def symbol_to_param(sym):
"""
Take a symbol or expression of symbols and returns the corresponding
Parameters
"""
from sympy.core.function import AppliedUndef
from codegeneration import sympycode
if sp is None:
error("sympy is needed for symbol_to_params to work.")
check_arg(sym, (sp.Symbol, AppliedUndef, sp.Derivative),
context=symbol_to_param)
param_str = sympycode(sym)
indexed = re.search(_indexed_format, param_str)
if indexed:
name, indices = indexed.groups()
# Get dict
param = _all_symbol_parameters.get(name)
if param is not None:
param = param.get(eval("({0})".format(indices)))
else:
param = _all_symbol_parameters.get(sympycode(sym))
if param is None:
value_error("No parameter with name '{0}' "\
"registered. Remember to declare Params which should be "\
"used in expression with names.".format(sympycode(sym)))
return param
def value_namespace(expr, include_derivatives=False):
"""
Create a value name space for the included symbols in the expression
"""
from codegeneration import sympycode
check_arg(expr, sp.Basic)
ns = {}
for sym in symbols_from_expr(expr, \
include_derivatives=include_derivatives):
# Get value
value = symbol_to_param(sym).value
# Check for indexed parameters
param_str = sympycode(sym)
indexed = re.search(_indexed_format, param_str)
if indexed:
name, indices = indexed.groups()
if name not in ns:
ns[name] = {}
ns[name][eval(indices)] = value
else:
ns[param_str] = value
return ns
return dict((sympycode(symbol), symbol_to_param(symbol).value) \
for symbol in symbols_from_expr(\
expr, include_derivatives=include_derivatives))
def symbol_to_param(sym):
"""
Take a symbol or expression of symbols and returns the corresponding
Parameters
"""
from sympy.core.function import AppliedUndef
from codegeneration import sympycode
if sp is None:
error("sympy is needed for symbol_to_params to work.")
check_arg(sym, (sp.Symbol, AppliedUndef, sp.Derivative),
context=symbol_to_param)
param_str = sympycode(sym)
indexed = re.search(_indexed_format, param_str)
if indexed:
name, indices = indexed.groups()
# Get dict
param = _all_symbol_parameters.get(name)
if param is not None:
param = param.get(eval("({0})".format(indices)))
else:
param = _all_symbol_parameters.get(sympycode(sym))
if param is None:
value_error("No parameter with name '{0}' "\
"registered. Remember to declare Params which should be "\
"used in expression with names.".format(sympycode(sym)))
return param
def symbol_param_value_namespace(expr):
"""
Create a value name space for the included symbols in the expression
"""
check_arg(expr, sp.Basic)
return dict((str(symbol_param), symbol_to_param(symbol_param).value) \
for symbol_param in iter_symbol_params_from_expr(expr))
def symbol_param_value_namespace(expr):
"""
Create a value name space for the included symbols in the expression
"""
check_arg(expr, sp.Basic)
return dict((str(symbol_param), symbol_to_param(symbol_param).value) \
for symbol_param in iter_symbol_params_from_expr(expr))
def setvalue(self, value):
"""
Set value of ArrayParameter
"""
# An initial slive for the whole array
index = slice(0,len(self._value)+1)
# Tuple means index assignment
# FIXME: Add support for slices
if isinstance(value, tuple):
if len(value) != 2:
value_error("expected a tuple of length 2 when assigning "\
"single items")
if not isinstance(value[0], integers):
value_error("expected first value in index assignment to be"\
" an integer")
if not isinstance(value[1], scalars):
value_error("expected second value in index assignment to be"\
" an scalar")
index = value[0]
value = value[1]
check_arg(value, nptypes, context=ArrayParam.setvalue)
if isinstance(value, np.ndarray):
if len(value) != len(self._value):
value_error("expected the passed array to be of "\
"size: '%d'"%len(self._value))
# Assign value
self._value[index] = self.check(value)
def store_symbol_parameter(param):
"""
Store a symbol parameter
"""
from codegeneration import sympycode
from parameters import ScalarParam
check_arg(param, ScalarParam)
sym = param.sym
#if str(sym) in _all_symbol_parameters:
# warning("Parameter with symbol name '%s' already "\
# "excist" % sym)
param_str = sympycode(sym)
indexed = re.search(_indexed_format, param_str)
if indexed:
name, indices = indexed.groups()
# Get dict
param_dict = _all_symbol_parameters.get(name)
if param_dict is None:
param_dict = {}
_all_symbol_parameters[name] = param_dict
param_dict[eval("({0})".format(indices))] = param
else:
_all_symbol_parameters[param_str] = param
def value_namespace(expr, include_derivatives=False):
"""
Create a value name space for the included symbols in the expression
"""
from codegeneration import sympycode
check_arg(expr, sp.Basic)
ns = {}
for sym in symbols_from_expr(expr, \
include_derivatives=include_derivatives):
# Get value
value = symbol_to_param(sym).value
# Check for indexed parameters
param_str = sympycode(sym)
indexed = re.search(_indexed_format, param_str)
if indexed:
name, indices = indexed.groups()
if name not in ns:
ns[name] = {}
ns[name][eval(indices)] = value
else:
ns[param_str] = value
return ns
return dict((sympycode(symbol), symbol_to_param(symbol).value) \
for symbol in symbols_from_expr(\
expr, include_derivatives=include_derivatives))
def store_symbol_parameter(param):
"""
Store a symbol parameter
"""
from codegeneration import sympycode
from parameters import ScalarParam
check_arg(param, ScalarParam)
sym = param.sym
#if str(sym) in _all_symbol_parameters:
# warning("Parameter with symbol name '%s' already "\
# "excist" % sym)
param_str = sympycode(sym)
indexed = re.search(_indexed_format, param_str)
if indexed:
name, indices = indexed.groups()
# Get dict
param_dict = _all_symbol_parameters.get(name)
if param_dict is None:
param_dict = {}
_all_symbol_parameters[name] = param_dict
param_dict[eval("({0})".format(indices))] = param
else:
_all_symbol_parameters[param_str] = param
def setvalue(self, value):
"""
Set value of ArrayParameter
"""
# An initial slive for the whole array
index = slice(0, len(self._value) + 1)
