def make_batch_logs(self, out, y, loss):
out, y = to_np(out), to_np(y)
metrics = self.compute_metrics(out, y)
metrics['loss'] = to_np(loss).item()
logs = {'stage': 'batch',
'metrics': metrics,
'opt_state': self.get_optimizer_parameters()}
return logs
def manifold(self, epoch):
save_dir = os.path.join(self.root, self.result_dir, self.dataset, self.model_name)
self.load(epoch)
self.G.eval()
self.E.eval()
self.FC.eval()
color_vec = []
Z = []
color = [
'Greys', 'Purples', 'Blues', 'Greens', 'Oranges', 'Reds',
'YlOrBr', 'YlOrRd', 'OrRd', 'PuRd', 'RdPu', 'BuPu',
'GnBu', 'PuBu', 'YlGnBu', 'PuBuGn', 'BuGn', 'YlGn']
for iter, (X, label) in enumerate(self.valid_loader):
X = utils.to_var(X)
label = utils.to_var(label)
z_mu, z_sigma = self.E(self.FC(X))
X_reconstruc = self.G(z_mu)
Z += [x for x in utils.to_np(z_mu)]
color_vec += [x for x in utils.to_np(label)]
self.G.train()
self.E.train()
self.FC.train()
Z = np.array(Z)
cmap = plt.get_cmap('gnuplot')
cmap = plt.cm.jet
cmaplist = [cmap(i) for i in range(cmap.N)]
cmap = cmap.from_list('Custom cmap', cmaplist, cmap.N)
fig, ax = plt.subplots(1, 1, figsize=(6, 6))
colors = [plt.cm.jet(float(i + 1) / 10) for i in range(10)]
# import matplotlib.cm as cm
# colors = cm.rainbow(np.linspace(0, 2, 20))
for k in range(10):
X = []
Y = []
for i, z in enumerate(Z):
if color_vec[i] == k:
X.append(z[0])
Y.append(z[1])
marker = ["*", "^"]
ax.scatter(X, Y, c=colors[k], marker=marker[k % 2], cmap=cmap, label=str(k), s=20)
# ax.scatter(Z[:5000, 0], Z[:5000, 1], c=color_vec[:5000], label= color_vec[:5000], marker='.', cmap=cmap, )
plt.legend(loc='upper right', ncol=1, borderaxespad=0.)
plt.xlim(-4, 4)
plt.ylim(-4, 4)
plt.xticks(fontsize=20)
plt.yticks(fontsize=20)
fig.savefig(os.path.join(save_dir, 'Z_mu' + '_epoch%03d' % epoch + '.png'), transparent=True)
plt.close()
def act(self, obs, sample=False, propensity=False):
obs = torch.FloatTensor(obs).to(self.device)
obs = obs.unsqueeze(0)
dist = self.actor(obs)
action = dist.sample() if sample else dist.mean
prob = dist.log_prob(action).sum(dim=-1, keepdim=True).exp()
action = action.clamp(*self.action_range)
assert action.ndim == 2 and action.shape[0] == 1
if propensity:
return utils.to_np(action[0]), utils.to_np(prob[0])
else:
return utils.to_np(action[0])
def log_weights_distribution(self, named_params, steps_completed):
if named_params is None:
return
for tag, value in named_params:
tag = tag.replace('.', '/')
if any(substring in tag for substring in self.logged_params):
self.tblogger.histogram_summary(tag, to_np(value),
steps_completed)
if self.log_gradients:
self.tblogger.histogram_summary(tag + '/grad',
to_np(value.grad),
steps_completed)
self.tblogger.sync_to_file()
def train(epoch_idx, net, train_loader, lr, logger, n_class):
net.cuda()
net.train()
base_params = list(map(id, net.base_net.parameters()))
top_params = filter(lambda p: id(p) not in base_params, net.parameters())
optimizer = torch.optim.SGD([{
'params': top_params
}, {
'params': net.base_net.parameters(),
'lr': lr * 0.1
}],
lr=lr,
momentum=0.9,
weight_decay=0.00004)
criterion = nn.CrossEntropyLoss(ignore_index=-1)
len_batch = len(train_loader)
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.cuda(), target.cuda()
optimizer.zero_grad()
score = net(data)
loss = criterion(score, target)
loss.backward()
optimizer.step()
_, predicted = score.max(1)
predicted, target = to_np(predicted), to_np(target)
acc, acc_cls, mean_iu = label_accuracy_score(target, predicted,
n_class)
info = {
'acc': acc,
'acc_cls': acc_cls,
'mean_iu': mean_iu,
'loss': loss.data[0]
}
for tag, value in info.items():
logger.scalar_summary(tag, value,
len_batch * epoch_idx + batch_idx + 1)
print(('train', batch_idx, epoch_idx))
if (epoch_idx + 1) % 10 == 0:
n = (epoch_idx + 1) / 10
state = net.state_dict()
torch.save(state, './deeplab_epoch_' + str(n) + '.pth')
def eval(
self,
mode=None,
batch_size=None,
output_dir=None
):
