def main():
if (FLAGS.vector_opt == 0):
z = np.zeros(shape=(FLAGS.batch_size, DIM))
elif (FLAGS.vector_opt == 1):
z = np.random.uniform(-0.5, 0.5, size=(FLAGS.batch_size, DIM))
elif (FLAGS.vector_opt == 2):
z = load_vector()
elif (FLAGS.vector_opt == 3):
z = gaussian_cf_sampler(np.load(FLAGS.vector))
elif (FLAGS.vector_opt == 4):
z = esm_cf_sampler(np.load(FLAGS.vector), 0.5)
else:
print(
"Invalid value for vector_opt argument. 0-4 are the only acceptable values."
)
print("Use -h or --help flags for more information.")
return
np.save(
"./prev_img_vector.npy",
z) # save our vectors to a file, so if we like one we can replicate
network, tf_session = restore_model()
if (FLAGS.gif):
base_images = [
np.load('./counterfactualGANs/base_vectors/%s' % i)
for i in BASE_VECTORS
]
utils.make_gif(base_images, 'bases.gif')
else:
generate_image(z, network, tf_session, VIZ_OPTION)
def main():
counter = 20
rn = "golf"
nz = 100
parallel = True
W = torch.load('runs%d/nets_%s/netZ_glo.pth' % (counter, rn))
W = W['emb.weight'].data.cpu().numpy()
netG = model_video_orig.netG_new(nz).cuda()
if torch.cuda.device_count() > 1:
parallel = True
print("Using", torch.cuda.device_count(), "GPUs!")
netG = nn.DataParallel(netG)
Zs = utils.sample_gaussian(torch.from_numpy(W), 10000)
Zs = Zs.data.cpu().numpy()
state_dict = torch.load('runs%d/nets_%s/netG_glo.pth' % (counter, rn))
netG.load_state_dict(state_dict) # load the weights for generator (GLO)
if parallel:
netG = netG.module
gmm = GaussianMixture(n_components=100,
covariance_type='full',
max_iter=100,
n_init=10)
gmm.fit(W)
z = torch.from_numpy(gmm.sample(100)[0]).float().cuda()
video = netG(z)
utils.make_gif(video, 'runs%d/ims_%s/sample' % (counter, rn), 16)
return video
def main():
counter = 21
rn = "golf"
nz = 100
parallel = True
W = torch.load('runs%d/nets_%s/netZ_glo.pth' % (counter, rn))
W = W['emb.weight'].data.cpu().numpy()
netG = model_video_orig.netG_new(nz).cuda()
if torch.cuda.device_count() > 1:
parallel = True
print("Using", torch.cuda.device_count(), "GPUs!")
netG = nn.DataParallel(netG)
state_dict = torch.load('runs%d/nets_%s/netG_glo.pth' % (counter, rn))
netG.load_state_dict(state_dict) # load the weights for generator (GLO)
if parallel:
netG = netG.module
d = 16
nepoch = 200
vaet = vae.VAETrainer(W, d)
vaet.train_vae(nepoch)
torch.save(vaet.vae.netVAE.state_dict(), 'runs%d/nets_%s/netVAE.pth' % (counter, rn))
z = vaet.vae.netVAE.decode(torch.randn(100, d).cuda())
video = netG(z)
print("video shape is", video.shape)
utils.make_gif(video, 'runs%d/ims_%s/sample' % (counter, rn), 16)
return video
def play(self, coord):
total_steps = 0
test_episode_count = 0
print("Starting agent " + str(self.thread_id))
while not coord.should_stop():
episode_regret = 0
episode_suboptimal_arm = 0
episode_reward = [0 for _ in range(FLAGS.nb_actions)]
episode_step_count = 0
episode_frames = []
d = False
r = 0
a = 0
t = 0
self.env.set(self.settings["envs"][test_episode_count])
while not d:
a = random.randint(0, FLAGS.nb_actions - 1)
r, d, t = self.env.pull_arm(a)
episode_frames.append(
set_image_bandit(episode_reward, self.env.get_bandit(), a,
t))
episode_regret += self.env.get_timestep_regret(a)
optimal_action = self.env.get_optimal_arm()
if optimal_action != a:
episode_suboptimal_arm += 1
episode_reward[a] += r
total_steps += 1
episode_step_count += 1
self.episode_rewards.append(np.sum(episode_reward))
self.episodes_suboptimal_arms.append(episode_suboptimal_arm)
self.episode_regrets.append(episode_regret)
self.images = np.array(episode_frames)
make_gif(self.images,
FLAGS.frames_test_dir + '/image' +
str(test_episode_count) + '.gif',
duration=len(self.images) * 0.1,
true_image=True)
if test_episode_count == FLAGS.nb_test_episodes - 1:
mean_regret = np.mean(self.episode_regrets)
mean_nb_suboptimal_arms = np.mean(
self.episodes_suboptimal_arms)
print("Mean regret for the model is {}".format(mean_regret))
print("Regret in terms of suboptimal arms is {}".format(
mean_nb_suboptimal_arms))
return 1
test_episode_count += 1
def train():
tf.keras.backend.clear_session()
with strategy.scope() if config['use_tpu'] else empty_context_mgr():
model = GAN_PG(**config)
# Define optimizers
optimizer_g = tf.train.AdamOptimizer(
learning_rate=config['learning_rate'], beta1=0.0)
optimizer_d = tf.train.AdamOptimizer(
learning_rate=config['learning_rate'], beta1=0.0)
# Compile the model
model.compile_model(optimizer_g=optimizer_g,
optimizer_d=optimizer_d,
loss=config['gan_mode'],
tpu_strategy=strategy,
resolver=resolver,
config=config['sess_config'])
if config['restore']:
model.load_weights('{}/weights'.format(config['folder']))
# Prepare inputs
inputs = (X_train, y_train) if config['conditional'] else X_train
# Train
for stage in config['train_stages']:
# Get training stage num
stage_num = config['train_stages'].index(stage)
print(
'\nProcessing stage: {} with image size {} =========================================='
.format(stage_num, stage['size']))
# Define schedulers
alpha_scheduler = Scheduler(stage['train_epochs'], [0, 0.5],
[0, 1.0])
learning_rate_scheduler = Scheduler(
stage['train_epochs'], [0, 0.5],
[stage['lr'] * 0.1, stage['lr']])
model.fit_stage(inputs,
config['batch_size'],
stage_num=stage_num,
alpha_scheduler=alpha_scheduler,
learning_rate_scheduler=learning_rate_scheduler,
folder=config['folder'],
save_epoch=config['save_epoch'],
seed_noise=seed_noise,
seed_labels=seed_labels)
make_gif(
glob.iglob('{}/progress/*.png'.format(config['folder'])),
'{}/progress/{}_{}.gif'.format(config['folder'], config['gan_mode'],
'progress'))
def main():
# initial values
x, y = 1, 2
gd = GradientDescent(x, y, parameters["SGD"])
for step in range(0, parameters["n_epochs"]):
gd.minimize()
# save animation
save_anim(parameters["n_epochs"], gd.x_hist, gd.y_hist, gd.z_hist, "SGD")
make_gif(parameters["n_epochs"])
def generate_random_walking(self):
"""Generate fonts with random walking
Generate fonts from random walking inputs.
Results are changed gradually.
"""
for c in self.embedding_chars:
dst_dir = os.path.join(self.dst_walk, c)
if not os.path.exists(dst_dir):
os.mkdir(dst_dir)
batch_n = (self.char_embedding_n * FLAGS.char_img_n) // self.batch_size
c_ids = self.real_dataset.get_ids_from_labels(self.embedding_chars)
for batch_i in tqdm(range(batch_n)):
style_id_start = batch_i
if batch_i == batch_n - 1:
style_id_end = 0
else:
style_id_end = batch_i + 1
generated_imgs = self.sess.run(
self.generated_imgs,
feed_dict={
self.style_ids_x:
np.ones(self.batch_size) * style_id_start,
self.style_ids_y:
np.ones(self.batch_size) * style_id_end,
self.style_ids_alpha:
np.repeat(
np.linspace(0., 1., num=self.walk_step,
endpoint=False), self.char_embedding_n),
self.char_ids_x:
np.tile(c_ids, self.batch_size // self.char_embedding_n),
self.char_ids_y:
np.tile(c_ids, self.batch_size // self.char_embedding_n),
self.char_ids_alpha:
np.zeros(self.batch_size)
})
for img_i in range(generated_imgs.shape[0]):
img = generated_imgs[img_i]
img = (img + 1.) * 127.5
pil_img = Image.fromarray(np.uint8(img))
pil_img.save(
os.path.join(
self.dst_walk,
str(self.embedding_chars[img_i %
self.char_embedding_n]),
'{:05d}.png'.format(
(batch_i * self.batch_size + img_i) //
self.char_embedding_n)))
print('making gif animations...')
