def zero_state(self, batch_size, dtype):
with tf.variable_scope('init', reuse=self.reuse):
# 读取权重的初始化
read_vector_list = [
expand(tf.tanh(learned_init(self.memory_vector_dim)),
dim=0,
N=batch_size) for i in range(self.read_head_num)
]
# 写入权重的初始化
w_list = [
expand(tf.nn.softmax(learned_init(self.memory_size)),
dim=0,
N=batch_size)
for i in range(self.read_head_num + self.write_head_num)
]
# RNN初始化
controller_init_state = self.controller.zero_state(
batch_size, dtype)
# 存储单元初始
M = expand(tf.get_variable(
'init_M', [self.memory_size, self.memory_vector_dim],
initializer=tf.constant_initializer(1e-6)),
dim=0,
N=batch_size)
# 前面定义的：NTMControllerState = collections.namedtuple(
# 'NTMControllerState', ('controller_state', 'read_vector_list',
# 'w_list', 'M'))
return NTMControllerState(controller_state=controller_init_state,
read_vector_list=read_vector_list,
w_list=w_list,
M=M)
def train(self):
self.costs = []
gradients = []
for i, (char, target) in enumerate(zip(self.text, self.text[1:])):
self.x = expand(np.eye(self.x_size)[self.char_to_i[char]])
self.target = expand(np.eye(self.x_size)[self.char_to_i[target]])
self.y = self.forward_pass()
gradients.append(self.backward_pass())
if (i + 1) % self.batch_size == 0:
dCdWxh, dCdWhh, dCdWhy, dCdBh, dCdBy = np.average(gradients,
axis=0)
self.W_xh -= self.learning_rate * dCdWxh
self.W_hh -= self.learning_rate * dCdWhh
self.W_hy -= self.learning_rate * dCdWhy
self.B_h -= self.learning_rate * dCdBh.T
self.B_y -= self.learning_rate * dCdBy.T
self.costs.append(cost(self.target, self.y))
self.diagnose(i)
plt.plot(self.costs)
plt.show(block=False)
time.sleep(1000)
def zero_state(self, batch_size, dtype):
with tf.compat.v1.variable_scope('init', reuse=self.reuse):
read_vector_list = [expand(tf.tanh(learned_init(self.memory_vector_dim)), dim=0, N=batch_size)
for i in range(self.read_head_num)]
w_list = [expand(tf.nn.softmax(learned_init(self.memory_size)), dim=0, N=batch_size)
for i in range(self.read_head_num + self.write_head_num)]
controller_init_state = self.controller.zero_state(batch_size, dtype)
if self.init_mode == 'learned':
M = expand(tf.tanh(
tf.reshape(
learned_init(self.memory_size * self.memory_vector_dim),
[self.memory_size, self.memory_vector_dim])
), dim=0, N=batch_size)
elif self.init_mode == 'random':
M = expand(
tf.tanh(tf.get_variable('init_M', [self.memory_size, self.memory_vector_dim],
initializer=tf.random_normal_initializer(mean=0.0, stddev=0.5))),
dim=0, N=batch_size)
elif self.init_mode == 'constant':
M = expand(
tf.get_variable('init_M', [self.memory_size, self.memory_vector_dim],
initializer=tf.constant_initializer(1e-6)),
dim=0, N=batch_size)
return NTMControllerState(
controller_state=controller_init_state,
read_vector_list=read_vector_list,
w_list=w_list,
M=M)
def system_generator(x0, z0, dt):
t = 0
x = x0
z = z0
while True:
xx = delta_matrix(x)
xx_norm = expand(norm(xx))
zz = delta_matrix(z)
zz_norm = expand(norm(zz))
xz = delta_matrix(x, z)
zx = -np.moveaxis(xz, 0, 1)
xz_norm = norm(xz)
zx_norm = xz_norm.T
xz_norm = expand(xz_norm)
zx_norm = expand(zx_norm)
with np.errstate(divide='ignore', invalid='ignore'):
vx = np.nansum(prey_social(xx_norm) / xx_norm * xx, 1) / N + \
np.nansum(prey_predator(xz_norm) / xz_norm * xz, 1) / N2
vz = np.nansum(predator_social(zz_norm) / zz_norm * zz, 1) / N2 + \
np.nansum(predator_prey(zx_norm) / zx_norm * zx, 1) / N
yield t, x, z, vx, vz
x = x + vx * dt
z = z + vz * dt
t += dt
def system_generator(x0, dt):
x = x0
t = 0
dB = 0
while True:
xx = delta_matrix(x)
r_xx = norm(xx, keepdims=True)
# Interaction among individuals
with np.errstate(divide='ignore', invalid='ignore'):
v = np.nansum(F(r_xx, a) / r_xx * xx, axis=1) / N
# Environment influence
if barr_type != NO_BARR:
# Add env
if gate_len != 0:
# Point repulsion
c = (0, 1.5)
xc = x - c
v += .5 / norm(xc, keepdims=True) ** 2 * xc
else:
# Gravity
v += (0, - .1)
# dx without barrier consideration
# Plus random diffusion
dx = v * dt + mu * dB
# Now consider different barrier
if barr_type != NO_BARR:
x_next = x + dx
# barrier with EPS (vague judgement, not so strict)
pr = x[:, 1]
nx = x_next[:, 1]
invalid = (pr > BARR) & (nx < BARR + EPS) # Upper to lower
invalid |= (pr < BARR) & (nx > BARR - EPS) # Lower to upper
if barr_type == ABSORB:
dx *= ~ expand(invalid)
else:
# Reflecting barrier
if gate_len != 0:
# Exclude those via gate: if before or after are within gate
exclude = (np.abs(x_next[:, 0] - 0) < gate_len /
2) | (np.abs(x[:, 0] - 0) < gate_len / 2)
invalid &= ~ exclude
bdd = np.hstack((np.tile(False, (N, 1)), expand(invalid)))
dx = np.where(bdd, - dx, dx)
v = dx / dt
yield t, x, v
t += dt
# Do not use +=, as it modifies the mutable variable x (by `__iadd__` method)
x = x + dx
if mu != 0:
dB = np.random.randn(N, 2)
def _build_model(self):
if args.mann == 'none':
def single_cell(num_units):
return tf.contrib.rnn.BasicLSTMCell(num_units, forget_bias=1.0)
cell = tf.contrib.rnn.OutputProjectionWrapper(
tf.contrib.rnn.MultiRNNCell([
single_cell(args.num_units) for _ in range(args.num_layers)
]),
args.num_bits_per_vector,