# Sets the module in evaluation mode.
self.NetG.eval()
self.NetD.eval()
if batch_size is None:
batch_size = self.data_loader.batch_size
nrows = batch_size // 8
viz_labels = np.array([num for _ in range(nrows) for num in range(8)])
viz_labels = torch.LongTensor(viz_labels).to(self.device)
with torch.no_grad():
if self.name == 'cgan':
viz_tensor = torch.randn(batch_size, self.latent_dim, device=self.device)
viz_sample = self.NetG(viz_tensor, viz_labels) # generated image from random noise
elif self.name == 'infogan':
viz_tensor = torch.randn(batch_size, self.latent_dim, device=self.device)
labels_onehot = self._to_onehot(viz_labels, dim=self.classes)
z_code = torch.zeros((batch_size, self.code_dim), device=self.device)
if mode is not None:
for i in range(batch_size):
z_code[i, mode] = 4. * i / batch_size - 2.
viz_sample = self.NetG(viz_tensor, labels_onehot, z_code)
viz_vector = utils.to_np(viz_tensor).reshape(batch_size, self.latent_dim)
cur_time = datetime.now().strftime("%Y%m%d-%H%M%S")
np.savetxt(os.path.join(output_dir, 'vec_{}.txt'.format(cur_time)), viz_vector)
vutils.save_image(viz_sample, os.path.join(output_dir, 'img_{}.png'.format(cur_time)), nrow=8, normalize=True)
logging.info(f'\nSaving evaluation image to {output_dir}...')
def eval(self,
mode=None,
batch_size=None):
self.netG.eval()
self.netD.eval()
if batch_size is None:
batch_size = self.data_loader.batch_size
nrows = batch_size // 8
viz_labels = np.array([num for _ in range(nrows) for num in range(8)])
viz_labels = torch.LongTensor(viz_labels).to(self.device)
with torch.no_grad():
if self.infogan:
viz_tensor = torch.randn(batch_size, self.latent_dim, device=self.device)
labels_onehot = self._to_onehot(viz_labels, dim=self.classes)
z_style = torch.zeros((batch_size, self.style_dim), device=self.device)
if mode is not None:
for i in range(batch_size):
z_style[i, mode] = 4. * i / batch_size - 2.
viz_sample = self.netG(viz_tensor, labels_onehot, z_style)
else:
viz_tensor = torch.randn(batch_size, self.latent_dim, 1, 1, device=self.device)
viz_sample = self.netG(viz_tensor, viz_labels)
viz_vector = utils.to_np(viz_tensor).reshape(batch_size, self.latent_dim)
cur_time = datetime.now().strftime("%Y%m%d-%H%M%S")
np.savetxt('vec_{}.txt'.format(cur_time), viz_vector)
vutils.save_image(viz_sample, 'img_{}.png'.format(cur_time), nrow=8, normalize=True)
def generateResults(encoder_decoder: EncoderDecoder, data_loader,
resultFilename, input_tokens_list):
idx_to_tok = encoder_decoder.lang.idx_to_tok
all_output_seqs = []
all_target_seqs = []
for batch_idx, (input_idxs, target_idxs, _,
_) in enumerate(tqdm(data_loader)):
input_lengths = (input_idxs != 0).long().sum(dim=1)
sorted_lengths, order = torch.sort(input_lengths, descending=True)
input_variable = Variable(input_idxs[order, :][:, :max(input_lengths)])
target_variable = Variable(target_idxs[order, :])
output_log_probs, output_seqs = encoder_decoder(
input_variable, list(sorted_lengths))
print(output_seqs.size())
all_output_seqs.extend(trim_seqs(output_seqs))
all_target_seqs.extend([list(seq[seq > 0])]
for seq in to_np(target_variable))
with open(resultFilename, 'w') as fo:
for seq, input_tokens in zip(all_output_seqs, input_tokens_list):
print(type(seq))
#seq = seq.data.view(-1)
eos_idx = seq.index(2) if 2 in seq else seq
string = seq_to_string(seq[:eos_idx + 1],
idx_to_tok,
input_tokens=None)
fo.write(string + '\n')
return None
def ss_preds_var(self, obs, next_obs, action):
