def __init__(self, labels, img_train, labels_train, img_test, labels_test,
**kwargs):
self.labels = labels
self.x_train = flatten(img_train)
self.y_train = to_one_hot(labels_train, len(labels))
self.x_test = flatten(img_test)
self.y_test = to_one_hot(labels_test, len(labels))
self.x_train, self.x_val, self.y_train, self.y_val = \
train_test_split(self.x_train, self.y_train, test_size=0.1)
self.n = self.x_train.shape[1]
self.m = self.y_train.shape[1]
self.k = kwargs.pop('hidden_layer_size', 256)
self.lr = kwargs.pop('learning_rate', 1e-3)
self.dropout = kwargs.pop('dropout', False)
self.regularization = kwargs.pop('regularization', False)
self.adaptive_lr = kwargs.pop('adaptive_lr', False)
self.history_path = kwargs.pop('results_path')
self.model_path = kwargs.pop('model_path')
self.init_nn()
self.saver = tf.train.Saver()
def fit(self, train_lbs):
_, loss = self.sess.run(
[self.train, self.loss], {
self.x: to_one_hot(add_swap_noise(train_lbs)),
self.y: to_one_hot(train_lbs)
})
return loss
def filter_pair(union_bbox, subject_bbox, object_bbox, subject_id, object_id,
img):
union_mask = torch.zeros(*img.shape[:2])
union_mask[union_bbox[0]:union_bbox[1], union_bbox[2]:union_bbox[3]] = 1
subject_mask = torch.zeros(*img.shape[:2])
subject_mask[subject_bbox[0]:subject_bbox[1],
subject_bbox[2]:subject_bbox[3]] = 1
object_mask = torch.zeros(*img.shape[:2])
object_mask[object_bbox[0]:object_bbox[1],
object_bbox[2]:object_bbox[3]] = 1
mask = torch.cat([
union_mask.unsqueeze(0),
subject_mask.unsqueeze(0),
object_mask.unsqueeze(0)
], 0)
mask = spatial_transform(mask).unsqueeze(0)
object_dist = torch.FloatTensor([to_one_hot(object_id, 100)])
subject_dist = torch.FloatTensor([to_one_hot(subject_id, 100)])
inputs = [mask, subject_dist, object_dist]
inputs = [torch.autograd.Variable(x.cuda(), volatile=True) for x in inputs]
output = model.net.forward(inputs)
_, result = output.max(1)
if result.data[0] == 1:
return True
else:
return False
def source_target_separate(datasets, sources, targets):
N1 = len(sources.keys())
N_domain = N1 + len(targets.keys())
domain_idx = 0
sets = {}
for key in sources.keys():
sources[key] = domain_idx
sets[key] = domain_idx
domain_idx = domain_idx + 1
for key in targets.keys():
targets[key] = domain_idx
sets[key] = domain_idx
domain_idx = domain_idx + 1
for key in datasets.keys():
datasets[key]['train']['domains'] = to_one_hot(
sets[key] * np.ones((datasets[key]['train']['images'].shape[0],
)).astype('float32'), N_domain)
datasets[key]['valid']['domains'] = to_one_hot(
sets[key] * np.ones((datasets[key]['valid']['images'].shape[0],
)).astype('float32'), N_domain)
datasets[key]['test']['domains'] = to_one_hot(
sets[key] * np.ones((
datasets[key]['test']['images'].shape[0], )).astype('float32'),
N_domain)
return datasets
def _data(data_pth, split_val=True, verbose=1):
data = np.load(data_pth, allow_pickle=True)
x, y = data['x'], data['y']
x = x[:, :, np.newaxis]
x_train, y_train, x_test, y_test = split_data(x, y)
class_weights_dict = calc_class_weights(y_train)
if split_val:
x_train, y_train, x_val, y_val = split_data(x_train, y_train)
y_train = to_one_hot(y_train, dimension=10)
y_test = to_one_hot(y_test, dimension=10)
y_val = to_one_hot(y_val, dimension=10)
if verbose:
print('\nx_train shape: %s'
'\ny_train shape: %s'
'\nx_test shape: %s'
'\ny_test shape: %s'
'\nx_val shape: %s'
'\ny_val shape: %s' %
(x_train.shape, y_train.shape, x_test.shape, y_test.shape,
x_val.shape, y_val.shape))
return x_train, y_train, x_test, y_test, x_val, y_val, class_weights_dict
else:
y_train = to_one_hot(y_train, dimension=10)
y_test = to_one_hot(y_test, dimension=10)
if verbose:
print('\nx_train shape: %s'
'\ny_train shape: %s'
'\nx_test shape: %s'
'\ny_test shape: %s' %
(x_train.shape, y_train.shape, x_test.shape, y_test.shape))
return x_train, y_train, x_test, y_test, class_weights_dict
def forward(self, states, action):
cuda = states.is_cuda
batch_size = states.size(0)
num_nodes = states.size(1)
