def get_flownet_feature(concat_inputs, trainable=True):
with tf.variable_scope('FlowNetS'):
with slim.arg_scope([slim.conv2d, slim.conv2d_transpose],
# Only backprop this network if trainable
trainable=trainable,
# He (aka MSRA) weight initialization
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=LeakyReLU,
# We will do our own padding to match the original Caffe code
padding='VALID'):
weights_regularizer = slim.l2_regularizer(LONG_SCHEDULE['weight_decay'])
with slim.arg_scope([slim.conv2d], weights_regularizer=weights_regularizer):
with slim.arg_scope([slim.conv2d], stride=2):
conv_1 = slim.conv2d(pad(concat_inputs, 3), 64, 7, scope='conv1')
conv_2 = slim.conv2d(pad(conv_1, 2), 128, 5, scope='conv2')
conv_3 = slim.conv2d(pad(conv_2, 2), 256, 5, scope='conv3')
conv3_1 = slim.conv2d(pad(conv_3), 256, 3, scope='conv3_1')
with slim.arg_scope([slim.conv2d], num_outputs=512, kernel_size=3):
conv4 = slim.conv2d(pad(conv3_1), stride=2, scope='conv4')
conv4_1 = slim.conv2d(pad(conv4), scope='conv4_1')
conv5 = slim.conv2d(pad(conv4_1), stride=2, scope='conv5')
conv5_1 = slim.conv2d(pad(conv5), scope='conv5_1')
conv6 = slim.conv2d(pad(conv5_1), 1024, 3, stride=2, scope='conv6')
conv6_1 = slim.conv2d(pad(conv6), 1024, 3, scope='conv6_1')
shape = conv6_1.shape.as_list()
conv6_1 = tf.reshape(conv6_1, [-1, shape[-3]*shape[-2]*shape[-1]])
return conv6_1
def ransac(keypoints1, keypoints2, matches, n_iters=200, threshold=20):
"""
Use RANSAC to find a robust affine transformation
1. Select random set of matches
2. Compute affine transformation matrix
3. Compute inliers
4. Keep the largest set of inliers
5. Re-compute least-squares estimate on all of the inliers
Args:
keypoints1: M1 x 2 matrix, each row is a point
keypoints2: M2 x 2 matrix, each row is a point
matches: N x 2 matrix, each row represents a match
[index of keypoint1, index of keypoint 2]
n_iters: the number of iterations RANSAC will run
threshold: the number of threshold to find inliers
Returns:
H: a robust estimation of affine transformation from keypoints2 to
keypoints 1
"""
# Copy matches array, to avoid overwriting it
orig_matches = matches.copy()
matches = matches.copy()
N = matches.shape[0]
print(N)
n_samples = int(N * 0.2)
p_matched1 = pad(keypoints1[matches[:, 0]])
p_matched2 = pad(keypoints2[matches[:, 1]])
matched1 = keypoints1[matches[:, 0]]
matched2 = keypoints2[matches[:, 1]]
max_inliers = np.zeros(N)
n_inliers = 0
# RANSAC iteration start
### YOUR CODE HERE
for i in range(n_iters):
rand_indices = np.random.randint(0, N, (n_samples, ))
rand_matched1 = matched1[rand_indices]
rand_matched2 = matched2[rand_indices]
H_est = fit_affine_matrix(rand_matched1, rand_matched2)
inliers = np.sqrt(
((p_matched2.dot(H_est) - p_matched1)**2).sum(axis=-1)) < threshold
curr_n_inliers = np.count_nonzero(inliers)
if curr_n_inliers > n_inliers:
n_inliers = curr_n_inliers
max_inliers = inliers
H = fit_affine_matrix(matched1[max_inliers], matched2[max_inliers])
### END YOUR CODE
print(H)
return H, orig_matches[max_inliers]
def fit_affine_matrix(p1, p2):
""" Fit affine matrix such that p2 * H = p1
Hint:
You can use np.linalg.lstsq function to solve the problem.
Args:
p1: an array of shape (M, P)
p2: an array of shape (M, P)
Return:
H: a matrix of shape (P * P) that transform p2 to p1.
"""
assert (p1.shape[0] == p2.shape[0]),\
'Different number of points in p1 and p2'
p1 = pad(p1)
p2 = pad(p2)
#####################################
# START YOUR CODE HERE #
#####################################
H = np.linalg.lstsq(p2, p1, rcond=None)[0]
######################################
# END OF YOUR CODE #
######################################
# Sometimes numerical issues cause least-squares to produce the last
# column which is not exactly [0, 0, 1]
H[:, 2] = np.array([0, 0, 1])
return H
def fit_affine_matrix(p1, p2):
""" Fit affine matrix such that p2 * H = p1
Hint:
You can use np.linalg.lstsq function to solve the problem.
Args:
p1: an array of shape (M, P)
p2: an array of shape (M, P)
Return:
H: a matrix of shape (P * P) that transform p2 to p1.
"""
assert (p1.shape[0] == p2.shape[0]),\
'Different number of points in p1 and p2'
p1 = pad(p1)
p2 = pad(p2)
### YOUR CODE HERE
H, _, _, _ = np.linalg.lstsq(p2, p1)
### END YOUR CODE
H[:, 2] = np.array([0, 0, 1])
return H
def make_dnn_feats(fpath, noise_path, snr, P, maxlen=1339):
speech = read(fpath)
noise = read(noise_path)
noise = pad_noise(speech, noise)
blend = generate_noisy(speech, noise, snr)
mel_clean = librosa.feature.melspectrogram(y=speech,
sr=16000,
n_fft=512,
hop_length=256,
n_mels=64)
mel_noisy = librosa.feature.melspectrogram(y=blend,
sr=16000,
n_fft=512,
hop_length=256,
n_mels=64)
feats = pad(window(mel_noisy, P), maxlen).T
target = np.log(mel_clean)
mask = pad(np.ones((target.shape[0], target.shape[1])), maxlen).T
target = pad(target, maxlen).T
return {
'x': torch.tensor(feats),
't': torch.tensor(target),
'mask': torch.tensor(mask)
}
def fit_affine_matrix(p1, p2):
""" Fit affine matrix such that p2 * H = p1
Hint:
You can use np.linalg.lstsq function to solve the problem.