for i in range(self.char_embedding_n):
make_gif(
os.path.join(self.dst_walk, self.embedding_chars[i]),
os.path.join(self.dst_walk, self.embedding_chars[i] + '.gif'))
def main():
#定义网络
net = models.LeNetWithAngle(classes_num)
if use_gpu:
net = net.cuda()
#定义优化器
optimizer = torch.optim.SGD(net.parameters(),
lr=model_lr,
weight_decay=1e-5,
nesterov=True,
momentum=0.9)
print("net and optimzer load succeed")
#定义数据加载
trainloader, testloader = dataloader.get_loader(batch_size=batch_size,
root_path="./data/MNIST")
print("data load succeed")
#定义logger
logger = utils.Logger(tb_path="./logs/tblog/")
#定义学习率调整器
scheduler = lr_sche.StepLR(optimizer, 30, 0.1)
#定义损失函数
criterion = a_softmax.AngleSoftmaxLoss(gamma=0)
best_acc = 0
#开始训练
for i in range(1, epochs + 1):
scheduler.step(epoch=i)
net.train()
train_acc,train_loss,all_feature,all_labels=\
train(net,optimizer,criterion,trainloader,i)
utils.plot_features(all_feature, all_labels, classes_num, i,
"./logs/images/train/train_{}.png")
net.eval()
test_acc, test_loss, all_feature, all_labels = test(
net, criterion, testloader, i)
utils.plot_features(all_feature, all_labels, classes_num, i,
"./logs/images/test/test_{}.png")
print("{} epoch end, train acc is {:.4f}, test acc is {:.4f}".format(
i, train_acc, test_acc))
content = {
"Train/acc": train_acc,
"Test/acc": test_acc,
"Train/loss": train_loss,
"Test/loss": test_loss
}
logger.log(step=i, content=content)
if best_acc < test_acc:
best_acc = test_acc
utils.save_checkpoints("./logs/weights/net_{}.pth",i,\
net.state_dict(),(best_acc==test_acc))
utils.make_gif("./logs/images/train/", "./logs/train.gif")
utils.make_gif("./logs/images/test/", "./logs/test.gif")
print("Traing finished...")
def play(self, coord):
total_steps = 0
test_episode_count = 0
print("Starting agent " + str(self.thread_id))
while not coord.should_stop():
episode_regret = 0
episode_suboptimal_arm = 0
episode_reward = [0 for _ in range(FLAGS.nb_actions)]
episode_step_count = 0
episode_frames = []
d = False
r = 0
a = 0
t = 0
self.env.set(self.settings["envs"][test_episode_count])
while not d:
a = random.randint(0, FLAGS.nb_actions - 1)
r, d, t = self.env.pull_arm(a)
episode_frames.append(set_image_bandit(episode_reward, self.env.get_bandit(), a, t))
episode_regret += self.env.get_timestep_regret(a)
optimal_action = self.env.get_optimal_arm()
if optimal_action != a:
episode_suboptimal_arm += 1
episode_reward[a] += r
total_steps += 1
episode_step_count += 1
self.episode_rewards.append(np.sum(episode_reward))
self.episodes_suboptimal_arms.append(episode_suboptimal_arm)
self.episode_regrets.append(episode_regret)
self.images = np.array(episode_frames)
make_gif(self.images, FLAGS.frames_test_dir + '/image' + str(test_episode_count) + '.gif',
duration=len(self.images) * 0.1, true_image=True)
if test_episode_count == FLAGS.nb_test_episodes - 1:
mean_regret = np.mean(self.episode_regrets)
mean_nb_suboptimal_arms = np.mean(self.episodes_suboptimal_arms)
print("Mean regret for the model is {}".format(mean_regret))
print("Regret in terms of suboptimal arms is {}".format(mean_nb_suboptimal_arms))
return 1
test_episode_count += 1
def visualize(self, epoch, opt_params):
Igen = self.netG(self.fixed_noise) # GLO on a noise
utils.make_gif(
Igen,
'runs%d/ims_%s/generations_epoch_%03d' % (counter, self.rn, epoch),
opt_params.batch_size)
z = utils.sample_gaussian(self.netZ.emb.weight.clone().cpu(),
self.vis_n)
Igauss = self.netG(z) # GLO on gaussian
utils.make_gif(
Igauss,
'runs%d/ims_%s/gaussian_epoch_%03d' % (counter, self.rn, epoch),
opt_params.batch_size)
def celeb():
""" Function that loads Celeb data, reshapes it to TF desired input (hight, width, color)
Then, the function defines the shape of the discriminator and generator
"""
# This just gets a list of filenames to be loaded in dynamically due to their large number and size
X = utils.load_Celeb()
dimensions = 64 # Assumes images are square - uses only 1 dimension
colors = 3
# Hyperparamters gathered from other official implementations that worked! Selected with hyper param optimisation techniques
# Hyperparameter keys:
# conv layer: (feature maps, filter size, stride=2, batch norm used?)
# dense layer: (hidden units, batch norm used?)
discriminator_sizes = {
'conv_layers': [(64, 5, 2, False), (128, 5, 2, True),
(256, 5, 2, True), (512, 5, 2, True)],
'dense_layers': []
}
# Hyperparameter keys:
# z : latent variable dimensionality (drawing uniform random samples from it)
# projection: initial number of feature maps (flat vector -> 3D image!)
# batchNorm_after_projection: flag, showing, if we want to use batchnorm after projecting the flat vector
# conv layer: (feature maps, filter size, stride=2, batch norm used?)
# dense layer: (hidden units, batch norm used?)
# output_action: activation function - using sigmoid since the Celeb data is scaled between {-1, 1} - This is recommended by GAN researchers
generator_sizes = {
'z':
100,
'projection':
512,
'batchNorm_after_projection':
True,
'conv_layers': [(256, 5, 2, True), (128, 5, 2, True), (64, 5, 2, True),
(colors, 5, 2, False)],
'dense_layers': [],
'output_activation':
tf.tanh,
}
# Create the DCGAN and fit it to the images
name = 'Celeb'
GAN = DCGAN(dimensions, colors, discriminator_sizes, generator_sizes)
GAN.fit(X, name)
utils.make_gif(name)
def mnist():
""" Function that loads MNIST, reshapes it to TF desired input (hight, width, color)
Then, the function defines the shape of the discriminator and generator
"""
X, _ = utils.load_MNIST()
X = X.reshape(len(X), 28, 28, 1)
dimensions = X.shape[
1] # Assumes images are square - uses only 1 dimension
colors = X.shape[-1]
# Hyperparamters gathered from other official implementations that worked! Selected with hyper param optimisation techniques
# Hyperparameter keys:
# conv layer: (feature maps, filter size, stride=2, batch norm used?)
# dense layer: (hidden units, batch norm used?)
discriminator_sizes = {
'conv_layers': [(2, 5, 2, False), (64, 5, 2, True)],
'dense_layers': [(1024, True)]
