypes are allowed in std")
return Array._new(
np.std(x._array, axis=axis, ddof=correction, keepdims=keepdims))
def sum(
x: Array,
/,
*,
axis: Optional[Union[int, Tuple[int, ...]]] = None,
dtype: Optional[Dtype] = None,
keepdims: bool = False,
) -> Array:
if x.dtype not in _numeric_dtypes:
raise TypeError("Only numeric dtypes are allowed in sum")
# Note: sum() and prod() always upcast integers to (u)int64 and float32 to
# float64 for dtype=None. `np.sum` does that too for integers, but not for
# float32, so we need to special-case it here
def forward_conv(A_previous, Filter, Bias, pad, stride,
function = 'identity', verbose = False):
'''
A forward convolution step.
Calcul output shape : ((x-f+2*pad)/stride)+1
Parameters
----------
A_previous : cp.array(examples, height, width, depth)
Input images from the previous layer.
Filter : cp.array(f, f, depth, number of filter)
Filter to convolve with the input image.
Bias : cp.array(1, 1, 1, number of filter)
Bias for each filter.
pad : int
Padding edge width.
stride : int
Stride number.
Returns
-------
Z : cp.array(examples, ((h-f+2*pad)/stride)+1, ((w-f+2*pad)/stride)+1), number of filter)
Output layer image.
'''
(m, n_H_prev, n_W_prev, n_C_prev) = A_previous.shape
(f, f, n_C_prev, n_C) = Filter.shape
mu = cp.mean(Filter)
s = cp.std(Filter)
Filter = (Filter-mu)/(s+1e-5)
n_H = int(((n_H_prev-f+2*pad)/stride)+1)
n_W = int(((n_W_prev-f+2*pad)/stride)+1)
Z = cp.zeros([m, n_H, n_W, n_C])
A_prev_pad = cp.pad(A_previous, ((0,0), (pad,pad), (pad,pad), (0,0),), mode='constant', constant_values = (0,0))
i0 = cp.repeat(cp.arange(f), f)
i1 = stride * cp.repeat(cp.arange(n_W), n_H)
j0 = cp.tile(cp.arange(f), f)
j1 = stride * cp.tile(cp.arange(n_H), n_W)
i = cp.reshape(i0, (-1, 1))+cp.reshape(i1, (1, -1))
j = cp.reshape(j0, (-1, 1))+cp.reshape(j1, (1, -1))
k = cp.reshape(cp.repeat(cp.arange(n_C_prev), f**2), (-1, 1))
Ztest = A_prev_pad[:, i, j, :]
weights = cp.reshape(Filter, (f**2, n_C_prev, n_C))
conV = cp.tensordot(weights, Ztest, ((0, 1), (1, 3)))
Z = cp.reshape(cp.transpose(conV, (1, 2, 0)), (m, n_H, n_W, n_C)) + Bias
Z = activation('forward', function, Z)
if(verbose):
print("Filter :")
print(Filter)
print("Weights :")
print(weights)
print("Z :")
print(Ztest)
print("Conv :")
print(conV)
print("Result :")
print(Z)
'''
for i in range(m):
a_prev_pad = A_prev_pad[i, :, :, :]
for h in range(n_H):
vert_start = h*stride
vert_end = h*stride+f
for w in range(n_W):
horiz_start = w*stride
horiz_end = w*stride+f
a_slice_prev = a_prev_pad[vert_start:vert_end, horiz_start:horiz_end, :]
for c in range(n_C):
Z[i, h, w, c] = cp.squeeze(cp.sum(a_slice_prev*Filter[:, :, :, c])+Bias[:, :, :, c])
'''
return Z
def forward_function(A_previous, W, mu, sigma, gamma, beta, function, dropout = 1):
'''
A forward function step
Parameters
----------
A_previous : cp.array(in_dim, examples)
Result of the previous layer.
W : cp.array(out_dim, in_dim)
Weight matrix.
mu : cp.array(out_dim, number of epochs)
Gather Z mean over epochs.
sigma : cp.array(out_dim, number of epochs)
Gather Z std over epochs.
gamma : cp.array(out_dim, 1)
Weight matrix.
beta : cp.array(out_dim, 1)
Bias matrix.
function : string
The desired activation function.
dropout : float (in (0,1)), optional
Percentage of disabled neurons. The default is 1.
Returns
-------
A : cp.array(out_dim, examples)
Output layer result.
z : cp.array(out_dim, examples)
Activation function input.
zhat : cp.array(out_dim, examples)
Normalized Z.
Z : cp.array(out_dim, examples)
Result after applying W weights.
mu : cp.array(out_dim, number of epochs)
Updated mu.
sigma : cp.array(out_dim, number of epochs)
Updated sigma.
D : cp.array(out_dim, 1)
Mask matrix for dropout (filled with ones if dropout = 1).
'''
m = A_previous.shape[1]
eps = 1e-8
Z = cp.dot(W, A_previous)
if mu.any():
mu = cp.concatenate((mu, cp.mean(Z, axis = 1, keepdims = True)), axis = 1)
sigma = cp.concatenate((sigma, cp.std(Z, axis = 1, keepdims = True)), axis = 1)
else :
mu = cp.mean(Z, axis = 1, keepdims = True)
sigma = cp.std(Z, axis = 1, keepdims = True)
zhat = ((Z-cp.mean(mu, axis = 1, keepdims = True))
/(((m/((m-1)+eps))*cp.mean(sigma, axis = 1, keepdims = True))+eps))
z = (gamma*zhat)+beta
A = activation('forward', function, z)
D = cp.random.rand(A.shape[0], A.shape[1])
D = (D <= dropout).astype(int)
A = A*D
A = A/dropout
return A, z, zhat, Z, mu, sigma, D