activation=None)
initial_state = tuple(
tf.contrib.rnn.LSTMStateTuple(
c=expand(tf.tanh(learned_init(args.num_units)),
dim=0,
N=args.batch_size),
h=expand(tf.tanh(learned_init(args.num_units)),
dim=0,
N=args.batch_size))
for _ in range(args.num_layers))
elif args.mann == 'ntm':
cell = NTMCell(args.num_layers,
args.num_units,
args.num_memory_locations,
args.memory_size,
args.num_read_heads,
args.num_write_heads,
addressing_mode='content_and_location',
shift_range=args.conv_shift_range,
reuse=False,
output_dim=args.num_bits_per_vector,
clip_value=args.clip_value,
init_mode=args.init_mode)
initial_state = cell.zero_state(args.batch_size, tf.float32)
output_sequence, _ = tf.nn.dynamic_rnn(cell=cell,
inputs=self.inputs,
time_major=False,
initial_state=initial_state)
if args.task == 'copy':
self.output_logits = output_sequence[:, self.max_seq_len + 1:, :]
elif args.task == 'associative_recall':
self.output_logits = output_sequence[:,
3 * (self.max_seq_len + 1) +
2:, :]
if args.task in ('copy', 'associative_recall'):
self.outputs = tf.sigmoid(self.output_logits)
def _refresh(self, uid, params):
s, api_url, params = prepare_request(params)
url = api_url + self.instance_url(uid=uid)
r = s.get(url, params=utils.expand(params))
r.raise_for_status()
self._refresh_from(r.json())
return self
def loss_fun(self, inputs, targets):
tanh_arg, softmax, h = {}, {}, {-1: self.h_prev}
dwxh, dwhh, dwhy = np.zeros_like(self.wxh), np.zeros_like(
self.whh), np.zeros_like(self.why)
dbh, dby = np.zeros_like(self.bh), np.zeros_like(self.by)
loss = 0
for t in range(len(inputs)):
tanh_arg[t] = inputs[t].dot(self.wxh) + h[t - 1].dot(
self.whh) + self.bh
h[t] = np.tanh(tanh_arg[t])
y = h[t].dot(self.why) + self.by
exp_y = np.exp(y)
softmax[t] = exp_y / exp_y.sum()
correct_scores = (softmax[t] * targets[t]).sum(1)
for t in reversed(range(len(inputs))):
df = softmax[t] - targets[t]
dby += df
dwhy += h[t].T.dot(df)
dh = self.why.dot(df.T)
dtanh_arg = tanh_prime(tanh_arg[t]) * dh.T
dbh += dtanh_arg
dwxh += expand(inputs[t]).dot(dtanh_arg)
dwhh += h[t - 1].T.dot(dtanh_arg)
self.h_prev = h[-1]
return dwxh, dwhh, dwhy, dbh, dby
def __BFS(self, verbose=False):
visited = []
max_depth = 0
max_queue_len = 0
queue = [
self.START_PUZZLE,
]
while queue:
v = queue.pop(0)
if v.status not in visited:
visited.append(v.status)
if v.status == GOAL_STATE:
break
queue.extend([i for i in expand(v) if i not in visited])
max_queue_len = max(max_queue_len, len(queue))
max_depth = max(max_depth, v.depth)
if verbose:
print('depth:', v.depth)
print('visited list:', visited,
'(visited len: {})'.format(len(visited)))
print('queue:', queue, '(len: {})'.format(len(queue)))
if len(visited) >= 100000:
print('search space exceeded.')
break
goal_puzzle = v
return self.START_PUZZLE, goal_puzzle, len(
visited), max_queue_len, max_depth
def __DFS(self, verbose=False):
visited = []
max_depth = 0
max_stack_len = 0
stack = [
self.START_PUZZLE,
]
while stack:
v = stack.pop()
if v.status not in visited:
visited.append(v.status)
if v.status == GOAL_STATE:
break
stack.extend([i for i in expand(v) if i not in visited])
max_stack_len = max(max_stack_len, len(stack))
max_depth = max(max_depth, v.depth)
if verbose:
print('depth:', v.depth)
print('visited list:', visited,
'(visited len: {})'.format(len(visited)))
print('stack:', stack, '(len: {})'.format(len(stack)))
if len(visited) >= 100000:
print('search space exceeded.')
break
goal_puzzle = v
return self.START_PUZZLE, goal_puzzle, len(
visited), max_stack_len, max_depth
def _refresh(self, uid, params):
s, api_url, params = prepare_request(params)
url = api_url + self.instance_url(uid=uid)
r = s.get(url, params=utils.expand(params))
r.raise_for_status()
self._refresh_from(r.json())
return self
def loss_fun(inputs, targets, h_prev):
tanh_arg, softmax, h = {}, {}, {}
dwxh, dwhh, dwhy = np.zeros_like(wxh), np.zeros_like(whh), np.zeros_like(
why)
dbh, dby = np.zeros_like(bh), np.zeros_like(by)
loss = 0
h[-1] = h_prev
for t in range(len(inputs)):
tanh_arg[t] = inputs[t].dot(wxh) + h[t - 1].dot(whh) + bh
h[t] = np.tanh(tanh_arg[t])
y = h[t].dot(why) + by
exp_y = np.exp(y)
softmax[t] = exp_y / exp_y.sum()
correct_scores = (softmax[t] * targets[t]).sum(1)
loss += np.sum(-np.log(correct_scores))
for t in reversed(range(len(inputs))):
df = softmax[t] - targets[t]
dby += df.sum(0, keepdims=True)
dwhy += h[t].T.dot(df)
dh = why.dot(df.T)
dtanh_arg = tanh_prime(tanh_arg[t]) * dh.T
dbh += dtanh_arg
dwxh += expand(inputs[t]).dot(dtanh_arg)
dwhh += h[t - 1].T.dot(dtanh_arg)
return loss, dwxh, dwhh, dwhy, dbh, dby, h[-1]
def genereal_search(puzzle, args):
if args.method == 1:
# init node of uniform cost search using its heuristic
start_node = UniformNode(0, puzzle, None, None)
elif args.method == 2:
# init node of A* with misplaces tile using its heuristic
start_node = MisplacedNode(0, puzzle, None, None)
elif args.method == 3:
# init node of A* with Manhattan distance using its heuristic
start_node = ManhattanNode(0, puzzle, None, None)
else:
raise ValueError("Not Implemented Method!")
