def init_seq_idx(conf_seq, data):
idx_dct, ext_dct, n_dct = get(data, "indexing", "extended", "n")
base_shape, diff_ii = get(idx_dct, "base_shape", "diff_ii")
caps, refs, cm = get(ext_dct, "caps", "refs", "cm")
n_conf, n_diff = get(n_dct, "n_conf", "n_diff")
ic = conf_seq.initial_conditions
# Set initial condition
# shape (n_conf, n_cf, n_diff,)
ic_ii = np.array(list(iic.ref_ii for iic in ic), dtype=int)
# shape (n_conf, n_cf, n_diff,)
cf_ii = conf_seq.configuration_sets[0].cf
m_refs_idx = (
Ellipsis,
cf_ii,
ic_ii,
diff_ii,
)
m_caps_idx = (
Ellipsis,
cf_ii,
diff_ii,
)
init_idxs = {"m_refs_idx": m_refs_idx, "m_caps_idx": m_caps_idx}
return init_idxs
def u_thres_to_code(u, thres, data):
s_thres = get(get(data, "std")[0], "s_thres")[0]
if s_thres > 0:
thres = np.random.normal(thres, s_thres, np.shape(thres))
# reinterpret shape (..., n_diff) to (..., n_thres, n_diff)
code = np.diff(u, axis=-1) if np.size(u, -1) == 2 else u
code = np.sum(code >= thres[..., np.newaxis, :], axis=-1)
return code
def recreate_cache(eff,
caps,
refs,
thres,
ins,
common_mode,
c_seq,
codes,
scalar=None):
fun = sims.Simulator
scalar = default(scalar, len(np.shape(eff)) == 0)
if scalar:
codes = codes[:, np.newaxis, ...]
data = fun.simulate_setup(eff,
caps,
refs,
thres,
ins,
common_mode,
c_seq,
scalar=scalar)
ext_dct = get(data, "extended")[0]
eff, thres, cm = get(ext_dct, "eff", "thres", "cm")
# shape (..., n_conf, n_diff,)
seq_idx = fun.init_seq_idx(c_seq, data)
set_data = None
du_idx = None
trans_idx = None
sets_cache = []
cache = {"data": data, "seq_idx": seq_idx, "sets": sets_cache}
for ii_set, c_set in enumerate(c_seq.configuration_sets):
set_data = fun.init_set(c_set, data, set_data, du_idx)
if set_data["previous"] is not None:
ds_offset = data["ds_offset"]
code = codes[ds_offset:ds_offset + 1, ...]
trans_idx = fun.transition_step_idx(c_set, set_data, data,
code)
ds_offset = data["ds_offset"]
code = codes[ds_offset:ds_offset + c_set.ds_samples, ...]
du_idx = fun.du_indexes(code, c_set, set_data, data)
sets_cache.append((
set_data,
trans_idx,
du_idx,
))
return cache
def du_indexes(code, conf_set, set_data, data):
n_samples = np.size(code, 0)
idx_dct, ext_dct, n_dct = get(data, "indexing", "extended", "n")
idx_set, n_set, ds_map = get(set_data, "indexing", "n", "ds_map")
refs, ins = get(ext_dct, "refs", "ins")
diff_ii, diff_ii_ext, cap_axis, base_ii, base_len = get(
idx_dct, "diff_ii", "diff_ii_ext", "cap_axis", "base_ii",
"base_len")
cs_ii, cs_ii_base, = get(idx_set, "cs_ii", "cs_ii_base")
ds_offset = data["ds_offset"]
in_ref_ii, in_ins_ii = conf_set.generate_in(n_samples, ds_offset)
data["ds_offset"] = ds_offset + n_samples
# Using take instead of indexing because for some reason it's faster
ref_ii = ds_map.take(code, axis=1)
ds_map_transpose = tuple(range(1, base_len + 1 + 2)) + (
0,
-1,
)
# shape = (n_samples, base_shape, n_conf, n_cs, n_diff)
ref_ii = np.transpose(ref_ii, ds_map_transpose)
set_data["indexing"]["ref_ii"] = ref_ii[-1, ...]
