class Actor():
""" Create an agent and set it up for Proximal Policy Optimization """
maxGradientNormilization = 0.5
clipParameter = 0.1
bufferSize = 2500
batchSize = 128
epoch = 10
def __init__(self,
outputModel,
inputModel,
discount=0.99,
framesPerState=8,
trainingMode=False,
hwDevice='cpu'):
""" The actor constructor
Parameters
-------
outputModel : string
the path to where the output model should be saved, excluding the file extension.
inputModel : string
the path, including file extension to the input model.
discount : float
The discount factor.
framesPerState : int
Number of frames per state.
trainingMode: : boolean
Whether this Actor is used for training or predicting.
hwDevice : string,
CPU or CUDA (whether to offload to GPU or use the CPU).
"""
self.hardwareDevice = hwDevice
self.trainingMode = trainingMode
self.inputPath = inputModel
self.outputPath = outputModel
self.discount = discount
self.transition = np.dtype([
('s', np.float64, (framesPerState, 96, 96)),
('matrix_a', np.float64, (3, )), ('coefficient', np.float64),
('slice_to_concat', np.float64),
('index_exp ', np.float64, (framesPerState, 96, 96))
])
self.nn = NN().double().to(self.hardwareDevice)
self.buffer = np.empty(self.bufferSize, dtype=self.transition)
self.optimizer = optim.Adam(self.nn.parameters(), lr=1e-3)
self.trainingStep = 0
self.counter = 0
def saveModel(self):
""" Save a model to a pytorch PKL file
Raises
-------
AssertionError
Raised if the output path hasn't been provided
Notes
-------
`self.outputPath` has to be provided WITHOUT a file extension.
"""
assert self.outputPath != None, "You haven't given an output path!"
path = f"{self.outputPath}"
while (True):
if (not os.path.exists(f'{path}.pkl')):
filename = f'{self.outputPath}.pkl'
torch.save(self.nn.state_dict(), filename)
break
else:
path = f'{path}-new'
def loadModel(self):
""" Load a model from a pytorch PKL file
Raises
-------
AssertionError
Raised if the given model path doesn't exist in the filesystem
Notes:
-------
`self.inputPath` is a path to a model file INCLUDING it's file extension (usually `.pkl`)
"""
if not self.inputPath:
print(
'No input model argument was given, starting point is now set to untrained.'
)
return
assert os.path.exists(
self.inputPath), "The given model path doesn't exist!"
self.nn.load_state_dict(torch.load(self.inputPath))
def chooseAction(self, state):
"""
Choose an action to perform on the track
Parameters
-------
state:
The current state of the car.
Returns
-------
action : np.ndarray
An action for the network to run on the track
Notes
-------
This function is only called when the --train flag is NOT provided.
"""
state = torch.from_numpy(state).double().to(
self.hardwareDevice).unsqueeze(0)
with torch.no_grad():
alpha, beta = self.nn(state)[0]
action = alpha / (alpha + beta)
return action.squeeze().cpu().numpy()
def chooseActionTrain(self, state):
""" Choose an action during training mode
Parameters
-------
state:
The current state of the car.
Returns
-------
action : np.ndarray
The actions to run on the track
coefficient : float
The logarithmic probability for an action
Notes
-------
This function is only called when the --train flag IS provided.
"""
state = torch.from_numpy(state).double().to(
self.hardwareDevice).unsqueeze(0)
with torch.no_grad():
alpha, beta = self.nn(state)[0]
dist = Beta(alpha, beta)
action = dist.sample()
coefficient = dist.log_prob(action).sum(dim=1)
action = action.squeeze().cpu().numpy()
coefficient = coefficient.item()
return action, coefficient
def storeInBuffer(self, transition):
""" Store a transition in a buffer
Parameters
-------
transition : dtype=self.transition
A transition element which is saved to the internal memory buffer
Returns
-------
Boolean
A boolean representing whether the buffer was SUCCESFULLY stored and didn't overflow.