if (start_node.state == goal_state).all(): # no need to expand if input is the goal state
print("Number of expanded nodes: ", 0)
print("Max queue size: ", 0)
return start_node
queue = []
max_queue = 0
heapq.heappush(queue, start_node)
explored_states = [start_node.state]
while len(queue) > 0:
max_queue = max(len(queue), max_queue)
expand_node = heapq.heappop(queue) # pop smallest cost node from queue to expand/explore
end_node = expand(queue, expand_node, goal_state, explored_states)
if end_node is not None:
print("Number of expanded nodes: ", len(explored_states))
print("Max queue size: ", max_queue)
print("Depth:", end_node.depth)
return end_node
return None
def table(self):
'''Generate a mysql like result table'''
# more pythonic approach for assigning value to multi vars?
header, body, sep = [''], [''], ['']
for key, value in self.items():
if isinstance(value, Dataset):
value = value.summarize()
value = str(value)
width = real_len(max(len(key), len(value)))
header.append(expand(key, width))
body.append(expand(value, width))
sep.append('-' * (width + TAB_WIDTH))
for i in (header, body, sep):
i.append('')
header = '|'.join(header)
body = '|'.join(body)
sep = '+'.join(sep)
return '\n'.join((sep, header, sep, body, sep))
def table(self):
'''Generate a mysql like result table'''
# more pythonic approach for assigning value to multi vars?
header, body, sep = [''], [''], ['']
for key, value in self.items():
if isinstance(value, Dataset):
value = value.summarize()
value = str(value)
width = real_len(max(len(key), len(value)))
header.append(expand(key, width))
body.append(expand(value, width))
sep.append('-' * (width + TAB_WIDTH))
for i in (header, body, sep):
i.append('')
header = '|'.join(header)
body = '|'.join(body)
sep = '+'.join(sep)
return '\n'.join((sep, header, sep, body, sep))
def table(self):
'''Generate a mysql like result table'''
header, body, sep = [''], [], ['']
#: find out the max length of word of each column
max_len = {}
for key in self.header:
width = len(str(key))
for line in self.body:
if isinstance(line[key], Dataset):
value = line[key].summarize()
else:
value = str(line[key])
width = max(width, len(value))
max_len[key] = real_len(width)
#: build the header and sep in the same time
header.append(expand(key, max_len[key]))
sep.append('-' * (max_len[key] + TAB_WIDTH))
for i in (header, sep):
i.append('')
header = '|'.join(header)
sep = '+'.join(sep)
#: build body
for line in self.body:
b = ['']
for key, value in line.items():
if isinstance(value, Dataset):
value = value.summarize()
else:
value = str(value)
width = max_len[key]
b.append(expand(value, width))
b.append('')
body.append('|'.join(b))
#: build table
table = [sep, header, sep]
for line in body:
table.append(line)
table.append(sep)
return '\n'.join(table)
def table(self):
'''Generate a mysql like result table'''
header, body, sep = [''], [], ['']
#: find out the max length of word of each column
max_len = {}
for key in self.header:
width = len(str(key))
for line in self.body:
if isinstance(line[key], Dataset):
value = line[key].summarize()
else:
value = str(line[key])
width = max(width, len(value))
max_len[key] = real_len(width)
#: build the header and sep in the same time
header.append(expand(key, max_len[key]))
sep.append('-' * (max_len[key] + TAB_WIDTH))
for i in (header, sep):
i.append('')
header = '|'.join(header)
sep = '+'.join(sep)
#: build body
for line in self.body:
b = ['']
for key, value in line.items():
if isinstance(value, Dataset):
value = value.summarize()
else:
value = str(value)
width = max_len[key]
b.append(expand(value, width))
b.append('')
body.append('|'.join(b))
#: build table
table = [sep, header, sep]
for line in body:
table.append(line)
table.append(sep)
return '\n'.join(table)
def barrier_dist(x):
bdd = np.abs(x[:, 1] - BARR) <= 2 * EPS
bdd = expand(bdd)
# e.g. [[nan, nan], [x_h, barrier], ...]
h = np.where(bdd, x, np.nan)[:, 0]
# Drop nan
dist = h[~ np.isnan(h)]
plt.figure()
sns.distplot(dist, hist=False, kde=True, bins=int(180 / 5), rug=True)
def __Astar(self, h_func, verbose=False):
if h_func == 'h1':
h_func = h1
elif h_func == 'h2':
h_func = h2
open_, closed_ = [], []
max_open_len = 0
max_depth = 0
start_puzzle = copy.deepcopy(self.START_PUZZLE)
start_puzzle.cost = 0 + h_func(self.START_PUZZLE)
open_.append(start_puzzle)
while True:
if not open_:
print('There is no solution.')