ext_idx = (
np.newaxis,
Ellipsis,
)
ext_base_ii = tuple(bb[ext_idx] for bb in base_ii)
m_ref = ext_base_ii + (
cs_ii_base[ext_idx],
ref_ii,
diff_ii_ext[ext_idx],
)
m_in_ref = (Ellipsis, ) + (
cs_ii[ext_idx],
in_ref_ii,
diff_ii[ext_idx],
)
m_in_ins = (Ellipsis, ) + (
in_ins_ii,
diff_ii[ext_idx],
)
ravel = np.ravel_multi_index
r_ref = ravel(m_ref, refs.shape, mode='wrap')
return {"r_ref": r_ref, "m_in_ref": m_in_ref, "m_in_ins": m_in_ins}
def du_compute(indexes, conf_set, set_data, data):
idx_dct, ext_dct, n_dct = get(data, "indexing", "extended", "n")
caps, refs, ins, cm = get(ext_dct, "caps", "refs", "ins", "cm")
cap_axis, base_len, base_ii, diff_ii = get(idx_dct, "cap_axis",
"base_len", "base_ii",
"diff_ii")
n_diff = get(n_dct, "n_diff")[0]
r_ref, m_in_ref, m_in_ins = get(indexes, "r_ref", "m_in_ref",
"m_in_ins")
cs_ii, cf_ii = get(get(set_data, "indexing")[0], "cs_ii", "cf_ii")
std = get(data, "std")[0]
s_ref, s_in = get(std, "s_ref", "s_in")
cs = caps[np.newaxis, ..., cs_ii, diff_ii] # used in du
cf = caps[np.newaxis, ..., cf_ii, diff_ii]
CF = np.sum(cf, axis=-2, keepdims=True) # used in du
# Using take instead of indexing because for some reason it's faster
ref = refs.ravel().take(r_ref)
if s_ref > 0:
ref = np.random.normal(ref, s_ref, size=np.shape(ref))
in_ref = refs[m_in_ref]
if s_ref > 0:
in_ref = np.random.normal(in_ref, s_ref, size=np.shape(in_ref))
in_ins = ins[m_in_ins]
if s_in > 0:
in_ins = np.random.normal(in_ins, s_in, size=np.shape(in_ins))
in_ref = np.transpose(in_ref, (-4, ) + tuple(range(base_len)) + (
-3,
-2,
-1,
))
in_ins = np.transpose(in_ins, (-4, ) + tuple(range(base_len)) + (
-3,
-2,
-1,
))
# sum done un numpy to prevent issue 79 of numexpr
du = np.sum(ne.evaluate("(cs/CF)*(in_ref + in_ins - ref)"),
cap_axis + 1)
if n_diff == 2:
du += cm[np.newaxis, ...] - np.mean(du, axis=-1, keepdims=True)
return du
def init_seq(indexes, data):
idx_dct, ext_dct, n_dct, std = get(data, "indexing", "extended", "n",
"std")
base_shape = get(idx_dct, "base_shape")[0]
caps, refs, cm = get(ext_dct, "caps", "refs", "cm")
n_conf, n_diff = get(n_dct, "n_conf", "n_diff")
s_ref = get(std, "s_ref")[0]
m_refs_idx, m_caps_idx = get(indexes, "m_refs_idx", "m_caps_idx")
u = np.zeros(base_shape + (
n_conf,
n_diff,
))
# shape (base_shape, n_conf, n_cf, n_diff,)
ic_refs = refs[m_refs_idx]
if s_ref > 0:
ic_refs = np.random.normal(ic_refs, s_ref, size=np.shape(ic_refs))
ic_cf = caps[m_caps_idx]
ic_g = ic_cf / np.sum(ic_cf, axis=-2, keepdims=True)
u += np.sum(ic_g * ic_refs, axis=-2)
if n_diff == 2:
u += cm - np.mean(u, axis=-1, keepdims=True)
u = u[np.newaxis, ...]