Args:
p1: an array of shape (M, P)
p2: an array of shape (M, P)
Return:
H: a matrix of shape (P, P) that transform p2 to p1.
"""
assert (p1.shape[0] == p2.shape[0]),\
'Different number of points in p1 and p2'
p1 = pad(p1)
p2 = pad(p2)
### YOUR CODE HERE
#np.linalg.lstsq参考以下文档：https://numpy.org/devdocs/reference/generated/numpy.linalg.lstsq.html#numpy.linalg.lstsq
H, r, rank, s = np.linalg.lstsq(p2, p1, rcond=None)
pass
### END YOUR CODE
# Sometimes numerical issues cause least-squares to produce the last
# column which is not exactly [0, 0, 1]
H[:, 2] = np.array([0, 0, 1])
return H
def _process(agent_skill, data):
# TODO: this could be more efficient
sub_skill_names = [None] + [
sub_skill.skill_name for sub_skill in agent_skill.sub_skills
]
iput = np.asarray([
utils.pad(step.arg, agent_skill.skill_model.arg_in_len) + [step.cnt] +
utils.one_hot(sub_skill_names.index(step.ret_name),
agent_skill.skill_model.num_sub) +
utils.pad(step.ret_val, agent_skill.skill_model.ret_in_len)
for step in data
])
sub = np.asarray([sub_skill_names.index(step.sub_name) for step in data])
arg = np.asarray([
utils.pad(step.sub_arg, agent_skill.skill_model.arg_out_len)
for step in data
])
return DictTree(
len=len(data),
iput=iput,
oput=DictTree(
sub=sub,
arg=arg,
),
)
def train(self, X, y, y_real, output_dir):
X_split = [normalize(t.split()) for t in X]
split_arrays = train_test_split(X_split, y, y_real, test_size=self.settings['test_size'], stratify=y_real)
X_train, X_test, y_train, y_test, y_real_train, y_real_test = split_arrays
text_field = data.Field()
text_field.build_vocab(X_train, max_size=10000)
# pad
X_train_pad = [pad(s, self.settings['max_seq_length']) for s in tqdm(X_train, desc='pad')]
X_test_pad = [pad(s, self.settings['max_seq_length']) for s in tqdm(X_test, desc='pad')]
# to index
X_train_index = [to_indexes(text_field.vocab, s) for s in tqdm(X_train_pad, desc='to index')]
X_test_index = [to_indexes(text_field.vocab, s) for s in tqdm(X_test_pad, desc='to index')]
train_dataset = self.to_dataset(X_train_index, y_train, y_real_train)
val_dataset = self.to_dataset(X_test_index, y_test, y_real_test)
model = self.model(text_field)
model.to(device())
self.full_train(model, train_dataset, val_dataset, output_dir)
torch.save(text_field, os.path.join(output_dir, self.vocab_name))
return model, text_field.vocab
def predict_turn(self, turn, ontology, args, threshold=0.5):
model, tokenizer = self.bert, self.tokenizer
batch_size = args.batch_size
was_training = model.training
self.eval()
preds = []
examples = turn_to_examples(turn, ontology, tokenizer)
for i in range(0, len(examples), batch_size):
batch = examples[i:i + batch_size]
slots, values, input_ids, token_type_ids, _ = list(zip(*batch))
# Padding and Convert to Torch Tensors
input_ids, input_masks = pad(input_ids, args.device)
token_type_ids = pad(token_type_ids, args.device)[0]
# Forward Pass
logits = model(input_ids, token_type_ids, input_masks)
probs = torch.softmax(logits, dim=-1)[:, 1].cpu().data.numpy()
# Update preds
for j in range(len(batch)):
if probs[j] >= threshold:
preds.append((slots[j], values[j]))
if was_training:
self.train()
return preds
def fit_affine_matrix(p1, p2):
"""
Fit affine matrix such that p2 * H = p1. First, pad the descriptor vectors
with a 1 using pad() to convert to homogeneous coordinates, then return
the least squares fit affine matrix in homogeneous coordinates.
Hint:
You can use np.linalg.lstsq function to solve the problem.
Explicitly specify np.linalg.lstsq's new default parameter rcond=None
to suppress deprecation warnings, and match the autograder.
Args:
p1: an array of shape (M, P) holding descriptors of size P about M keypoints
p2: an array of shape (M, P) holding descriptors of size P about M keypoints
Return:
H: a matrix of shape (P+1, P+1) that transforms p2 to p1 in homogeneous
coordinates
"""
assert (p1.shape[0] == p2.shape[0]),\
'Different number of points in p1 and p2'
p1 = pad(p1)
p2 = pad(p2)
### YOUR CODE HERE
pass
### END YOUR CODE
# Sometimes numerical issues cause least-squares to produce the last
# column which is not exactly [0, 0, 1]
H[:, 2] = np.array([0, 0, 1])
return H
def _build_frange_part(start, stop, stride, zfill=0):
"""
Private method: builds a proper and padded
:class:`fileseq.framerange.FrameRange` string.