return None
while True:
v = open_.pop(0) # lowest cost
max_depth = max(max_depth, v.depth)
if verbose:
print(v.status, v.cost)
if closed_:
if v.status not in [i.status for i in closed_]:
break
else:
for i in closed_:
if v.status == i.status:
if v.cost <= i.cost:
closed_.remove(i)
break
else:
break
if v.status == GOAL_STATE:
goal_puzzle = v
return start_puzzle, goal_puzzle, len(
closed_), max_open_len, max_depth
closed_.append(v)
children = [i for i in expand(v)]
for i in children:
i.cost = i.depth + h_func(i)
open_.extend(children)
max_open_len = max(max_open_len, len(open_))
open_ = sorted(open_, key=lambda i: i.cost)
closed_ = sorted(closed_, key=lambda i: i.cost)
def diagnose(self, i):
if i % 100 == 0:
print('Iteration: {}. Cost: {}'.format(
i, np.average(self.costs[-10:])))
x = np.random.randint(self.x_size)
for j in range(200):
print(self.i_to_char[x], end='')
self.x = expand(np.eye(self.x_size)[x])
y = self.forward_pass()
y = normalize(y.reshape(self.x_size, ))
x = np.argmax(y)
#x = np.random.choice(np.arange(self.x_size), p=y)
print('\n')
def get_resnet_result(model, dataset, result_save_folder):
if not os.path.exists(result_save_folder):
os.makedirs(result_save_folder)
model.eval()
if CUDA:
model.cuda()
transform = utils.transform_factory(split='test', dataset=dataset)
cleansing_video_dataset = VideoDataset_(
root=os.path.join(args.dataset, 'test_dataset'),
dataset=dataset,
special_list=[],
filter_type='not_in',
feature_folder=feature_floder,
ground_truth_folder='annotation_folder')
for video_index in range(cleansing_video_dataset.__len__()):
video_imgs, video_labels, video_name = cleansing_video_dataset.__getitem__(
video_index, video_index + 1)
video_dataset = VideoDataset(video_imgs, video_labels, transform)
video_loader = data.DataLoader(dataset=video_dataset,
shuffle=False,
batch_size=1,
drop_last=False) # tackle one by one
video_name = video_name[0]
lines = [i for i in range(len(video_labels[0]) + 1)]
lines[0] = 'Frame\tPhase\n'
result_save_path = os.path.join(result_save_folder,
video_name + '_pred.txt')
with torch.no_grad():
for iter, (imgs, labels, img_names) in enumerate(video_loader):
if CUDA:
imgs, labels = imgs.cuda(), labels.cuda()
bs, ncrops, c, h, w = imgs.size()
feature, res = model(imgs.view(-1, c, h, w))
avg_res = res.view(bs, ncrops, -1).mean(1)
avg_res = F.softmax(avg_res, dim=1)
predict = avg_res.max(1)[1].item()
lines[iter + 1] = '{}\t{}\n'.format(iter, predict)
print('testing video {} done!'.format(video_name, ))
lines = utils.expand(
lines, dataset=dataset,
timeF=5)[0:config[dataset]['testset_len'][video_name]]
with open(result_save_path, 'w') as f:
for line in lines:
f.write(str(line))
def put_on(img, el, condition, bg_color=1):
# Center of the structural element
x_c = el.size[0] // 2
y_c = el.size[1] // 2
# Additional emptiness to add to img along x and y
dx_min = x_c
dx_max = el.size[0] - x_c - 1
dy_min = y_c
dy_max = el.size[1] - y_c - 1
element_pixels = Image2ll_binary(el)
image_pixels = expand(Image2ll_binary(img),
dx_min, dx_max, dy_min, dy_max, bg_color)
new_image_pixels = [[
1 for idx_y in range(img.size[1] + dy_min + dy_max)]
for idx_x in range(img.size[0] + dx_min + dx_max)]
black_count = 0
for idx_x in range(dx_min, dx_min + img.size[0]):
for idx_y in range(dy_min, dy_min + img.size[1]):
if condition == 'any' and image_pixels[idx_x][idx_y] == 0:
for dx in range(-dx_min, dx_max + 1):
for dy in range(-dy_min, dy_max + 1):
if element_pixels[x_c + dx][y_c + dy] == 0:
if new_image_pixels[idx_x + dx][idx_y + dy] != 0:
black_count += 1
new_image_pixels[idx_x + dx][idx_y + dy] = 0
elif condition == 'all':
if image_pixels[idx_x][idx_y] == 0:
all_match = True
for dx in range(-dx_min, dx_max + 1):
for dy in range(-dy_min, dy_max + 1):
if (image_pixels[idx_x + dx][idx_y + dy] != 0
and element_pixels[x_c + dx][y_c + dy] == 0):
all_match = False
break
if not all_match:
break
if all_match:
new_image_pixels[idx_x][idx_y] = 0
black_count += 1
print('black:', black_count,
'white:', img.size[0] * img.size[1] - black_count)
return ll2Image_binary(cut(new_image_pixels, dx_min, dx_max, dy_min, dy_max))
def get_labels(self, images, bboxes):
"""
Method for assigning labels to bounding boxes in images
Arguments:
---
images: list of PIL images
bboxes: list of bounding boxes for images
(bbox config: [x1, y1, x2, y2])
Returns:
---
label_list: list of labels assigned to bounding boxes in images
prob_list: confidence values for labels assigned
"""
label_list = []
prob_list = []
for index, img in enumerate(images):
boxes = bboxes[index]
data = []
for i in range(len(boxes)):
im = img.resize([224, 224], box=expand(boxes[i], img))
data.append(self.transform(im))
data = torch.stack(data).to(self.device)
outputs = self.model(data)
labels = [torch.max(op, 1)[1] for op in outputs]
labels = {
self.tasks[i]: label.cpu().numpy()
for i, label in enumerate(labels)
}
probs = {}
for idx, task in enumerate(self.tasks):
task_probs = []
for i in range(len(labels[task])):
p = outputs[idx][i][labels[task][i]].item()
task_probs.append(p)
probs[task] = task_probs
label_list.append(labels)
prob_list.append(probs)
return label_list, prob_list
for batch_i, batch in enumerate(chunker(list(profiles.keys()),
FLAGS.batch_size)):
size = min(len(batch), FLAGS.batch_size)
# create needed binary vectors
bin_profiles = {}
masks = {}
# only consider the movie that users have iteracted
for userid in batch:
user_profile = np.array([0.] * len(all_movies))
mask = [0] * (len(all_movies) * 5)
for movie_id, rat in profiles[userid]:
user_profile[all_movies.index(movie_id)] = rat
for _i in range(5):
mask[5 * all_movies.index(movie_id) + _i] = 1
example = expand(np.array([user_profile])).astype('float32')
bin_profiles[userid] = example
masks[userid] = mask
profile_batch = [bin_profiles[el] for el in batch]
masks_batch = [masks[id] for id in batch]
# train_batch = np.array(profile_batch).reshape(size,
# len(all_movies * 5))
train_batch = np.array(profile_batch).reshape(size, len(all_movies) * 5)
train_masks = np.array(masks_batch).reshape(size,
len(all_movies) * 5)
# _ = sess.run([rbm.optimizer], feed_dict={rbm.input: train_batch, rbm.mask : masks_batch})
_ = sess.run([rbm.optimizer], feed_dict={rbm.input: train_batch, rbm.mask : train_masks})
sys.stdout.write('.')