# TODO (chongyi zheng):
# do we need next_obs (forward) or action (inverse) - measure the prediction error,
# or we just need to predictions - measure the prediction variance?
# task identity inference - threshold or statistical hypothesis testing like: https://arxiv.org/abs/1902.09434
assert obs.shape == next_obs.shape and obs.shape[0] == next_obs.shape[0] == action.shape[0], \
"invalid transitions shapes!"
# TODO (chongyi zheng): Do we need to set agent mode to evaluation before prediction?
with torch.no_grad():
obs = torch.FloatTensor(obs).to(self.device) \
if not isinstance(obs, torch.Tensor) else obs.to(self.device)
next_obs = torch.FloatTensor(next_obs).to(self.device) \
if not isinstance(next_obs, torch.Tensor) else next_obs.to(self.device)
action = torch.FloatTensor(action).to(self.device) \
if not isinstance(action, torch.Tensor) else action.to(self.device)
if len(obs.size()) == 3 or len(obs.size()) == 1:
obs = obs.unsqueeze(0)
next_obs = next_obs.unsqueeze(0)
action = action.unsqueeze(0)
# prediction variances
if self.use_fwd:
preds = self.ss_fwd_pred_ensem(obs, action)
if self.use_inv:
# (chongyi zheng): we compute logits variance here
preds = self.ss_inv_pred_ensem(obs, next_obs)
# (chongyi zheng): the same as equation (1) in https://arxiv.org/abs/1906.04161
preds = torch.stack(preds.chunk(self.num_ensem_comps, dim=0))
preds_var = torch.var(preds, dim=0).sum(dim=-1)
return utils.to_np(preds_var)
def main():
model = Deeplab()
dataset = VOC2012ClassSeg('./dataset', split='train', transform=True)
val_loader = torch.utils.data.DataLoader(dataset,
batch_size=1,
shuffle=False,
num_workers=1,
pin_memory=True)
# n_class = len(dataset.class_names)
# model_file = ''
# moda_data = torch.load(model_file)
# try:
# model.load_state_dict(model_data)
# except Exception:
# model.load_state_dict(model_data['model_state_dict'])
# if torch.cuda.is_available():
# model.cuda()
model.eval()
label_trues, label_preds = [], []
for batch_idx, (data, target) in enumerate(val_loader):
# if torch.cuda.is_available():
# data, target = data.cuda(), target.cuda()
data, target = Variable(data, volatile=True), Variable(target)
score = model(data)
_, predicted = score.max(1)
predicted = to_np(predicted)
target = to_np(target)
for lt, lp in zip(target, predicted):
label_trues.append(lt)
label_preds.append(lp)
if batch_idx == 5:
break
n_class = 21
print(len(label_preds))
metrics = label_accuracy_score(label_trues, label_preds, n_class=n_class)
metrics = np.array(metrics)
metrics *= 100
print(metrics)
def show2(epoch):
encoder.load_state_dict(
torch.load(model_dir + 'encoder_{:d}.pkl'.format(epoch)))
decoder.load_state_dict(
torch.load(model_dir + 'decoder_{:d}.pkl'.format(epoch)))
encoder.train()
decoder.train()
cover, yuv = getCoverExample('../pics_lfw/1.JPEG', params.input_size)
secret = getSecretExample('../pics_lfw/4.JPEG', params.input_size)
cover = transform(cover)
secret = transform(secret)
cover.resize_(1, 1, params.input_size, params.input_size)
secret.resize_(1, 1, params.input_size, params.input_size)
concat = torch.cat([secret, cover], dim=1)
if params.use_cuda:
