def run_training(model, run_name, epochs, plot=None):
global optimizer
model.cuda()
optimizer = optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum)
if plot is None:
plot = PlotLearning('./plots/cifar/', 10, prefix=run_name)
for epoch in range(1, epochs + 1):
accu_train, loss_train = train(epoch)
accu_test, loss_test = test()
logs = {}
logs['acc'] = accu_train
logs['val_acc'] = accu_test
logs['loss'] = loss_train
logs['val_loss'] = loss_test
plot.plot(logs)
return plot
def run_training(model, run_name, epochs, plot=None):
visdom_log = {
'model_type': run_name,
'epoch': [],
'train_accuracy': [],
'test_accuracy': [],
'train_loss': [],
'test_loss': [],
'top_train_data': {
'epoch': 0,
'accuracy': 0.0,
'loss': 0.0
},
'top_test_data': {
'epoch': 0,
'accuracy': 0.0,
'loss': 0.0
}
}
model.cuda()
if plot is None:
plot = PlotLearning('./plots/cifar/', 10, plot_name=run_name)
for epoch in range(1, epochs + 1):
accu_train, loss_train = train(model, epoch)
accu_test, loss_test = test(model)
logs = {}
logs['acc'] = accu_train
logs['val_acc'] = accu_test
logs['loss'] = loss_train
logs['val_loss'] = loss_test
plot.plot(logs)
visdom_log['epoch'].append(epoch)
visdom_log['train_accuracy'].append(accu_train)
visdom_log['test_accuracy'].append(accu_test)
visdom_log['train_loss'].append(loss_train)
visdom_log['test_loss'].append(loss_test)
return plot, visdom_log
best_test["test_accuracy"] = accu_test
best_test["validation_accuarcy"] = accu_train
# print "========================================="
# print "the epoch with best test accuracy is:",
# print best_test
# print "the accuracy reached with net2net in",
# print acc
# print "========================================="
return plot, best_test, log
if __name__ == "__main__":
logs = []
visdom_plot = PlotLearning('./plots/cifar/',
10,
prefix='Net2Net Implementation',
plot_name=args.plot_name + '_wider')
print("\n\n > Teacher (Base Network) training ... ")
model = Net()
model.double()
model.cuda()
print model
criterion = nn.NLLLoss()
plot, _, log_base = run_training(model,
'Teacher',
args.epochs,
plot=visdom_plot,
color='red')
logs.append(log_base)
eps_end=0.01,
inp_dims=[8],
lr=0.001)
scores, eps_history = [], []
n_games = 500
for i in range(n_games):
score = 0
done = False
observation = env.reset()
while not done:
action = agent.choose_action(observation)
observation_, reward, done, info = env.step(action)
score += reward
agent.store_transition(observation, action, reward, observation_,
done)
agent.learn()
observation = observation_
scores.append(score)
eps_history.append(agent.epsilon)
avg_score = np.mean(scores[-100:])
print('episode ', i, 'score %.2f' % score,
'average score %.2f' % avg_score, 'epsilon %.2f' % agent.epsilon)
x = [i + 1 for i in range(n_games)]
filename = 'lunar_lander.png'
PlotLearning(x, scores, eps_history, filename)
# https://github.com/user432/DeepRL-in-PyTorch/
start_t = time.time()
print("\n\n > Wider teacher training ... ")
colors.append('green')
trace_names.extend(['Wider Random Train', 'Wider Random Test'])
model_ = ConvNet(net_dataset=CIFAR10)
del model
model = model_
model.define_wider(widening_factor=2)
model.cuda()
print model
_, log_random_init = run_training(model, 'Wider_teacher_', args.epochs)
print(" >> Time taken {}".format(time.time() - start_t))
logs.append(log_random_init)
# # wider deeper teacher training
# print("\n\n > Wider+Deeper teacher training ... ")
# start_t = time.time()
# model_ = Net()
#
# del model
# model = model_
# model.define_wider_deeper()
# run_training(model, 'Wider_Deeper_teacher_', args.epochs + 1)
# print(" >> Time taken {}".format(time.time() - start_t))
visdom_plot_final = PlotLearning('./plots/cifar/',
10,
plot_name=args.plot_name)
visdom_plot_final.plot_logs(logs, trace_names, colors)
#
# visdom_plot = PlotLearning('./plots/cifar/', 10, prefix='Net2Net Implementation',
# plot_name=args.plot_name + "_deeper")
# visdom_plot.plot_logs(logs, args.plot_name)
# For Wider + Deeper model training
#
# del logs[1:]
#
# # wider deeper model training from scratch
# print("\n\n > Wider+Deeper Model training from scratch... ")
# model_wider_deeper = Net()
# model_wider_deeper.define_wider_deeper()
# _, _, log = run_training(model_wider_deeper, 'Wider_Deeper_teacher_', args.epochs)
# logs.append(log)
#
# # wider + deeper model training from Net2Net
# print("\n\n > Wider+Deeper Student training with Net2Net... ")
# model_wider_deeper_net2net = Net()
# model_wider_deeper_net2net.net2net_wider()
# model_wider_deeper_net2net = copy.deepcopy(model_wider_Net2Net)
# model_wider_deeper_net2net.net2net_deeper()
# _, _, log = run_training(model_wider_deeper_net2net, 'WiderDeeper_student_', args.epochs)
# logs.append(log)
visdom_plot = PlotLearning('./plots/cifar/',
10,
prefix='Net2Net Implementation',
plot_name=args.plot_name + "_deeper_wider")
visdom_plot.plot_logs(logs, args.plot_name)
if plot is not None:
win_accuracy, win_loss = plot_live_data(plot, win_accuracy,
win_loss, run_name,
**live_data)
return visdom_log, win_accuracy, win_loss
if __name__ == "__main__":
logs = []
colors = []
trace_names = []
if args.plot_name is not None:
visdom_live_plot = PlotLearning('./plots/cifar/',
10,
plot_name=args.plot_name,
env_name=args.env_name)
else:
visdom_live_plot = None
start_t = time.time()
print("\n\n > Teacher training ... ")
colors.append('orange')
trace_names.extend(['Teacher Train', 'Teacher Test'])
model = Net()
model.cuda()
criterion = nn.NLLLoss()
log_base, win_accuracy, win_loss = run_training(model, 'Teacher_',
args.epochs,
visdom_live_plot)
logs.append(log_base)
if __name__ == '__main__':
agent = Agent(ALPHA=0.001,
inp_dims=[8],
GAMMA=0.99,
n_actions=4,
l1_size=128,
l2_size=128)
env = gym.make('LunarLander-v2')
score_history = []
score = 0
num_episodes = 2500
env = wrappers.Monitor(env,
"tmp/lunar-lander",
video_callable=lambda episode_id: True,
force=True)
for i in range(num_episodes):
print('episode: ', i, 'score: ', score)
done = False
score = 0
observation = env.reset()
while not done:
action = agent.choose_action(observation)
observation_, reward, done, info = env.step(action)
agent.store_rewards(reward)
observation = observation_
score += reward
score_history.append(score)
agent.learn()
filename = 'lunar-lander-alpha001-128x128fc-newG.png'
PlotLearning(score_history, filename=filename, window=25)