sys.stdout.flush()
def compute(self, context, *expressions, **kwargs):
if not expressions:
raise TypeError("groupby() takes at least 1 argument")
# TODO: allow lists/tuples of arguments to group by the combinations
# of keys
for expr in expressions:
if isinstance(expr, (bool, int, float)):
raise TypeError("groupby() does not work with constant "
"arguments")
if isinstance(expr, (tuple, list)):
raise TypeError("groupby() takes expressions as arguments, "
"not a list of expressions")
# On python 3, we could clean up this code (keyword only arguments).
expr = kwargs.pop('expr', None)
if expr is None:
expr = Count()
# by = kwargs.pop('by', None)
filter_value = kwargs.pop('filter', None)
percent = kwargs.pop('percent', False)
possible_values = kwargs.pop('pvalues', None)
totals = kwargs.pop('totals', True)
expr_vars = [v.name for v in collect_variables(expr)]
labels = [str(e) for e in expressions]
columns = [expr_eval(e, context) for e in expressions]
columns = [expand(c, context_length(context)) for c in columns]
if filter_value is not None:
filtered_columns = [col[filter_value] for col in columns]
# FIXME: use the actual filter_expr instead of not_hashable
filtered_context = context.subset(filter_value, expr_vars,
not_hashable)
else:
filtered_columns = columns
filtered_context = context
if possible_values is None:
possible_values = [np.unique(col) for col in filtered_columns]
# We pre-filtered columns instead of passing the filter to partition_nd
# because it is a bit faster this way. The indices are still correct,
# because we use them on a filtered_context.
groups = partition_nd(filtered_columns, True, possible_values)
if not groups:
return LabeledArray([], labels, possible_values)
# evaluate the expression on each group
# we use not_hashable to avoid storing the subset in the cache
contexts = [filtered_context.subset(indices, expr_vars, not_hashable)
for indices in groups]
data = [expr_eval(expr, c) for c in contexts]
# TODO: use group_indices_nd directly to avoid using np.unique
# this is twice as fast (unique is very slow) but breaks because
# the rest of the code assumes all combinations are present
# if self.filter is not None:
# filter_value = expr_eval(self.filter, context)
# else:
# filter_value = True
#
# d = group_indices_nd(columns, filter_value)
# pvalues = sorted(d.keys())
# ndim = len(columns)
# possible_values = [[pv[i] for pv in pvalues]
# for i in range(ndim)]
# groups = [d[k] for k in pvalues]
# groups is a (flat) list of list.
# the first variable is the outer-most "loop",
# the last one the inner most.
# add total for each row
len_pvalues = [len(vals) for vals in possible_values]
if percent:
totals = True
if totals:
width = len_pvalues[-1]
height = prod(len_pvalues[:-1])
rows_indices = [np.concatenate([groups[y * width + x]
for x in range(width)])
for y in range(height)]
cols_indices = [np.concatenate([groups[y * width + x]
for y in range(height)])
for x in range(width)]
cols_indices.append(np.concatenate(cols_indices))
# evaluate the expression on each "combined" group (ie compute totals)
row_ctxs = [filtered_context.subset(indices, expr_vars, not_hashable)
for indices in rows_indices]
row_totals = [expr_eval(expr, ctx) for ctx in row_ctxs]
col_ctxs = [filtered_context.subset(indices, expr_vars, not_hashable)
for indices in cols_indices]
col_totals = [expr_eval(expr, ctx) for ctx in col_ctxs]
else:
row_totals = None
col_totals = None
if percent:
# convert to np.float64 to get +-inf if total_value is int(0)
# instead of Python's built-in behaviour of raising an exception.
# This can happen at least when using the default expr (count())
# and the filter yields empty groups
total_value = np.float64(col_totals[-1])
data = [100.0 * value / total_value for value in data]
row_totals = [100.0 * value / total_value for value in row_totals]
col_totals = [100.0 * value / total_value for value in col_totals]