return u
def init_set(conf_set, data, prev_set_data, prev_du_idx):
idx_dct, ext_dct, n_dct = get(data, "indexing", "extended", "n")
n_refs, n_codes = get(n_dct, "n_refs", "n_codes")
base_len, diff_ii = get(idx_dct, "base_len", "diff_ii")
n_cs = conf_set.n_cs
cs_ii = conf_set.cs
cf_ii = conf_set.cf
cs_ii_base = cs_ii[(np.newaxis, ) * base_len + (Ellipsis, )]
meta = conf_set.meta
ds_map = gen.ds_map(n_cs, n_refs, n_codes, meta.differential)
if prev_du_idx is None:
prev_dct = None
else:
prev_dct = {
"r_ref": prev_du_idx["r_ref"],
"cs_ii": prev_set_data["indexing"]["cs_ii"],
"cf_ii": prev_set_data["indexing"]["cf_ii"]
}
return {
"indexing": {
"cs_ii": cs_ii,
"cf_ii": cf_ii,
"cs_ii_base": cs_ii_base,
},
"n": {
"n_cs": n_cs
},
"previous": prev_dct,
"ds_map": ds_map
}
def recreate(eff,
caps,
refs,
thres,
ins,
common_mode,
c_seq,
cache,
scalar=None,
limit_samples=None):
fun = sims.Simulator
scalar = default(scalar, len(np.shape(eff)) == 0)
limit_samples = default(limit_samples, c_seq.samples) + 1
samples = 0
cache = dict(cache)
cache["data"] = fun.simulate_setup(eff,
caps,
refs,
thres,
ins,
common_mode,
c_seq,
scalar=scalar)
data = cache["data"]
seq_idx = cache["seq_idx"]
u_history = []
# shape (..., n_conf, n_diff,)
u = fun.init_seq(seq_idx, data)
u_history.append(u)
samples += 1
# Simulate each configuration set
dct_idx, dct_ext, dct_n = get(data, "indexing", "extended", "n")
n_conf, n_diff = get(dct_n, "n_conf", "n_diff")
eff, cm = get(dct_ext, "eff", "cm")
base_shape, base_len = get(dct_idx, "base_shape", "base_len")
c_sets = c_seq.configuration_sets
def multi_idx_filter(idx_tuple, sub_idx):
return tuple(ii[sub_idx] if hasattr(ii, "__getitem__") else ii
for ii in idx_tuple)
for ii_set, c_set, data_trans_du in zip(range(len(c_sets)), c_sets,
cache["sets"]):
set_data, trans_idx, du_idx = data_trans_du
if set_data["previous"] is not None:
u = fun.transition_step(trans_idx, u[-1:, ...], set_data, data)
u_history.append(u)
samples += 1
r_du = (1 - eff) * cm
n_u = np.empty((c_set.ds_samples, ) + base_shape + (
n_conf,
n_diff,
))
local_du_idx = dict(du_idx)
# used in u
du = fun.du_compute(du_idx, c_set, set_data, data)
for idx in cartesian(*tuple(range(ss) for ss in base_shape)):
local_idx = tuple(slice(ii, ii + 1)
for ii in idx) + (Ellipsis, )
ext_local_idx = (slice(None), ) + local_idx
local_du_idx["r_ref"] = local_du_idx["r_ref"][ext_local_idx]
local_du_idx["m_in_ref"] = multi_idx_filter(
local_du_idx["m_in_ref"], ext_local_idx)
local_du_idx["m_in_ins"] = multi_idx_filter(
local_du_idx["m_in_ins"], ext_local_idx)
zi = u[(slice(-1, None), ) + local_idx] * eff[local_idx]
local_u = lfilter([1], [1, -eff[local_idx].item()],
du[ext_local_idx] +
r_du[(np.newaxis, ) + local_idx],
zi=zi,
axis=0)[0]
n_u[ext_local_idx] = local_u
u = n_u
u_history.append(u)
samples += np.size(u, 0)
if samples >= limit_samples:
break
u_history = np.concatenate(u_history, axis=0)
u_history = u_history[:limit_samples, ...]
if scalar:
assert np.size(u_history, 1) == 1
u_history = u_history[:, 0, ...]