if self.immovable_bit:
# add movable/immovable bit (the first two objects are immovable, this is hardcoded for now)
tmp = torch.zeros_like(states[:, :, 0:1])
tmp[:, :2, :] = 1.0
states = torch.cat([states, tmp], dim=2)
# states: [batch_size (B), num_objects, embedding_dim]
# node_attr: Flatten states tensor to [B * num_objects, embedding_dim]
node_attr = states.view(-1, self.input_dim)
edge_attr = None
edge_index = None
if num_nodes > 1:
# edge_index: [B * (num_objects*[num_objects-1]), 2] edge list
edge_index = self._get_edge_list_fully_connected(
batch_size, num_nodes, cuda)
row, col = edge_index
edge_attr = self._edge_model(
node_attr[row], node_attr[col], edge_attr, source_indices=(row % self.num_objects).cpu().numpy(),
target_indices=(col % self.num_objects).cpu().numpy())
if not self.ignore_action:
if self.copy_action:
action_vec = utils.to_one_hot(
action, self.action_dim).repeat(1, self.num_objects)
action_vec = action_vec.view(-1, self.action_dim)
else:
if not self.factored_continuous_action:
action_vec = utils.to_one_hot(action, self.action_dim * num_nodes)
else:
action_vec = action
action_vec = action_vec.view(-1, self.action_dim)
# Attach action to each state
node_attr = torch.cat([node_attr, action_vec], dim=-1)
node_attr = self._node_model(
node_attr, edge_index, edge_attr)
# [batch_size, num_nodes, hidden_dim]
node_attr = node_attr.view(batch_size, num_nodes, -1)
if self.immovable_bit:
# object embeddings have an additional bit for movable/immovable objects
# we do not need to predict that
node_attr = node_attr[:, :, :self.input_dim - 1]
return node_attr
def fit(self, mov_ims, fix_ims, mov_lbs, fix_lbs):
_, loss = self.sess.run(
[self.train, self.loss], {
self.x: mov_ims,
self.y: fix_ims,
self.xlabel: to_one_hot(mov_lbs),
self.ylabel: to_one_hot(fix_lbs)
})
return loss
def forward(self, states, action):
cuda = states.is_cuda
batch_size = states.size(0)
num_nodes = states.size(1)
# states: [batch_size (B), num_objects, embedding_dim]
# node_attr: Flatten states tensor to [B * num_objects, embedding_dim]
node_attr = states.view(-1, self.input_dim)
edge_attr = None
edge_index = None
if num_nodes > 1:
# edge_index: [B * (num_objects*[num_objects-1]), 2] edge list
edge_index = self._get_edge_list_fully_connected(
batch_size, num_nodes, cuda)
row, col = edge_index
edge_attr = self._edge_model(node_attr[row], node_attr[col],
edge_attr)
if not self.ignore_action:
if self.action_type == "action_one_hot":
if self.copy_action:
action_vec = utils.to_one_hot(action,
self.action_dim).repeat(
1, self.num_objects)
action_vec = action_vec.view(-1, self.action_dim)
else:
action_vec = utils.to_one_hot(action,
self.action_dim * num_nodes)
action_vec = action_vec.view(-1, self.action_dim)
else:
assert self.action_type == "action_image"
assert not self.copy_action # here each object's action vec is learned separately
# action_vec = self.act_extractor(action)
action_vec = self.act_encoder(action)
action_vec = action_vec.view(-1, self.action_dim).repeat(
self.num_objects, 1)
# Attach action to each state
node_attr = torch.cat([node_attr, action_vec], dim=-1)
node_attr = self._node_model(node_attr, edge_index, edge_attr)
# [batch_size, num_nodes, hidden_dim]
return node_attr.view(batch_size, num_nodes, -1)
def cw(model, X, y, verbose=False, params={}):
C = params.get('C', 0.0001)
niters = params.get('niters', 50)
step_size = params.get('step_size', 0.01)
confidence = params.get('confidence', 0.0001)
img_min = params.get('img_min', 0.0)
img_max = params.get('img_max', 1.0)
Xt = arctanh_rescale(X, img_min, img_max)
Xt_adv = Variable(Xt.data, requires_grad=True)
y_onehot = to_one_hot(y, model.num_actions).float()
optimizer = optim.Adam([Xt_adv], lr=step_size)
for i in range(niters):
logits = model.forward(tanh_rescale(Xt_adv, img_min, img_max))
real = (y_onehot * logits).sum(dim=1)
other = ((1.0 - y_onehot) * logits -
(y_onehot * TARGET_MULT)).max(1)[0]
loss1 = torch.clamp(real - other + confidence, min=0.)
