def __init__(self, state_dim, action_dim, non_linearity=F.relu, hidden_layers=1, hidden_dim=20, output_non_linearity=F.sigmoid): super(MlpCritic, self).__init__() input_dim = state_dim + action_dim # Q1 FA self._Q1 = MLP(input_dim=input_dim, output_dim=1, output_non_linearity=output_non_linearity, hidden_dim=hidden_dim, hidden_non_linearity=non_linearity, hidden_layers=hidden_layers) # Q2 FA self._Q2 = MLP(input_dim=input_dim, output_dim=1, output_non_linearity=output_non_linearity, hidden_dim=hidden_dim, hidden_non_linearity=non_linearity, hidden_layers=hidden_layers)
class MlpCritic(nn.Module): def __init__(self, state_dim, action_dim, non_linearity=F.relu, hidden_layers=1, hidden_dim=20, output_non_linearity=F.sigmoid): super(MlpCritic, self).__init__() input_dim = state_dim + action_dim # Q1 FA self._Q1 = MLP(input_dim=input_dim, output_dim=1, output_non_linearity=output_non_linearity, hidden_dim=hidden_dim, hidden_non_linearity=non_linearity, hidden_layers=hidden_layers) # Q2 FA self._Q2 = MLP(input_dim=input_dim, output_dim=1, output_non_linearity=output_non_linearity, hidden_dim=hidden_dim, hidden_non_linearity=non_linearity, hidden_layers=hidden_layers) def forward(self, s, a): s = tt(s) a = tt(a) if len(s.shape) == 1: x = torch.cat((s, a)) else: x = torch.cat((s, a), dim=1) q1 = self._Q1(x) q2 = self._Q2(x) return q1, q2 def Q1(self, s, a): s = tt(s) a = tt(a) if len(s.shape) == 1: x = torch.cat((s, a)) else: x = torch.cat((s, a), dim=1) q1 = self._Q1(x) return q1 def reset_parameters(self): self._Q1.reset_parameters() self._Q2.reset_parameters()
for snt_id in fea_dev.keys():
spk_id=snt_id.split('_')[0]
fea_pase_dev[snt_id]=(fea_pase_dev[snt_id]-mean_spk_dev[spk_id])#/std_spk_dev[spk_id]
# Label file reading
with open(lab_file, 'rb') as handle:
lab = pickle.load(handle)
with open(lab_file_dev, 'rb') as handle:
lab_dev = pickle.load(handle)
# Network initialization
nnet=MLP(options,inp_dim)
nnet.to(device)
cost=nn.NLLLoss()
# Optimizer initialization
optimizer = optim.SGD(nnet.parameters(), lr=lr, momentum=0.0)
# Seeds initialization
np.random.seed(seed)
torch.manual_seed(seed)
# Batch creation (train)
fea_lst=[]
lab_lst=[]
# apply speaker normalization
for snt_id in fea_dev.keys():
spk_id = snt_id.split('_')[0]
# /std_spk_dev[spk_id]
fea_pase_dev[snt_id] = (fea_pase_dev[snt_id] -
mean_spk_dev[spk_id])
# Label file reading
with open(lab_file, 'rb') as handle:
lab = pickle.load(handle)
with open(lab_file_dev, 'rb') as handle:
lab_dev = pickle.load(handle)
# Network initialization
nnet = MLP(options, inp_dim)
nnet.to(device)
cost = nn.NLLLoss()
# Optimizer initialization
optimizer = optim.SGD(nnet.parameters(), lr=lr, momentum=0.0)
# Seeds initialization
np.random.seed(seed)
torch.manual_seed(seed)
# Batch creation (train)
fea_lst = []
lab_lst = []
fea_pase[snt_id] = pase(fea[snt_id])
fea_pase[snt_id] = fea_pase[snt_id].view(
fea_pase[snt_id].shape[1], fea_pase[snt_id].shape[2]).transpose(0, 1)
inp_dim = fea_pase[snt_id].shape[1] * (left + right + 1)
# Computing pase features for test
fea_pase_dev = {}
for snt_id in fea_dev.keys():
fea_pase_dev[snt_id] = pase(fea_dev[snt_id]).detach()
fea_pase_dev[snt_id] = fea_pase_dev[snt_id].view(
fea_pase_dev[snt_id].shape[1],
fea_pase_dev[snt_id].shape[2]).transpose(0, 1)
# Network initialization
nnet = MLP(options, inp_dim)
nnet.to(device)
cost = nn.NLLLoss()
# Optimizer initialization
optimizer = optim.SGD(list(nnet.parameters()) + list(pase.parameters()),
lr=lr,
momentum=0.0)
# Seeds initialization
np.random.seed(seed)
torch.manual_seed(seed)
# Batch creation (train)
fea_lst = []
lab_lst = []
mean_spk_dev[spk_id]),
dim=0)
std_spk_dev[spk_id] = torch.mean(torch.stack(std_spk_dev[spk_id]),
dim=0)
# apply speaker normalization
for snt_id in fea_dev.keys():
spk_id = snt_id.split('_')[0]
fea_pase_dev[snt_id] = (fea_pase_dev[snt_id] - mean_spk_dev[spk_id]
) #/std_spk_dev[spk_id]
fea_pase_dev[snt_id] = context_window(fea_pase_dev[snt_id], left,
right)
# Network initialization
inp_dim = fea_pase_dev[snt_id].shape[1] * (left + right + 1)
nnet = MLP(options, inp_dim)
nnet.to(device)
nnet.load_state_dict(torch.load(model_file))
nnet.eval()
post_file = open_or_fd(ark_file, output_folder, 'wb')
for snt_id in fea_dev.keys():
pout = nnet(torch.from_numpy(fea_pase_dev[snt_id]).to(device).float())
# TO DO IT!!
#pout=pout-log_counts
write_mat(output_folder, post_file, pout.data.cpu().numpy(), snt_id)
# doing decoding
print('Decoding...')