# Tuple means index assignment
# FIXME: Add support for slices
if isinstance(value, tuple):
if len(value) != 2:
value_error("expected a tuple of length 2 when assigning "\
"single items")
if not isinstance(value[0], integers):
value_error("expected first value in index assignment to be"\
" an integer")
if not isinstance(value[1], scalars):
value_error("expected second value in index assignment to be"\
" an scalar")
index = value[0]
value = value[1]
check_arg(value, nptypes, context=ArrayParam.setvalue)
if isinstance(value, np.ndarray):
if len(value) != len(self._value):
value_error("expected the passed array to be of "\
"size: '%d'"%len(self._value))
# Assign value
self._value[index] = self.check(value)
def __init__(self, value, ge=None, le=None, gt=None, lt=None, \
unit="1", name="", description=""):
"""
Creating a ScalarParam
Arguments
---------
value : scalar
The initial value of this parameter
gt : scalar (optional)
Greater than, range control of argument
ge : scalar (optional)
Greater than or equal, range control of argument
lt : scalar (optional)
Lesser than, range control of argument
le : scalar (optional)
Lesser than or equal, range control of argument
unit : str (optional, if sympy is available)
The unit of the scalar parameter
name : str (optional)
The name of the parameter. Used in pretty printing
description : str (optional)
A description associated with the Parameter
"""
check_arg(value, scalars, 0, ScalarParam)
super(ScalarParam, self).__init__(value, name, description)
self._range = Range(ge, le, gt, lt)
self._in_range = self._range._in_range
check_kwarg(unit, "unit", str)
self._unit = unit
# Define some string used for pretty print
self._in_str = self._range._in_str
self._not_in_str = self._range._not_in_str
# Create symbol
if name == "":
self._sym = dummy_sym
elif sp is None:
self._sym = None
else:
self._sym = sp.Symbol(name,
real=True,
imaginary=False,
commutative=True,
hermitian=True,
complex=True)
# Store parameter
store_symbol_parameter(self)
# Set the value using the check functionality
# (Only if not called from derived class)
if type(self) == ScalarParam:
self.setvalue(value)
def iter_symbol_params_from_expr(expr):
"""
Return an iterator over sp.Symbols from expr
"""
check_arg(expr, sp.Basic)
# Filter out dummy symbols
return (atom for atom in expr.atoms() if isinstance(atom, sp.Symbol) \
and not isinstance(atom, sp.Dummy) and atom.name)
def iter_symbol_params_from_expr(expr):
"""
Return an iterator over sp.Symbols from expr
"""
check_arg(expr, sp.Basic)
# Filter out dummy symbols
return (atom for atom in expr.atoms() if isinstance(atom, sp.Symbol) \
and not isinstance(atom, sp.Dummy) and atom.name)
def __init__(self, value, ge=None, le=None, gt=None, lt=None, \
unit="1", name="", description=""):
"""
Creating a ScalarParam
Arguments
---------
value : scalar
The initial value of this parameter
gt : scalar (optional)
Greater than, range control of argument
ge : scalar (optional)
Greater than or equal, range control of argument
lt : scalar (optional)
Lesser than, range control of argument
le : scalar (optional)
Lesser than or equal, range control of argument
unit : str (optional, if sympy is available)
The unit of the scalar parameter
name : str (optional)
The name of the parameter. Used in pretty printing
description : str (optional)
A description associated with the Parameter
"""
check_arg(value, scalars, 0, ScalarParam)
super(ScalarParam, self).__init__(value, name, description)
self._range = Range(ge, le, gt, lt)
self._in_range = self._range._in_range
check_kwarg(unit, "unit", str)
self._unit = unit
# Define some string used for pretty print
self._in_str = self._range._in_str
self._not_in_str = self._range._not_in_str
# Create symbol
if name == "":
self._sym = dummy_sym
elif sp is None:
self._sym = None
else:
self._sym = sp.Symbol(name, real=True, imaginary=False,
commutative=True, hermitian=True, complex=True)
# Store parameter
store_symbol_parameter(self)
# Set the value using the check functionality
# (Only if not called from derived class)
if type(self) == ScalarParam:
self.setvalue(value)
def __init__(self, value, options, name="", description=""):
"""
Initialize the OptionParam
Arguments
---------
value : scalars or str
The initial value of this parameter. The type of value is stored
for future type checks
options : list
A list of acceptable values for the parameter
name : str (optional)
The name of the parameter. Used in pretty printing
description : str (optional)
A description associated with the Parameter
"""
check_arg(options, list)
if len(options) < 2:
value_error(
"expected the options argument to be at least of length 2")
super(OptionParam, self).__init__(value, name, description)
# Check valid types for an 'option check'
for option in options:
if not isinstance(option, option_types):
type_error("options can only be 'str' and scalars got: '%s'" % \
type(option).__name__)
# Define a 'check function'
self._in_range = lambda value: value in options
# Define some string used for pretty print
self._in_str = "%%s \xe2\x88\x88 %s" % repr(options)
self._not_in_str = "%%s \xe2\x88\x89 %s" % repr(options)
# Define a 'repr string'
#self._repr_str = "OptionParam(%%s, %s)" % repr(options)
# Set the value using the check functionality
self.setvalue(value)
# Check that all values in options has the same type
for val in options:
if not isinstance(val, self.value_type):
type_error("All values of the 'option check' " +\
"need to be of type: '%s'" % type(self._value).__name__)
# Store options
self._options = options
def __init__(self, value, options, name="", description=""):
"""
Initialize the OptionParam
Arguments
---------
value : scalars or str
The initial value of this parameter. The type of value is stored
for future type checks
options : list
A list of acceptable values for the parameter
name : str (optional)
The name of the parameter. Used in pretty printing
description : str (optional)