loss2 = torch.sum((X - tanh_rescale(Xt_adv, img_min, img_max)).pow(2),
dim=[1, 2, 3])
loss = loss1 + loss2 * C
optimizer.zero_grad()
model.features.zero_grad()
loss.backward()
optimizer.step()
# if verbose:
# print('loss1: {}, loss2: {}'.format(loss1, loss2))
return tanh_rescale(Xt_adv, img_min, img_max).data
def do_test(model, data, measures):
start_time = time.time()
input_size = len(_g.vocab)
if not _g.args.quiet:
print('Testing...')
criterion = nn.NLLLoss(ignore_index=_g.vocab.stoi[_g.padding_symbol])
losses = None
so_far = 0
try:
for i, batch in zip(range(len(data)), data): # TODO necessary for now to do it this way
loss = _t.evaluate(model, criterion, _u.to_one_hot(batch.before, input_size), batch.after, measures)
loss = loss.unsqueeze(dim=1)
losses = loss if losses is None else torch.cat((losses, loss), dim=1)
so_far = i+1
if not _g.args.quiet:
print('Testing done successfully')
except KeyboardInterrupt:
print('\nExiting earlier than expected. Wait a moment!')
losses = losses.mean(dim=1)
text = 'Test {} elements in {}.'.format(so_far * data.batch_size, _u.pretty_print_time(time.time() - start_time))
eval_measures = _u.to_builtin({n: (x,y) for n,x,y in
zip(['loss'] + list(measures.keys()), losses[::2], losses[1::2])})
for i, j in eval_measures.items():
text += ' ' + i + ' {:5.6f}({:5.6f}).'.format(j[0], j[1])
if not _g.args.quiet:
print(text)
def test_text_disciminator():
d_batch = 4
d_max_seq_len = 52
quora_default_args = torch.load(
QUORA_PARAPHRASE_PRETRAINED_DEFAULT_CONFIG_PATH)
quora_word_id_to_word = torch.load(
QUORA_PARAPHRASE_PRETRAINED_WORD_ID_TO_WORD_PATH)
word_id_to_word = quora_word_id_to_word
d_vocab = len(word_id_to_word)
args = edict(quora_default_args)
d_text_feature = args.txtSize
dis_dropout = args.drop_prob_lm
d_dis_cnn = args.cnn_dim
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
text_dis = TextDiscriminator(d_vocab=d_vocab,
d_text_feature=d_text_feature,
dis_dropout=dis_dropout,
d_dis_cnn=d_dis_cnn).to(device).train()
text_dis.load_state_dict(torch.load('faster_text_dis_v1.pth'))
assert text_dis
texts = torch.randint(low=0, high=d_vocab,
size=(d_batch, d_max_seq_len)).to(device)
text_features, valids = text_dis(to_one_hot(texts, d_vocab))
assert text_features.size() == (
d_batch, d_text_feature) and text_features.dtype == torch.float
assert valids.size() == (d_batch, ) and valids.dtype == torch.float
text_log_probs = -torch.rand(d_batch, d_max_seq_len, d_vocab).to(device)
text_features, valids = text_dis(text_log_probs)
assert text_features.size() == (
d_batch, d_text_feature) and text_features.dtype == torch.float
assert valids.size() == (d_batch, ) and valids.dtype == torch.float
def train_one_epoch(self, dataloader):
self.net.train()
self.hook.flag_hook = False
for mb, (x, y) in enumerate(dataloader):
x, y = x.to(self.device), y.to(self.device)
y_one_hot = utils.to_one_hot(y, self.c_dim)
if self.cfg.train_random > 0 and (mb + 1) % 10 == 0:
with torch.no_grad():
x_rand = torch.randn(size=x.shape).to(self.device)
logits_rand = self.net(x_rand)
entropy_rand = metrics.entropy(
utils.logits_to_probs(logits_rand))
if torch.mean(entropy_rand).item() <= self.thresh_entropy:
print(
'training on random inputs & random labels for minibatch {}'
.format(mb + 1))
x = torch.randn(size=x.shape).to(self.device)