}
# Hyperparameter keys:
# z : latent variable dimensionality (drawing uniform random samples from it)
# projection: initial number of feature maps (flat vector -> 3D image!)
# batchNorm_after_projection: flag, showing, if we want to use batchnorm after projecting the flat vector
# conv layer: (feature maps, filter size, stride=2, batch norm used?)
# dense layer: (hidden units, batch norm used?)
# output_action: activation function - using sigmoid since MNIST varies between {0, 1}
generator_sizes = {
'z': 100,
'projection': 128,
'batchNorm_after_projection': False,
'conv_layers': [(128, 5, 2, True), (colors, 5, 2, False)],
'dense_layers': [(1024, True)],
'output_activation': tf.sigmoid,
}
# Create the DCGAN and fit it to the images
name = 'MNIST'
GAN = DCGAN(dimensions, colors, discriminator_sizes, generator_sizes)
GAN.fit(X, name)
utils.make_gif(name)
def build_weixin_message(weixin):
items = []
index = 0
message_title = ''
if weixin.cover.name != u'' and weixin.title != u'' and weixin.recommended_reason != u'':
message_title = weixin.title
items.append(WeixinMessageItem(image=weixin.cover.path, title=weixin.title, digest=weixin.title, content=weixin.recommended_reason))
index += 1
for news in weixin.news.all():
title = u'游戏情报站 - %s' % news.brief_comment
if index == 0:
message_title = title
items.append(WeixinMessageItem(image=make_gif(news), title=title, digest=title, content=news.recommended_reason + u'<br><br><font color="gray">点击“阅读原文”查看更多</font>', sourceurl=u'http://cow.bestgames7.com/d/%s/' % news.id))
index += 1
for game in weixin.games.all():
title = u'%s - %s' % (game.name, game.brief_comment)
if index == 0:
message_title = title
if index > 0:
items.append(WeixinMessageItem(image=make_gif(game), title=title, digest=title, content=_normalize_content(game), sourceurl=u'http://cow.bestgames7.com/d/%s/' % game.id))
else:
items.append(WeixinMessageItem(image=make_gif(game), title=title, digest=title, content=_normalize_content(game), sourceurl=u'http://cow.bestgames7.com/d/%s/' % game.id))
index += 1
for player in weixin.players.all():
title = u'我是玩家 - %s' % player.brief_comment
if index == 0:
message_title = title
items.append(WeixinMessageItem(image=player.image_url.path, title=title, digest=title, content=player.recommended_reason))
index += 1
for puzzle in weixin.puzzles.all():
title = u'趣味答题 - %s' % puzzle.title
if index == 0:
message_title = title
content = puzzle.description + '<br>'
content += 'A.' + puzzle.option1 + '<br>'
content += 'B.' + puzzle.option2 + '<br>'
content += 'C.' + puzzle.option3 + '<br>'
content += 'D.' + puzzle.option4 + '<br></br>'
content += u'<font color="gray">回复"答题"，参与答题得积分换礼品的活动吧</font><br>'
items.append(WeixinMessageItem(image=puzzle.image_url.path, title=title, digest=title, content=content))
index += 1
return WeixinMessage(weixin.id, message_title, items)
def get_optimized_model_camera(mesh, camera, config):
model = ModelCamera(mesh.vertices, mesh.faces, mesh.regions, config.CAMERA.REF_SILHOUETTE,
camera.x, camera.y, camera.z, min_distance=config.CAMERA.MIN_DISTANCE,
max_rotation_l_r=config.CAMERA.MAX_ROTATION_LR, max_rotation_u_d=config.CAMERA.MAX_ROTATION_UD,
use_anchor_points=config.CAMERA.USE_ANCHOR_POINTS,
silhouette_nose=config.CAMERA.ANCHOR_NOSE_IMG, silhouette_mouth=config.CAMERA.ANCHOR_MOUTH_IMG)
model.cuda()
optimizer = torch.optim.Adam(model.parameters(), lr=config.OPT.LR_CAMERA)
dir_imgs = os.path.join(config.PATH.TMP, "camera")
if not os.path.isdir(dir_imgs):
os.mkdir(dir_imgs)
loop = tqdm(range(config.OPT.ITER_CAMERA))
for i in loop:
loop.set_description(f"Optimizing camera")
optimizer.zero_grad()
loss = model()
loss.backward()
optimizer.step()
images, _, _ = model.renderer(model.vertices, model.faces, torch.tanh(model.textures))
image = images.detach().cpu().numpy()[0].transpose(1, 2, 0)
imsave(os.path.join(dir_imgs, "%04d.png" % i), (255 * image).astype(np.uint8))
# Save output igms and gif
make_gif(os.path.join(config.PATH.OUT, "camera.gif"), dir_imgs)
# Silhouette faces
image = model.renderer(model.vertices, model.faces, mode='silhouettes')
image = image.detach().cpu().numpy().transpose(1, 2, 0)
imsave(os.path.join(config.PATH.OUT, "silhouette_camera.png"), (255 * image).astype(np.uint8))
# Silhouette nose
image = model.renderer(model.vertices, model.triangles_nose, mode='silhouettes')
image = image.detach().cpu().numpy().transpose(1, 2, 0)
imsave(os.path.join(config.PATH.OUT, "silhouette_nose.png"), (255 * image).astype(np.uint8))
# Silhouette mouth
image = model.renderer(model.vertices, model.triangles_mouth, mode='silhouettes')
image = image.detach().cpu().numpy().transpose(1, 2, 0)
imsave(os.path.join(config.PATH.OUT, "silhouette_mouth.png"), (255 * image).astype(np.uint8))
return model
def post_process_and_save_he(sols,
results_dir="./data/HE",
name_simulation="HE_sim_default_name",
save_plots=False):
try:
os.makedirs(results_dir)
except:
pass
np.save(os.path.join(results_dir, name_simulation + "_array"), sols)
if save_plots:
make_gif(sols,
os.path.join(results_dir, name_simulation + "gif" + ".gif"))
ampsn = []
for sol in sols:
ampsn.append(np.sum(sol))
ampsn = np.array(ampsn) / np.size(sol)
amps_max = []
for sol in sols:
amps_max.append(np.max(sol))
amps_max = np.array(amps_max)
fig = plt.figure(figsize=(15, 15))
plt.plot(np.arange(0, len(ampsn), 1), np.abs(ampsn), label="avg amp")
plt.legend()
fig.savefig(
os.path.join(results_dir, name_simulation + "amplitude_avg.png"))
fig = plt.figure(figsize=(15, 15))
plt.plot(np.arange(0, len(amps_max), 1),
np.abs(amps_max),
label="amp max")
plt.legend()
fig.savefig(
os.path.join(results_dir, name_simulation + "amplitude_max.png"))
def post_process_and_save_ac(sols,
results_dir="./data/AC",
name_simulation="AC_sim_default_name",
save_plots=False,
duration=0.05):
try:
os.makedirs(results_dir)
except:
pass
np.save(os.path.join(results_dir, name_simulation + "_array"), sols)
if save_plots:
make_gif(sols,
os.path.join(results_dir, name_simulation + "gif" + ".gif"),
duration=duration)
amps = []
for sol in sols:
amps.append(np.sum(np.abs(sol)))
amps = np.array(amps) / np.size(sol)
ampsn = []
for sol in sols:
ampsn.append(np.sum(sol))
ampsn = np.array(ampsn) / np.size(sol)
fig = plt.figure(figsize=(15, 15))
plt.plot(np.arange(0, len(amps), 1), amps, label="avg abs phase")
plt.plot(np.arange(0, len(ampsn), 1),
np.abs(ampsn),
label="abs avg phase")
plt.legend()
plt.savefig(
os.path.join(results_dir, name_simulation + "avg_phases_time.png"))
def on_epoch_end(self, trainer, model):
#epoch = trainer.current_epoch
self._epoch += 1
epoch = int(self._epoch / 2) #dirty fix
if epoch % self._plot_every == 0:
plt.close("all")
fig = plot_2D_comparison_analytical(model,
self._val_fun,
title="epoch {}".format(epoch))
array = fig_to_array(fig)
self._arrays.append(array)
if epoch % self._save_every == 0:
make_gif(self._arrays,
self._results_dir + "/{}.gif".format(self._name),
size=self._size,
duration=0.5)
def get_optimized_model_morphing(mesh, camera, config):
model = ModelMorphing(mesh.vertices, mesh.faces, config.CAMERA.REF_SILHOUETTE, camera.x, camera.y, camera.z)
model.cuda()
optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=config.OPT.LR_MORPHING)
dir_imgs = os.path.join(config.PATH.TMP, "morphing")
if not os.path.isdir(dir_imgs):
os.mkdir(dir_imgs)
loop = tqdm(range(config.OPT.ITER_MORPHING))
for i in loop:
loop.set_description(f"Morphing model")
optimizer.zero_grad()
loss = model()
loss.backward()
optimizer.step()
images = model.renderer(model.vertices, model.faces, mode='silhouettes')
image = images.detach().cpu().numpy().transpose(1, 2, 0)
imsave(os.path.join(dir_imgs, "%04d.png" % i), (255 * image).astype(np.uint8))
make_gif(os.path.join(config.PATH.OUT, "morphing.gif"), dir_imgs)
return model
def play_episode(model, env, memory, save_path):
memory.clear()
frame = env.reset()
done = False
cumulative_reward = 0
frames = []
while not done:
frames.append(frame)
with torch.no_grad():
action = model(
memory.get_state(frame)).max(1)[1][0].cpu() # .numpy()
next_frame, reward, done, info = env.step(action)
cumulative_reward += reward
memory.add_experience(frame, action, reward, next_frame, done)
frame = next_frame
save_path = Path(
save_path.parent,
save_path.stem + '_r' + str(int(cumulative_reward)) + save_path.suffix)
# save_path = save_path.name + '_r' + cumulative_reward + save_path.suffix
make_gif(frames, save_path)
return cumulative_reward
def play(self, sess, coord, saver):
episode_count = sess.run(self.global_episode)
total_steps = 0
if not FLAGS.train:
test_episode_count = 0
print("Starting agent " + str(self.thread_id))
with sess.as_default(), sess.graph.as_default():
while not coord.should_stop():
if FLAGS.train and episode_count > FLAGS.max_nb_episodes_train:
return 0
sess.run(self.update_local_vars)
episode_buffer = []
# if not FLAGS.train:
# print("Episode {}".format(test_episode_count))
episode_regret = 0
episode_suboptimal_arm = 0
episode_values = []
episode_frames = []
episode_reward = [0 for _ in range(FLAGS.nb_actions)]
episode_step_count = 0
d = False
if FLAGS.meta:
r = 0
a = 0
t = 0
if not FLAGS.resume and FLAGS.train:
self.env.reset()
else:
#print(test_episode_count)
self.env.set(self.settings["envs"][test_episode_count])
rnn_state = self.local_AC.state_init
while not d:
if FLAGS.meta:
feed_dict = {
self.local_AC.prev_rewards: [[r]],
self.local_AC.timestep: [[t]],
self.local_AC.prev_actions: [a],
self.local_AC.state_in[0]: rnn_state[0],
self.local_AC.state_in[1]: rnn_state[1]
}
else:
feed_dict = {
self.local_AC.timestep: [[t]],
self.local_AC.prev_actions: [a],
self.local_AC.state_in[0]: rnn_state[0],
self.local_AC.state_in[1]: rnn_state[1]
}
pi, v, rnn_state_new = sess.run([
self.local_AC.policy, self.local_AC.value,
self.local_AC.state_out
],
feed_dict=feed_dict)
a = np.random.choice(pi[0])
a = np.argmax(pi == a)
