# the out here should be the shape: data_size * nclasses
gmm_mdl_class_hat = np.argmax(gmm_mdl_out, axis=0) + 1
istrue_gmm_mdl = gmm_mdl_class_hat == int(true_class)
print(
"GMM-HMM -- ", data_file, "Done ...",
"{}/{}".format(str(istrue_gmm_mdl.sum()),
str(istrue_gmm_mdl.shape[0])))
#######################################################
# Get the predicted class values for NVP-HMM
#######################################################
# zero pad data for batch training
max_len_ = max(l)
x_padded = pad_data(X, max_len_)
batchdata = DataLoader(dataset=TheDataset(
x_padded, lengths=l, device=nvp_mdl_loaded.hmms[0].device),
batch_size=batch_size_,
shuffle=True)
nvp_mdl_out_list = [nvp_mdl_loaded.forward(x) for x in batchdata]
nvp_mdl_out = torch.cat(nvp_mdl_out_list, dim=1)
nvp_mdl_class_hat = torch.argmax(nvp_mdl_out, dim=0) + 1
istrue_nvp_mdl = nvp_mdl_class_hat == int(true_class)
print(
"NVP-HMM -- ", data_file, "Done ...",
"{}/{}".format(str(istrue_nvp_mdl.sum().cpu().numpy()),
str(istrue_nvp_mdl.shape[0])))
get_freer_gpu())))
print("Try to push to device: {}".format(device))
mdl.device = device
mdl.pushto(mdl.device)
break
except:
# if push error (maybe memory overflow, try again)
print("Push to device cuda:{} fail, try again ...")
continue
print("epoch:{}\tclass:{}\tPush model to {}. Done.".format(
epoch_str, iclass_str, mdl.device),
file=sys.stdout)
# zero pad data for batch training
max_len_ = max([x.shape[0] for x in xtrain])
xtrain_padded = pad_data(xtrain, max_len_)
traindata = DataLoader(dataset=TheDataset(xtrain_padded,
lengths=l,
device=mdl.device),
batch_size=options["Train"]["batch_size"],
shuffle=True)
# niter counts the number of em steps before saving a model checkpoint
niter = options["Train"]["niter"]
# add number of training data in model
mdl.number_training_data = len(xtrain)
# set model into train mode
mdl.train()