return u_history
def transition_step(indexes, u, set_data, data):
idx_dct, ext_dct, n_dct = get(data, "indexing", "extended", "n")
cap_axis = get(idx_dct, "cap_axis")[0]
n_diff = get(n_dct, "n_diff")[0]
eff, caps, refs, thres, ins, cm = get(ext_dct, "eff", "caps", "refs",
"thres", "ins", "cm")
m_cf_cs_idx, m_cs_cf_idx, m_cs_cs_idx, m_cf_cf_idx = get(
indexes, "m_cf_cs_idx", "m_cs_cf_idx", "m_cs_cs_idx",
"m_cf_cf_idx")
r_this_r_ref, r_prev_r_ref = get(indexes, "r_this_r_ref",
"r_prev_r_ref")
m_this_in_ref, m_this_in_ins = get(indexes, "m_this_in_ref",
"m_this_in_ins")
std = get(data, "std")[0]
s_ref, s_in = get(std, "s_ref", "s_in")
cf_cs = caps[m_cf_cs_idx]
cs_cf = caps[m_cs_cf_idx]
cs_cs = caps[m_cs_cs_idx] # used in du
cf_cf = caps[m_cf_cf_idx]
prev_ref = refs.ravel().take(r_prev_r_ref)
if s_ref > 0:
prev_ref = np.random.normal(prev_ref,
s_ref,
size=np.shape(prev_ref))
this_ref = refs.ravel().take(r_this_r_ref)
if s_ref > 0:
this_ref = np.random.normal(this_ref,
s_ref,
size=np.shape(this_ref))
this_in_ref = refs[m_this_in_ref] # used in du
if s_ref > 0:
this_in_ref = np.random.normal(this_in_ref,
s_ref,
size=np.shape(this_in_ref))
this_in_ins = ins[m_this_in_ins] # used in du
if s_in > 0:
this_in_ins = np.random.normal(this_in_ins,
s_in,
size=np.shape(this_in_ins))
# used in du
u_gain = (np.sum(cf_cf, axis=-2) + np.sum(cf_cs, axis=-2)) / (
np.sum(cf_cf, axis=-2) + np.sum(cs_cf, axis=-2))
# Sum on next_cs shaped
du_stmt = ("sum(cs_cs*(this_in_ref + this_in_ins)"
" - (cf_cs + cs_cs)*(this_ref), axis={})").format(cap_axis)
du = ne.evaluate(du_stmt)
# Sum on prev_cs shaped
du_stmt = ("sum(cs_cf*prev_ref, axis={})").format(cap_axis)
du += ne.evaluate(du_stmt)
assert np.size(u, 0) == 1, "Only one sample."
CF = np.sum(cf_cf, axis=cap_axis) + np.sum(cs_cf, axis=cap_axis)
# Apply gain and charge loss
u = ne.evaluate("u*u_gain*eff + (1-eff)*cm + du/CF")
# common mode feedback
if n_diff == 2:
u += cm - np.mean(u, axis=-1, keepdims=True)[np.newaxis, ...]
return u
def transition_step_idx(conf_set, set_data, data, code):
assert np.size(code, 0) == 1
idx_dct, ext_dct, n_dct = get(data, "indexing", "extended", "n")
idx_set, pre_set, n_set, ds_map = get(set_data, "indexing", "previous",
"n", "ds_map")
cs_ii, cf_ii, cs_ii_base = get(idx_set, "cs_ii", "cf_ii", "cs_ii_base")
prev_cs_ii, prev_cf_ii = get(pre_set, "cs_ii", "cf_ii")
prev_r_ref = get(pre_set, "r_ref")[0][-1, ...]
diff_ii, diff_ii_ext, cap_axis = get(idx_dct, "diff_ii", "diff_ii_ext",
"cap_axis")
base_shape, base_len, base_ii = get(idx_dct, "base_shape", "base_len",
"base_ii")
eff, caps, refs, thres, ins, cm = get(ext_dct, "eff", "caps", "refs",
"thres", "ins", "cm")
n_conf, n_diff = get(n_dct, "n_conf", "n_diff")
# Compute transition
cs_cf_ii, cf_cs_ii, cs_cs_ii, cf_cf_ii = transition_cx_change(
prev_cs_ii, prev_cf_ii, cs_ii, cf_ii)
# shape(n_cs, ..., n_conf, n_diff)
this_ref_ii = ds_map[:, code, :]
# shape(..., n_conf, n_cs, n_diff)
ds_map_transpose = tuple(range(1, base_len + 1 + 2)) + (
0,
-1,
)
# shape = (n_samples, base_shape, n_conf, n_cs, n_diff) (before idx)
this_ref_ii = np.transpose(this_ref_ii, ds_map_transpose)[0, ...]