:type start: int
:param start: first frame
:type stop: int
:param stop: last frame
:type stride: int
:param stride: increment
:type zfill: int
:param zfill: width for zero padding
:rtype: str
"""
if stop is None:
return ''
pad_start = pad(start, zfill)
pad_stop = pad(stop, zfill)
if stride is None or start == stop:
return '{0}'.format(pad_start)
elif abs(stride) == 1:
return '{0}-{1}'.format(pad_start, pad_stop)
else:
return '{0}-{1}x{2}'.format(pad_start, pad_stop, stride)
def cal_dev_loss(model, dev_data, batch_size, sent_vocab, tag_vocab, device):
""" Calculate loss on the development data
Args:
model: the model being trained
dev_data: development data
batch_size: batch size
sent_vocab: sentence vocab
tag_vocab: tag vocab
device: torch.device on which the model is trained
Returns:
the average loss on the dev data
"""
is_training = model.training
model.eval()
loss, n_sentences = 0, 0
with torch.no_grad():
for sentences, tags in utils.batch_iter(dev_data,
batch_size,
shuffle=False):
sentences, sent_lengths = utils.pad(sentences,
sent_vocab[sent_vocab.PAD],
device)
tags, _ = utils.pad(tags, tag_vocab[sent_vocab.PAD], device)
batch_loss = model(sentences, tags, sent_lengths) # shape: (b,)
loss += batch_loss.sum().item()
n_sentences += len(sentences)
model.train(is_training)
return loss / n_sentences
def fit_affine_matrix(p1, p2):
""" Fit affine matrix such that p2 * H = p1
Hint:
You can use np.linalg.lstsq function to solve the problem.
Args:
p1: an array of shape (M, P)
p2: an array of shape (M, P)
Return:
H: a matrix of shape (P * P) that transform p2 to p1.
"""
assert (p1.shape[0] == p2.shape[0]),\
'Different number of points in p1 and p2'
p1 = pad(p1)
p2 = pad(p2)
result = linalg.lstsq(p2, p1) #求解从p1变换到p2的最优变换矩阵
H = result[0]
#变换矩阵X是一个3×3的矩阵
# Sometimes numerical issues cause least-squares to produce the last
# column which is not exactly [0, 0, 1]
H[:, 2] = np.array([0, 0, 1])
return H
def visualize(self, lookahead):
source_energy = utils.pad(self.source_energy, lookahead)
assigned_energy = utils.pad(self.assigned_energy, lookahead)
# available_energy = source_energy - assigned_energy
planned_energy = utils.pad(self.planned_energy, lookahead)
consumed_energy = assigned_energy + planned_energy
fig, ax = plt.subplots()
ax.set_xlabel('ticks')
ax.set_ylabel('energy')
ax.set_ylim((-0.6, 1.25))
ax.plot(source_energy, color='green', label='available')
ax.plot(assigned_energy, color='red', label='assigned')
ax.plot(planned_energy, color='blue', label='planned')
ax.plot(consumed_energy, color='black', label='consumed')
for job in self.running_jobs:
offset = 0
duration = len(job.profile)
rect = patches.Rectangle((offset, -0.2-0.1*job.request_id), duration, 0.1, linewidth=1, edgecolor='red', facecolor='pink')
ax.add_patch(rect)
for request in self.waiting_requests:
offset = self.plan[request.request_id]
duration = len(request.profile)
rect = patches.Rectangle((offset, -0.2-0.1*request.request_id), duration, 0.1, linewidth=1, edgecolor='blue', facecolor='lightblue')
ax.add_patch(rect)
ax.legend(loc='upper right')
return fig
def ransac(keypoints1, keypoints2, matches, n_iters=200, threshold=20):
"""
Use RANSAC to find a robust affine transformation
1. Select random set of matches
2. Compute affine transformation matrix
3. Compute inliers
4. Keep the largest set of inliers
5. Re-compute least-squares estimate on all of the inliers
Args:
keypoints1: M1 x 2 matrix, each row is a point
keypoints2: M2 x 2 matrix, each row is a point
matches: N x 2 matrix, each row represents a match
[index of keypoint1, index of keypoint 2]
n_iters: the number of iterations RANSAC will run
threshold: the number of threshold to find inliers
Returns:
H: a robust estimation of affine transformation from keypoints2 to
keypoints 1
"""
# Copy matches array, to avoid overwriting it
orig_matches = matches.copy()
matches = matches.copy()
N = matches.shape[0]
n_samples = int(N * 0.2)
matched1 = keypoints1[matches[:, 0]]
matched2 = keypoints2[matches[:, 1]]
padded1 = pad(matched1)
padded2 = pad(matched2)
#max_inliers = np.zeros(N)
max_inliers = []
n_inliers = 0
# RANSAC iteration start
### YOUR CODE HERE
for i in range(n_iters):
m = [p for p in range(N)]
np.random.shuffle(m)
m = m[:n_samples]
H = fit_affine_matrix(matched1[m], matched2[m])
ins = []
for j in range(N):
trans = padded2[j] @ H
trans = np.array([trans[0] / trans[2], trans[1] / trans[2]])
di = np.linalg.norm(trans - matched1[j])
if di < threshold:
ins.append(j)
if len(ins) > n_inliers:
n_inliers = len(ins)
max_inliers = ins
H = fit_affine_matrix(matched1[max_inliers], matched2[max_inliers])
### END YOUR CODE
return H, orig_matches[max_inliers]
def _build_frange_part(start, stop, stride, zfill=0):
"""
Private method: builds a proper and padded
:class:`fileseq.framerange.FrameRange` string.