y_one_hot = torch.ones(size=(x.shape[0], self.c_dim)).to(
self.device) / self.c_dim
self.optimizer.zero_grad()
logits = self.net(x)
losses = self.criterion(logits, y_one_hot)
torch.mean(losses).backward()
self.optimizer.step()
def predict(self,
thetas,
x_test,
y_test,
modeldir='../model/LR',
Onsave=True):
x = np.insert(x_test, 0, 1, axis=1)
y_real = y_test
y_pred = [
np.argmax([self._sigmoid(xi @ theta) for theta in thetas])
for xi in x
]
p_real = to_one_hot(y_test)
p_pred = [[self._sigmoid(xi @ theta) for theta in thetas] for xi in x]
#
# save thetas, p_pred of each bi-classifiers
#
if Onsave:
try:
np.savez('%s/test_rst.npz' % modeldir,
y_real=y_real,
y_pred=y_pred,
p_real=p_real,
p_pred=p_pred)
np.savez('%s/thetas.npz' % modeldir, thetas=thetas)
except:
print('save thetas and test_rst failed')
return y_pred
def test_simple(self):
mask: torch.Tensor = torch.randint(low=0, high=2, size=(2, 15, 10))
num_classes: int = 3
oh_mask: torch.Tensor = utils.to_one_hot(mask, num_classes)
self.assertEqual(oh_mask.size(), torch.Size((2, num_classes, 15, 10)))
self.assertTrue(_is_same_tensor(oh_mask.argmax(dim=1), mask))
def forward(self, text_features):
'''
encoded : (batch_size, feat_size)
seq: (batch_size, seq_len)
lengths: (batch_size, )
'''
d_batch = text_features.size(0)
device = text_features.device
d_max_seq_len = self.d_max_seq_len
text_log_probs = torch.zeros(d_batch, d_max_seq_len, self.d_vocab, device=device)
text_log_probs[:, 0, :] = torch.log(to_one_hot(torch.tensor(self.start_token), self.d_vocab))[None, :]
texts = torch.full((d_batch, d_max_seq_len), self.pad_token, dtype=torch.long, device=device)
texts[:, 0] = self.start_token
hs = None
for i in range(1, d_max_seq_len):
if i > 1:
word_embeddings = self.embed(words) # (d_batch, 1) -> (d_batch, 1, d_text_feature)
else:
word_embeddings = text_features.unsqueeze(1) # (d_batch, d_text_feature) -> (d_batch, 1, d_text_feature)
log_probs, hs = self.step(word_embeddings, hs)
log_probs = log_probs.squeeze(1) # (d_batch, 1, d_vocab) -> (d_batch, 1, d_vocab)
text_log_probs[:, i] = log_probs
words = torch.multinomial(torch.exp(log_probs), 1)
texts[:, i] = words.squeeze(1) # (d_batch, 1) -> (d_batch,)
texts, text_log_probs, text_lens = mask_in_place_and_calc_length(texts, text_log_probs, self.end_token, self.pad_token)
return texts, text_log_probs, text_lens
def validate(val_loader, model, log):
'''evaluate trained model'''
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
# Switch to evaluate mode
model.eval()
for i, (input, target) in enumerate(val_loader):
if args.use_cuda:
input = input.cuda()
target = target.cuda()
with torch.no_grad():
output = model(input)
target_reweighted = to_one_hot(target, args.num_classes)
loss = bce_loss(softmax(output), target_reweighted)
# Measure accuracy and record loss
prec1, prec5 = accuracy(output.data, target, topk=(1, 5))
losses.update(loss.item(), input.size(0))
top1.update(prec1.item(), input.size(0))
top5.update(prec5.item(), input.size(0))
print_log(
'**Test ** Prec@1 {top1.avg:.2f} Prec@5 {top5.avg:.2f} Error@1 {error1:.2f} Loss: {losses.avg:.3f} '
.format(top1=top1, top5=top5, error1=100 - top1.avg,
losses=losses), log)
return top1.avg, losses.avg
def forward_n_layers(self,
x,
target=None,
mixup=False,
mixup_hidden=False,
mixup_alpha=None,