ds_offset = data["ds_offset"]
in_ref_ii, in_ins_ii = conf_set.generate_in(1, ds_offset)
data["ds_offset"] = ds_offset + 1
m_this_ref_idx = base_ii + (
cs_ii_base,
this_ref_ii,
diff_ii_ext,
)
r_this_ref_idx = np.ravel_multi_index(m_this_ref_idx, refs.shape)
transition_idx = {
"m_cf_cs_idx": (
Ellipsis,
cf_cs_ii,
diff_ii,
),
"m_cs_cf_idx": (
Ellipsis,
cs_cf_ii,
diff_ii,
),
"m_cs_cs_idx": (
Ellipsis,
cs_cs_ii,
diff_ii,
),
"m_cf_cf_idx": (
Ellipsis,
cf_cf_ii,
diff_ii,
),
"r_this_r_ref": r_this_ref_idx,
"r_prev_r_ref": prev_r_ref,
"m_this_in_ref": (
Ellipsis,
cs_ii,
in_ref_ii[0, ...],
diff_ii,
),
"m_this_in_ins": (
Ellipsis,
in_ins_ii[0, ...],
diff_ii,
)
}
return transition_idx
def simulate(self,
eff,
caps,
refs,
thres,
ins,
common_mode,
c_seq,
scalar=None,
raise_=False):
scalar = default(scalar, len(np.shape(eff)) == 0)
meta = c_seq.meta
data = self.simulate_setup(eff,
caps,
refs,
thres,
ins,
common_mode,
c_seq,
scalar=scalar)
self._standard_deviations(meta, data)
idx_dct, ext_dct, n_dct = get(data, "indexing", "extended", "n")
eff, thres, cm = get(ext_dct, "eff", "thres", "cm")
n_diff = get(n_dct, "n_diff")[0]
u_history = []
with self._random_state as _:
# shape (..., n_conf, n_diff,)
seq_idx = self.init_seq_idx(c_seq, data)
u = self.init_seq(seq_idx, data)
if self.u_history:
u_history.append(u)
# Simulate each configuration set
codes = []
du_idx = None
set_data = None
n_sets = len(c_seq.configuration_sets)
for ii_set, c_set in enumerate(c_seq.configuration_sets):
set_data = self.init_set(c_set, data, set_data, du_idx)
if set_data["previous"] is not None:
code = self.u_thres_to_code(u[-1, ...], thres, data)
code = code[np.newaxis, ...]
trans_idx = self.transition_step_idx(
c_set, set_data, data, code)
u = self.transition_step(trans_idx, u[-1:, ...], set_data,
data)
codes.append(code)
if self.u_history:
u_history.append(u)
for sample in range(c_set.ds_samples):
# print(" {}/{} Sample {}/{} ({:0.2f}%)".format(
# ii_set+1, n_sets, sample+1, c_set.ds_samples, 100*sample/c_set.ds_samples),
# end='\r')
code = self.u_thres_to_code(u[-1, ...], thres, data)
code = code[np.newaxis, ...]
codes.append(code)
du_idx = self.du_indexes(code, c_set, set_data, data)
# used in u
du = self.du_compute(du_idx, c_set, set_data, data)
u = ne.evaluate("u*eff + (1-eff)*cm + du")
# common mode feedback
if n_diff == 2:
u += cm[np.newaxis, ...] - np.mean(
u, axis=-1, keepdims=True)
if self.u_history:
u_history.append(u)
codes = np.concatenate(codes, axis=0)
if self.u_history:
u_history = np.concatenate(u_history, axis=0)
if scalar:
assert np.size(codes, 1) == 1
assert np.size(u_history, 1) == 1
codes = codes[:, 0, ...]
u_history = u_history[:, 0, ...]
if ((u_history < meta.fsr[0] - meta.lsb).any()
or (u_history > meta.fsr[1] + meta.lsb).any()):
message = "Residual out of range."
if raise_:
raise ValueError(message)
else:
warnings.warn(message)
return codes, u_history