rnn_state = rnn_state_new
r, d, t = self.env.pull_arm(a)
# if not FLAGS.train:
episode_regret += self.env.get_timestep_regret(a)
optimal_action = self.env.get_optimal_arm()
if optimal_action != a:
episode_suboptimal_arm += 1
episode_buffer.append([a, r, t, d, v[0, 0]])
episode_values.append(v[0, 0])
if not FLAGS.game == '11arms':
episode_frames.append(
set_image_bandit(episode_reward,
self.env.get_bandit(), a, t))
else:
episode_frames.append(
set_image_bandit_11_arms(
episode_reward, self.env.get_optimal_arm(), a,
t))
episode_reward[a] += r
total_steps += 1
episode_step_count += 1
self.episode_rewards.append(np.sum(episode_reward))
self.episodes_suboptimal_arms.append(episode_suboptimal_arm)
self.episode_regrets.append(episode_regret)
# if not FLAGS.train:
# print("Episode total reward was: {} vs optimal reward {}".format(np.sum(episode_reward),
# episode_rewards_for_optimal_arm))
# print("Regret is {}".format(max(episode_rewards_for_optimal_arm - np.sum(episode_reward), 0)))
# print("Suboptimal arms in the episode: {}".format(episode_suboptimal_arm))
self.episode_lengths.append(episode_step_count)
self.episode_mean_values.append(np.mean(episode_values))
if len(episode_buffer) != 0 and FLAGS.train == True:
if episode_count % FLAGS.summary_interval == 0 and episode_count != 0:
l, v_l, p_l, e_l, g_n, v_n, ms = self.train(
episode_buffer, sess, 0.0, self.settings, True)
else:
self.train(episode_buffer, sess, 0.0, self.settings,
False)
if not FLAGS.train and test_episode_count == FLAGS.nb_test_episodes - 1:
# episode_regret = [max(o - r, 0) for (o, r) in
# zip(self.episode_optimal_rewards, self.episode_rewards)]
mean_regret = np.mean(self.episode_regrets)
mean_nb_suboptimal_arms = np.mean(
self.episodes_suboptimal_arms)
if self.settings["mode"] == "val":
with open(FLAGS.results_val_file, "a+") as f:
f.write(
"Model: game={} lr={} gamma={} mean_regret={} mean_nb_subopt_arms={}\n"
.format(self.settings["game"],
self.settings["lr"],
self.settings["gamma"], mean_regret,
mean_nb_suboptimal_arms))
elif self.settings["mode"] == "test":
self.images = np.array(episode_frames)
#make_gif(self.images,
# self.settings["frames_dir"] + '/image' + str(test_episode_count) + '.gif',
# duration=len(self.images) * 0.1, true_image=True)
with open(FLAGS.results_test_file, "a+") as f:
f.write(
"Model: game={} lr={} gamma={} mean_regret={} mean_nb_subopt_arms={}\n"
.format(self.settings["game"],
self.settings["lr"],
self.settings["gamma"], mean_regret,
mean_nb_suboptimal_arms))
else:
with open(FLAGS.results_eval_file, "a+") as f:
f.write("{} ".format(mean_regret))
print("Mean regret for the model is {}".format(
mean_regret))
print(
"Regret in terms of suboptimal arms is {}".format(
mean_nb_suboptimal_arms))
return 1
if not FLAGS.train:
self.images = np.array(episode_frames)
make_gif(self.images,
FLAGS.frames_test_dir + '/image' +
str(episode_count) + '.gif',
duration=len(self.images) * 0.1,
true_image=True)
if FLAGS.train == True and episode_count % FLAGS.checkpoint_interval == 0 and episode_count != 0:
saver.save(sess,
self.model_path + '/model-' +
str(episode_count) + '.cptk',
global_step=self.global_episode)
print("Saved Model at {}".format(self.model_path +
'/model-' +
str(episode_count) +
'.cptk'))
if FLAGS.train and episode_count % FLAGS.summary_interval == 0 and episode_count != 0:
# if episode_count % FLAGS.frames_interval == 0 and self.name == 'agent_0':
# self.images = np.array(episode_frames)
# make_gif(self.images, self.settings.frames_dir + '/image' + str(episode_count) + '.gif',
# duration=len(self.images) * 0.1, true_image=True)
mean_reward = np.mean(
self.episode_rewards[-FLAGS.summary_interval:])
mean_length = np.mean(
self.episode_lengths[-FLAGS.summary_interval:])
mean_value = np.mean(
self.episode_mean_values[-FLAGS.summary_interval:])
episode_regret = [
max(o - r, 0) for (o, r) in zip(
self.
episode_optimal_rewards[-FLAGS.summary_interval:],
self.episode_rewards[-50:])
]
mean_regret = np.mean(episode_regret)
mean_nb_suboptimal_arms = np.mean(
self.episodes_suboptimal_arms[-FLAGS.summary_interval:]
)
self.summary.value.add(tag='Perf/Reward',
simple_value=float(mean_reward))
self.summary.value.add(tag='Perf/Length',
simple_value=float(mean_length))
self.summary.value.add(tag='Perf/Value',
simple_value=float(mean_value))
self.summary.value.add(tag='Mean Regret',
simple_value=float(mean_regret))
self.summary.value.add(
tag='Mean NSuboptArms',
simple_value=float(mean_nb_suboptimal_arms))
self.summary.value.add(tag='Losses/Total Loss',
simple_value=float(l))
self.summary.value.add(tag='Losses/Value Loss',
simple_value=float(v_l))
self.summary.value.add(tag='Losses/Policy Loss',
simple_value=float(p_l))
self.summary.value.add(tag='Losses/Entropy',
simple_value=float(e_l))
self.summary.value.add(tag='Losses/Grad Norm',
simple_value=float(g_n))
self.summary.value.add(tag='Losses/Var Norm',
simple_value=float(v_n))
summaries = tf.Summary().FromString(ms)
sub_summaries_dict = {}
for value in summaries.value:
value_field = value.WhichOneof('value')
value_ifo = sub_summaries_dict.setdefault(
value.tag, {
'value_field': None,
'values': []
})
if not value_ifo['value_field']:
value_ifo['value_field'] = value_field
else:
assert value_ifo['value_field'] == value_field
value_ifo['values'].append(getattr(value, value_field))
for name, value_ifo in sub_summaries_dict.items():
summary_value = self.summary.value.add()
summary_value.tag = name
if value_ifo['value_field'] == 'histo':
values = value_ifo['values']
summary_value.histo.min = min(
[x.min for x in values])
summary_value.histo.max = max(
[x.max for x in values])
summary_value.histo.num = sum(
[x.num for x in values])
summary_value.histo.sum = sum(
[x.sum for x in values])
summary_value.histo.sum_squares = sum(
[x.sum_squares for x in values])
for lim in values[0].bucket_limit:
summary_value.histo.bucket_limit.append(lim)
for bucket in values[0].bucket:
summary_value.histo.bucket.append(bucket)
else:
print(
'Warning: could not aggregate summary of type {}'
.format(value_ifo['value_field']))
self.summary_writer.add_summary(self.summary,
episode_count)
self.summary_writer.flush()
if self.name == 'agent_0':
sess.run(self.increment_global_episode)
if not FLAGS.train:
# if self.settings["mode"] == "test":
# self.images = np.array(episode_frames)
# make_gif(self.images, self.settings["frames_dir"] + '/image' + str(test_episode_count) + '.gif',
# duration=len(self.images) * 0.1, true_image=True)
test_episode_count += 1
episode_count += 1
W = W['emb.weight'].data.cpu().numpy()
netG = modelvideo.netG_new(nz).cuda()
if torch.cuda.device_count() > 1:
parallel = True
print("Using", torch.cuda.device_count(), "GPUs!")
netG = nn.DataParallel(netG)
state_dict = torch.load('runs%d/nets_%s/netG_glo.pth' % (counter, rn))
netG.load_state_dict(state_dict) # load the weights for generator (GLO)
if parallel:
netG = netG.module
# d is the dimension of noise vector (e)
d = 16
nepoch = 50
icpt = icp.ICPTrainer(W, d)
icpt.train_icp(nepoch)
torch.save(icpt.icp.netT.state_dict(), 'runs%d/nets_%s/netT_nag.pth' % (counter, rn)) #saves the param of the netT
#Prediction
z = icpt.icp.netT(torch.randn(100, d).cuda())
print("z shape", z.shape)
video = netG(z)
print("video shape is", video.shape)
# utils.make_gif(utils.denorm(video), 'runs3/ims_%s/sample' % (rn), 1)
utils.make_gif(video, 'runs%d/ims_%s/sample' % (counter, rn), 5)
#!/usr/bin/env python
import argparse
from glob import glob
from utils import make_gif
parser = argparse.ArgumentParser()
parser.add_argument("--model_name",
type=str,
default='ptb_2018-08-22_11-44-47')
parser.add_argument("--max_frame", type=int, default=5)
parser.add_argument("--output", type=str, default="sampe.gif")
parser.add_argument("--title", type=str, default="")
if __name__ == "__main__":
args = parser.parse_args()
paths = glob(f"./logs/{args.model_name}/networks/*.png")
make_gif(paths,
args.output,
max_frame=args.max_frame,
prefix=f"{args.title}\n" if args.title else "")
'''
n_updates += 1
counter += 1
if (batch_index + 1) % 10 == 0:
save_img(
flow[0].data.cpu(), DIR_TO_SAVE + 'original_flow_%s_%s_a.jpg' %
(current_epoch, batch_index))
save_img(
gen_flow[0].data.cpu(), DIR_TO_SAVE + 'fake_flow_%s_%s_a.jpg' %
(current_epoch, batch_index))
make_gif(
denorm(videos.data.cpu()[0]), DIR_TO_SAVE +
'original_gifs_%s_%s_b.gif' % (current_epoch, batch_index))
save_img(
denorm(first_frame[0].data.cpu()), DIR_TO_SAVE +
'fake_gifs_%s_%s_a.jpg' % (current_epoch, batch_index))
make_gif(
denorm(fake_videos.data.cpu()[0]), DIR_TO_SAVE +
'fake_gifs_%s_%s_b.gif' % (current_epoch, batch_index))
make_gif(
denorm(mask.data.cpu()[0]), DIR_TO_SAVE + 'mask__%s_%s_b.gif' %
(current_epoch, batch_index))
text_logger.info('Gifs saved at epoch: %d, batch_index: %d' %
(current_epoch, batch_index))
if current_epoch % 100 == 0:
torch.save(generator.state_dict(), './generator1.pkl')
# avg loss values for plot
D_avg_losses.append(D_avg_loss)
G_avg_losses.append(G_avg_loss)
# Show result for test image
gen_image = G(Variable(test_input)) #.cuda()))
gen_image = gen_image.cpu().data
utils.plot_test_result(test_input,
test_target,
gen_image,
epoch,
save=True,
save_dir=save_dir)
np.savetxt("D_losses.csv", D_losses, delimiter=",")
np.savetxt("G_losses.csv", G_losses, delimiter=",")
# Plot average losses
utils.plot_loss(D_losses,
G_losses,
params.num_epochs,
save=True,
save_dir=save_dir)
# Make gif
utils.make_gif(params.dataset, params.num_epochs, save_dir=save_dir)
# Save trained parameters of model
torch.save(G.state_dict(), model_dir + 'generator_param.pkl')
torch.save(D.state_dict(), model_dir + 'discriminator_param.pkl')
col=10,
figsize=(50, 50),
dataformat='CHW')
fig.savefig(PICTURE_DIR + "%s_%d.png" %
(str(count).zfill(3), index))
plt.close(fig)
logger.info(
"Epoch({}/{}) D_loss : {}, G_loss : {}, Q_loss : {}".