A description associated with the Parameter
"""
check_arg(options, list)
if len(options) < 2:
value_error("expected the options argument to be at least of length 2")
super(OptionParam, self).__init__(value, name, description)
# Check valid types for an 'option check'
for option in options:
if not isinstance(option, option_types):
type_error("options can only be 'str' and scalars got: '%s'" % \
type(option).__name__)
# Define a 'check function'
self._in_range = lambda value : value in options
# Define some string used for pretty print
self._in_str = "%%s \xe2\x88\x88 %s" % repr(options)
self._not_in_str = "%%s \xe2\x88\x89 %s" % repr(options)
# Define a 'repr string'
#self._repr_str = "OptionParam(%%s, %s)" % repr(options)
# Set the value using the check functionality
self.setvalue(value)
# Check that all values in options has the same type
for val in options:
if not isinstance(val, self.value_type):
type_error("All values of the 'option check' " +\
"need to be of type: '%s'" % type(self._value).__name__)
# Store options
self._options = options
def __init__(self, opts, nfilter_max=512, **kwargs):
super(EngineGenerator, self).__init__()
self.opts = opts
self.z_dim = opts.z
self.num_components = opts.num_components
self.hdim = opts.hidden_dim
self.expand_dim = opts.nfilterG
if opts.do_memory:
self.base_dim = opts.memory_dim
else:
self.base_dim = opts.hidden_dim
state_dim_multiplier = 1
if utils.check_arg(self.opts, 'state_dim_multiplier'):
state_dim_multiplier = self.opts.state_dim_multiplier
state_dim = self.base_dim * state_dim_multiplier
# Memory Module
if self.opts.do_memory:
self.memory = Memory(opts, state_dim)
# Dynamics Engine
self.engine = EngineModule(opts, state_dim)
self.simple_enc = model_utils.choose_netG_encoder(basechannel=state_dim, opts=self.opts)
# Rendering Engine
self.graphics_renderer = RenderingEngine(G_ch=opts.nfilterG, opts=opts, resolution=self.opts.img_size[0])
self.num_components = self.opts.num_components
def run_warmup(self,
zdist,
states,
actions,
warm_up,
train=True,
M=None,
prev_alpha=None,
prev_read_v=None,
force_sharp=False):
'''
Run warm-up phase
'''
batch_size = states[0].size(0)
prev_state = None
h, c = self.engine.init_hidden(batch_size)
h = utils.check_gpu(self.opts.gpu, h)
c = utils.check_gpu(self.opts.gpu, c)
outputs, maps, zs, alphas, alpha_logits = [], [], [], [], []
init_maps = []
if utils.check_arg(self.opts, 'do_memory'):
# initialize memory and alpha
if M is None:
M = self.memory.init_memory(batch_size)
if prev_alpha is None:
prev_alpha = utils.check_gpu(
self.opts.gpu, torch.zeros(batch_size,
self.memory.num_mem))
mem_wh = int(math.sqrt(prev_alpha.size(1)))
prev_alpha[:, mem_wh * (mem_wh // 2) + mem_wh // 2] = 1.0
if prev_read_v is None:
prev_read_v = utils.check_gpu(
self.opts.gpu, torch.zeros(batch_size,
self.opts.memory_dim))
alpha_losses = 0
base_imgs_all = []
hiddens = []
for i in range(warm_up):
input_state = states[i]
prev_state, m, prev_alpha, alpha_loss, z, M, prev_read_v, h, c, init_map, base_imgs, _, cur_hidden = self.run_step(
input_state, h, c, actions[i], \
batch_size, prev_read_v, prev_alpha, M, zdist, step=i, force_sharp=force_sharp)
outputs.append(prev_state)
maps.append(m)
alphas.append(prev_alpha)
alpha_losses += alpha_loss
zs.append(z)
base_imgs_all.append(base_imgs)
hiddens.append(cur_hidden)
init_maps.append(init_map)
warm_up_state = [h, c]
if prev_state is None:
prev_state = states[
0] # warm_up is 0, the initial screen is always used
return prev_state, warm_up_state, M, prev_read_v, prev_alpha, outputs, maps, alphas, alpha_losses, zs, base_imgs_all, 0, \
hiddens, init_maps
def _set_name(self, name):
"""
Set the name. Can only be done if not set during instantiation
"""
check_arg(name, str)
super(ScalarParam, self)._set_name(name)
if sp is None:
return
# Create a new symbol with the updated name
self._sym = sp.Symbol(name, real=True, imaginary=False,
commutative=True, hermitian=True, complex=True)
# Store parameter
store_symbol_parameter(self)
def __getitem__(self, idx):
data = np.load(self.samples[idx], encoding='latin1')
states, actions, neg_actions = [], [], []
ep_len = len(data[0]['np_img_state']) - self.opts.num_steps
start_pt = random.randint(0, max(ep_len - 1, 0))
i = 0
samples = []
cur_sample = 0
while i < self.opts.num_steps:
try:
if start_pt + i >= len(data[0]['np_img_state']) - 2:
cur_s = data[0]['np_img_state'][
len(data[0]['np_img_state']) - 2]
cur_a = np.zeros(self.opts.action_space)
cur_a[0] = 1
else:
cur_s = data[0]['np_img_state'][start_pt + i]
cur_a = data[0]['np_action'][start_pt + i]
except:
import pdb
pdb.set_trace()
if self.opts.img_size[0] != cur_s.shape[1] or self.opts.img_size[
1] != cur_s.shape[0]:
cur_s = cv2.resize(
cur_s.astype('float32'),
dsize=(self.opts.img_size[0], self.opts.img_size[1]),
)
s_t = (np.transpose(cur_s, axes=(2, 0, 1)) /
255.).astype('float32')
s_t = (s_t - 0.5) / 0.5
if utils.check_arg(self.opts, 'normalize_mean'):
s_t = s_t - np.array([-0.9219, -0.9101, -0.8536]).reshape(
(3, 1, 1)).astype('float32')
s_t = s_t / 1.9219
a_t = np.copy(cur_a).astype('float32')
action_idx = cur_a.tolist().index(1)
# false action
false_a_idx = random.randint(0, 4)
while false_a_idx == action_idx:
false_a_idx = random.randint(0, 4)
false_a_t = np.zeros(cur_a.shape).astype('float32')
false_a_t[false_a_idx] = 1
# save
states.append(s_t)
actions.append(a_t)
neg_actions.append(false_a_t)
samples.append(cur_sample)
i = i + 1
return states, actions, neg_actions
def __getattr__(self, key):
check_arg(key, str, 0, ParameterDict.__getattr__)
# Fix for newer ipython
if key in [
"__dict__", "__methods__", "trait_names", "_getAttributeNames"
]:
return
if not dict.__contains__(self, key):
error("'%s' is not an item in this ParameterDict." % key, \
exception=AttributeError)
value = dict.__getitem__(self, key)
if isinstance(value, Param):
value = value.getvalue()
return value
def update(self, other):
"""
A recursive update that handles parameter subsets
correctly unlike dict.update.