"""
self.buffer[self.counter] = transition
self.counter += 1
if not self.bufferSize == self.counter:
return False
self.counter = 0
return True
def update(self):
""" Run an update on the network """
self.trainingStep += 1
sliceToConcat = torch.tensor(self.buffer['slice_to_concat'],
dtype=torch.double).to(
self.hardwareDevice).view(-1, 1)
matrixA = torch.tensor(self.buffer['matrix_a'],
dtype=torch.double).to(self.hardwareDevice)
indexExp = torch.tensor(self.buffer['index_exp'],
dtype=torch.double).to(self.hardwareDevice)
s = torch.tensor(self.buffer['s'],
dtype=torch.double).to(self.hardwareDevice)
old_coefficient = torch.tensor(self.buffer['coefficient'],
dtype=torch.double).to(
self.hardwareDevice).view(-1, 1)
with torch.no_grad():
target = sliceToConcat + self.discount * self.nn(indexExp)[1]
advantage = target - self.nn(s)[1]
for _ in range(self.epoch):
for index in BatchSampler(
SubsetRandomSampler(range(self.bufferSize)),
self.batchSize, False):
alpha, beta = self.nn(s[index])[0]
distance = Beta(alpha, beta)
coefficient = distance.log_prob(matrixA[index]).sum(
dim=1, keepdim=True)
relativeAdvantage = torch.exp(coefficient -
old_coefficient[index])
s1 = relativeAdvantage * advantage[index]
s2 = torch.clamp(ratio, 1.0 - self.clipParameter,
1.0 + self.clipParameter) * advantage[index]
# Loss on an action
aLoss = -torch.min(s1, s2).mean()
# Loss on the value
vLoss = F.smooth_l1_loss(self.nn(s[index])[1], target[index])
# Total loss calculation
loss = aLoss + (vLoss * 2.0)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
def train(fold_ind, dataset_dir, type_name, setting_symbol, device_id=None):
if device_id is not None:
device_ind = device_id
else:
device_ind = (fold_ind - 1) % 4
device = torch.device(
f'cuda:{device_ind}' if torch.cuda.is_available() else 'cpu')
print(device)
# loss
if args.classifier == "nn":
crit = nn.NLLLoss()
elif args.classifier == "dm" or args.classifier == "ip":
crit = nn.BCELoss()
# read dataset
dataset = torch.load(
os.path.join(dataset_dir, f'dataset_fold_{fold_ind}.pt'))
num_node_features = dataset.num_node_features
dataset = dataset.to(device)
print(dataset)
# read train and test
with open(
os.path.join(dataset_dir, f"train_test_split_fold_{fold_ind}.pkl"),
"rb") as f:
positive_train, positive_test = pickle.load(f)
with open(
os.path.join(
dataset_dir,
f"../10/negative_train_test_split_fold_{fold_ind}.pkl"),
"rb") as f:
negative_train, negative_test = pickle.load(f)
# generate data loader
if args.classifier == "nn":
train_loader, test_loader = read_train_test_nn(positive_train,
positive_test,
negative_train,
negative_test, args.bs)
elif args.classifier == "dm" or args.classifier == "ip":
train_loader, test_loader = read_train_test_dm(positive_train,
positive_test,
negative_train,
negative_test, args.bs)
# read omim and test (use for eval_all)
with open(f"preprocess/edgelist_result/union_omim_gene_list.pkl",
"rb") as f:
omim_list, gene_list = pickle.load(f)
with open(os.path.join(dataset_dir, f"id_conversion_fold_{fold_ind}.pkl"),
"rb") as f:
node_name_to_contiguous_id, contiguous_id_to_node_name = pickle.load(f)
refine_gene_list = []
for gene in gene_list:
if gene in node_name_to_contiguous_id:
refine_gene_list.append(node_name_to_contiguous_id[gene])
else:
refine_gene_list.append(-1)
add_links = []
if "S" in type_name:
sc_add_name = "sc_add_topall20.txt"
with open(
os.path.join("preprocess/edgelist_result/disease_gene/",
sc_add_name), "r") as f:
for line in f.read().strip().split("\n"):
d, g = line.strip().split("\t")
d, g = "omim:" + str(d), "gene:" + str(g)
if d in omim_list and g in gene_list:
add_links.append([
node_name_to_contiguous_id[d],
node_name_to_contiguous_id[g]
])
# embedding gcn
gcn = GCN_EMB(num_node_features, args.gcn_dim, args.conv_type,
args.dropout_ratio).to(device)
# classifier
if args.classifier == "nn":
classifier = NN(args.nn_dim, device).to(device)
elif args.classifier == "dm":
classifier = DM(args.dm_dim).to(device)
elif args.classifier == "ip":
classifier = IP().to(device)
params = list(gcn.parameters()) + list(classifier.parameters())
optimizer = torch.optim.SGD(params,
lr=args.lr,
momentum=0.9,
weight_decay=5e-4)
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer,
milestones=args.milestone,
gamma=0.1)
# train
best_auc = 0.0
tqdm_train = tqdm(range(args.epoch_num))
for epoch in tqdm_train:
t1 = time.time()
gcn.train()
classifier.train()
epoch_loss = 0.0
count = 0
# calculate loss based on train and test split
delta_time = 0.0
for batch_ind, batch_target in train_loader:
optimizer.zero_grad()