layer_num=None):
#print x.shape
if self.per_img_std:
x = per_image_standardization(x)
out = x
if target is not None:
target_reweighted = to_one_hot(target, self.num_classes)
out = self.conv1(out)
out = self.layer1(out)
if layer_num == 1:
return out, target_reweighted
out = self.layer2(out)
if layer_num == 2:
return out, target_reweighted
out = self.layer3(out)
if layer_num == 3:
return out, target_reweighted
out = act(self.bn1(out))
out = F.avg_pool2d(out, 8)
out = out.view(out.size(0), -1)
out = self.linear(out)
return out, target_reweighted
def get_latents(self, encodings, probs_b):
"""Read out latents (z) form input encodings for a single segment."""
readout_mask = probs_b[:, 1:, None] # Offset readout by 1 to left.
readout = (encodings[:, :-1] * readout_mask).sum(1)
hidden = F.relu(self.head_z_1(readout))
logits_z = self.head_z_2(hidden)
# Gaussian latents.
if self.latent_dist == 'gaussian':
if self.training:
mu, log_var = torch.split(logits_z, self.latent_dim, dim=1)
sample_z = utils.gaussian_sample(mu, log_var)
else:
sample_z = logits_z[:, :self.latent_dim]
# Concrete / Gumbel softmax latents.
elif self.latent_dist == 'concrete':
if self.training:
sample_z = utils.gumbel_softmax_sample(logits_z,
temp=self.temp_z)
else:
sample_z_idx = torch.argmax(logits_z, dim=1)
sample_z = utils.to_one_hot(sample_z_idx, logits_z.size(1))
else:
raise ValueError('Invalid argument for `latent_dist`.')
return logits_z, sample_z
def update_sr(sr, sequence, discount, learning_rate):
"""Update SR matrix.
Args:
sr: n_state x n_state matrix
sequence: iterable containing sequence of state inds
discount: scalar discount factor between [0 inclusive, 1 exclusive)
learning_rate: scalar learning rate between [0 inclusive, 1 inclusive]
Returns:
updated sr matrix
"""
n_states = sr.shape[0]
for state_ind, state_ind_next in zip(sequence[:-1], sequence[1:]):
if (state_ind is not None) and (state_ind_next is not None):
state_vec = to_one_hot(state_ind, n_states)
# compute successor prediction error:
# state observed + discount * SR at next state, minus previous estimate
pred_err = state_vec + discount * sr[state_ind_next, :] - sr[
state_ind, :]
# use prediction error to update
sr[state_ind, :] = sr[state_ind, :] + learning_rate * pred_err
return sr
def forward(self, x):
state, action = x
action = utils.to_one_hot(action, self.action_size)
# flatten state
batch_size = state.size(0)
obj_size = state.size(1)
state_r = state.reshape(batch_size * obj_size, state.size(2))
# create keys and queries
key_r = self.fc_key(state_r)
query = self.fc_query(action)
value = self.fc_value(action)
key = key_r.reshape(batch_size, obj_size, self.key_query_size)
# compute a vector of attention weights, one for each object slot
if self.sqrt_scale:
weights = F.softmax((key * query[:, None]).sum(dim=2) * (1 / np.sqrt(self.key_query_size)), dim=-1)
else:
weights = F.softmax((key * query[:, None]).sum(dim=2), dim=-1)
if self.ablate_weights:
# set uniform weights to check if they provide any benefit
weights = torch.ones_like(weights) / weights.shape[1]
# create a separate action for each object slot
# weights: [|B|, |O|], value: [|B|, value_size]
# => [|B|, |O|, value_size]
return weights[:, :, None] * value[:, None, :]
def __init__(self, labels, img_train, labels_train, img_test, labels_test,
