def predict_qs(model_name, terminal, voltages):
workspace.ResetWorkspace()
# requires voltages is an numpy array of size
# (batch size, input_dimension)
# the first dimension is Vg and the second dimenstion is Vd
# preprocess the origin input and create adjoint input
preproc_param = pickle.load(
open(model_name + '_' + terminal + '_preproc_param.p', "rb"))
dummy_qs = np.zeros(voltages[0].shape[0])
voltages, dummy_qs = preproc.ac_qv_preproc(voltages, dummy_qs,
preproc_param['scale'],
preproc_param['vg_shift'])
adjoint_input = np.ones((voltages.shape[0], 1))
# Expand dimensions of input and set data type of inputs
origin_input = np.expand_dims(voltages, axis=1)
origin_input = origin_input.astype(np.float32)
adjoint_input = adjoint_input.astype(np.float32)
workspace.FeedBlob('DBInput_train/origin_input', voltages)
workspace.FeedBlob('DBInput_train/adjoint_input', adjoint_input)
pred_net = exporter.load_net(model_name + '_init', model_name + '_predict')
workspace.RunNet(pred_net)
qs = np.squeeze(workspace.FetchBlob('origin/NanCheck/origin_pred'))
gradients = np.squeeze(workspace.FetchBlob('adjoint/fc0/output'))
restore_integral_func, restore_gradient_func = preproc.get_restore_q_func(
preproc_param['scale'], preproc_param['vg_shift'])
original_qs = restore_integral_func(qs)
original_gradients = restore_gradient_func(gradients)
return original_qs, original_gradients
def predict_ids(model_name, vg, vd):
workspace.ResetWorkspace()
# preproc the input
vg = vg.astype(np.float32)
vd = vd.astype(np.float32)
#if len(self.preproc_param) == 0:
preproc_param = pickle.load(open(model_name + '_preproc_param.p', "rb"))
dummy_ids = np.zeros(len(vg))
preproc_data_arrays = preproc.dc_iv_preproc(
vg,
vd,
dummy_ids,
preproc_param['scale'],
preproc_param['vg_shift'],
)
# print(preproc_data_arrays[0].shape)
_preproc_data_arrays = [
np.expand_dims(x, axis=1) if len(x.shape) < 2 else x
for x in preproc_data_arrays
]
workspace.FeedBlob('DBInput_train/sig_input', _preproc_data_arrays[0])
workspace.FeedBlob('DBInput_train/tanh_input', _preproc_data_arrays[1])
pred_net = exporter.load_net(model_name + '_init', model_name + '_predict')
workspace.RunNet(pred_net)
_ids = np.squeeze(workspace.FetchBlob('prediction'))
restore_id_func, get_restore_id_grad_func = preproc.get_restore_id_func(
preproc_param['scale'])
ids = restore_id_func(_ids)
return ids
def predict_ids_grads(model_name, vg, vd):
workspace.ResetWorkspace()
# preproc the input
vg = vg.astype(np.float32)
vd = vd.astype(np.float32)
#if len(self.preproc_param) == 0:
preproc_param = pickle.load(open(model_name + '_preproc_param.p', "rb"))
dummy_ids = np.zeros(len(vg))
preproc_data_arrays = preproc.dc_iv_preproc(
vg,
vd,
dummy_ids,
preproc_param['scale'],
preproc_param['vg_shift'],
)
_preproc_data_arrays = [
np.expand_dims(x, axis=1) for x in preproc_data_arrays
]
workspace.FeedBlob('DBInput_train/sig_input', _preproc_data_arrays[0])
workspace.FeedBlob('DBInput_train/tanh_input', _preproc_data_arrays[1])
adjoint_input = np.ones((_preproc_data_arrays[0].shape[0], 1))
workspace.FeedBlob('adjoint_input', adjoint_input)
pred_net = exporter.load_net(model_name + '_init', model_name + '_predict')
workspace.RunNet(pred_net)
_ids = np.squeeze(workspace.FetchBlob('origin/Mul/origin_pred'))
_sig_grad = np.squeeze(
workspace.FetchBlob('adjoint/sig_fc_layer_0/output'))
_tanh_grad = np.squeeze(
workspace.FetchBlob('adjoint/tanh_fc_layer_0/output'))
restore_id_func, get_restore_id_grad_func = preproc.get_restore_id_func(
preproc_param['scale'])
ids = restore_id_func(_ids)
sig_grad, tanh_grad = get_restore_id_grad_func(_sig_grad, _tanh_grad)
return ids, sig_grad, tanh_grad
