def build_model(input_var,ExptDict):
# Unpack necessary variables
model = ExptDict["model"]["model_id"]
n_loc = ExptDict["task"]["n_loc"]
n_out = ExptDict["task"]["n_out"]
batch_size = ExptDict["batch_size"]
n_in = ExptDict["n_in"]
n_hid = ExptDict["n_hid"]
out_nonlin = ExptDict["task"]["out_nonlin"]
if model == 'LeInitRecurrent':
diag_val = ExptDict["model"]["diag_val"]
offdiag_val = ExptDict["model"]["offdiag_val"]
l_out, l_rec = models.LeInitRecurrent(input_var, batch_size=batch_size,
n_in=(n_loc+1)*n_in, n_out=n_out,
n_hid=n_hid, diag_val=diag_val,
offdiag_val=offdiag_val,
out_nlin=out_nonlin)
elif model == 'OrthoInitRecurrent':
init_val = ExptDict["model"]["init_val"]
l_out, l_rec = models.OrthoInitRecurrent(input_var, batch_size=batch_size,
n_in=(n_loc+1)*n_in, n_out=n_out,
n_hid=n_hid, init_val=init_val,
out_nlin=out_nonlin)
elif model == 'ResidualRecurrent':
leak_inp = ExptDict["model"]["leak_inp"]
leak_hid = ExptDict["model"]["leak_hid"]
l_out, l_rec = models.ResidualRecurrent(input_var, batch_size=batch_size,
n_in=(n_loc+1)*n_in, n_out=n_out,
n_hid=n_hid, leak_inp=leak_inp,
leak_hid=leak_hid,
out_nlin=out_nonlin)
elif model == 'GRURecurrent':
diag_val = ExptDict["model"]["diag_val"]
offdiag_val = ExptDict["model"]["offdiag_val"]
l_out, l_rec = models.GRURecurrent(input_var, batch_size=batch_size,
n_in=(n_loc+1)*n_in, n_out=n_out, n_hid=n_hid,
diag_val=diag_val, offdiag_val=offdiag_val,
out_nlin=out_nonlin)
return l_out, l_rec
def build_model(input_var,ExptDict):
# Unpack necessary variables
model = ExptDict["model"]["model_id"]
n_loc = ExptDict["task"]["n_loc"]
n_out = ExptDict["task"]["n_out"]
batch_size = ExptDict["batch_size"]
n_in = ExptDict["n_in"]
n_hid = ExptDict["n_hid"]
out_nonlin = ExptDict["task"]["out_nonlin"]
if model == 'LeInitRecurrent':
diag_val = ExptDict["model"]["diag_val"]
offdiag_val = ExptDict["model"]["offdiag_val"]
l_out, l_rec = models.LeInitRecurrent(input_var, batch_size=batch_size,
n_in=n_loc*n_in, n_out=n_out,
n_hid=n_hid, diag_val=diag_val,
offdiag_val=offdiag_val,
out_nlin=out_nonlin)
elif model == 'TanhRecurrent':
wscale = ExptDict["model"]["wscale"]
l_out, l_rec = models.TanhRecurrent(input_var, batch_size=batch_size,
n_in=n_loc*n_in, n_out=n_out,
n_hid=n_hid, wscale=wscale,
out_nlin=out_nonlin)
elif model == 'LeInitRecurrentWithFastWeights':
diag_val = ExptDict["model"]["diag_val"]
offdiag_val = ExptDict["model"]["offdiag_val"]
gamma = ExptDict["model"]["gamma"]
l_out, l_rec = models.LeInitRecurrentWithFastWeights(input_var, batch_size=batch_size,
n_in=n_loc*n_in, n_out=n_out,
n_hid=n_hid, diag_val=diag_val,
offdiag_val=offdiag_val,
out_nlin=out_nonlin, gamma=gamma)
elif model == 'OrthoInitRecurrent':
init_val = ExptDict["model"]["init_val"]
l_out, l_rec = models.OrthoInitRecurrent(input_var, batch_size=batch_size,
n_in=n_loc*n_in, n_out=n_out,
n_hid=n_hid, init_val=init_val,
out_nlin=out_nonlin)
elif model == 'ResidualRecurrent':
leak_inp = ExptDict["model"]["leak_inp"]
leak_hid = ExptDict["model"]["leak_hid"]
wscale = ExptDict["model"]["wscale"]
l_out, l_rec = models.ResidualRecurrent(input_var, batch_size=batch_size,
n_in=n_loc*n_in, n_out=n_out,
n_hid=n_hid, wscale=wscale, leak_inp=leak_inp,
leak_hid=leak_hid,
out_nlin=out_nonlin)
elif model == 'GRURecurrent':
diag_val = ExptDict["model"]["diag_val"]
offdiag_val = ExptDict["model"]["offdiag_val"]
l_out, l_rec = models.GRURecurrent(input_var, batch_size=batch_size,
n_in=n_loc*n_in, n_out=n_out, n_hid=n_hid,
diag_val=diag_val, offdiag_val=offdiag_val,
out_nlin=out_nonlin)
elif model == 'LeInitRecurrentWithLayerNorm':
diag_val = ExptDict["model"]["diag_val"]
offdiag_val = ExptDict["model"]["offdiag_val"]
l_out, l_rec = models.LeInitRecurrentWithLayerNorm(input_var, batch_size=batch_size,
n_in=n_loc*n_in, n_out=n_out,
n_hid=n_hid, diag_val=diag_val,
offdiag_val=offdiag_val,
out_nlin=out_nonlin)
return l_out, l_rec
l_out, l_rec = models.LeInitRecurrent(
input_var,
mask_var=mask_var,
batch_size=generator.batch_size,
n_in=generator.n_loc * generator.n_in,
n_out=generator.n_out,
n_hid=n_hid,
diag_val=diag_val,
offdiag_val=offdiag_val,
out_nlin=lasagne.nonlinearities.sigmoid)
elif model == 'ResidualRecurrent':
l_out, l_rec = models.ResidualRecurrent(
input_var,
mask_var=mask_var,
batch_size=generator.batch_size,
n_in=generator.n_loc * generator.n_in,
n_out=generator.n_out,
n_hid=n_hid,
leak_inp=1.0,
leak_hid=1.0)
elif model == 'GRURecurrent':
l_out, l_rec = models.GRURecurrent(input_var,
mask_var=mask_var,
batch_size=generator.batch_size,
n_in=generator.n_loc *
generator.n_in,
n_out=generator.n_out,
n_hid=n_hid)
# The generated output variable and the loss function
pred_var = T.clip(lasagne.layers.get_output(l_out), 1e-6, 1.0 - 1e-6)