**kwargs):
self.labels = labels
self.x_train = np.reshape(img_train, img_train.shape + (1, ))
self.y_train = to_one_hot(labels_train, len(labels))
self.x_test = np.reshape(img_test, img_test.shape + (1, ))
self.y_test = to_one_hot(labels_test, len(labels))
self.n = self.x_train.shape[1:]
self.x_train, self.x_val, self.y_train, self.y_val = \
train_test_split(self.x_train, self.y_train, test_size=0.1)
self.lr = kwargs.pop('learning_rate', 1e-3)
self.history_path = kwargs.pop('results_path')
self.model_path = kwargs.pop('model_path')
self.init_nn()
def _get_files_and_labels(img_dir):
files = utils.list_files(img_dir)
labels = [CLASS_MAP[x.split('_')[2][:-4]] for x in files]
files, labels = utils.resample_unbalanced_data(files, labels)
unique, counts = np.unique(labels, return_counts=True)
print '{} labels has counts as: {}'.format(unique, counts)
return [os.path.join(img_dir, x)
for x in files], [utils.to_one_hot(x, 3) for x in labels]
def train_tcn(in_path,nn_path,n_epochs=5):
dataset=data.seqs.read_seqs(in_path)
train,test=dataset.split()
X,y,params=to_dataset(train)
if("n_cats" in params ):
y=utils.to_one_hot(y,params["n_cats"])
model=make_tcn(params)
model.fit(X,y,epochs=n_epochs,batch_size=32)
model.save_weights(nn_path)
def forward(self, ins):
obj_ids = torch.arange(self.num_objects)
obj_ids = utils.to_one_hot(obj_ids, self.num_objects).unsqueeze(0)
obj_ids = obj_ids.repeat((ins.size(0), 1, 1)).to(ins.get_device())
h = torch.cat((ins, obj_ids), -1)
h = self.act1(self.fc1(h))
h = self.act2(self.fc2(h))
h = self.fc3(h).sum(1)
return h.view(-1, self.output_size[0], self.output_size[1],
self.output_size[2])
def preprocess_save_to_queue(preprocess_fn, q, list_of_lists, output_files,
segs_from_prev_stage, classes, transpose_forward):
errors_in = []
for i, l in enumerate(list_of_lists):
try:
output_file = output_files[i]
print("preprocessing", output_file)
d, _, dct = preprocess_fn(l)
print(output_file, dct)
if segs_from_prev_stage[i] is not None:
assert isfile(
segs_from_prev_stage[i]
) and segs_from_prev_stage[i].endswith(
".nii.gz"
), "segs_from_prev_stage must point to a segmentation file"
seg_prev = sitk.GetArrayFromImage(
sitk.ReadImage(segs_from_prev_stage[i]))
# check to see if shapes match
img = sitk.GetArrayFromImage(sitk.ReadImage(l[0]))
assert all([i == j for i, j in zip(seg_prev.shape, img.shape)]), "image and segmentation from previous " \
"stage don't have the same pixel array " \
"shape! image: %s, seg_prev: %s" % \
(l[0], segs_from_prev_stage[i])
seg_prev = seg_prev.transpose(transpose_forward)
seg_reshaped = resize_segmentation(seg_prev,
d.shape[1:],
order=1,
cval=0)
seg_reshaped = to_one_hot(seg_reshaped, classes)
d = np.vstack((d, seg_reshaped)).astype(np.float32)
print(d.shape)
if np.prod(d.shape) > (
2e9 / 4 * 0.85
): # *0.85 just to be save, 4 because float32 is 4 bytes
print(
"This output is too large for python process-process communication. "
"Saving output temporarily to disk")
np.save(output_file[:-7] + ".npy", d)
d = output_file[:-7] + ".npy"
q.put((output_file, (d, dct)))
except KeyboardInterrupt:
raise KeyboardInterrupt
except Exception as e:
print("error in", l)
print(e)
q.put("end")
if len(errors_in) > 0:
print("There were some errors in the following cases:", errors_in)
print("These cases were ignored.")