"""
check_arg(other, dict, 0, ParameterDict.update)
for key in dict.iterkeys(other):
if key not in self:
continue
self_value = self[key]
other_value = other[key]
if isinstance(self_value, dict):
# Update my own subdict with others subdict
self_value.update(other_value)
elif isinstance(self_value, Param):
# Set my own value to others value
self_value.setvalue(other_value)
else:
self[key] = other_value
def _set_name(self, name):
"""
Set the name. Can only be done if not set during instantiation
"""
check_arg(name, str)
super(ScalarParam, self)._set_name(name)
if sp is None:
return
# Create a new symbol with the updated name
self._sym = sp.Symbol(name,
real=True,
imaginary=False,
commutative=True,
hermitian=True,
complex=True)
# Store parameter
store_symbol_parameter(self)
def __setattr__(self, key, value):
check_arg(key, str, 0, ParameterDict.__setattr__)
if key == "_members":
dict.__setattr__(self, key, value)
return
# Check if key is a registered parameter
if not dict.__contains__(self, key):
error("'%s' is not an item in this ParameterDict." % key, \
exception=AttributeError)
# Get the original value, used for checks
org_value = dict.__getitem__(self, key)
if isinstance(org_value, ParameterDict):
type_error("cannot overwrite a ParameterDict")
# Set the new value
if isinstance(org_value, Param):
org_value.setvalue(value)
else:
dict.__setitem__(self, key, value)
def __init__(self, opts, set_type=0, permute_color=False, datadir=''):
self.opts = opts
self.set_type = set_type
self.permute_color = permute_color
self.samples = []
num_data = len(os.listdir(datadir))
if set_type == 0:
sample_list = list(range(0, int(num_data * 0.9)))
else:
sample_list = list(range(int(num_data * 0.9), num_data))
for el in sample_list:
self.samples.append('%s/%d.npy' % (datadir, el))
self.end_bias = 0
if utils.check_arg(self.opts, 'end_bias'):
self.end_bias = self.opts.end_bias
def add_pair_to_subs(subs, old, new):
"""
Add a pair of old and new symbols to subs. If a subs with old as a
key already excist it will be removed before insertion.
"""
check_arg(subs, list, 0)
check_arg(old, sp.Basic, 1)
check_arg(new, sp.Basic, 2)
for ind, (old0, new0) in enumerate(subs):
if old0 == old:
subs.pop(ind)
break
subs.append((old, new))
def add_pair_to_subs(subs, old, new):
"""
Add a pair of old and new symbols to subs. If a subs with old as a
key already excist it will be removed before insertion.
"""
check_arg(subs, list, 0)
check_arg(old, sp.Basic, 1)
check_arg(new, sp.Basic, 2)
for ind, (old0, new0) in enumerate(subs):
if old0 == old:
subs.pop(ind)
break
subs.append((old, new))
def __init__(self, opts, set_type=0, permute_color=False, datadir=''):
self.opts = opts
self.set_type = set_type
self.permute_color = permute_color
self.samples = []
files = os.listdir(datadir)
num_data = len(files)
if set_type == 0:
sample_list = files[:int(num_data * 0.9)]
else:
sample_list = files[int(num_data * 0.9):]
# Here's the difference - we're using gzipped pickle
# (additionally checking if teh file exists)
for file in sample_list:
path = f'{datadir}/{file}'
self.samples.append(path)
# Bias
self.end_bias = self.opts.end_bias if utils.check_arg(
self.opts, 'end_bias') else 0.5
def __init__(self, opts, state_dim):
super(EngineModule, self).__init__()
self.hdim = opts.hidden_dim
self.opts = opts
action_space = 10 if not utils.check_arg(self.opts, 'action_space') else self.opts.action_space
self.a_to_input = nn.Sequential(nn.Linear(action_space, self.hdim),
nn.LeakyReLU(0.2))
self.z_to_input = nn.Sequential(nn.Linear(opts.z, self.hdim),
nn.LeakyReLU(0.2))
# engine module input network
e_input_dim = self.hdim*2
if self.opts.do_memory:
e_input_dim += self.opts.memory_dim
self.f_e = nn.Sequential(nn.Linear(e_input_dim, e_input_dim),
nn.LeakyReLU(0.2))
self.rnn_e = ActionLSTM(e_input_dim,
hidden_size=self.hdim, opts=opts)
self.state_bias = nn.Sequential(nn.Linear(state_dim, self.hdim * 4))
def run_step(self, state, h, c, action, batch_size, prev_read_v, prev_alpha, M, zdist,
read_only=False, step=0, decode=True, play=False, force_sharp=False):
'''
Run the model one time step
'''
# encode the image input
if self.opts.input_detach:
state = state.detach()
s = self.simple_enc(state)
# sample a noise
z = utils.check_gpu(self.opts.gpu, zdist.sample((batch_size,)))
# run dynamics engine
prev_hidden = h[0].clone()
h, c, cur_hidden = self.engine(h, c, s, action, z, prev_read_v=prev_read_v, step=step)
# run memory module
if self.opts.do_memory:
base, M, alpha, prev_read_v = self.memory(cur_hidden, action, prev_hidden, prev_alpha, M, c=c[0], read_only=read_only, force_sharp=force_sharp)
prev_alpha = alpha
bases = base
else:
base = cur_hidden
bases = [base] * self.num_components
# run the rendering engine
alpha_loss = 0
out, m, eloss, init_maps, base_imgs = self.graphics_renderer(bases, num_components=self.num_components)
if utils.check_arg(self.opts, 'alpha_loss_multiplier') and self.opts.alpha_loss_multiplier > 0:
# memory regularization
for i in range(1, len(m)):
alpha_loss += (m[i].abs().sum() / batch_size)
prev_state = out
return prev_state, m, prev_alpha, alpha_loss, z, M, prev_read_v, h, c, init_maps, base_imgs, 0, cur_hidden
you must specify sourcename and device identifier
"""
elif sys.argv[0] == "list":
for file in os.listdir(config.map["global"]["flow_dir"]):
if file.endswith(".flow"):
name = file[file.find("/") + 1 : -5]
print "Flow name: " + name
fmap = config.read_struct(config.map["global"]["flow_dir"] + "/" + file)
print "\tSource type(s): [" + ",".join(fmap["source_types"]) + "]"
print "\tTasks:\n\t\t" + "\n\t\t".join(fmap["tasks"])
print ""
elif sys.argv[0] == "run":
fromtime = datetime.now() - timedelta(weeks=1)
totime = datetime.now()
ret, vals, sys.argv = utils.check_arg(sys.argv, "--from", 1)
if ret:
try:
fromtime = utils.str_to_date(vals[0])
except ValueError, msg:
print "Bad from time: " + str(msg)
