def get_model(num_timesteps, num_input_vars, num_output_vars, p_params=None):
params = dict(archtype='rnn',optimizer='rmsprop',discount=0.9, hidden_layer_dims=[])
if p_params:
params.update( p_params )
#params.update( dict(macro_dims=500, archtype='RNN',hidden_activation='leakyrelu') )
#params.update( dict(macro_dims=500, hidden_activation='leakyrelu') ) # gets better than mean perfornace on ieee300
#params.update( dict(macro_dims=trajs_trn.shape[2], archtype='RNNIdentity', hidden_activation='leakyrelu') ) # gets better than mean perfornace on ieee300
#params.update( dict(macro_dims=trajs_trn.shape[2], archtype='RNNIdentity', hidden_activation='linear') ) # gets better than mean perfornace on ieee300
#params.update( dict(macro_dims=500, hidden_activation='srelu') )# , hidden_layer_dims=[500,])
model_name = "models/" + "_".join(["%s"%x[1] for x in sorted(params.items()) if x[1]]) + \
"-" + "-".join(map(str, [num_timesteps, num_input_vars, num_output_vars]))
if False:
# elsewhere...
from keras.models import model_from_json
with timeIt("Loading model object"):
model = model_from_json(open('%s.json'%model_name).read())
model.load_weights('%s_weights.h5' % model_name)
else:
with timeIt("Creating model object"):
model = trainrnn.get_rnn_model(num_timesteps, num_input_vars, num_output_vars=num_output_vars, output_type='real', **params)
model.model_name = model_name
return model
def get_model(num_timesteps, num_input_vars, num_output_vars, p_params=None):
params = dict(archtype='rnn',
optimizer='rmsprop',
discount=0.9,
hidden_layer_dims=[])
if p_params:
params.update(p_params)
#params.update( dict(macro_dims=500, archtype='RNN',hidden_activation='leakyrelu') )
#params.update( dict(macro_dims=500, hidden_activation='leakyrelu') ) # gets better than mean perfornace on ieee300
#params.update( dict(macro_dims=trajs_trn.shape[2], archtype='RNNIdentity', hidden_activation='leakyrelu') ) # gets better than mean perfornace on ieee300
#params.update( dict(macro_dims=trajs_trn.shape[2], archtype='RNNIdentity', hidden_activation='linear') ) # gets better than mean perfornace on ieee300
#params.update( dict(macro_dims=500, hidden_activation='srelu') )# , hidden_layer_dims=[500,])
model_name = "models/" + "_".join(["%s"%x[1] for x in sorted(params.items()) if x[1]]) + \
"-" + "-".join(map(str, [num_timesteps, num_input_vars, num_output_vars]))
if False:
# elsewhere...
from keras.models import model_from_json
with timeIt("Loading model object"):
model = model_from_json(open('%s.json' % model_name).read())
model.load_weights('%s_weights.h5' % model_name)
else:
with timeIt("Creating model object"):
model = trainrnn.get_rnn_model(num_timesteps,
num_input_vars,
num_output_vars=num_output_vars,
output_type='real',
**params)
model.model_name = model_name
return model
def get_data(DIRNAME='ieee300'):
#DIRNAME = 'alpha1.0r0.0'
TOPNUM = None
with timeIt():
df = get_df(DIRNAME, TOPNUM)
df['PertId'] = df.PertId.apply(
lambda x: tuple(sorted(str(x).split(','))))
def split_mx(df, unique_perts=False):
done_perts = set()
clocs = None
pertDF = df[df.PertId != -1]
def emit(clocs):
cPerts = pertDF.PertId.iloc[clocs[0]]
if unique_perts and cPerts in done_perts:
return None
done_perts.add(cPerts)
return np.vstack(pertDF.Effs2.iloc[clocs]), np.vstack(
pertDF.Eff.iloc[clocs])
for ndx, t in enumerate(pertDF.t):
if t == 0:
if clocs is not None:
if len(clocs) == 15:
d = emit(clocs)
if d is not None:
yield d
clocs = []
clocs.append(ndx)
if len(clocs) == 15:
d = emit(clocs)
if d is not None:
yield d
with timeIt("Loading trajectories"):
mxs = zip(*split_mx(df, unique_perts=True))
trajs = np.stack(mxs[0])
#trajs = trajs[0:1000]
trajs_trn = trajs.copy()
trajs_trn[:, 1:, :] = 0.0
if False:
num_mostvaried = 1
#mostvaried=np.argsort(trajs[:,-1,:].var(axis=0))[-num_mostvaried:]
#observable = trajs[:,:,mostvaried]
observable = trajs - trajs.mean(axis=0)[None, :, :]
print trajs.shape, observable.shape
observable = trajs
#observable = np.stack(mxs[1])*1000
v = ((observable - observable.mean(axis=0)[None, :, :])**2)
print "Mean error for avg : %0.7f" % v.mean()
print "Mean error for avg (last10%%) : %0.7f" % v[int(len(v) *
.9):].mean()
print "Mean error for avg (frst10%%) : %0.7f" % v[:int(len(v) *
.1)].mean()
return trajs, trajs_trn, observable
def get_data(DIRNAME="ieee300"):
# DIRNAME = 'alpha1.0r0.0'
TOPNUM = None
with timeIt():
df = get_df(DIRNAME, TOPNUM)
df["PertId"] = df.PertId.apply(lambda x: tuple(sorted(str(x).split(","))))
def split_mx(df, unique_perts=False):
done_perts = set()
clocs = None
pertDF = df[df.PertId != -1]
def emit(clocs):
cPerts = pertDF.PertId.iloc[clocs[0]]
if unique_perts and cPerts in done_perts:
return None
done_perts.add(cPerts)
return np.vstack(pertDF.Effs2.iloc[clocs]), np.vstack(pertDF.Eff.iloc[clocs])
for ndx, t in enumerate(pertDF.t):
if t == 0:
if clocs is not None:
if len(clocs) == 15:
d = emit(clocs)
if d is not None:
yield d
clocs = []
clocs.append(ndx)
if len(clocs) == 15:
d = emit(clocs)
if d is not None:
yield d
with timeIt("Loading trajectories"):
mxs = zip(*split_mx(df, unique_perts=True))
trajs = np.stack(mxs[0])
# trajs = trajs[0:1000]
trajs_trn = trajs.copy()
trajs_trn[:, 1:, :] = 0.0
if False:
num_mostvaried = 1
# mostvaried=np.argsort(trajs[:,-1,:].var(axis=0))[-num_mostvaried:]
# observable = trajs[:,:,mostvaried]
observable = trajs - trajs.mean(axis=0)[None, :, :]
print trajs.shape, observable.shape
observable = trajs
# observable = np.stack(mxs[1])*1000
v = (observable - observable.mean(axis=0)[None, :, :]) ** 2
print "Mean error for avg : %0.7f" % v.mean()
print "Mean error for avg (last10%%) : %0.7f" % v[int(len(v) * 0.9) :].mean()
print "Mean error for avg (frst10%%) : %0.7f" % v[: int(len(v) * 0.1)].mean()
return trajs, trajs_trn, observable