concat = concat.cuda()
concat = Variable(concat)
stego = encoder(concat)
stego = to_np(stego)
stego = (((stego - stego.min()) * 255) /
(stego.max() - stego.min())).astype(np.uint8)
print(np.shape(stego))
stego = stego[0][0]
stego = Image.fromarray(stego, 'L')
stego.show()
yuv[:, :, 0] = stego
img = Image.fromarray(yuv, 'YCbCr')
stego = img.convert('RGB')
stego.save('./lfw_mssim_results/stego.bmp')
stego, _ = getCoverExample('./lfw_mssim_results/stego.bmp',
params.input_size)
stego = transform(stego)
stego.resize_(1, 1, params.input_size, params.input_size)
stego = stego.cuda()
stego = Variable(stego)
secret2 = decoder(stego)
secret2 = to_np(secret2)
secret2 = (((secret2 - secret2.min()) * 255) /
(secret2.max() - secret2.min())).astype(np.uint8)
secret2 = secret2[0][0]
secret2 = Image.fromarray(secret2, 'L')
secret2.show()
secret2.save('./lfw_mssim_results/secret.png')
def act(self, obs, sample=False):
obs = torch.FloatTensor(obs).to(self.device)
obs = obs.unsqueeze(0)
dist = self.actor(obs)
action = dist.sample() if sample else dist.mean
action = action.clamp(*self.action_range)
assert action.ndim == 2 and action.shape[0] == 1
return utils.to_np(action[0])
def act(self, obs, sample=False):
grid_state = torch.from_numpy(obs).unsqueeze(0).long().to(self.device)
text_state = torch.from_numpy(get_text_state(grid_state, self.indexed_embedding_map)).float().to(self.device)
dist = self.actor(self.fusion((grid_state, text_state)))
action = dist.sample() if sample else dist.mean
action = action.clamp(*self.action_range)
# assert action.ndim == 2 and action.shape[0] == 1
return utils.to_np(action[0])
def evaluate(encoder_decoder: EncoderDecoder, data_loader):
loss_function = torch.nn.NLLLoss(
ignore_index=0, reduce=False
) # what does this return for ignored idxs? same length output?
losses = []
all_output_seqs = []
all_target_seqs = []
for batch_idx, (input_idxs, target_idxs, _,
_) in enumerate(tqdm(data_loader)):
input_lengths = (input_idxs != 0).long().sum(dim=1)
sorted_lengths, order = torch.sort(input_lengths, descending=True)
input_variable = Variable(input_idxs[order, :][:, :max(input_lengths)],
volatile=True)
target_variable = Variable(target_idxs[order, :], volatile=True)
batch_size = input_variable.shape[0]
output_log_probs, output_seqs = encoder_decoder(
input_variable, list(sorted_lengths))
all_output_seqs.extend(trim_seqs(output_seqs))
all_target_seqs.extend([list(seq[seq > 0])]
for seq in to_np(target_variable))
flattened_log_probs = output_log_probs.view(
batch_size * encoder_decoder.decoder.max_length, -1)
batch_losses = loss_function(flattened_log_probs,
target_variable.contiguous().view(-1))
losses.extend(list(to_np(batch_losses)))
mean_loss = len(losses) / sum(losses)
"""
for i in range(20):
print(all_target_seqs[i])
print(all_output_seqs[i])
print('*'*80)
"""
bleu_score = corpus_bleu(all_target_seqs,
all_output_seqs,
smoothing_function=SmoothingFunction().method2)
print('BLEU SCORE: ' + str(bleu_score))
return mean_loss, bleu_score
def log_weights_filter_magnitude(self, model, epoch, multi_graphs=False):
"""Log the L1-magnitude of the weights tensors.