:type start: int
:param start: first frame
:type stop: int
:param stop: last frame
:type stride: int
:param stride: increment
:type zfill: int
:param zfill: width for zero padding
:rtype: str
"""
if stop is None:
return ''
pad_start = pad(start, zfill)
pad_stop = pad(stop, zfill)
if stride is None or start == stop:
return '{0}'.format(pad_start)
elif abs(stride) == 1:
return '{0}-{1}'.format(pad_start, pad_stop)
else:
return '{0}-{1}x{2}'.format(pad_start, pad_stop, stride)
def __init__(self, dataframe, test=False):
"""
dataframe: pandas.DataFrame object containing the dataset
test: boolean values describing whether or not this is the test set
"""
self.samples = [] # Will store the final data
self.test = test # Whether or not the dataset is a test set
# Convert the strings to token indices
input1, input2 = get_tokenized(dataframe)
# Pad data according to the longest sentence
x1, x1_mask = pad(input1)
x2, x2_mask = pad(input2)
# Build the tuple structure to be used in training and testing
for idx, _ in enumerate(tqdm(range(len(x1)), desc="Loading Data", leave=False)):
sample = {
"x1": x1[idx],
"x1_mask": x1_mask[idx],
"x2": x2[idx],
"x2_mask": x2_mask[idx],
}
# The score feature is only available for test and dev sets
if not self.test:
sample["score"] = dataframe.iloc[idx].scores
self.samples.append(sample)
def backward(self, previousGrad):
X = self.input
batch_size = np.shape(self.input)[-1]
dX = np.zeros(np.shape(self.input))
self.grads = np.zeros(np.shape(self.parameters))
X_pad = utils.pad(X, self.pad_w, self.pad_h)
dX_pad = utils.pad(dX, self.pad_w, self.pad_h)
for h in range(self.n_H):
for w in range(self.n_W):
y_start = self.stride * h
y_end = y_start + self.kernel_shape[1]
x_start = self.stride * w
x_end = x_start + self.kernel_shape[0]
x_slice = X_pad[x_start:x_end, y_start:y_end, :, :]
dX_pad[x_start:x_end, y_start:y_end, :, :] += np.dot(
self.parameters[:, :, :, :], previousGrad[w, h, :, :])
self.gradient[:, :, :, :] += np.dot(x_slice,
previousGrad[w, h, :, :].T)
self.gradient = self.gradient / batch_size
for parent in self.parents:
parent.sonsVisited += 1
parent.backward(dX)
def make(cls, example, tokenizer, max_len):
sys_utt_tokens, usr_utt_tokens = example.utt
joint_sys_utt_label = np.array(
[x for x, _, _ in example.label_dict.values()]).any(axis=0)
joint_usr_utt_label = np.array(
[x for _, x, _ in example.label_dict.values()]).any(axis=0)
class_label_id_dict, start_pos_dict, end_pos_dict = {}, {}, {}
for slot, (sys_utt_label, usr_utt_label,
class_type) in example.label_dict.items():
sys_tokens, sys_token_labels = \
sub_tokenize(sys_utt_tokens, sys_utt_label, joint_sys_utt_label, tokenizer)
usr_tokens, usr_token_labels = \
sub_tokenize(usr_utt_tokens, usr_utt_label, joint_usr_utt_label, tokenizer)
# TO DO : truncate
token_labels = utils.pad(max_len, 0, [0] + sys_token_labels + [0] +
usr_token_labels)
class_label_id_dict[slot] = class_types.index(class_type)
start_pos_dict[slot], end_pos_dict[slot] = \
get_start_end_pos(class_type, token_labels, max_len)
tokens = ['[CLS]'] + sys_tokens + ['[SEP]'] + usr_tokens + ['[SEP]']
input_ids = tokenizer.convert_tokens_to_ids(tokens)
input_mask = [1] * len(input_ids)
segment_ids = [0] * (2 + len(sys_tokens)) + [1] * (1 + len(usr_tokens))
utils.pad(max_len, 0, input_ids, input_mask, segment_ids)
return cls(example.guid, tokens, input_ids, input_mask, segment_ids,
start_pos_dict, end_pos_dict, class_label_id_dict)
def __init__(self, in_channel, hidden_size, dim):
'''
:param in_channel: 表示输入的特征通道数
:param hidden_size: 表示隐藏层的神经元个数
:param dim: 表示光谱维的长度
'''
super().__init__()
self.dim = dim
m = list()
padding_0 = pad(dim, 5, 3)
m.append(
nn.Conv3d(in_channel,
hidden_size, (5, 3, 3), (3, 1, 1),
padding=(padding_0, 0, 0)))
m.append(nn.ReLU())
m.append(nn.BatchNorm3d(hidden_size))
dim_1 = (dim - 5 + 2 * padding_0) // 3 + 1
padding_1 = pad(dim_1, 5, 3)
m.append(
nn.Conv3d(hidden_size,
hidden_size, (5, 3, 3), (3, 1, 1),
padding=(padding_1, 0, 0)))
m.append(nn.ReLU())
m.append(nn.BatchNorm3d(hidden_size))
# dim_2 = (dim_1 - 5 + 2 * padding_1) // 3 + 1
# padding_2 = pad(dim_2, 5, 3)
# m.append(nn.Conv3d(hidden_size, hidden_size, (5, 3, 3), (3, 1, 1), padding=(padding_2, 0, 0)))
# m.append(nn.ReLU())
# m.append(nn.BatchNorm3d(hidden_size))
self.out_dim = (dim_1 - 5 + 2 * padding_1) // 3 + 1
self.encoder = nn.Sequential(*m)
def __init__(self, out_channel, hidden_size, dim):
super().__init__()
padding_0 = pad(dim, 5, 3)
dim_1 = (dim - 5 + 2 * padding_0) // 3 + 1
padding_1 = pad(dim_1, 5, 3)
dim_2 = (dim_1 - 5 + 2 * padding_1) // 3 + 1
padding_2 = pad(dim_2, 5, 3)
m = list()
m.append(
nn.ConvTranspose3d(hidden_size,
hidden_size, (5, 3, 3), (3, 1, 1),
padding=(padding_2, 0, 0)))
m.append(nn.ReLU())
m.append(nn.BatchNorm3d(hidden_size))
m.append(
nn.ConvTranspose3d(hidden_size,
hidden_size, (5, 3, 3), (3, 1, 1),
padding=(padding_1, 0, 0)))
m.append(nn.ReLU())
m.append(nn.BatchNorm3d(hidden_size))
m.append(
nn.ConvTranspose3d(hidden_size,
out_channel, (5, 3, 3), (3, 1, 1),
padding=(padding_0, 0, 0)))
m.append(nn.ReLU())
m.append(nn.BatchNorm3d(out_channel))
self.decoder = nn.Sequential(*m)