batch_target = batch_target.to(device)
# forward
emb = gcn(dataset)
# distmult
batch_output = classifier(emb, batch_ind)
# loss
loss = crit(batch_output, batch_target)
epoch_loss += loss.item() * batch_target.size(0)
count += batch_target.size(0)
loss.backward()
optimizer.step()
epoch_loss /= count
# eval acc
if args.classifier == "nn":
acc, test_loss, roc_auc = eval_test_nn(gcn, classifier,
test_loader, device,
dataset, crit)
elif args.classifier == "dm" or args.classifier == "ip":
acc, test_loss, roc_auc = eval_test_dm(gcn, classifier,
test_loader, device,
dataset, crit)
tqdm_train.set_description(
f"fold {fold_ind}, epoch {epoch+1}, average loss: {round(epoch_loss, 4)}, test accuracy: {round(acc, 4)} , roc auc: {round(roc_auc, 4)} , test loss: {round(test_loss, 4)}"
)
if roc_auc >= best_auc:
best_auc = roc_auc
torch.save(
gcn.state_dict(),
f"best_auc_gcn_{type_name}_{setting_symbol}_fold_{fold_ind}.pth"
)
torch.save(
classifier.state_dict(),
f"best_auc_classifier_{type_name}_{setting_symbol}_fold_{fold_ind}.pth"
)
scheduler.step()
# final evaluation
# eval on last
if args.classifier == "nn":
mean_auc, mean_rank, fpr_list, tpr_list = eval_all_nn(
gcn, classifier, device, dataset, refine_gene_list,
positive_train + add_links, positive_test, negative_train)
elif args.classifier == "dm" or args.classifier == "ip":
mean_auc, mean_rank, fpr_list, tpr_list = eval_all_dm(
gcn, classifier, device, dataset, refine_gene_list,
positive_train + add_links, positive_test, negative_train)
print(
f"====== \n {type_name}_{setting_symbol}, fold {fold_ind}, eval on last \n mean auc is: {round(mean_auc, 4)} \n mean rank is: {round(mean_rank, 4)} \n======"
)
torch.save(gcn.state_dict(),
f"last_gcn_{type_name}_{setting_symbol}_fold_{fold_ind}.pth")
torch.save(
classifier.state_dict(),
f"last_classifier_{type_name}_{setting_symbol}_fold_{fold_ind}.pth")
# save fold result
each_mean_auc_on_last[fold_ind - 1].append(mean_auc)
each_mean_rank_on_last[fold_ind - 1].append(mean_rank)
each_fpr_list_on_last[fold_ind - 1].extend(fpr_list)
each_tpr_list_on_last[fold_ind - 1].extend(tpr_list)
# eval on best auc
gcn.load_state_dict(
torch.load(
f"best_auc_gcn_{type_name}_{setting_symbol}_fold_{fold_ind}.pth"))
classifier.load_state_dict(
torch.load(
f"best_auc_classifier_{type_name}_{setting_symbol}_fold_{fold_ind}.pth"
))
if args.classifier == "nn":
mean_auc, mean_rank, fpr_list, tpr_list = eval_all_nn(
gcn, classifier, device, dataset, refine_gene_list,
positive_train + add_links, positive_test, negative_train)
elif args.classifier == "dm" or args.classifier == "ip":
mean_auc, mean_rank, fpr_list, tpr_list = eval_all_dm(
gcn, classifier, device, dataset, refine_gene_list,
positive_train + add_links, positive_test, negative_train)
print(
f"====== \n {type_name}_{setting_symbol}, fold {fold_ind}, eval on best auc \n mean auc is: {round(mean_auc, 4)} \n mean rank is: {round(mean_rank, 4)} \n======"
)
# save fold result
each_mean_auc[fold_ind - 1].append(mean_auc)
each_mean_rank[fold_ind - 1].append(mean_rank)
each_fpr_list[fold_ind - 1].extend(fpr_list)
each_tpr_list[fold_ind - 1].extend(tpr_list)
dataset,loader,test_dataset,test_loader = loadDatasets(args.train_dataset,
args.test_dataset,args.window_size,args.batch_size,shuffle=True,vocab_size=args.vocab_size,compute_dicts = args.embeddings_path is None)
print("Number of training examples / test examples: {} / {}".format(len(dataset),len(test_dataset)))
if args.embeddings_path is not None:
embLoader = EmbeddingsLoader(args.embeddings_path)
print("Embeddings loaded, vocab size: {} embedding dims: {}".format(embLoader.get_dimensions()[0],embLoader.get_dimensions()[1]))
args.vocab_size,args.embedding_size = embLoader.get_dimensions()
model = NN(vocab_size=args.vocab_size,number_tags=dataset.get_num_tags(),
batch_size=args.batch_size,gpu=args.gpu,window_size=args.window_size,
include_cap=args.capitalization,number_suffix=dataset.get_num_suffix(),
embedding_dim=args.embedding_size,activation=args.activation)
model.load_state_dict(torch.load(args.trained_model))
if args.gpu and torch.cuda.is_available():
model.cuda()
tagged_file = open(args.output_file,"w")
if args.embeddings_path is not None:
word_to_index = embLoader.word_to_index
else:
word_to_index = dataset.word_to_index
for i,(words_input,tags) in enumerate(test_loader):
input_tensor,target_tensor,cap_tensor,suf_tensor = model.get_train_tensors(words_input,tags,
word_to_index,dataset.tag_to_index,suffix2id=dataset.suffix_to_index)
output = model(Variable(input_tensor),Variable(cap_tensor),Variable(suf_tensor))