"""
for name, param in model.state_dict().items():
if param.dim() in [4]:
self.tblogger.list_summary('magnitude/filters/' + name,
list(to_np(norm_filters(param))),
epoch, multi_graphs)
self.tblogger.sync_to_file()
def act(self, obs, sample=False):
obs = torch.FloatTensor(obs).to(self.device)
obs = obs.unsqueeze(0)
dist = self.actor(obs)
# using discrete action space
action = dist.sample() if sample else dist.probs.argmax(dim=-1,
keepdim=True)
# action = action.clamp(*self.action_range)
# assert action.ndim == 2 and action.shape[0] == 1
return utils.to_np(action[0])
def act(self, observation, desired_goal, sample=False):
observation = torch.FloatTensor(observation).to(
self.device).unsqueeze(0)
desired_goal = torch.FloatTensor(desired_goal).to(
self.device).unsqueeze(0)
dist = self.actor(observation, desired_goal)
action = dist.sample() if sample else dist.mean
action = action.clamp(*self.action_range)
assert action.ndim == 2 and action.shape[0] == 1
return utils.to_np(action[0])
def get_mse(self, epoch):
# self.load(epoch)
self.G.eval()
self.E.eval()
self.FC.eval()
self.C.eval()
critirion = nn.MSELoss()
count = 0
test_pred = []
test_true = []
for X, labels in self.valid_loader:
count += 1
X = utils.to_var(X)
labels = utils.to_var(labels)
mu, sigma = self.E(self.FC(X))
z_class = self.C(mu)
X_hat = self.G(mu)
loss = (X_hat.view(X_hat.size(0), -1).cpu().data.numpy() - X.view(X.size(0), -1).cpu().data.numpy()) ** 2
loss = np.mean(loss, 1)
_, test_argmax = torch.max(z_class, 1)
if count == 1:
final_loss = loss
test_pred = utils.to_np(test_argmax.squeeze())
print(test_pred)
test_true = utils.to_np(labels.squeeze())
print(test_true)
else:
final_loss = np.concatenate((final_loss, loss), 0)
test_pred = np.concatenate((test_pred, utils.to_np(test_argmax.squeeze())), axis=0)
test_true = np.concatenate((test_true, utils.to_np(labels.squeeze())), axis=0)
print(final_loss.shape)
test_accuracy = metrics.accuracy_score(test_true, test_pred)
print("Final mse mean is %.5f, std is %.5f" % (np.mean(final_loss), np.std(final_loss)))
print("Accuracy is %.5f" % (test_accuracy))
def test(epoch_idx, net, test_loader, logger, n_class):
net.cuda()
net.eval()
len_batch = len(test_loader)
visualizations = []
hist = np.zeros((n_class, n_class))
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(test_loader):
inputs, targets = inputs.cuda(), targets.cuda()
output = net(inputs)
_, predicted = output.max(1)
predicted, targets = to_np(predicted), to_np(targets)
print(('test', batch_idx, epoch_idx))
hist += fast_hist(targets.flatten(), predicted.flatten(), n_class)
miou = np.diag(hist) / (hist.sum(1) + hist.sum(0) - np.diag(hist))
miou = np.sum(miou) / len(miou)
logger.scalar_summary('Mean iou', miou, epoch_idx)
print(('Mean iou: ', miou))
def act(self, obs, sample=False):
pix_obs = torch.FloatTensor(obs['pix_obs']).to(self.device)
pix_obs = pix_obs.unsqueeze(0)
state_low_obs = None
if self.lstate_shape != 0:
state_low_obs = torch.FloatTensor(obs['state_low_obs']).to(
self.device)
state_low_obs = torch.unsqueeze(state_low_obs, 0)
dist = self.actor(pix_obs, state_low_obs)
action = dist.sample() if sample else dist.mean
action = action.clamp(*self.action_range)
assert action.ndim == 2 and action.shape[0] == 1
return utils.to_np(action[0])
def multi_step(self,
agent,
n_steps=None,
single_episode=False,
video_file_suffix=None):
"""
Performs multiple steps (either n_steps or a single episode) in the environment
and returns tensors with all the (s, a, ns) transitions. Either n_steps or single_episode
must be specified.