def ransac(keypoints1, keypoints2, matches, n_iters=200, threshold=20):
"""
Use RANSAC to find a robust affine transformation
1. Select random set of matches
2. Compute affine transformation matrix
3. Compute inliers
4. Keep the largest set of inliers
5. Re-compute least-squares estimate on all of the inliers
Args:
keypoints1: M1 x 2 matrix, each row is a point
keypoints2: M2 x 2 matrix, each row is a point
matches: N x 2 matrix, each row represents a match
[index of keypoint1, index of keypoint 2]
n_iters: the number of iterations RANSAC will run
threshold: the number of threshold to find inliers
Returns:
H: a robust estimation of affine transformation from keypoints2 to
keypoints 1
"""
N = matches.shape[0]
n_samples = int(N * 0.2)
matched1 = pad(keypoints1[matches[:, 0]])
matched2 = pad(keypoints2[matches[:, 1]])
max_inliers = np.zeros(N)
n_inliers = 0
# RANSAC iteration start
### YOUR CODE HERE
for o in np.arange(n_iters):
# step1:
random_select_matches = matches[
np.random.choice(np.arange(N), n_samples, replace=False), :]
# step2:
p1 = keypoints1[random_select_matches[:, 0]]
p2 = keypoints2[random_select_matches[:, 1]]
H = fit_affine_matrix(p1, p2)
# step3:
diff = matched2.dot(H) - matched1
temp_inliers = np.linalg.norm(diff, axis=1)**2 < threshold
temp_n_inliers = temp_inliers.sum()
# step4:
if temp_n_inliers > n_inliers:
n_inliers = temp_n_inliers
max_inliers = temp_inliers
max_matches = matches[max_inliers, :]
p1 = keypoints1[max_matches[:, 0]]
p2 = keypoints2[max_matches[:, 1]]
H = fit_affine_matrix(p1, p2)
### END YOUR CODE
return H, matches[max_inliers]
def ransac(keypoints1, keypoints2, matches, n_iters=200, threshold=20):
"""
Use RANSAC to find a robust affine transformation
1. Select random set of matches
2. Compute affine transformation matrix
3. Compute inliers
4. Keep the largest set of inliers
5. Re-compute least-squares estimate on all of the inliers
Args:
keypoints1: M1 x 2 matrix, each row is a point
keypoints2: M2 x 2 matrix, each row is a point
matches: N x 2 matrix, each row represents a match
[index of keypoint1, index of keypoint 2]
n_iters: the number of iterations RANSAC will run
threshold: the number of threshold to find inliers
Returns:
H: a robust estimation of affine transformation from keypoints2 to
keypoints 1
"""
N = matches.shape[0]
n_samples = int(N * 0.2)
matched1 = pad(keypoints1[matches[:,0]])
matched2 = pad(keypoints2[matches[:,1]])
max_inliers = []
n_inliers = 0
# RANSAC iteration start
H = None
for n in range(n_iters):
matches_indices = np.random.randint(N, size = n_samples)
p1 = keypoints1[matches[matches_indices,0]]
p2 = keypoints2[matches[matches_indices,1]]
affine = fit_affine_matrix(p1, p2)
dot_temp = np.dot(matched2, affine)
temp_inliers = []
for i in range(N):
point_trans = dot_temp[i]
point_orig = matched1[i]
distance = np.linalg.norm(point_trans - point_orig)
if distance < threshold:
temp_inliers.append(i)
if len(temp_inliers) > len(max_inliers):
max_inliers = temp_inliers
H = affine
return H, matches[max_inliers]
def ransac(keypoints1, keypoints2, matches, n_iters=200, threshold=20):
"""
Use RANSAC to find a robust affine transformation
1. Select random set of matches
2. Compute affine transformation matrix
3. Compute inliers
4. Keep the largest set of inliers
5. Re-compute least-squares estimate on all of the inliers
Args:
keypoints1: M1 x 2 matrix, each row is a point
keypoints2: M2 x 2 matrix, each row is a point
matches: N x 2 matrix, each row represents a match
[index of keypoint1, index of keypoint 2]
n_iters: the number of iterations RANSAC will run
threshold: the number of threshold to find inliers
Returns:
H: a robust estimation of affine transformation from keypoints2 to
keypoints 1
"""
N = matches.shape[0]
n_samples = int(N * 0.2)
matched1 = pad(keypoints1[matches[:, 0]])
matched2 = pad(keypoints2[matches[:, 1]])
max_inliers = np.zeros(N)
n_inliers = 0
# RANSAC iteration start
### YOUR CODE HERE
t = 0
while t < n_iters:
nn_inliers = 0
index_sample = [random.randint(0, N - 1) for _ in range(n_samples)]
p1 = matched1[index_sample, :]
p2 = matched2[index_sample, :]
H = np.linalg.lstsq(p2, p1)[0]
H[:, 2] = np.array([0, 0, 1])
test1 = np.dot(matched2, H)
temp = (test1 - matched1)**2
distance = np.sqrt(temp[:, 0] + temp[:, 1])
print(distance)
for i in range(distance.shape[0]):
if distance[i] <= threshold:
nn_inliers += 1
if nn_inliers > n_inliers:
n_inliers = nn_inliers
max_inliers = np.where(distance <= threshold)[0]
t += 1
p1 = matched1[max_inliers]
p2 = matched2[max_inliers]
H = np.linalg.lstsq(p2, p1)[0]
H[:, 2] = np.array([0, 0, 1])
pass
### END YOUR CODE
return H, matches[max_inliers]
def ransac(keypoints1, keypoints2, matches, n_iters=200, threshold=20):
"""
Use RANSAC to find a robust affine transformation
1. Select random set of matches
2. Compute affine transformation matrix
3. Compute inliers
4. Keep the largest set of inliers
5. Re-compute least-squares estimate on all of the inliers
Args:
keypoints1: M1 x 2 matrix, each row is a point
keypoints2: M2 x 2 matrix, each row is a point
matches: N x 2 matrix, each row represents a match
[index of keypoint1, index of keypoint 2]
n_iters: the number of iterations RANSAC will run