# if self.by or self.percent:
# if self.percent:
# total_value = data[-1]
# divisors = [total_value for _ in data]
# else:
# num_by = len(self.by)
# inc = prod(len_pvalues[-num_by:])
# num_groups = len(groups)
# num_categories = prod(len_pvalues[:-num_by])
#
# categories_groups_idx = [range(cat_idx, num_groups, inc)
# for cat_idx in range(num_categories)]
#
# divisors = ...
#
# data = [100.0 * value / divisor
# for value, divisor in izip(data, divisors)]
# convert to a 1d array. We don't simply use data = np.array(data),
# because if data is a list of ndarray (for example if we use
# groupby(a, expr=id), *and* all the ndarrays have the same length,
# the result is a 2d array instead of an array of ndarrays like we
# need (at this point).
arr = np.empty(len(data), dtype=type(data[0]))
arr[:] = data
data = arr
# and reshape it
data = data.reshape(len_pvalues)
return LabeledArray(data, labels, possible_values,
row_totals, col_totals)
def boundary_condition(bias):
eps = .01
return (bias <= expand(np.min(bias, axis=-1))) & (bias < eps)
def boundary_condition(bias):
eps = .01
return (bias <= expand(np.min(bias, axis=-1))) & (bias < eps)
fig, ax = plt.subplots()
ax.autoscale_view()
camera = Camera(fig)
for t in progressbar(range(int(T / dt))):
plt.plot(*(corner.T), color='r')
plt.scatter(*(x.T), color='black', s=5)
plt.scatter(*z, color='blue', s=5)
camera.snap()
m = delta_matrix(x)
m_norm = expand(norm(m, 2))
with np.errstate(divide='ignore', invalid='ignore'):
social = (1 / m_norm - a) * m
social = np.nan_to_num(social)
# axis 1 is the dummy varible
social = social.sum(axis=1) / N
zx = x - np.tile(z, (N, 1))
vx = social + b * zx / expand(norm(zx, 2))
vz = c / N * (zx / expand(norm(zx, p))).sum(axis=0)
# Boundary condition
eps = 0.01
# If at door: |x[0]| <= 1 and |x[1]| < eps
def run(name, dataset, config, all_users, all_movies, tests, initial_v, sep):
config_name = config['name']
number_hidden = config['number_hidden']
epochs = config['epochs']
ks = config['ks']
momentums = config['momentums']
l_w = config['l_w']
l_v = config['l_v']
l_h = config['l_h']
decay = config['decay']
batch_size = config['batch_size']
config_result = config.copy()
config_result['results'] = []
vis = T.matrix()
vmasks = T.matrix()
rbm = CFRBM(len(all_movies) * 5, number_hidden)
profiles = defaultdict(list)
with open(dataset, 'rt') as data:
for i, line in enumerate(data):
uid, mid, rat, timstamp = line.strip().split(sep)
profiles[uid].append((mid, float(rat)))
print("Users and ratings loaded")
for j in range(epochs):
def get_index(col):
if j/(epochs/len(col)) < len(col):
return j/(epochs/len(col))
else:
return -1
index = get_index(ks)
mindex = get_index(momentums)
icurrent_l_w = get_index(l_w)
icurrent_l_v = get_index(l_v)
icurrent_l_h = get_index(l_h)
k = ks[index]
momentum = momentums[mindex]
current_l_w = l_w[icurrent_l_w]
current_l_v = l_v[icurrent_l_v]
current_l_h = l_h[icurrent_l_h]
train = rbm.cdk_fun(vis,
vmasks,
k=k,
w_lr=current_l_w,
v_lr=current_l_v,
h_lr=current_l_h,
decay=decay,
momentum=momentum)
predict = rbm.predict(vis)
for batch_i, batch in enumerate(chunker(profiles.keys(),
batch_size)):
size = min(len(batch), batch_size)
# create needed binary vectors
bin_profiles = {}
masks = {}
for userid in batch:
user_profile = [0.] * len(all_movies)
mask = [0] * (len(all_movies) * 5)
for movie_id, rat in profiles[userid]:
user_profile[all_movies.index(movie_id)] = rat
for _i in range(5):
mask[5 * all_movies.index(movie_id) + _i] = 1
example = expand(np.array([user_profile])).astype('float32')
bin_profiles[userid] = example
masks[userid] = mask
profile_batch = [bin_profiles[id] for id in batch]
masks_batch = [masks[id] for id in batch]
train_batch = np.array(profile_batch).reshape(size,
len(all_movies) * 5)
train_masks = np.array(masks_batch).reshape(size,
len(all_movies) * 5)
train_masks = train_masks.astype('float32')
train(train_batch, train_masks)
sys.stdout.write('.')
sys.stdout.flush()
ratings = []
predictions = []
for batch in chunker(tests.keys(), batch_size):
size = min(len(batch), batch_size)
# create needed binary vectors
bin_profiles = {}
masks = {}
for userid in batch:
user_profile = [0.] * len(all_movies)
mask = [0] * (len(all_movies) * 5)
for movie_id, rat in profiles[userid]:
user_profile[all_movies.index(movie_id)] = rat
for _i in range(5):
mask[5 * all_movies.index(movie_id) + _i] = 1
example = expand(np.array([user_profile])).astype('float32')
bin_profiles[userid] = example
masks[userid] = mask
positions = {profile_id: pos for pos, profile_id
in enumerate(batch)}
profile_batch = [bin_profiles[el] for el in batch]
test_batch = np.array(profile_batch).reshape(size,
len(all_movies) * 5)
user_preds = revert_expected_value(predict(test_batch))
for profile_id in batch:
test_movies = tests[profile_id]
try:
for movie, rating in test_movies:
current_profile = user_preds[positions[profile_id]]
predicted = current_profile[all_movies.index(movie)]
rating = float(rating)
ratings.append(rating)
predictions.append(predicted)
except Exception:
pass
vabs = np.vectorize(abs)
distances = np.array(ratings) - np.array(predictions)
mae = vabs(distances).mean()
rmse = sqrt((distances ** 2).mean())
iteration_result = {
'iteration': j,
'k': k,
'momentum': momentum,
'mae': mae,
'rmse': rmse,
'lrate': current_l_w
}
config_result['results'].append(iteration_result)
print(iteration_str.format(j, k, current_l_w, momentum, mae, rmse))
with open('experiments/{}_{}.json'.format(config_name, name), 'wt') as res_output:
res_output.write(json.dumps(config_result, indent=4))
W,V,H = rbm.get_weights()
print H
kernel_size = int(8*std+1) #SIFT
localWorkSizeX= 16 #for shared local memory
localWorkSizeY= 16
mode = "mirror" #borders handling : mirror, reflect, nearest, wrap
if ((kernel_size % 2) == 0):
print("Error: the filter size is not odd !")
exit(1)
'''
***** Pre-allocating *****
'''
l2=lena().astype("float32")
l = utils.expand(l2, 4*std, mode=mode)
#l=numpy.copy(l[0:32,0:32]) #memory layout is not contiguous...