else:
print("This worker has ended successfully, no errors to report")
def forward(self, encoder_outputs, inputs, final_encoder_hidden, targets=None, keep_prob=1.0, teacher_forcing=0.0):
batch_size = encoder_outputs.data.shape[0]
seq_length = encoder_outputs.data.shape[1]
#hidden = Variable(torch.zeros(1, batch_size, self.hidden_size))
hidden = final_encoder_hidden
if next(self.parameters()).is_cuda:
hidden = hidden.cuda()
else:
hidden = hidden
# every decoder output seq starts with <SOS>
sos_output = Variable(torch.zeros((batch_size, self.embedding.num_embeddings + seq_length)))
sampled_idx = Variable(torch.ones((batch_size, 1)).long())
if next(self.parameters()).is_cuda:
sos_output = sos_output.cuda()
sampled_idx = sampled_idx.cuda()
sos_output[:, 1] = 1.0 # index 1 is the <SOS> token, one-hot encoding
decoder_outputs = [sos_output]
sampled_idxs = [sampled_idx]
if keep_prob < 1.0:
dropout_mask = (Variable(torch.rand(batch_size, 1, 2 * self.hidden_size + self.embedding.embedding_dim)) < keep_prob).float() / keep_prob
else:
dropout_mask = None
selective_read = Variable(torch.zeros(batch_size, 1, self.hidden_size))
one_hot_input_seq = to_one_hot(inputs, len(self.lang.tok_to_idx) + seq_length)
if next(self.parameters()).is_cuda:
selective_read = selective_read.cuda()
one_hot_input_seq = one_hot_input_seq.cuda()
for step_idx in range(1, self.max_length):
if targets is not None and teacher_forcing > 0.0 and step_idx < targets.shape[1]:
# replace some inputs with the targets (i.e. teacher forcing)
teacher_forcing_mask = Variable((torch.rand((batch_size, 1)) < teacher_forcing), requires_grad=False)
if next(self.parameters()).is_cuda:
teacher_forcing_mask = teacher_forcing_mask.cuda()
sampled_idx = sampled_idx.masked_scatter(teacher_forcing_mask, targets[:, step_idx-1:step_idx])
sampled_idx, output, hidden, selective_read = self.step(sampled_idx, hidden, encoder_outputs, selective_read, one_hot_input_seq, dropout_mask=dropout_mask)
decoder_outputs.append(output)
sampled_idxs.append(sampled_idx)
decoder_outputs = torch.stack(decoder_outputs, dim=1)
sampled_idxs = torch.stack(sampled_idxs, dim=1)
return decoder_outputs, sampled_idxs
def deploy(self, dir_path, test_lbs):
z = to_hard_seg(self.sess.run(self.z, {self.x: to_one_hot(test_lbs)}))
if dir_path is not None:
temp_path = dir_path + '/{:02d}_{}.png'
for i in range(z.shape[0]):
save_image(temp_path.format(i + 1, 'x'),
test_lbs[i, :, :, 0],
is_integer=True)
save_image(temp_path.format(i + 1, 'y'),
z[i, :, :, 0],
is_integer=True)
return z
def new_canvas(self, goal):
self.goal = goal
self.canvas = np.zeros(self.dim)
self.previous_score = None
self.stroke_count = 0
self.terminal = False
# print(self.goal)
# print(self.classifier.number_of_goals)
return U.to_one_hot(
self.goal,
self.classifier.number_of_goals), self.canvas, 0, self.terminal
def test_text_encoder():
d_batch = 2
d_max_seq_len = 26
d_vocab = 27699
d_text_feature = 512
text_enc_dropout = 0.5
d_text_enc_cnn = 512
text_enc = TextEncoder(d_vocab=d_vocab, d_text_feature=d_text_feature, text_enc_dropout=text_enc_dropout, d_text_enc_cnn=d_text_enc_cnn)
text_enc.load_state_dict(torch.load('new_text_enc.pth'))
texts = torch.randint(low=0, high=d_vocab, size=(d_batch, d_max_seq_len))
text_features = text_enc(to_one_hot(texts, d_vocab))
assert text_features.size() == (d_batch, d_text_feature) and text_features.dtype == torch.float
def run_test_with_mixup(cuda, C, loader, mix_rate, mix_layer, num_trials=1):
correct = 0
total = 0
loss = 0.0
softmax = torch.nn.Softmax()
bce_loss = torch.nn.CrossEntropyLoss() #torch.nn.BCELoss()
lam = np.array(mix_rate)
lam = Variable(torch.from_numpy(np.array([lam]).astype('float32')).cuda())
for i in range(0, num_trials):