ret, vals, sys.argv = utils.check_arg(sys.argv, "--to", 1)
if ret:
try:
totime = utils.str_to_date(vals[0])
except ValueError, msg:
print "Bad to time: " + str(msg)
pretend, j, sys.argv = utils.check_arg(sys.argv, "--pretend")
local, j, sys.argv = utils.check_arg(sys.argv, "--local")
#!/usr/bin/python3
import sys
from utils import check_arg, get_values, print_res
from bubble import bubble
from selection import selection
from insertion import insertion
from merge import merge
from quicksort import quicksort
def main(args):
numbers = get_values(args[0])
if (len(numbers) <= 0):
sys.exit(84)
print(len(numbers), " element", sep="", end="")
print("s" if len(numbers) > 1 else "")
selection(numbers[::])
insertion(numbers[::])
bubble(numbers[::])
print_res("Quicksort:", 0 if (len(numbers) <= 1) else quicksort(numbers[::])[1])
print_res("Merge sort:", merge(numbers[::])[1])
try:
check_arg(sys.argv)
main(sys.argv[1:])
except:
sys.exit(84)
entries = learningdb.getHMMGaussianEmissions(id);
gparm = learningdb.computeGlobalHMMGaussianParameters(entries);
print "\tPlugload \"%s\" Total Learning Sets: %d"%(plname,len(entries))
for g in gparm:
print "\t\tState %-3d: N=%-7d mean=%20.10e variance=%20.10e"%(g.state_id,g.counts,g.mean,g.variance)
print "\n";
elif (op == 'insert'):
import devicedb,postgresops
devicedb.connect();
fromtime = None
totime = None
dtype = "GHMM"
ret,vals,sys.argv = utils.check_arg(sys.argv,'--from',1)
if ret:
try:
fromtime = utils.str_to_date(vals[0])
except ValueError, msg:
print "Bad from time: "+str(msg)
ret,vals,sys.argv = utils.check_arg(sys.argv,'--to',1)
if ret:
try:
totime = utils.str_to_date(vals[0])
except ValueError, msg:
print "Bad to time: "+str(msg)
ret,vals,sys.argv = utils.check_arg(sys.argv,'--type',1)
if ret:
def __init__(self, opts, nfilter=32, nfilter_max=1024):
super(Discriminator, self).__init__()
self.opts = opts
self.simple_blocks = utils.check_arg(self.opts, 'simple_blocks')
# single frame discrimiantor
f_size = 4
if self.opts.img_size[0] == 84:
f_size = 5
elif self.simple_blocks:
f_size = 3
expand_dim = opts.nfilterD * 4
if opts.img_size[0] == 128:
self.ds = nn.Sequential(
SN(nn.Conv2d(3, expand_dim, 4, 2)), nn.LeakyReLU(0.2),
SN(nn.Conv2d(expand_dim, expand_dim * 2, 3, 2)),
nn.LeakyReLU(0.2),
SN(nn.Conv2d(expand_dim * 2, expand_dim * 4, 3, 2)),
nn.LeakyReLU(0.2),
SN(nn.Conv2d(expand_dim * 4, expand_dim * 4, 3, 2)),
nn.LeakyReLU(0.2),
SN(nn.Conv2d(expand_dim * 4, expand_dim * 4, 3, 2)),
nn.LeakyReLU(0.2), View((-1, expand_dim * 4, 3, 3)))
else:
self.ds = nn.Sequential(
SN(nn.Conv2d(3, expand_dim, 4, 2)), nn.LeakyReLU(0.2),
SN(nn.Conv2d(expand_dim, expand_dim * 2, 3, 2)),
nn.LeakyReLU(0.2),
SN(nn.Conv2d(expand_dim * 2, expand_dim * 4, 3, 2)),
nn.LeakyReLU(0.2),
SN(nn.Conv2d(expand_dim * 4, expand_dim * 4, 3, 2)),
nn.LeakyReLU(0.2), View((-1, expand_dim * 4, 3, 3)))
# patch level logits
self.single_frame_discriminator_patch = nn.Sequential(
SN(nn.Conv2d(expand_dim * 4, expand_dim * 4, 2, 1, 1)),
nn.LeakyReLU(0.2), SN(nn.Conv2d(expand_dim * 4, 1, 2, 1)),
View((-1, 1, 3, 3)))
# single logit for entire image
self.single_frame_discriminator_all = nn.Sequential(
SN(nn.Conv2d(expand_dim * 4, expand_dim * 4, 2, 1)),
nn.LeakyReLU(0.2), SN(nn.Conv2d(expand_dim * 4, 1, 2, 1)),
View((-1, 1)))
conv3d_dim = self.opts.nfilterD_temp
else:
# bigGAN discriminator architecture
self.ds = DiscriminatorSingle(D_ch=opts.nfilterD,
opts=opts,
resolution=self.opts.img_size[0])
conv3d_dim = self.opts.nfilterD_temp
# temporal discriminator
self.temporal_window = self.opts.config_temporal
self.conv3d, self.conv3d_final = \
model_utils.choose_netD_temporal(
self.opts, conv3d_dim, window=self.temporal_window
)
self.conv3d = nn.ModuleList(self.conv3d)
self.conv3d_final = nn.ModuleList(self.conv3d_final)
self.which_conv = functools.partial(layers.SNConv2d,
kernel_size=f_size,
padding=0,
num_svs=1,
num_itrs=1,
eps=1e-12)
self.which_linear = functools.partial(layers.SNLinear,
num_svs=1,
num_itrs=1,
eps=1e-12)
# action-conditioned discriminator
self.trans_conv = self.which_conv(opts.nfilterD * 16 * 2, 256)
self.action_linear1 = self.which_linear(512, 512)
self.action_linear_out = self.which_linear(512, 1)
action_space = 10 if not utils.check_arg(
self.opts, 'action_space') else self.opts.action_space
self.action_to_feat = nn.Linear(action_space, 256)
self.to_transition_feature = nn.Sequential(self.trans_conv,
nn.LeakyReLU(0.2),
View((-1, 256)))
self.action_discriminator = nn.Sequential(self.action_linear1,
nn.LeakyReLU(0.2),
self.action_linear_out)
self.reconstruct_action_z = nn.Sequential(
nn.Linear(256, action_space + self.opts.z), ) # 4, 1, 0),
def _set_name(self, name):
check_arg(name, str)
if self._name:
value_error("Cannot set name attribute of %s, it is already set "\
"to '%s'" % (self.__class__.__name__, self._name))
self._name = name
def __init__(self, value, size=None, ge=None, le=None, gt=None, lt=None, \
unit="1", name="", description=""):
"""
Creating an ArrayParam
Arguments
---------
value : scalar, np.ndarray
The initial value of this parameter
size : integer (optional)
Set the size of the ns.array. If set value must be a scalar
gt : scalar (optional)
Greater than, range control of argument
ge : scalar (optional)
Greater than or equal, range control of argument
lt : scalar (optional)
Lesser than, range control of argument
le : scalar (optional)
Lesser than or equal, range control of argument
unit : str (optional, if sympy is available)
The unit of the scalar parameter
name : str (optional)
The name of the parameter. Used in pretty printing
description : str (optional)
A description associated with the Parameter
"""
if np is None:
error("numpy is not installed so ArrayParam is not available")
# If setting value using size
if size is not None:
# If a size is provided a scalar is expected for the value
check_kwarg(size, "size", integers, ArrayParam, ge=1)
check_arg(value, scalars, 0, ArrayParam)