format(epoch + 1, train_epochs, D_loss, G_loss, Q_loss))
# Save model
saver = tf.train.Saver(max_to_keep=10)
saver.save(self.sess,
os.path.join(SAVE_DIR, 'model'),
global_step=epoch + 1)
logger.info("Model save in %s" % SAVE_DIR)
if __name__ == '__main__':
create_dir(SAVE_DIR)
create_dir(PICTURE_DIR)
infogan = InfoGAN()
infogan.initialize()
infogan.train(100000)
for index in range(2):
images_path = glob.glob(os.path.join(PICTURE_DIR, '*_%d.png' % index))
gif_path = os.path.join(PICTURE_DIR, '%d.gif' % index)
make_gif(sorted(images_path), gif_path)
def play(self, sess, coord, saver):
episode_count = sess.run(self.global_episode)
total_steps = 0
if not FLAGS.train:
test_episode_count = 0
print("Starting agent " + str(self.thread_id))
with sess.as_default(), sess.graph.as_default():
while not coord.should_stop():
if FLAGS.train and episode_count > FLAGS.max_nb_episodes_train:
return 0
sess.run(self.update_local_vars)
episode_buffer = []
# if not FLAGS.train:
# print("Episode {}".format(test_episode_count))
episode_regret = 0
episode_suboptimal_arm = 0
episode_values = []
episode_frames = []
episode_reward = [0 for _ in range(FLAGS.nb_actions)]
episode_step_count = 0
d = False
if FLAGS.meta:
r = 0
a = 0
t = 0
if not FLAGS.resume and FLAGS.train:
self.env.reset()
else:
#print(test_episode_count)
self.env.set(self.settings["envs"][test_episode_count])
rnn_state = self.local_AC.state_init
while not d:
if FLAGS.meta:
feed_dict = {
self.local_AC.prev_rewards: [[r]],
self.local_AC.timestep: [[t]],
self.local_AC.prev_actions: [a],
self.local_AC.state_in[0]: rnn_state[0],
self.local_AC.state_in[1]: rnn_state[1]}
else:
feed_dict = {
self.local_AC.timestep: [[t]],
self.local_AC.prev_actions: [a],
self.local_AC.state_in[0]: rnn_state[0],
self.local_AC.state_in[1]: rnn_state[1]}
pi, v, rnn_state_new = sess.run(
[self.local_AC.policy, self.local_AC.value, self.local_AC.state_out], feed_dict=feed_dict)
a = np.random.choice(pi[0])
a = np.argmax(pi == a)
rnn_state = rnn_state_new
r, d, t = self.env.pull_arm(a)
# if not FLAGS.train:
episode_regret += self.env.get_timestep_regret(a)
optimal_action = self.env.get_optimal_arm()
if optimal_action != a:
episode_suboptimal_arm += 1
episode_buffer.append([a, r, t, d, v[0, 0]])
episode_values.append(v[0, 0])
if not FLAGS.game == '11arms':
episode_frames.append(set_image_bandit(episode_reward, self.env.get_bandit(), a, t))
else:
episode_frames.append(
set_image_bandit_11_arms(episode_reward, self.env.get_optimal_arm(), a, t))
episode_reward[a] += r
total_steps += 1
episode_step_count += 1
self.episode_rewards.append(np.sum(episode_reward))
self.episodes_suboptimal_arms.append(episode_suboptimal_arm)
self.episode_regrets.append(episode_regret)
# if not FLAGS.train:
# print("Episode total reward was: {} vs optimal reward {}".format(np.sum(episode_reward),
# episode_rewards_for_optimal_arm))
# print("Regret is {}".format(max(episode_rewards_for_optimal_arm - np.sum(episode_reward), 0)))
# print("Suboptimal arms in the episode: {}".format(episode_suboptimal_arm))
self.episode_lengths.append(episode_step_count)
self.episode_mean_values.append(np.mean(episode_values))
if len(episode_buffer) != 0 and FLAGS.train == True:
if episode_count % FLAGS.summary_interval == 0 and episode_count != 0:
l, v_l, p_l, e_l, g_n, v_n, ms = self.train(episode_buffer, sess, 0.0, self.settings, True)
else:
self.train(episode_buffer, sess, 0.0, self.settings, False)
if not FLAGS.train and test_episode_count == FLAGS.nb_test_episodes - 1:
# episode_regret = [max(o - r, 0) for (o, r) in
# zip(self.episode_optimal_rewards, self.episode_rewards)]
mean_regret = np.mean(self.episode_regrets)
mean_nb_suboptimal_arms = np.mean(self.episodes_suboptimal_arms)
if self.settings["mode"] == "val":
with open(FLAGS.results_val_file, "a+") as f:
f.write("Model: game={} lr={} gamma={} mean_regret={} mean_nb_subopt_arms={}\n".format(
self.settings["game"],
self.settings["lr"],
self.settings["gamma"],
mean_regret,
mean_nb_suboptimal_arms))
elif self.settings["mode"] == "test":
self.images = np.array(episode_frames)
#make_gif(self.images,
# self.settings["frames_dir"] + '/image' + str(test_episode_count) + '.gif',
# duration=len(self.images) * 0.1, true_image=True)
with open(FLAGS.results_test_file, "a+") as f:
f.write("Model: game={} lr={} gamma={} mean_regret={} mean_nb_subopt_arms={}\n".format(
self.settings["game"],
self.settings["lr"],
self.settings["gamma"],
mean_regret,
mean_nb_suboptimal_arms))
else:
with open(FLAGS.results_eval_file, "a+") as f:
f.write("{} ".format(mean_regret))
print("Mean regret for the model is {}".format(mean_regret))
print("Regret in terms of suboptimal arms is {}".format(mean_nb_suboptimal_arms))
return 1
if not FLAGS.train:
self.images = np.array(episode_frames)
make_gif(self.images, FLAGS.frames_test_dir + '/image' + str(episode_count) + '.gif',
duration=len(self.images) * 0.1, true_image=True)
if FLAGS.train == True and episode_count % FLAGS.checkpoint_interval == 0 and episode_count != 0:
saver.save(sess, self.model_path + '/model-' + str(episode_count) + '.cptk',
global_step=self.global_episode)
print("Saved Model at {}".format(self.model_path + '/model-' + str(episode_count) + '.cptk'))
if FLAGS.train and episode_count % FLAGS.summary_interval == 0 and episode_count != 0:
# if episode_count % FLAGS.frames_interval == 0 and self.name == 'agent_0':
# self.images = np.array(episode_frames)
# make_gif(self.images, self.settings.frames_dir + '/image' + str(episode_count) + '.gif',
# duration=len(self.images) * 0.1, true_image=True)
mean_reward = np.mean(self.episode_rewards[-FLAGS.summary_interval:])
mean_length = np.mean(self.episode_lengths[-FLAGS.summary_interval:])
mean_value = np.mean(self.episode_mean_values[-FLAGS.summary_interval:])
episode_regret = [max(o - r, 0) for (o, r) in
zip(self.episode_optimal_rewards[-FLAGS.summary_interval:],
self.episode_rewards[-50:])]
mean_regret = np.mean(episode_regret)
mean_nb_suboptimal_arms = np.mean(self.episodes_suboptimal_arms[-FLAGS.summary_interval:])
self.summary.value.add(tag='Perf/Reward', simple_value=float(mean_reward))
self.summary.value.add(tag='Perf/Length', simple_value=float(mean_length))
self.summary.value.add(tag='Perf/Value', simple_value=float(mean_value))
self.summary.value.add(tag='Mean Regret', simple_value=float(mean_regret))
self.summary.value.add(tag='Mean NSuboptArms', simple_value=float(mean_nb_suboptimal_arms))
self.summary.value.add(tag='Losses/Total Loss', simple_value=float(l))
self.summary.value.add(tag='Losses/Value Loss', simple_value=float(v_l))
self.summary.value.add(tag='Losses/Policy Loss', simple_value=float(p_l))
self.summary.value.add(tag='Losses/Entropy', simple_value=float(e_l))
self.summary.value.add(tag='Losses/Grad Norm', simple_value=float(g_n))
self.summary.value.add(tag='Losses/Var Norm', simple_value=float(v_n))
summaries = tf.Summary().FromString(ms)
sub_summaries_dict = {}
for value in summaries.value:
value_field = value.WhichOneof('value')
value_ifo = sub_summaries_dict.setdefault(value.tag,
{'value_field': None, 'values': []})
if not value_ifo['value_field']:
value_ifo['value_field'] = value_field
else:
assert value_ifo['value_field'] == value_field
value_ifo['values'].append(getattr(value, value_field))
for name, value_ifo in sub_summaries_dict.items():
summary_value = self.summary.value.add()
summary_value.tag = name
if value_ifo['value_field'] == 'histo':
values = value_ifo['values']
summary_value.histo.min = min([x.min for x in values])
summary_value.histo.max = max([x.max for x in values])
summary_value.histo.num = sum([x.num for x in values])
summary_value.histo.sum = sum([x.sum for x in values])
summary_value.histo.sum_squares = sum([x.sum_squares for x in values])
for lim in values[0].bucket_limit:
summary_value.histo.bucket_limit.append(lim)
for bucket in values[0].bucket:
summary_value.histo.bucket.append(bucket)
else:
print(
'Warning: could not aggregate summary of type {}'.format(value_ifo['value_field']))
self.summary_writer.add_summary(self.summary, episode_count)
self.summary_writer.flush()
if self.name == 'agent_0':
sess.run(self.increment_global_episode)
if not FLAGS.train:
# if self.settings["mode"] == "test":
# self.images = np.array(episode_frames)
# make_gif(self.images, self.settings["frames_dir"] + '/image' + str(test_episode_count) + '.gif',
# duration=len(self.images) * 0.1, true_image=True)
test_episode_count += 1
episode_count += 1
PATH = sys.argv[1] if len(sys.argv) == 2 else str(Path(__file__).parent.absolute())[:-3]+'imgs/'
NAME = 'huge'
fst = Param(
CLUSTERS = 12**2,
CLUST_SIDE_LEN = 7,
DAYS = 100,
INFECTION_TIME = 20,
DEATH_RATE = 0.4,
INFECT_RATE = 0.3,
MIGRATIONS_PER_DAY = 100,
FEAR_RATE = 0.4,
HEALTHCARE_CAPACITY = 2500
)
start = time()
cumulative, active, healed, dead, migrations, real_mig = run_simulation(fst, PATH)
print('time elapsed:', time() - start, flush=True)
make_gif(PATH, NAME)
# plt.plot(range(len(migrations)), migrations, label='migrations', color='blue')
plt.plot(range(len(real_mig)), real_mig, label='migrations', color='blue', ls=':')
make_plot(fst, cumulative, active, healed, dead, label='', color='red')
plt.savefig(PATH+NAME+'.png')
print("The visualization of the simulation is available at './imgs/'", flush=True)
[0.65, 0.65, 2.5],
[-0.65, 0.65, 2.5],
])
# # PRIMEIRA QUESTÃO
q1(cube_points, parallel_points, pyramid_points, tronco_points)
# SEGUNDA QUESTÃO
cube_points, parallel_points, pyramid_points, tronco_points = q2(
cube_points, parallel_points, pyramid_points, tronco_points)
# TERCEIRA QUESTÃO
c_points, p_points = q3(copy(cube_points), copy(pyramid_points),
copy(parallel_points), copy(tronco_points),
np.array([3, -3, 3]), True)
# QUARTA QUESTÃO
q4(c_points, p_points, True)
# OPCIONAL
eyes = [np.array([i, -3, 3]) for i in np.arange(-2, 8, .5)]
for i, eye in enumerate(eyes):
# TERCEIRA QUESTÃO
c_points, p_points = q3(copy(cube_points), copy(pyramid_points),
copy(parallel_points), copy(tronco_points),
eye)
# QUARTA QUESTÃO
q4(c_points, p_points, show=False)
make_gif()
def train_agent_model_free(agent, env, params):
update_timestep = params['update_every_n_steps']
seed = params['seed']
log_interval = 1000
gif_interval = 500000
n_random_actions = params['n_random_actions']
n_evals = params['n_evals']
n_collect_steps = params['n_collect_steps']
use_statefilter = params['obs_filter']
save_model = params['save_model']
assert n_collect_steps > agent.batchsize, "We must initially collect as many steps as the batch size!"
avg_length = 0
time_step = 0
cumulative_timestep = 0
cumulative_log_timestep = 0
n_updates = 0
i_episode = 0
log_episode = 0
samples_number = 0
episode_rewards = []
episode_steps = []
if use_statefilter:
state_filter = MeanStdevFilter(env.env.observation_space.shape[0])
else:
state_filter = None
random.seed(seed)
torch.manual_seed(seed)
np.random.seed(seed)
env.seed(seed)
env.action_space.np_random.seed(seed)
max_steps = env.spec.max_episode_steps
writer = SummaryWriter()
while samples_number < 3e7:
time_step = 0
episode_reward = 0
i_episode += 1
log_episode += 1
state = env.reset()
if state_filter:
state_filter.update(state)
done = False
while (not done):
cumulative_log_timestep += 1
cumulative_timestep += 1
time_step += 1
samples_number += 1
if samples_number < n_random_actions:
action = env.action_space.sample()
else:
action = agent.get_action(state, state_filter=state_filter)
nextstate, reward, done, _ = env.step(action)
# if we hit the time-limit, it's not a 'real' done; we don't want to assign low value to those states
real_done = False if time_step == max_steps else done
agent.replay_pool.push(
Transition(state, action, reward, nextstate, real_done))
state = nextstate
if state_filter:
state_filter.update(state)
episode_reward += reward
# update if it's time
if cumulative_timestep % update_timestep == 0 and cumulative_timestep > n_collect_steps:
q1_loss, q2_loss, pi_loss, a_loss = agent.optimize(
update_timestep, state_filter=state_filter)
n_updates += 1
# logging
if cumulative_timestep % log_interval == 0 and cumulative_timestep > n_collect_steps:
writer.add_scalar('Loss/Q-func_1', q1_loss, n_updates)
writer.add_scalar('Loss/Q-func_2', q2_loss, n_updates)
writer.add_scalar('Loss/policy', pi_loss, n_updates)
writer.add_scalar('Loss/alpha', a_loss, n_updates)
writer.add_scalar('Values/alpha',
np.exp(agent.log_alpha.item()), n_updates)
avg_length = np.mean(episode_steps)
running_reward = np.mean(episode_rewards)
eval_reward = evaluate_agent(env,
agent,
state_filter,
n_starts=n_evals)
writer.add_scalar('Reward/Train', running_reward,
cumulative_timestep)
writer.add_scalar('Reward/Test', eval_reward,
cumulative_timestep)
print(
'Episode {} \t Samples {} \t Avg length: {} \t Test reward: {} \t Train reward: {} \t Number of Policy Updates: {}'
.format(i_episode, samples_number, avg_length, eval_reward,
running_reward, n_updates))
episode_steps = []
episode_rewards = []
if cumulative_timestep % gif_interval == 0:
make_gif(agent, env, cumulative_timestep, state_filter)
if save_model:
make_checkpoint(agent, cumulative_timestep, params['env'])
episode_steps.append(time_step)
episode_rewards.append(episode_reward)
def build_weixin_message(weixin):
items = []
index = 0
message_title = ''
if weixin.cover.name != u'' and weixin.title != u'' and weixin.recommended_reason != u'':
message_title = weixin.title
items.append(
WeixinMessageItem(image=weixin.cover.path,
title=weixin.title,
digest=weixin.title,
content=weixin.recommended_reason))
index += 1
for news in weixin.news.all():
title = u'游戏情报站 - %s' % news.brief_comment
if index == 0:
message_title = title
items.append(
WeixinMessageItem(
image=make_gif(news),
title=title,
digest=title,
content=news.recommended_reason +
u'<br><br><font color="gray">点击“阅读原文”查看更多</font>',
sourceurl=u'http://cow.bestgames7.com/d/%s/' % news.id))
index += 1
for game in weixin.games.all():
title = u'%s - %s' % (game.name, game.brief_comment)
if index == 0:
message_title = title
if index > 0:
items.append(
WeixinMessageItem(
image=make_gif(game),
title=title,
digest=title,
content=_normalize_content(game),
sourceurl=u'http://cow.bestgames7.com/d/%s/' % game.id))
else:
items.append(
WeixinMessageItem(