threshold: the number of threshold to find inliers
Returns:
H: a robust estimation of affine transformation from keypoints2 to
keypoints 1
"""
# Copy matches array, to avoid overwriting it
orig_matches = matches.copy()
matches = matches.copy()
N = matches.shape[0]
print(N)
n_samples = int(N * 0.2)
matched1 = pad(keypoints1[matches[:, 0]])
matched2 = pad(keypoints2[matches[:, 1]])
max_inliers = np.zeros(N)
n_inliers = 0
# RANSAC iteration start
### YOUR CODE HERE
for i in range(n_iters):
temp_max = np.zeros(N, dtype=np.int32)
temp_n = 0
idx = np.random.choice(N, n_samples, replace=False)
p1 = matched1[idx, :]
p2 = matched2[idx, :]
H = np.linalg.lstsq(p2, p1, rcond=None)[0]
H[:, 2] = np.array([0, 0, 1])
temp_max = np.linalg.norm(matched2.dot(H) - matched1,
axis=1)**2 < threshold
temp_n = np.sum(temp_max)
if temp_n > n_inliers:
max_inliers = temp_max.copy()
n_inliers = temp_n
H = np.linalg.lstsq(matched2[max_inliers],
matched1[max_inliers],
rcond=None)[0]
H[:, 2] = np.array([0, 0, 1])
### END YOUR CODE
print(H)
return H, orig_matches[max_inliers]
def ransac(keypoints1, keypoints2, matches, n_iters=200, threshold=20):
"""
Use RANSAC to find a robust affine transformation
1. Select random set of matches
2. Compute affine transformation matrix
3. Compute inliers
4. Keep the largest set of inliers
5. Re-compute least-squares estimate on all of the inliers
Args:
keypoints1: M1 x 2 matrix, each row is a point
keypoints2: M2 x 2 matrix, each row is a point
matches: N x 2 matrix, each row represents a match
[index of keypoint1, index of keypoint 2]
n_iters: the number of iterations RANSAC will run
threshold: the number of threshold to find inliers
Returns:
H: a robust estimation of affine transformation from keypoints2 to
keypoints 1
"""
N = matches.shape[0]
n_samples = int(N * 0.2)
matched1 = pad(keypoints1[matches[:, 0]])
matched2 = pad(keypoints2[matches[:, 1]])
max_inliers = np.zeros(N)
n_inliers = 0
H = None
# RANSAC iteration start
### YOUR CODE HERE
pass
for i in range(0, n_iters):
#1. Select random set of matches
random_indices = random.sample(list(range(N)), n_samples)
chosen_kp1 = matched1[random_indices]
chosen_kp2 = matched2[random_indices]
#2. Compute affine transformation matrix
affine_transformation_matrix = np.linalg.lstsq(chosen_kp2,
chosen_kp1)[0]
affine_transformation_matrix[:, 2] = [0, 0, 1]
#3. Compute number of inlier
real_matched1 = np.dot(matched2, affine_transformation_matrix)
diff = real_matched1 - matched1
dists = np.linalg.norm(diff, axis=1)
curr_inliers = len(np.argwhere(dists < threshold))
curr_max_inliers = np.where(dists < threshold)
#4. Keep the largest set of inliers
if curr_inliers > n_inliers:
n_inliers = curr_inliers
H = affine_transformation_matrix
max_inliers = curr_max_inliers
#5. Re-compute least-squares estimate on all of the inliers
H = np.linalg.lstsq(matched2[max_inliers], matched1[max_inliers])[0]
H[:, 2] = np.array([0, 0, 1])
### END YOUR CODE
return H, matches[max_inliers]
def ransac(keypoints1, keypoints2, matches, n_iters=200, threshold=20):
"""
Use RANSAC to find a robust affine transformation
1. Select random set of matches
2. Compute affine transformation matrix
3. Compute inliers
4. Keep the largest set of inliers
5. Re-compute least-squares estimate on all of the inliers
Args:
keypoints1: M1 x 2 matrix, each row is a point
keypoints2: M2 x 2 matrix, each row is a point
matches: N x 2 matrix, each row represents a match
[index of keypoint1, index of keypoint 2]
n_iters: the number of iterations RANSAC will run
threshold: the number of threshold to find inliers
Returns:
H: a robust estimation of affine transformation from keypoints2 to
keypoints 1
"""
N = matches.shape[0]
choices = list(range(N))
n_samples = int(N * 0.2)
matched1 = pad(keypoints1[matches[:,0]])
matched2 = pad(keypoints2[matches[:,1]])
max_inliers = np.zeros(N)
n_inliers = 0
# RANSAC iteration start
### YOUR CODE HERE
for i in range(n_iters):
# 1 select random set of matches
sample = np.random.choice(choices, n_samples, replace=False)
# 2 fitting the the set of p1 to p2
p1 = matched1[sample,:]
p2 = matched2[sample,:]
#fitt the matrix
mat= np.linalg.lstsq(p2, p1)[0]
#compute the image of rest of the point
image_matched1= matched2.dot(mat)
# compute the number of inlinier
dists = np.linalg.norm(image_matched1-matched1,axis=1)
inliers= (dists < threshold)
if(inliers.sum()> n_inliers):
n_inliers = inliers.sum()
H= mat
max_inliers=inliers
### END YOUR CODE
return H, matches[max_inliers]
def ransac(keypoints1, keypoints2, matches, n_iters=200, threshold=20):
"""
Use RANSAC to find a robust affine transformation:
1. Select random set of matches
2. Compute affine transformation matrix
3. Compute inliers via Euclidean distance
4. Keep the largest set of inliers
5. Re-compute least-squares estimate on all of the inliers
Update max_inliers as a boolean array where True represents the keypoint
at this index is an inlier, while False represents that it is not an inlier.
Hint:
You can use np.linalg.lstsq function to solve the problem.
Explicitly specify np.linalg.lstsq's new default parameter rcond=None
to suppress deprecation warnings, and match the autograder.