# Create numpy array based on the passed value
# Use intc and float_ to be compatible with c code.
value = np.array([value]*size, dtype=np.intc \
if isinstance(value, integers) \
else np.float_)
# If setting value using only value argument
else:
# Allow using list of scalars
if isinstance(value, list):
check_arg(value, list, 0, ArrayParam, scalars)
if len(value) == 0:
value_error("expected a list with at least 1 element")
value = np.fromiter(value, dtype=type(value[0]))
check_arg(value, nptypes, 0, ArrayParam)
# Fist check any scalars passed
if isinstance(value, integers):
value = np.array([value], dtype=np.intc)
elif isinstance(value, scalars):
value = np.array([value], dtype=np.float_)
# Then check passed NumPy arrays
elif value.dtype in integers:
value = value.astype(np.intc)
elif value.dtype in scalars:
value = value.astype(np.float_)
else:
type_error("expected a scalar or a scalar valued np.ndarray"
"or list as value argument.")
# Init super class with dummy value
super(ArrayParam, self).__init__(value[0], ge, le, gt, lt, unit, \
name, description)
# Assign value
self._value = value
self.value_type = nptypes
# Use setvalue to set value using the range
self.setvalue(value)
def _set_name(self, name):
check_arg(name, str)
if self._name:
value_error("Cannot set name attribute of %s, it is already set "\
"to '%s'" % (self.__class__.__name__, self._name))
self._name = name
def __init__(self, value, size=None, ge=None, le=None, gt=None, lt=None, \
unit="1", name="", description=""):
"""
Creating an ArrayParam
Arguments
---------
value : scalar, np.ndarray
The initial value of this parameter
size : integer (optional)
Set the size of the ns.array. If set value must be a scalar
gt : scalar (optional)
Greater than, range control of argument
ge : scalar (optional)
Greater than or equal, range control of argument
lt : scalar (optional)
Lesser than, range control of argument
le : scalar (optional)
Lesser than or equal, range control of argument
unit : str (optional, if sympy is available)
The unit of the scalar parameter
name : str (optional)
The name of the parameter. Used in pretty printing
description : str (optional)
A description associated with the Parameter
"""
if np is None:
error("numpy is not installed so ArrayParam is not available")
# If setting value using size
if size is not None:
# If a size is provided a scalar is expected for the value
check_kwarg(size, "size", integers, ArrayParam, ge=1)
check_arg(value, scalars, 0, ArrayParam)
# Create numpy array based on the passed value
# Use intc and float_ to be compatible with c code.
value = np.array([value]*size, dtype=np.intc \
if isinstance(value, integers) \
else np.float_)
# If setting value using only value argument
else:
# Allow using list of scalars
if isinstance(value, list):
check_arg(value, list, 0, ArrayParam, scalars)
if len(value)==0:
value_error("expected a list with at least 1 element")
value = np.fromiter(value, dtype=type(value[0]))
check_arg(value, nptypes, 0, ArrayParam)
# Fist check any scalars passed
if isinstance(value, integers):
value = np.array([value], dtype=np.intc)
elif isinstance(value, scalars):
value = np.array([value], dtype=np.float_)
# Then check passed NumPy arrays
elif value.dtype in integers:
value = value.astype(np.intc)
elif value.dtype in scalars:
value = value.astype(np.float_)
else:
type_error("expected a scalar or a scalar valued np.ndarray"
"or list as value argument.")