image=make_gif(game),
title=title,
digest=title,
content=_normalize_content(game),
sourceurl=u'http://cow.bestgames7.com/d/%s/' % game.id))
index += 1
for player in weixin.players.all():
title = u'我是玩家 - %s' % player.brief_comment
if index == 0:
message_title = title
items.append(
WeixinMessageItem(image=player.image_url.path,
title=title,
digest=title,
content=player.recommended_reason))
index += 1
for puzzle in weixin.puzzles.all():
title = u'趣味答题 - %s' % puzzle.title
if index == 0:
message_title = title
content = puzzle.description + '<br>'
content += 'A.' + puzzle.option1 + '<br>'
content += 'B.' + puzzle.option2 + '<br>'
content += 'C.' + puzzle.option3 + '<br>'
content += 'D.' + puzzle.option4 + '<br></br>'
content += u'<font color="gray">回复"答题"，参与答题得积分换礼品的活动吧</font><br>'
items.append(
WeixinMessageItem(image=puzzle.image_url.path,
title=title,
digest=title,
content=content))
index += 1
return WeixinMessage(weixin.id, message_title, items)
def train(self):
# Seed
np.random.seed(self.manual_seed)
random.seed(self.manual_seed)
torch.manual_seed(self.manual_seed)
# For fast training
cudnn.benchmark = True
# For BatchNorm
self.G.train()
self.D.train()
# Fixed noise for sampling from G
fixed_noise = torch.randn(self.batch_size,
self.z_dim,
device=self.device)
if self.num_of_classes < self.batch_size:
fixed_labels = torch.from_numpy(
np.tile(np.arange(self.num_of_classes),
self.batch_size // self.num_of_classes +
1)[:self.batch_size]).to(self.device)
else:
fixed_labels = torch.from_numpy(np.arange(self.batch_size)).to(
self.device)
# For gan loss
label = torch.full((self.batch_size, ), 1, device=self.device)
ones = torch.full((self.batch_size, ), 1, device=self.device)
# Losses file
log_file_name = os.path.join(self.save_path, 'log.txt')
log_file = open(log_file_name, "wt")
# Init
start_time = time.time()
G_losses = []
D_losses_real = []
D_losses_fake = []
D_losses = []
D_xs = []
D_Gz_trainDs = []
D_Gz_trainGs = []
# Instance noise - make random noise mean (0) and std for injecting
inst_noise_mean = torch.full(
(self.batch_size, 3, self.imsize, self.imsize),
0,
device=self.device)
inst_noise_std = torch.full(
(self.batch_size, 3, self.imsize, self.imsize),
self.inst_noise_sigma,
device=self.device)
# Start training
for self.step in range(self.start, self.total_step):
# Instance noise std is linearly annealed from self.inst_noise_sigma to 0 thru self.inst_noise_sigma_iters
inst_noise_sigma_curr = 0 if self.step > self.inst_noise_sigma_iters else (
1 - self.step /
self.inst_noise_sigma_iters) * self.inst_noise_sigma
inst_noise_std.fill_(inst_noise_sigma_curr)
# ================== TRAIN D ================== #
for _ in range(self.d_steps_per_iter):
# Zero grad
self.reset_grad()
# TRAIN with REAL
# Get real images & real labels
real_images, real_labels = self.get_real_samples()
# Get D output for real images & real labels
inst_noise = torch.normal(mean=inst_noise_mean,
std=inst_noise_std).to(self.device)
d_out_real = self.D(real_images + inst_noise, real_labels)
# Compute D loss with real images & real labels
if self.adv_loss == 'hinge':
d_loss_real = torch.nn.ReLU()(ones - d_out_real).mean()
elif self.adv_loss == 'wgan_gp':
d_loss_real = -d_out_real.mean()
else:
label.fill_(1)
d_loss_real = self.criterion(d_out_real, label)
# Backward
d_loss_real.backward()
# TRAIN with FAKE
# Create random noise
z = torch.randn(self.batch_size,
self.z_dim,
device=self.device)
# Generate fake images for same real labels
fake_images = self.G(z, real_labels)
# Get D output for fake images & same real labels
inst_noise = torch.normal(mean=inst_noise_mean,
std=inst_noise_std).to(self.device)
d_out_fake = self.D(fake_images.detach() + inst_noise,
real_labels)
# Compute D loss with fake images & real labels
if self.adv_loss == 'hinge':
d_loss_fake = torch.nn.ReLU()(ones + d_out_fake).mean()
elif self.adv_loss == 'dcgan':
label.fill_(0)
d_loss_fake = self.criterion(d_out_fake, label)
else:
d_loss_fake = d_out_fake.mean()
# Backward
d_loss_fake.backward()
# If WGAN_GP, compute GP and add to D loss
if self.adv_loss == 'wgan_gp':
d_loss_gp = self.lambda_gp * self.compute_gradient_penalty(
real_images, real_labels, fake_images.detach())
d_loss_gp.backward()
# Optimize
self.D_optimizer.step()
# ================== TRAIN G ================== #
for _ in range(self.g_steps_per_iter):
# Zero grad
self.reset_grad()
# Get real images & real labels (only need real labels)
real_images, real_labels = self.get_real_samples()
# Create random noise
z = torch.randn(self.batch_size, self.z_dim).to(self.device)
# Generate fake images for same real labels
fake_images = self.G(z, real_labels)
# Get D output for fake images & same real labels
inst_noise = torch.normal(mean=inst_noise_mean,
std=inst_noise_std).to(self.device)
g_out_fake = self.D(fake_images + inst_noise, real_labels)
# Compute G loss with fake images & real labels
if self.adv_loss == 'dcgan':
label.fill_(1)
g_loss = self.criterion(g_out_fake, label)
else:
g_loss = -g_out_fake.mean()
# Backward + Optimize
g_loss.backward()
self.G_optimizer.step()
# Print out log info
if self.step % self.log_step == 0:
G_losses.append(g_loss.mean().item())
D_losses_real.append(d_loss_real.mean().item())
D_losses_fake.append(d_loss_fake.mean().item())
D_loss = D_losses_real[-1] + D_losses_fake[-1]
if self.adv_loss == 'wgan_gp':
D_loss += d_loss_gp.mean().item()
D_losses.append(D_loss)
D_xs.append(d_out_real.mean().item())
D_Gz_trainDs.append(d_out_fake.mean().item())
D_Gz_trainGs.append(g_out_fake.mean().item())
curr_time = time.time()
curr_time_str = datetime.datetime.fromtimestamp(
curr_time).strftime('%Y-%m-%d %H:%M:%S')
elapsed = str(
datetime.timedelta(seconds=(curr_time - start_time)))
log = (
"[{}] : Elapsed [{}], Iter [{} / {}], G_loss: {:.4f}, D_loss: {:.4f}, D_loss_real: {:.4f}, D_loss_fake: {:.4f}, D(x): {:.4f}, D(G(z))_trainD: {:.4f}, D(G(z))_trainG: {:.4f}\n"
.format(curr_time_str, elapsed, self.step, self.total_step,
G_losses[-1], D_losses[-1], D_losses_real[-1],
D_losses_fake[-1], D_xs[-1], D_Gz_trainDs[-1],
D_Gz_trainGs[-1]))
print(log)
log_file.write(log)
log_file.flush()
utils.make_plots(G_losses, D_losses, D_losses_real,
D_losses_fake, D_xs, D_Gz_trainDs,
D_Gz_trainGs, self.log_step, self.save_path)
# Sample images
if self.step % self.sample_step == 0:
self.G.eval()
fake_images = self.G(fixed_noise, fixed_labels)
self.G.train()
sample_images = utils.denorm(
fake_images.detach()[:self.save_n_images])
# Save batch images
vutils.save_image(
sample_images,
os.path.join(self.sample_path,
'fake_{:05d}.png'.format(self.step)))
# Save gif
utils.make_gif(
sample_images[0].cpu().numpy().transpose(1, 2, 0) * 255,
self.step,
self.sample_path,
self.name,
max_frames_per_gif=self.max_frames_per_gif)
# Save model
if self.step % self.model_save_step == 0:
utils.save_ckpt(self)
disc_loss_fake.append(d_loss_f)
gen_loss = np.mean(np.asarray(gen_loss))
disc_loss_real = np.mean(np.asarray(disc_loss_real))
disc_loss_fake = np.mean(np.asarray(disc_loss_fake))
if (np.isnan(gen_loss) or np.isnan(disc_loss_real)
or np.isnan(disc_loss_fake)):
print("Something broke.")