Args:
agent (object): agent with get_action(state) method returning an action for the agent
n_steps (int, optional, default=None): number of steps to take in the environment
single_episode (boolean, optional, default=False): whether to perform only one episode
video_file_suffix (string, optional, default=None): Suffix added to the end of the video file name
Returns:
(s, a, s') staked transitions
staked all_old_states (torch Tensor[n_steps, d_state])
staked all_actions (torch Tensor[n_steps, d_action])
staked all_next_states (torch Tensor[n_steps, d_state])
"""
assert (n_steps is None) ^ (single_episode is False)
all_old_states = []
all_next_states = []
all_actions = []
if single_episode:
n_steps = sys.maxsize
for i in range(1, n_steps + 1):
# FIXME: are the view-numel things necessary?
action = agent.get_action(self.state.view(
1, self.state.numel())).to('cpu')
if self.torch_np_conversion:
state, next_state, done = self.step(
to_np(action), video_file_suffix=video_file_suffix)
else:
state, next_state, done = self.step(
action, video_file_suffix=video_file_suffix)
all_old_states.extend(state.view(1, state.numel()))
all_actions.extend(action.view(1, action.numel()))
all_next_states.extend(next_state.view(1, next_state.numel()))
if single_episode and done:
break
return torch.stack(all_old_states), torch.stack(
all_actions), torch.stack(all_next_states)
def __getitem__(self, index):
image_name = self.image_list[index]
label_name = self.label_list[index]
image = Image.open(os.path.join(self.image_path, image_name))
label = Image.open(os.path.join(self.label_path, label_name))
# Transform image, label in forms
if self.transform is not None:
image = self.transform(image)
label = to_tensor(to_np(label)).long()
return image, label
def act(self, obs, exploration=False):
# TODO (chongyi zheng)
if exploration and np.random.rand() < self.exploration_rate:
action = np.random.randint(self.action_shape)
else:
with torch.no_grad():
obs = torch.FloatTensor(obs).to(self.device)
obs = obs.unsqueeze(0)
prob, _ = self.q_net(obs)
q_values = (prob * self.atoms).sum(-1)
# greed action
action = utils.to_np(q_values.argmax(dim=-1))
return action
def _request_transformation(self, _from, _to):
topic = 'FrameTransformation.{}.{}'.format(_from, _to)
self.subscription.subscribe(topic)
try:
msg = self.channel.consume(timeout=5.0)
self.subscription.unsubscribe(topic)
except:
self.subscription.unsubscribe(topic)
return False
transformation = msg.unpack(FrameTransformation)
if _from not in self.transformations:
self.transformations[_from] = {}
self.transformations[_from][_to] = to_np(transformation.tf)
return True
def act(self, obs, sample=False):
obs = torch.FloatTensor(obs).to(self.device)
obs = obs.unsqueeze(0)
if self.aug_type == "crop":
obs = self.center_crop(obs)
elif self.aug_type == "translate":
pad = (self.image_size - obs.shape[-1]) // 2
obs = F.pad(obs, [pad, pad, pad, pad])
else:
raise ValueError(self.aug_type)
dist = self.actor(obs)
action = dist.sample() if sample else dist.mean
action = action.clamp(*self.action_range)
assert action.ndim == 2 and action.shape[0] == 1
return utils.to_np(action[0])
def fit_self(self, X):
logging.debug(
f'Reconstructor fit_self() called. Begin to fit {X.shape}')
X = normalize(X, axis=1) # unit length
num_sample = len(X)
dim = X.shape[1]
all_idx = np.arange(num_sample)
# the coef matrix to return
result_matrix = np.zeros([num_sample, num_sample])
# the coef matrix in torch
result_matrix_t = torch.zeros(num_sample,
num_sample).to(self.device)
# the X in torch
X_t = torch.from_numpy(X).to(self.device)
# for each column ticked out
for i1 in tnrange(num_sample, desc='fit_self', leave=False):
# idx ticked
this_idx = np.delete(all_idx, i1).tolist()
if dim >= num_sample:
coef, _ = self.gels(X_t[this_idx, :],
X_t[i1, :].unsqueeze(0))
coef.squeeze_()
else:
coef, _ = self.lrfit(X[this_idx, :],
X[i1, :].reshape(1, dim))
coef = torch.from_numpy(coef).squeeze().to(self.device)
# get the largest ones
val, idx = coef.abs().topk(K)
# reshape
tmp = torch.zeros_like(coef)
tmp.scatter_(0, idx, val * coef[idx].sign())
# assign
result_matrix_t[i1, this_idx] = tmp
# as np
result_matrix = utils.to_np(result_matrix_t)
return result_matrix
def visualize_results(self, epoch, N=0):
""" visualize the sampled results and save them"""
save_path = os.path.join(self.save_path, 'visualization')
if not os.path.exists(save_path):
os.makedirs(save_path)
samples = self.model.sample_fixed() if N == 0 else self.model.sampleN(
N)
samples_grid = self._make_grid(samples)
plt.imshow(utils.to_np(samples_grid).transpose(1, 2, 0))
plt.show()
torchvision.utils.save_image(samples_grid,
filename=os.path.join(
save_path,
'epoch%03d' % epoch + '.png'))
logging.info('Sampled images saved.')