# Init super class with dummy value
super(ArrayParam, self).__init__(value[0], ge, le, gt, lt, unit, \
name, description)
# Assign value
self._value = value
self.value_type = nptypes
# Use setvalue to set value using the range
self.setvalue(value)
def choose_netD_temporal(opts, conv3d_dim, window=[]):
'''
temporal discriminator
'''
in_dim = opts.nfilterD * 16
extractors, finals = [], []
# temporarily hand-designed for steps 6 / 12 / 18
first_spatial_filter = 2
if opts.img_size[0] == 84:
first_spatial_filter = 3
if utils.check_arg(opts, 'simple_blocks'):
net1 = nn.Sequential(
SN(nn.Conv3d(in_dim, conv3d_dim, (2, 2, 2), (1, 1, 1))),
nn.LeakyReLU(0.2),
SN(nn.Conv3d(conv3d_dim, conv3d_dim * 2, (3, 2, 2), (2, 1, 1))),
nn.LeakyReLU(0.2))
head1 = nn.Sequential(
SN(nn.Conv3d(conv3d_dim * 2, 1, (2, 1, 1), (2, 1, 1))), )
extractors.append(net1)
finals.append(head1)
if window > 6: # 18
net2 = nn.Sequential(
SN(
nn.Conv3d(conv3d_dim * 2, conv3d_dim * 4, (3, 1, 1),
(2, 1, 1))),
nn.LeakyReLU(0.2),
)
head2 = nn.Sequential(SN(nn.Conv3d(conv3d_dim * 4, 1,
(2, 1, 1))), )
extractors.append(net2)
finals.append(head2)
if window > 18: # 32
net3 = nn.Sequential(
SN(
nn.Conv3d(conv3d_dim * 4, conv3d_dim * 8, (3, 1, 1),
(2, 1, 1))),
nn.LeakyReLU(0.2),
)
head3 = nn.Sequential(SN(nn.Conv3d(conv3d_dim * 8, 1,
(3, 1, 1))), )
extractors.append(net3)
finals.append(head3)
elif 'sn' in opts.D_temp_mode:
net1 = nn.Sequential(
SN(
nn.Conv3d(in_dim, conv3d_dim,
(2, first_spatial_filter, first_spatial_filter),
(1, 1, 1))), nn.LeakyReLU(0.2),
SN(nn.Conv3d(conv3d_dim, conv3d_dim * 2, (3, 3, 3), (2, 1, 1))),
nn.LeakyReLU(0.2))
head1 = nn.Sequential(
SN(nn.Conv3d(conv3d_dim * 2, 1, (2, 1, 1), (1, 1, 1))), )
extractors.append(net1)
finals.append(head1)
if window >= 12: # 12
net2 = nn.Sequential(
SN(
nn.Conv3d(conv3d_dim * 2, conv3d_dim * 4, (3, 1, 1),
(1, 1, 1))),
nn.LeakyReLU(0.2),
)
head2 = nn.Sequential(SN(nn.Conv3d(conv3d_dim * 4, 1,
(3, 1, 1))), )
extractors.append(net2)
finals.append(head2)
if window >= 18: # 18
net3 = nn.Sequential(
SN(
nn.Conv3d(conv3d_dim * 4, conv3d_dim * 8, (3, 1, 1),
(2, 1, 1))),
nn.LeakyReLU(0.2),
)
if window == 18 or window == 28:
head3 = nn.Sequential(
SN(nn.Conv3d(conv3d_dim * 8, 1, (2, 1, 1), (1, 1, 1))), )
else:
head3 = nn.Sequential(
SN(nn.Conv3d(conv3d_dim * 8, 1, (4, 1, 1), (2, 1, 1))), )
extractors.append(net3)
finals.append(head3)
else:
net1 = nn.Sequential(
nn.Conv3d(in_dim, conv3d_dim, (2, 3, 3), (1, 1, 1)),
nn.LeakyReLU(0.2),
nn.Conv3d(conv3d_dim, conv3d_dim * 2, (3, 3, 3), (2, 1, 1)),
nn.LeakyReLU(0.2))
head1 = nn.Sequential(
nn.Conv3d(conv3d_dim * 2, 1, (2, 1, 1), (1, 1, 1)), )
extractors.append(net1)
finals.append(head1)
if window >= 12: # 12
net2 = nn.Sequential(
nn.Conv3d(conv3d_dim * 2, conv3d_dim * 4, (3, 1, 1),
(1, 1, 1)),
nn.LeakyReLU(0.2),
)
head2 = nn.Sequential(nn.Conv3d(conv3d_dim * 4, 1, (3, 1, 1)), )
extractors.append(net2)
finals.append(head2)
if window >= 18: # 18
net3 = nn.Sequential(
nn.Conv3d(conv3d_dim * 4, conv3d_dim * 8, (2, 1, 1),
(2, 1, 1)),
nn.LeakyReLU(0.2),
)
head3 = nn.Sequential(nn.Conv3d(conv3d_dim * 8, 1, (3, 1, 1)), )
extractors.append(net3)
finals.append(head3)
return extractors, finals
def __init__(self, G_ch=64, dim_z=512, bottom_width=8, resolution=64,
G_kernel_size=3, G_attn='64', n_classes=1000,
num_G_SVs=1, num_G_SV_itrs=1,
G_shared=True, shared_dim=0, hier=False,
cross_replica=False, mybn=False,
G_activation=nn.ReLU(inplace=False),
G_lr=5e-5, G_B1=0.0, G_B2=0.999, adam_eps=1e-8,
BN_eps=1e-5, SN_eps=1e-12, G_mixed_precision=False, G_fp16=False,
G_init='ortho', skip_init=False, no_optim=False,
G_param='SN', norm_style='bn', opts=None,
**kwargs):
super(RenderingEngine, self).__init__()
# Channel width mulitplier
self.ch = G_ch
# Dimensionality of the latent space
self.dim_z = opts.hidden_dim if not opts.do_memory else opts.memory_dim
# The initial spatial dimensions
if resolution == 84 or utils.check_arg(opts, 'simple_blocks'):