You can compute elementwise boolean operations between two numpy arrays,
and use boolean arrays to select array elements by index:
https://numpy.org/doc/stable/reference/arrays.indexing.html#boolean-array-indexing
Args:
keypoints1: M1 x 2 matrix, each row is a point
keypoints2: M2 x 2 matrix, each row is a point
matches: N x 2 matrix, each row represents a match
[index of keypoint1, index of keypoint 2]
n_iters: the number of iterations RANSAC will run
threshold: the number of threshold to find inliers
Returns:
H: a robust estimation of affine transformation from keypoints2 to
keypoints 1
"""
# Copy matches array, to avoid overwriting it
orig_matches = matches.copy()
matches = matches.copy()
N = matches.shape[0]
print(N)
n_samples = int(N * 0.2)
matched1 = pad(keypoints1[matches[:,0]])
matched2 = pad(keypoints2[matches[:,1]])
max_inliers = np.zeros(N, dtype=bool)
n_inliers = 0
# RANSAC iteration start
# Note: while there're many ways to do random sampling, please use `np.random.shuffle()`
# followed by slicing out the first `n_samples` matches here in order to align with the auto-grader.
### YOUR CODE HERE
pass
### END YOUR CODE
print(H)
return H, orig_matches[max_inliers]
def ransac(keypoints1, keypoints2, matches, n_iters=200, threshold=20):
"""
Use RANSAC to find a robust affine transformation
1. Select random set of matches
2. Compute affine transformation matrix
3. Compute inliers
4. Keep the largest set of inliers
5. Re-compute least-squares estimate on all of the inliers
Args:
keypoints1: M1 x 2 matrix, each row is a point
keypoints2: M2 x 2 matrix, each row is a point
matches: N x 2 matrix, each row represents a match
[index of keypoint1, index of keypoint 2]
n_iters: the number of iterations RANSAC will run
threshold: the number of threshold to find inliers
Returns:
H: a robust estimation of affine transformation from keypoints2 to
keypoints 1
"""
N = matches.shape[0]
n_samples = int(N * 0.2)
matched1 = pad(keypoints1[matches[:,0]])
matched2 = pad(keypoints2[matches[:,1]])
max_inliers = np.zeros(N, dtype=bool)
n_inliers = 0
# RANSAC iteration start
### YOUR CODE HERE
allIdxs = np.arange(N)
for iterTime in range(n_iters):
np.random.shuffle(allIdxs)
sampleIdxs = allIdxs[0:n_samples]
outIdxs = allIdxs[n_samples:]
sample_p1s = matched1[sampleIdxs]
sample_p2s = matched2[sampleIdxs]
out_p1s = matched1[outIdxs]
out_p2s = matched2[outIdxs]
tempH = np.linalg.lstsq(sample_p2s,sample_p1s)[0]
tempH[:,2] = np.array([0, 0, 1])
errors = np.sum( (np.dot(out_p2s,tempH) - out_p1s)**2 ,axis = 1)
otherInliers = outIdxs[errors < threshold]
totalInliers = len(otherInliers) + n_samples
if totalInliers > n_inliers:
max_inliers[:] = False
n_inliers = totalInliers
recomputeIdxs = np.concatenate( (sampleIdxs,otherInliers) )
max_inliers[recomputeIdxs] = True
recompute_p1s = matched1[recomputeIdxs]
recompute_p2s = matched2[recomputeIdxs]
H = np.linalg.lstsq(recompute_p2s,recompute_p1s)[0]
H[:,2] = np.array([0, 0, 1])
### END YOUR CODE
return H, matches[max_inliers]
def ransac(keypoints1, keypoints2, matches, n_iters=200, threshold=20):
"""
Use RANSAC to find a robust affine transformation
1. Select random set of matches
2. Compute affine transformation matrix
3. Compute inliers
4. Keep the largest set of inliers
5. Re-compute least-squares estimate on all of the inliers
Args:
keypoints1: M1 x 2 matrix, each row is a point
keypoints2: M2 x 2 matrix, each row is a point
matches: N x 2 matrix, each row represents a match
[index of keypoint1, index of keypoint 2]
n_iters: the number of iterations RANSAC will run
threshold: the number of threshold to find inliers
Returns:
H: a robust estimation of affine transformation from keypoints2 to
keypoints 1
"""
N = matches.shape[0]
n_samples = int(N * 0.2)
matched1 = pad(keypoints1[matches[:, 0]])
matched2 = pad(keypoints2[matches[:, 1]])
max_inliers = np.zeros(N)
n_inliers = 0
# RANSAC iteration start
### YOUR CODE HERE
for i in range(n_iters):
inliersArr = np.zeros(N)
idx = np.random.choice(N, n_samples, replace=False)
p1 = matched1[idx, :]
p2 = matched2[idx, :]
H, residuals, rank, s = np.linalg.lstsq(p2, p1, rcond=None)
H[:, 2] = np.array([0, 0, 1])
output = np.dot(matched2, H)
inliersArr = np.linalg.norm(output - matched1, axis=1)**2 < threshold
inliersCount = np.sum(inliersArr)
if inliersCount > n_inliers:
max_inliers = inliersArr.copy() #还是注意深拷贝和浅拷贝啊
n_inliers = inliersCount
# 迭代完成，拿最大数目的匹配点对进行估计变换矩阵
H, residuals, rank, s = np.linalg.lstsq(matched2[max_inliers],
matched1[max_inliers],
rcond=None)
H[:, 2] = np.array([0, 0, 1])
### END YOUR CODE
return H, matches[max_inliers]
def __getitem__(self, idx):
img_path = self.img_paths[idx]
mask_path = self.mask_paths[idx]
image = imread(img_path)
mask = imread(mask_path)
image = image.astype('float32') / 255
if 'train' in img_path:
mask = (mask>65535/2).astype('float32')
else:
if self.args.loss == 'LovaszHingeLoss':
mask = (mask>127).astype('float32')
else:
mask = mask.astype('float32') / 255
if not self.args.pad:
image = cv2.resize(image, (self.args.img_size, self.args.img_size))
mask = cv2.resize(mask, (self.args.img_size, self.args.img_size))
else:
image = pad(image, self.args.img_size)
mask = pad(mask, self.args.img_size)
if self.aug:
if self.args.aug == 1:
image, mask = random_fliplr(image, mask)
elif self.args.aug == 2:
image, mask = random_fliplr(image, mask)
image = random_erase(image)
elif self.args.aug == 3:
image, mask = random_fliplr(image, mask)
image, mask = random_shift(image, mask, wrg=0.1, hrg=0.1, fill_mode='nearest')
image = random_erase(image)
if 'Res' in self.args.arch:
means = [0.485, 0.456, 0.406]
stds = [0.229, 0.224, 0.225]
for i in range(3):
image[:,:,i] = (image[:,:,i] - means[i]) / stds[i]
if 'Inception' in self.args.arch:
means = [0.5, 0.5, 0.5]
stds = [0.5, 0.5, 0.5]
for i in range(3):
image[:,:,i] = (image[:,:,i] - means[i]) / stds[i]
if self.args.depth:
image = depth_encode(image)
if self.args.coord_conv:
image = coord_conv(image)
image = image.transpose((2, 0, 1))
mask = mask[:,:,np.newaxis]
mask = mask.transpose((2, 0, 1))
return image, mask
def _do_pad(match):
"""
Substitutes padded for unpadded frames.