[imwidth, imheight] = l.shape
output = numpy.zeros_like(l)
x=numpy.arange(kernel_size,dtype="float64")
shift = kernel_size//2.0
v = (x-shift)/std
g1 = numpy.exp(-v*v/2.0)
g=numpy.outer(g1,g1).astype("float32")
g/=g.sum()
#plt.imshow(g)
#plt.show()
patterns = [('iwidth',imwidth),('iheight',imheight),('fsize',kernel_size),('hsize',kernel_size//2),('tsize',(kernel_size//2)*2),('hisize',(kernel_size//2)*imwidth)]
opts = "-D IMAGE_W=iwidth -D IMAGE_H=iheight -D FILTER_SIZE=fsize -D HALF_FILTER_SIZE=hsize -D TWICE_HALF_FILTER_SIZE=tsize -D HALF_FILTER_SIZE_IMAGE_W=hisize"
for (p,r) in patterns:
def list(cls, **params):
s, api_url, params = prepare_request(params)
url = api_url + cls._class_url()
r = s.get(url, params=utils.expand(params))
r.raise_for_status()
return convert_to_sense_object(None, r.json())
def run(name, dataset, config, all_users, all_movies, tests, initial_v, sep):
config_name = config['name']
number_hidden = config['number_hidden']
epochs = config['epochs']
ks = config['ks']
momentums = config['momentums']
l_w = config['l_w']
l_v = config['l_v']
l_h = config['l_h']
decay = config['decay']
batch_size = config['batch_size']
config_result = config.copy()
config_result['results'] = []
vis = T.matrix()
vmasks = T.matrix()
rbm = CFRBM(len(all_movies) * 5, number_hidden)
profiles = defaultdict(list)
with open(dataset, 'rt') as data:
for i, line in enumerate(data):
uid, mid, rat, timstamp = line.strip().split(sep)
profiles[uid].append((mid, float(rat)))
print("Users and ratings loaded")
for j in range(epochs):
def get_index(col):
if j/(epochs/len(col)) < len(col):
return j/(epochs/len(col))
else:
return -1
index = get_index(ks)
mindex = get_index(momentums)
icurrent_l_w = get_index(l_w)
icurrent_l_v = get_index(l_v)
icurrent_l_h = get_index(l_h)
k = ks[index]
momentum = momentums[mindex]
current_l_w = l_w[icurrent_l_w]
current_l_v = l_v[icurrent_l_v]
current_l_h = l_h[icurrent_l_h]
train = rbm.cdk_fun(vis,
vmasks,
k=k,
w_lr=current_l_w,
v_lr=current_l_v,
h_lr=current_l_h,
decay=decay,
momentum=momentum)
predict = rbm.predict(vis)
for batch_i, batch in enumerate(chunker(profiles.keys(),
batch_size)):
size = min(len(batch), batch_size)
# create needed binary vectors
bin_profiles = {}
masks = {}
for userid in batch:
user_profile = [0.] * len(all_movies)
mask = [0] * (len(all_movies) * 5)
for movie_id, rat in profiles[userid]:
user_profile[all_movies.index(movie_id)] = rat
for _i in range(5):
mask[5 * all_movies.index(movie_id) + _i] = 1
example = expand(np.array([user_profile])).astype('float32')
bin_profiles[userid] = example
masks[userid] = mask
profile_batch = [bin_profiles[id] for id in batch]
masks_batch = [masks[id] for id in batch]
train_batch = np.array(profile_batch).reshape(size,
len(all_movies) * 5)
train_masks = np.array(masks_batch).reshape(size,
len(all_movies) * 5)
train_masks = train_masks.astype('float32')
train(train_batch, train_masks)
sys.stdout.write('.')
sys.stdout.flush()
ratings = []
predictions = []
for batch in chunker(tests.keys(), batch_size):
size = min(len(batch), batch_size)
# create needed binary vectors
bin_profiles = {}
masks = {}
for userid in batch:
user_profile = [0.] * len(all_movies)
mask = [0] * (len(all_movies) * 5)
for movie_id, rat in profiles[userid]:
user_profile[all_movies.index(movie_id)] = rat
for _i in range(5):
mask[5 * all_movies.index(movie_id) + _i] = 1
example = expand(np.array([user_profile])).astype('float32')
bin_profiles[userid] = example
masks[userid] = mask
positions = {profile_id: pos for pos, profile_id
in enumerate(batch)}
profile_batch = [bin_profiles[el] for el in batch]
test_batch = np.array(profile_batch).reshape(size,
len(all_movies) * 5)
user_preds = revert_expected_value(predict(test_batch))
for profile_id in batch:
test_movies = tests[profile_id]
try:
for movie, rating in test_movies:
current_profile = user_preds[positions[profile_id]]
predicted = current_profile[all_movies.index(movie)]
rating = float(rating)
ratings.append(rating)
predictions.append(predicted)
except Exception:
pass
vabs = np.vectorize(abs)
distances = np.array(ratings) - np.array(predictions)
mae = vabs(distances).mean()
rmse = sqrt((distances ** 2).mean())
iteration_result = {
'iteration': j,
'k': k,
'momentum': momentum,
'mae': mae,
'rmse': rmse,
'lrate': current_l_w
}
config_result['results'].append(iteration_result)
print(iteration_str.format(j, k, current_l_w, momentum, mae, rmse))
with open('{}_{}.json'.format(config_name, name), 'wt') as res_output:
res_output.write(json.dumps(config_result, indent=4))
def run(name, dataset, user_info, config, all_users, all_movies, all_occupations, all_sex, all_ages, tests, initial_v, sep):
config_name = config['name']
number_hidden = config['number_hidden']
epochs = config['epochs']
ks = config['ks']
momentums = config['momentums']
l_w = config['l_w']
l_v = config['l_v']
l_h = config['l_h']
decay = config['decay']
batch_size = config['batch_size']
config_result = config.copy()
config_result['results'] = []
vis_x = T.matrix()
vis_o = T.matrix()
vis_s = T.matrix()
vis_a = T.matrix()
vmasks_x = T.matrix()
vmasks_o = T.matrix()
vmasks_s = T.matrix()
vmasks_a = T.matrix()
rbm = CFRBM(len(all_movies) * 5, len(all_occupations), 1, len(all_ages), number_hidden)
profiles = defaultdict(list)
with open(dataset, 'rt') as data:
for i, line in enumerate(data):
uid, mid, rat, timstamp = line.strip().split(sep)
profiles[uid].append((mid, float(rat)))
print("Users and ratings loaded")
user_occ = defaultdict(list)
user_sex = defaultdict(list)
user_age = defaultdict(list)
r = csv.reader(open(user_info, 'rb'), delimiter='|')