return self
def visualize_results(self, epoch):
# self.load(199)
print("visulize..")
self.G.eval()
self.E.eval()
self.FC.eval()
save_dir = os.path.join(self.root, self.result_dir, 'mixed_gaussian',
self.model_name, str(self.args.seed_random))
print(save_dir)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
# Store results
Recon = []
Original = []
Z = []
Random = []
color_vec = []
for iter, (X, label) in enumerate(self.valid_loader):
z = utils.to_var(torch.randn(self.batch_size, self.z_dim))
X = utils.to_var(X)
label = utils.to_var(label)
z_mu, z_sigma = self.E(self.FC(X))
X_reconstruc = self.G(z_mu)
X_random = self.G(z)
Original += [x for x in utils.to_np(X)]
Recon += [x for x in utils.to_np(X_reconstruc)]
Z += [x for x in utils.to_np(z_mu)]
Random += [x for x in utils.to_np(X_random)]
color_vec += [x for x in utils.to_np(label)]
self.G.train()
self.E.train()
self.FC.train()
Original = np.array(Original)
Recon = np.array(Recon)
Z = np.array(Z)
Random = np.array(Random)
self.count(Random[:2500])
cmap = plt.get_cmap('gnuplot')
cmap = plt.cm.jet
cmaplist = [cmap(i) for i in range(cmap.N)]
cmap = cmap.from_list('Custom cmap', cmaplist, cmap.N)
fig, ax = plt.subplots(1, 1, figsize=(6, 6))
plt.xlim(-6, 6)
plt.ylim(-6, 6)
plt.xticks(fontsize=20)
plt.yticks(fontsize=20)
ax.scatter(Original[:, 0],
Original[:, 1],
c=color_vec,
marker='.',
cmap=cmap,
alpha=0.3)
fig.savefig(os.path.join(save_dir,
'X_original' + '_epoch%03d' % epoch + '.png'),
transparent=True)
plt.close()
fig, ax = plt.subplots(1, 1, figsize=(6, 6))
plt.xlim(-6, 6)
plt.ylim(-6, 6)
plt.xticks(fontsize=20)
plt.yticks(fontsize=20)
ax.scatter(Recon[:10000, 0],
Recon[:10000, 1],
c=color_vec[:10000],
marker='.',
cmap=cmap,
alpha=0.3)
fig.savefig(os.path.join(
save_dir, 'X_reconstruc' + '_epoch%03d' % epoch + '.png'),
transparent=True)
plt.close()
fig, ax = plt.subplots(1, 1, figsize=(6, 6))
plt.xlim(-6, 6)
plt.ylim(-6, 6)
plt.xticks(fontsize=20)
plt.yticks(fontsize=20)
ax.scatter(Random[:10000, 0],
Random[:10000, 1],
color='black',
marker='.',
alpha=0.3)
fig.savefig(os.path.join(save_dir,
'X_random' + '_epoch%03d' % epoch + '.png'),
transparent=True)
plt.close()
fig, ax = plt.subplots(1, 1, figsize=(6, 6))
plt.xlim(-3, 3)
plt.ylim(-3, 3)
ax.set_xticks([-3, -2, -1, 0, 1, 2, 3])
ax.set_yticks([-3, -2, -1, 0, 1, 2, 3])
plt.xticks(fontsize=20)
plt.yticks(fontsize=20)
ax.scatter(Z[:, 0],
Z[:, 1],
c=color_vec,
marker='.',
cmap=cmap,
alpha=0.3)
fig.savefig(os.path.join(save_dir,
'Z_mu' + '_epoch%03d' % epoch + '.png'),
transparent=True)
plt.close()
def train(model, data_loader, optimizer, epoch, writer):
"""
Train CapsuleNet model on training set
Args:
model: The CapsuleNet model.
data_loader: An interator over the dataset. It combines a dataset and a sampler.
optimizer: Optimization algorithm.
epoch: Current epoch.