break
manager.save()
generate_and_save_images(gen, ep + 1, seed, out_path)
print("Time for epoch:", time.time() - start)
print("Gen loss=", gen_loss)
print("Disc loss real=", disc_loss_real)
print("Disc loss fake=", disc_loss_fake)
if __name__ == "__main__":
args = parse_args()
epochs = args.epochs
batch_size = args.batch
ckpt_path = args.checkpoint
imgs_path = args.imgs
out_path = args.output
lr = args.lr
train(epochs, batch_size, ckpt_path, imgs_path, lr, out_path)
make_gif(out_path)
if (batch_index + 1) % 5 == 0:
text_logger.info("Epoch [%d/%d], Step[%d/%d], d_loss: %.4f, g_loss: %.4f, \
g_val_loss: %.4f, time: %4.4f" \
% (current_epoch, num_epoch, batch_index+1, num_batch, \
d_loss.data[0], g_loss.data[0], g_val_loss.data[0], time.time()-start_time))
counter += 1
if (batch_index + 1) % 100 == 0:
gen_out = generator(sample_input)
save_img(
sample_input.data.cpu(), DIR_TO_SAVE +
'fake_gifs_sample_%s_%s_a.jpg' % (current_epoch, batch_index))
make_gif(
denorm(gen_out.data.cpu()[0]), DIR_TO_SAVE +
'fake_gifs_sample__%s_%s_b.gif' % (current_epoch, batch_index))
save_img(
first_frame[0].data.cpu(), DIR_TO_SAVE +
'fake_gifs_%s_%s_a.jpg' % (current_epoch, batch_index))
make_gif(
denorm(fake_videos.data.cpu()[0]), DIR_TO_SAVE +
'fake_gifs_%s_%s_b.gif' % (current_epoch, batch_index))
text_logger.info('Gifs saved at epoch: %d, batch_index: %d' %
(current_epoch, batch_index))
if (batch_index + 1) % 1000 == 0:
torch.save(generator.state_dict(), './generator.pkl')
torch.save(discriminator.state_dict(), './discriminator.pkl')
def main(_):
pp.pprint(flags.FLAGS.__flags)
if not os.path.exists(FLAGS.checkpoint_dir):
os.makedirs(FLAGS.checkpoint_dir)
if not os.path.exists(FLAGS.sample_dir):
os.makedirs(FLAGS.sample_dir)
with tf.Session() as sess:
if FLAGS.dataset == 'mnist':
dcgan = DCGAN(sess, image_size=FLAGS.image_size, batch_size=FLAGS.batch_size, y_dim=10,
dataset_name=FLAGS.dataset, is_crop=FLAGS.is_crop, checkpoint_dir=FLAGS.checkpoint_dir)
else:
dcgan = DCGAN(sess, image_size=FLAGS.image_size, batch_size=FLAGS.batch_size,
dataset_name=FLAGS.dataset, is_crop=FLAGS.is_crop, checkpoint_dir=FLAGS.checkpoint_dir)
if FLAGS.is_train:
dcgan.train(FLAGS)
else:
dcgan.load(FLAGS.checkpoint_dir)
to_json("./web/js/layers.js", [dcgan.h0_w, dcgan.h0_b, dcgan.g_bn0],
[dcgan.h1_w, dcgan.h1_b, dcgan.g_bn1],
[dcgan.h2_w, dcgan.h2_b, dcgan.g_bn2],
[dcgan.h3_w, dcgan.h3_b, dcgan.g_bn3],
[dcgan.h4_w, dcgan.h4_b, None])
# Below is codes for visualization
OPTION = 2
if OPTION == 0:
z_sample = np.random.uniform(-0.5, 0.5, size=(FLAGS.batch_size, dcgan.z_dim))
samples = sess.run(dcgan.sampler, feed_dict={dcgan.z: z_sample})
save_images(samples, [8, 8], './samples/test_%s.png' % strftime("%Y-%m-%d %H:%M:%S", gmtime()))
elif OPTION == 1:
values = np.arange(0, 1, 1./FLAGS.batch_size)
for idx in xrange(100):
print(" [*] %d" % idx)
z_sample = np.zeros([FLAGS.batch_size, dcgan.z_dim])
for kdx, z in enumerate(z_sample):
z[idx] = values[kdx]
samples = sess.run(dcgan.sampler, feed_dict={dcgan.z: z_sample})
save_images(samples, [8, 8], './samples/test_arange_%s.png' % (idx))
elif OPTION == 2:
values = np.arange(0, 1, 1./FLAGS.batch_size)
for idx in [random.randint(0, 99) for _ in xrange(100)]:
print(" [*] %d" % idx)
z = np.random.uniform(-0.2, 0.2, size=(dcgan.z_dim))
z_sample = np.tile(z, (FLAGS.batch_size, 1))
#z_sample = np.zeros([FLAGS.batch_size, dcgan.z_dim])
for kdx, z in enumerate(z_sample):
z[idx] = values[kdx]
samples = sess.run(dcgan.sampler, feed_dict={dcgan.z: z_sample})
make_gif(samples, './samples/test_gif_%s.gif' % (idx))
elif OPTION == 3:
values = np.arange(0, 1, 1./FLAGS.batch_size)
for idx in xrange(100):
print(" [*] %d" % idx)
z_sample = np.zeros([FLAGS.batch_size, dcgan.z_dim])
for kdx, z in enumerate(z_sample):
z[idx] = values[kdx]
samples = sess.run(dcgan.sampler, feed_dict={dcgan.z: z_sample})
make_gif(samples, './samples/test_gif_%s.gif' % (idx))
elif OPTION == 4:
image_set = []
values = np.arange(0, 1, 1./FLAGS.batch_size)
for idx in xrange(100):
print(" [*] %d" % idx)
z_sample = np.zeros([FLAGS.batch_size, dcgan.z_dim])
for kdx, z in enumerate(z_sample): z[idx] = values[kdx]
image_set.append(sess.run(dcgan.sampler, feed_dict={dcgan.z: z_sample}))
make_gif(image_set[-1], './samples/test_gif_%s.gif' % (idx))
new_image_set = [merge(np.array([images[idx] for images in image_set]), [10, 10]) for idx in range(64) + range(63, -1, -1)]
make_gif(new_image_set, './samples/test_gif_merged.gif', duration=8)
elif OPTION == 5:
image_set = []
values = np.arange(0, 1, 1./FLAGS.batch_size)
z_idx = [[random.randint(0,99) for _ in xrange(5)] for _ in xrange(200)]
for idx in xrange(200):
print(" [*] %d" % idx)
#z_sample = np.zeros([FLAGS.batch_size, dcgan.z_dim])
z = np.random.uniform(-1e-1, 1e-1, size=(dcgan.z_dim))
z_sample = np.tile(z, (FLAGS.batch_size, 1))
for kdx, z in enumerate(z_sample):
for jdx in xrange(5):
z_sample[kdx][z_idx[idx][jdx]] = values[kdx]
image_set.append(sess.run(dcgan.sampler, feed_dict={dcgan.z: z_sample}))
make_gif(image_set[-1], './samples/test_gif_%s.gif' % (idx))
new_image_set = [merge(np.array([images[idx] for images in image_set]), [10, 20]) for idx in range(64) + range(63, -1, -1)]
make_gif(new_image_set, './samples/test_gif_random_merged.gif', duration=4)
elif OPTION == 6:
image_set = []
values = np.arange(0, 1, 1.0/FLAGS.batch_size).tolist()
z_idx = [[random.randint(0,99) for _ in xrange(10)] for _ in xrange(100)]
for idx in xrange(100):
print(" [*] %d" % idx)
z = np.random.uniform(-0.2, 0.2, size=(dcgan.z_dim))
z_sample = np.tile(z, (FLAGS.batch_size, 1))
for kdx, z in enumerate(z_sample):
for jdx in xrange(10):
z_sample[kdx][z_idx[idx][jdx]] = values[kdx]
image_set.append(sess.run(dcgan.sampler, feed_dict={dcgan.z: z_sample}))
save_images(image_set[-1], [8, 8], './samples/test_random_arange_%s.png' % (idx))
new_image_set = [merge(np.array([images[idx] for images in image_set]), [10, 10]) for idx in range(64) + range(63, -1, -1)]
make_gif(new_image_set, './samples/test_gif_merged_random.gif', duration=4)
elif OPTION == 7:
for _ in xrange(50):
z_idx = [[random.randint(0,99) for _ in xrange(10)] for _ in xrange(8)]
zs = []
for idx in xrange(8):
z = np.random.uniform(-0.2, 0.2, size=(dcgan.z_dim))
zs.append(np.tile(z, (8, 1)))
z_sample = np.concatenate(zs)
values = np.arange(0, 1, 1/8.)
for idx in xrange(FLAGS.batch_size):
for jdx in xrange(8):
z_sample[idx][z_idx[idx/8][jdx]] = values[idx%8]
samples = sess.run(dcgan.sampler, feed_dict={dcgan.z: z_sample})
save_images(samples, [8, 8], './samples/multiple_testt_%s.png' % strftime("%Y-%m-%d %H:%M:%S", gmtime()))
elif OPTION == 8:
counter = 0
for _ in xrange(50):
import scipy.misc
z_idx = [[random.randint(0,99) for _ in xrange(10)] for _ in xrange(8)]
zs = []
for idx in xrange(8):
z = np.random.uniform(-0.2, 0.2, size=(dcgan.z_dim))
zs.append(np.tile(z, (8, 1)))
z_sample = np.concatenate(zs)
values = np.arange(0, 1, 1/8.)
for idx in xrange(FLAGS.batch_size):
for jdx in xrange(8):
z_sample[idx][z_idx[idx/8][jdx]] = values[idx%8]
samples = sess.run(dcgan.sampler, feed_dict={dcgan.z: z_sample})
for sample in samples:
scipy.misc.imsave('./samples/turing/%s.png' % counter, sample)
counter += 1
else:
import scipy.misc
from glob import glob
samples = []
fnames = glob("/Users/carpedm20/Downloads/x/1/*.png")
fnames = sorted(fnames, key = lambda x: int(x.split("_")[1]) * 10000 + int(x.split('_')[2].split(".")[0]))
for f in fnames:
samples.append(scipy.misc.imread(f))
make_gif(samples, './samples/training.gif', duration=8, true_image=True)