bottom_width = 7
self.bottom_width = bottom_width
# Resolution of the output
self.resolution = resolution
# Kernel size?
self.kernel_size = G_kernel_size
# Attention?
self.attention = G_attn
# number of classes, for use in categorical conditional generation
self.n_classes = n_classes
# Use shared embeddings?
self.G_shared = G_shared
# Dimensionality of the shared embedding? Unused if not using G_shared
self.shared_dim = shared_dim if shared_dim > 0 else dim_z
# Hierarchical latent space?
self.hier = hier
# Cross replica batchnorm?
self.cross_replica = cross_replica
# Use my batchnorm?
self.mybn = mybn
# nonlinearity for residual blocks
self.activation = G_activation
# Initialization style
self.init = G_init
# Parameterization style
self.G_param = G_param
# Normalization style
self.norm_style = norm_style
# Epsilon for BatchNorm?
self.BN_eps = BN_eps
# Epsilon for Spectral Norm?
self.SN_eps = SN_eps
# fp16?
self.fp16 = G_fp16
# Architecture dict
self.arch = G_arch(self.ch, self.attention)[resolution]
self.opts = opts
# If using hierarchical latents, adjust z
if self.hier:
# Number of places z slots into
self.num_slots = len(self.arch['in_channels']) + 1
self.z_chunk_size = (self.dim_z // self.num_slots)
# Recalculate latent dimensionality for even splitting into chunks
self.dim_z = self.z_chunk_size * self.num_slots
else:
self.num_slots = 1
self.z_chunk_size = 0
# Which convs, batchnorms, and linear layers to use
if self.G_param == 'SN':
self.which_conv = functools.partial(layers.SNConv2d,
kernel_size=3, padding=1,
num_svs=num_G_SVs, num_itrs=num_G_SV_itrs,
eps=self.SN_eps)
self.which_conv = functools.partial(layers.SNConv2d,
kernel_size=3, padding=1,
num_svs=num_G_SVs, num_itrs=num_G_SV_itrs,
eps=self.SN_eps)
self.which_conv_ker2 = functools.partial(layers.SNConv2d,
kernel_size=2, padding=0,
num_svs=num_G_SVs, num_itrs=num_G_SV_itrs,
eps=self.SN_eps)
self.get_map_conv = functools.partial(layers.SNConv2d,
kernel_size=1, padding=0,
num_svs=num_G_SVs, num_itrs=num_G_SV_itrs,
eps=self.SN_eps)
self.which_linear = functools.partial(layers.SNLinear,
num_svs=num_G_SVs, num_itrs=num_G_SV_itrs,
eps=self.SN_eps)
else:
self.which_conv = functools.partial(nn.Conv2d, kernel_size=3, padding=1)
self.which_linear = nn.Linear
# We use a non-spectral-normed embedding here regardless;
# For some reason applying SN to G's embedding seems to randomly cripple G
self.which_embedding = nn.Embedding
# Prepare model
# First linear layer
self.linear = []
self.get_map = []
self.all_blocks = []
self.output_layer = []
self.spade_layers = []
self.out_to_one_dim = []
self.repeat = opts.num_components
self.free_dynamic_component = utils.check_arg(self.opts, 'free_dynamic_component')
for ind in range(self.repeat):
if self.repeat > 1:
in_dim = self.dim_z
att_dim = self.opts.att_dim
self.get_map.append(nn.Sequential(
nn.Linear(self.dim_z, (1 + att_dim) * (self.bottom_width ** 2)),
View((-1, 1 + att_dim, self.bottom_width, self.bottom_width))
))
if self.opts.spade_index > -1:
if self.opts.spade_index == 0:
spade_in_chan = self.opts.fixed_v_dim
else:
spade_in_chan = self.arch['in_channels'][self.opts.spade_index]
self.spade_layers.append(SPADE(spade_in_chan, att_dim))
if (ind == 1 and not self.free_dynamic_component) or (ind==0):
self.linear.append(nn.Sequential(self.which_linear(in_dim,self.opts.fixed_v_dim)))
else:
self.linear.append(self.which_linear(in_dim,
self.arch['in_channels'][0] * (self.bottom_width ** 2)))
else:
in_dim = self.dim_z
self.linear.append(self.which_linear(in_dim,
self.arch['in_channels'][0] * (self.bottom_width ** 2)))
# self.blocks is a doubly-nested list of modules, the outer loop intended
# to be over blocks at a given resolution (resblocks and/or self-attention)
# while the inner loop is over a given block
self.blocks = []
for index in range(len(self.arch['out_channels'])):
upsample_factor = 2
if index == 0:
self.in_dim = self.arch['in_channels'][index] if (ind == 1 and self.free_dynamic_component) or (self.repeat < 2) else self.opts.fixed_v_dim
in_dim = self.in_dim
if resolution == 84:
upsample_factor = 3
else:
in_dim = self.arch['in_channels'][index]
if utils.check_arg(opts, 'simple_blocks'):
self.blocks += [[layers.SimpleGBlock(in_channels=in_dim,
out_channels=self.arch['out_channels'][index],
num_conv = 2 if index == 0 else 1
)]]
else:
self.blocks += [[layers.GBlock(in_channels=in_dim,
out_channels=self.arch['out_channels'][index],
which_conv=self.which_conv,
activation=self.activation,
upsample=(functools.partial(F.interpolate, scale_factor=upsample_factor)
if self.arch['upsample'][index] else None))]]
# If attention on this block, attach it to the end
if self.arch['attention'][self.arch['resolution'][index]] and (self.repeat <= 1 or not utils.check_arg(self.opts, 'no_attention')):
if not utils.check_arg(opts, 'simple_blocks'):
print('Adding attention layer in G at resolution %d' % self.arch['resolution'][index])
self.blocks[-1] += [layers.Attention(self.arch['out_channels'][index], self.which_conv)]
# Turn self.blocks into a ModuleList so that it's all properly registered.
self.all_blocks.append(nn.ModuleList([nn.ModuleList(block) for block in self.blocks]))
# output layer: batchnorm-relu-conv.
# Consider using a non-spectral conv here
last_conv = self.which_conv
if utils.check_arg(self.opts, 'simple_blocks'):
last_conv = self.which_conv_ker2
if self.opts.no_in:
self.output_layer.append(nn.Sequential(self.activation,
last_conv(self.arch['out_channels'][-1], 3)))
else:
self.output_layer.append(nn.Sequential(nn.InstanceNorm2d(self.arch['out_channels'][-1]),
self.activation,
last_conv(self.arch['out_channels'][-1], 3)))
if self.repeat > 1:
self.out_to_one_dim.append(nn.Sequential(self.activation,
last_conv(self.arch['out_channels'][-1], 1),
nn.LeakyReLU(0.2)))
self.linear = nn.ModuleList(self.linear)
self.all_blocks = nn.ModuleList(self.all_blocks)
self.output_layer = nn.ModuleList(self.output_layer)
if self.repeat > 1:
if self.opts.spade_index > -1:
self.spade_layers = nn.ModuleList(self.spade_layers)
self.get_map = nn.ModuleList(self.get_map)
self.out_to_one_dim = nn.ModuleList(self.out_to_one_dim)
self.fine_mask = nn.Sequential(nn.Conv2d(opts.num_components, 512, 1, 1),
nn.LeakyReLU(0.2),
nn.Conv2d(512, opts.num_components, 1))
else:
if self.opts.do_memory:
self.out_to_one_dim = nn.Sequential(self.activation,
self.which_conv(self.arch['out_channels'][-1], 1),
nn.LeakyReLU(0.2))
self.fine_mask = nn.Sequential(nn.Conv2d(2, 512, 1, 1),
nn.LeakyReLU(0.2),
nn.Conv2d(512, 2, 1))
self.base_temperature = self.opts.temperature
if utils.check_arg(self.opts, 'base_temperature'):
self.base_temperature = self.opts.base_temperature
# Initialize weights. Optionally skip init for testing.
if not skip_init:
self.init_weights()