"""
print('===> Training mode')
num_batches = len(data_loader) # iteration per epoch. e.g: 469
total_step = args.epochs * num_batches
epoch_tot_acc = 0
# Switch to train mode
model.train()
if args.cuda:
# When we wrap a Module in DataParallel for multi-GPUs
model = model.module
start_time = timer()
for batch_idx, (data, target) in enumerate(tqdm(data_loader,
unit='batch')):
batch_size = data.size(0)
global_step = batch_idx + (epoch * num_batches) - num_batches
labels = target
target_one_hot = utils.one_hot_encode(target, length=args.num_classes)
assert target_one_hot.size() == torch.Size([batch_size, 10])
data, target = Variable(data), Variable(target_one_hot)
if args.cuda:
data = data.cuda()
target = target.cuda()
# Train step - forward, backward and optimize
optimizer.zero_grad()
output = model(data) # output from DigitCaps (out_digit_caps)
loss, margin_loss, recon_loss = model.loss(data, output, target)
loss.backward()
optimizer.step()
# Calculate accuracy for each step and average accuracy for each epoch
acc = utils.accuracy(output, labels, args.cuda)
epoch_tot_acc += acc
epoch_avg_acc = epoch_tot_acc / (batch_idx + 1)
# TensorBoard logging
# 1) Log the scalar values
writer.add_scalar('train/total_loss', loss.data[0], global_step)
writer.add_scalar('train/margin_loss', margin_loss.data[0],
global_step)
if args.use_reconstruction_loss:
writer.add_scalar('train/reconstruction_loss', recon_loss.data[0],
global_step)
writer.add_scalar('train/batch_accuracy', acc, global_step)
writer.add_scalar('train/accuracy', epoch_avg_acc, global_step)
# 2) Log values and gradients of the parameters (histogram)
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
writer.add_histogram(tag, utils.to_np(value), global_step)
writer.add_histogram(tag + '/grad', utils.to_np(value.grad),
global_step)
# Print losses
if batch_idx % args.log_interval == 0:
template = 'Epoch {}/{}, ' \
'Step {}/{}: ' \
'[Total loss: {:.6f},' \
'\tMargin loss: {:.6f},' \
'\tReconstruction loss: {:.6f},' \
'\tBatch accuracy: {:.6f},' \
'\tAccuracy: {:.6f}]'
tqdm.write(
template.format(
epoch, args.epochs, global_step, total_step, loss.data[0],
margin_loss.data[0],
recon_loss.data[0] if args.use_reconstruction_loss else 0,
acc, epoch_avg_acc))
# Print time elapsed for an epoch
end_time = timer()
print('Time elapsed for epoch {}: {:.0f}s.'.format(epoch,
end_time - start_time))
optimizer.step()
training_loss += loss.cpu().data.numpy()[0] * float(inputs.size(0))
train_acc = counter.acc()
counter.flush()
test_loss = 0
net.eval()
utils.set_strategy(net, 'running')
for _, (inputs, labels) in enumerate(testloader):
inputs, labels = Variable(inputs.cuda(async=True)), Variable(
labels.cuda(async=True))
outputs = net(inputs)
loss = criterion(outputs, labels)
test_loss += utils.to_np(loss) * float(inputs.size(0))
counter.add(utils.to_np(outputs), utils.to_np(labels))
print(
' -- Epoch %d | time: %.4f | loss: %.4f | training acc: %.4f validation accuracy: %.4f | lr %.6f --'
% (epoch, time() - t0, training_loss, train_acc, counter.acc(), lr))
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict':
net.module.state_dict() if use_cuda else net.state_dict(),
'test_accuracy': counter.acc(),
'optimizer': optimizer.state_dict(),
'name': args.model,
'model_args': model_args,