"""
result = list(match.groups())
result[1] = pad(result[1], zfill)
if result[4]:
result[4] = pad(result[4], zfill)
return ''.join((i for i in result if i))
def framesToFrameRange(frames, sort=True, zfill=0, compress=False):
"""
Converts an iterator of frames into a
:class:`fileseq.framerange.FrameRange`.
:type frames: iterable
:param frames: sequence of frames to process
:type sort: bool
:param sort: sort the sequence before processing
:type zfill: int
:param zfill: width for zero padding
:type compress: bool
:param compress: remove any duplicates before processing
:rtype: str
"""
if compress:
frames = unique(set(), frames)
frames = list(frames)
if not frames:
return ''
if len(frames) == 1:
return pad(frames[0], zfill)
if sort:
frames.sort()
return ','.join(FrameSet.framesToFrameRanges(frames, zfill))
def guess(self, words):
text = ' '.join(words)
text = utils.pad(text, self.maxlen)
if self.invert:
text = text[::-1]
input_vec = self.input_codec.encode(text)
preds = self.model.predict_classes(np.array([input_vec]), verbose=0)
return self.codec.decode(preds[0], calc_argmax=False).strip('\x00')
def print_raid(used_devices):
output = [[' Raid '], []]
raids = get_raids()
for raid_name, raid in raids.iteritems():
used_devices.update({'/dev/' + n: 'R' for n in raid['disks']})
output.append([raid_name, '[' + raid['personality'] + ']', str(raid['status']['non_degraded_disks']) + '/' + str(raid['status']['raid_disks'])])
for d in raid['disks']:
output.append([' ' + '/dev/' + d])
return pad(output)
def print_mounts(mounts):
output = [' Mounts ', '']
mnts = []
for fs_name, mount in sorted(mounts.iteritems(), key=lambda k: k[0]):
mnts.append([fs_name, mount['fs_type'], mount['total'], mount['used']])
mnts.append([' ' + mount['mount_point']])
output += pad(mnts)
return output
def infer(data_dir, model_name, sentence=None):
"""
"""
# Load components
with open(os.path.join(basedir, data_dir, 'char2index.json'), 'r') as f:
char2index = json.load(f)
with open(os.path.join(basedir, data_dir, 'index2class.json'), 'r') as f:
index2class = json.load(f)
# Enter the sentence
print ("Classes:", index2class.values())
if not sentence:
sentence = input("Please enter the sentence: ")
# Normalize the sentece
sentence = normalize_string(sentence)
# Convert sentence(s) to indexes
input_ = convert_sentence(
sentence=sentence,
char2index=char2index,
)
# Convert to model input
input_, _ = pad(
inputs=[input_],
char2index=char2index,
max_length=len(input_),
)
# Convert to Variable
X = Variable(torch.FloatTensor(input_), requires_grad=False)
# Load the model
model = torch.load(os.path.join(
basedir, "data", data_dir.split("/")[-1], model_name))
# Feed through model
model.eval()
scores = model(X)
probabilities = F.softmax(scores)
# Sorted probabilities
sorted_, indices = torch.sort(probabilities, descending=True)
for index in indices[0]:
print ("%s - %i%%" % (
index2class[str(index.data[0])],
100.0*probabilities.data[0][index.data[0]]))
def print_devices(used_devices):
output = [' Devices ', '']
for dev in sorted(parted.getAllDevices(), key=lambda x: x.path):
output.append(dev.path + ' ' + dev.model + ' ' + human_number(dev.getSize(), 'MiB'))
try:
partitions = []
disk = parted.Disk(dev)
for partition in disk.partitions:
used = ' ' + used_devices.get(partition.path, '') + ' '
partitions.append([used, partition.path, human_number(partition.getLength() * partition.geometry.device.sectorSize)])
if partitions:
output += pad(partitions)
except parted.DiskLabelException:
pass
return output
def print_lvm(mounts, used_devices):
output = [' LVM ', '']
lvm = LVM()
for vg in sorted(lvm.vgscan(), key=lambda x: x.name):
output.append(' '.join([vg.name, human_number(vg.size('KiB'), 'KiB'), '(' + str(int((1 - vg.free_size() / vg.size()) * 100)) + '%)']))
vs = []
for pv in sorted(vg.pvscan(), key=lambda x: x.name):
name = pv.name
used_devices[name] = 'L'
vs.append([' P ', name, human_number(pv.size('KiB'), 'KiB'), '(' + str(int((1 - pv.free() / pv.size()) * 100)) + '%)'])
for lv in sorted(vg.lvscan(), key=lambda x: x.name):
used = mounts.get('/dev/mapper/' + vg.name + '-' + lv.name, {'used': ''})['used']
vs.append([' L ', lv.name, human_number(lv.size('KiB'), 'KiB'), used])
output += pad(vs)
return output