for batch_idx, (data, target) in enumerate(loader):
if cuda:
data, target = data.cuda(), target.cuda()
data, target = Variable(data, volatile=True), Variable(target)
output, reweighted_target = C(data,
lam=lam,
target=target,
layer_mix=mix_layer)
'''take original with probability lam. First goal is to recover the target indices for the other batch. '''
pred = output.data.max(
1, keepdim=True)[1] # get the index of the max log-probability
correct += pred.eq(target.data.view_as(pred)).cpu().numpy().sum()
total += target.size(0)
'''These are the original targets in a one-hot space'''
target1_onehot = to_one_hot(target, 10)
#print lam
#print reweighted_target[0:3]
target2 = (reweighted_target - target1_onehot * (lam)).max(1)[1]
#print "reweighted target should put probability", lam, "on first set of indexed-values"
#print target[0:3]
#print target2[0:3]
loss += mixup_criterion(target, target2, lam)(
bce_loss, output) * target.size(0)
#t_loss /= total
t_accuracy = 100. * correct / total
average_loss = (loss / total)
#print "Test with mixup", mix_rate, "on layer", mix_layer, ', loss: ', average_loss
#print "Accuracy", t_accuracy
return t_accuracy, average_loss
def compute_y(self,x,targets,mixup=False, visible_mixup=False):
target_soft = to_one_hot(targets,10)
if visible_mixup:
lam = Variable(torch.from_numpy(np.array([np.random.beta(0.5,0.5)]).astype('float32')).cuda())
x, target_soft = mixup_process(x, target_soft, lam=lam)
h = self.h1(x)
if mixup:
lam = Variable(torch.from_numpy(np.array([np.random.beta(0.5,0.5)]).astype('float32')).cuda())
h, target_soft = mixup_process(h, target_soft, lam=lam)
y = self.sm(self.h2(h))
return y, target_soft
def run_test_with_mixup(cuda, C, loader,mix_rate,mix_layer,num_trials=1):
correct = 0
total = 0
loss = 0.0
softmax = torch.nn.Softmax()
bce_loss = torch.nn.CrossEntropyLoss()#torch.nn.BCELoss()
lam = np.array(mix_rate)
lam = Variable(torch.from_numpy(np.array([lam]).astype('float32')).cuda())
for i in range(0,num_trials):
for batch_idx, (data, target) in enumerate(loader):
if cuda:
data, target = data.cuda(), target.cuda()
data, target = Variable(data, volatile=True), Variable(target)
output,reweighted_target = C(data, lam=lam, target=target, layer_mix=mix_layer)
'''take original with probability lam. First goal is to recover the target indices for the other batch. '''
pred = output.data.max(1, keepdim=True)[1] # get the index of the max log-probability
correct += pred.eq(target.data.view_as(pred)).cpu().sum()
total += target.size(0)
'''These are the original targets in a one-hot space'''
target1_onehot = to_one_hot(target,10)
target2 = (reweighted_target - target1_onehot*(lam)).max(1)[1]
loss += mixup_criterion(target, target2, lam)(bce_loss,output) * target.size(0)
t_accuracy = 100. * correct / total
average_loss = (loss / total)
return t_accuracy, average_loss
# trian_dataset's length
print(len(train_data))
# test_dataset's length
print(len(test_data))
# decoding a sequence
print(decoding_newswires(reuters, train_data[0]))
# the label is an integer between 0 and 45.
# preparing the data
x_train = vectorize_sequences(train_data)
x_test = vectorize_sequences(test_data)
# set the labels to one_hot encoding
one_hot_train_labels = to_one_hot(train_labels)
one_hot_test_labels = to_one_hot(test_labels)
# there is a built-in way to do this in keras
one_hot_train_labels_ = to_categorical(train_labels)
one_hot_test_labels_ = to_categorical(test_labels)
# model definition
def model(x, y, x_val, y_val):
model_ = models.Sequential()
model_.add(layers.Dense(64, activation='relu', input_shape=(10000,)))
model_.add(layers.Dense(64, activation='relu'))
model_.add(layers.Dense(46, activation='softmax'))
# compile
model_.compile(optimizer=optimizers.Adam(lr=0.001),
