def train():
app = Net()
app = app.create_model()
images, labels = read_tfrecord()
images_val, labels_val = read_tfrecord(type_='val')
app.fit(images, labels, 128, 100, validation_data=(images_val, labels_val))
app.save('../model/net.h5')
def run_model(data, test_data, args):
### INITIALIZE SETTINGS ###
G = Global()
G.set_debug_true() ####
if data is None:
return 0
### DEFINE MODEL ###
n, m = data.shape
out_dim = args['output-dimension']
hidden_model = DeepNet
hidden_sizes = args['hidden-dims']
index = Net(hidden_model, m, out_dim, hidden_sizes).to(device)
### TRAIN MODEL ###
index = index.fit(data, args)
### TEST MODEL ###
results = index.test(data, test_data)
if args['to_plot']:
index.create_plot(data, filename='train')
## Plot test data ##
index.create_plot(test_data, name='Test Data', filename='test')
## Save plots ##
name = args['filename_prefix']
G.save_figs(name)
## Release Space ##
del G, index
## Return neighbors ##
return results
def generate_rollout(self,
itration_number,
iter_train,
iter_dev,
verbose=False):
self.log_probs = []
self.actions = []
self.entropies = []
self.reward = None
state = torch.zeros(self.hidden_size)
terminated = False
self.reward = 0
while not terminated:
log_prob, state, terminated = self.step(state)
self.log_probs.append(log_prob)
if verbose:
print('\nGenerated network:')
print(self.actions)
net = Net(self.actions)
accuracy = net.fit(iter_train, iter_dev)
self.reward += accuracy
return self.reward, net
def generate_rollout(self, iter_train, iter_dev, verbose=False):
self.log_probs = []
self.actions = []
self.entropies = []
self.reward = None
self.adv = 0
state = None
input = torch.zeros(self.hidden_size)
is_embed = True
terminate = False
self.reward = 0
while not terminate:
logits, state = self(input, state, is_embed)
idx = torch.distributions.Categorical(logits=logits).sample().detach()
probs = F.softmax(logits, dim=-1)
log_probs = torch.log(probs)
self.log_probs.append(log_probs[idx])
action = self.index_to_action[int(idx)]
self.actions.append(action)
entropy = -(log_probs * probs).sum(dim=-1)
self.entropies.append(entropy)
terminate = (action == 'EOS') or (len(self.actions) == self.max_depth)
is_embed = False
input = Variable(torch.LongTensor([idx]), requires_grad=False)
if verbose:
print('\nGenerated network:')
print(self.actions)
net = Net(self.actions)
accuracy = net.fit(iter_train, iter_dev)
self.reward += accuracy
# moving average baseline
if self.baseline is None:
self.baseline = self.reward
else:
self.baseline = self.decay * self.baseline + (1 - self.decay) * self.reward
self.adv = self.reward - self.baseline
return self.reward, self.adv