for row in r:
user_age[row[0]] = [int(x) for x in row[1:7]]
user_sex[row[0]] = [int(row[7])]
user_occ[row[0]] = [int(x) for x in row[8:]]
print("User info loaded")
for j in range(epochs):
def get_index(col):
if j/(epochs/len(col)) < len(col):
return j/(epochs/len(col))
else:
return -1
index = get_index(ks)
mindex = get_index(momentums)
icurrent_l_w = get_index(l_w)
icurrent_l_v = get_index(l_v)
icurrent_l_h = get_index(l_h)
k = ks[index]
momentum = momentums[mindex]
current_l_w = l_w[icurrent_l_w]
current_l_v = l_v[icurrent_l_v]
current_l_h = l_h[icurrent_l_h]
train = rbm.cdk_fun(vis_x,
vis_o,
vis_s,
vis_a,
vmasks_x,
vmasks_o,
vmasks_s,
vmasks_a,
k=k,
w_lr=current_l_w,
v_lr=current_l_v,
h_lr=current_l_h,
decay=decay,
momentum=momentum)
predict = rbm.predict(vis_x, vis_o, vis_s, vis_a)
start_time = time.time()
for batch_i, batch in enumerate(chunker(profiles.keys(),
batch_size)):
size = min(len(batch), batch_size)
# create needed binary vectors
bin_profiles = {}
occ_profiles = {}
sex_profiles = {}
age_profiles = {}
masks_x = {}
masks_o = {}
masks_s = {}
masks_a = {}
for userid in batch:
user_profile = [0.] * len(all_movies)
occ_profile = [0.] * len(all_occupations)
sex_profile = [0.] * 1
age_profile = [0.] * len(all_ages)
mask_x = [0] * (len(all_movies) * 5)
mask_o = [1] * (len(all_occupations))
mask_s = [1] * (1)
mask_a = [1] * (len(all_ages))
for movie_id, rat in profiles[userid]:
user_profile[all_movies.index(movie_id)] = rat
for _i in range(5):
mask_x[5 * all_movies.index(movie_id) + _i] = 1
mask_o = [1] * len(all_occupations)
mask_s = [1] * 1
mask_a = [1] * len(all_ages)
example_x = expand(np.array([user_profile])).astype('float32')
example_o = expand(np.array([occ_profile]), k=1).astype('float32')
example_s = expand(np.array([sex_profile]), k=1).astype('float32')
example_a = expand(np.array([age_profile]), k=1).astype('float32')
bin_profiles[userid] = example_x
occ_profiles[userid] = example_o
sex_profiles[userid] = example_s
age_profiles[userid] = example_a
masks_x[userid] = mask_x
masks_o[userid] = mask_o
masks_s[userid] = mask_s
masks_a[userid] = mask_a
profile_batch = [bin_profiles[id] for id in batch]
occ_batch = [occ_profiles[id] for id in batch]
sex_batch = [sex_profiles[id] for id in batch]
age_batch = [age_profiles[id] for id in batch]
masks_x_batch = [masks_x[id] for id in batch]
masks_o_batch = [masks_o[id] for id in batch]
masks_s_batch = [masks_s[id] for id in batch]
masks_a_batch = [masks_a[id] for id in batch]
train_batch_x = np.array(profile_batch).reshape(size,
len(all_movies) * 5)
train_batch_o = np.array(occ_batch).reshape(size,
len(all_occupations))
train_batch_s = np.array(sex_batch).reshape(size,
1)
train_batch_a = np.array(age_batch).reshape(size,
len(all_ages))
train_masks_x = np.array(masks_x_batch).reshape(size,
len(all_movies) * 5)
train_masks_o = np.array(masks_o_batch).reshape(size,
len(all_occupations))
train_masks_s = np.array(masks_s_batch).reshape(size,
1)
train_masks_a = np.array(masks_a_batch).reshape(size,
len(all_ages))
train_masks_x = train_masks_x.astype('float32')
train_masks_o = train_masks_o.astype('float32')
train_masks_s = train_masks_s.astype('float32')
train_masks_a = train_masks_a.astype('float32')
train(train_batch_x, train_batch_o, train_batch_s, train_batch_a, train_masks_x, train_masks_o, train_masks_s, train_masks_a)
sys.stdout.write('.')
sys.stdout.flush()
end_time = time.time()
train_time = end_time - start_time
ratings = []
predictions = []
start_time = time.time()
for batch in chunker(tests.keys(), batch_size):
size = min(len(batch), batch_size)
# create needed binary vectors
bin_profiles = {}
occ_profiles = {}
sex_profiles = {}
age_profiles = {}
masks_x = {}
masks_o = {}
masks_s = {}
masks_a = {}
for userid in batch:
user_profile = [0.] * len(all_movies)
occ_profile = [0.] * len(all_occupations)
sex_profile = [0.] * 1
age_profile = [0.] * len(all_ages)
mask_x = [0] * (len(all_movies) * 5)
mask_o = [1] * (len(all_occupations))
mask_s = [1] * (1)
mask_a = [1] * (len(all_ages))
for movie_id, rat in profiles[userid]:
user_profile[all_movies.index(movie_id)] = rat
for _i in range(5):
mask_x[5 * all_movies.index(movie_id) + _i] = 1
mask_o = [1] * len(all_occupations)
mask_s = [1] * 1
mask_a = [1] * len(all_ages)
example_x = expand(np.array([user_profile])).astype('float32')
example_o = expand(np.array([occ_profile]), k=1).astype('float32')
example_s = expand(np.array([sex_profile]), k=1).astype('float32')
example_a = expand(np.array([age_profile]), k=1).astype('float32')
bin_profiles[userid] = example_x
occ_profiles[userid] = example_o
sex_profiles[userid] = example_s
age_profiles[userid] = example_a
masks_x[userid] = mask_x
masks_o[userid] = mask_o
masks_s[userid] = mask_s
masks_a[userid] = mask_a
positions = {profile_id: pos for pos, profile_id
in enumerate(batch)}
profile_batch = [bin_profiles[el] for el in batch]
occ_batch = [occ_profiles[el] for el in batch]
sex_batch = [sex_profiles[el] for el in batch]
age_batch = [age_profiles[el] for el in batch]
test_batch_x = np.array(profile_batch).reshape(size,
len(all_movies) * 5)
test_batch_o = np.array(occ_batch).reshape(size,
len(all_occupations))
test_batch_s = np.array(sex_batch).reshape(size,
1)
test_batch_a = np.array(age_batch).reshape(size,
len(all_ages))
user_preds = revert_expected_value(predict(test_batch_x, test_batch_o, test_batch_s, test_batch_a))
for profile_id in batch:
test_movies = tests[profile_id]
try:
for movie, rating in test_movies:
current_profile = user_preds[positions[profile_id]]
predicted = current_profile[all_movies.index(movie)]
rating = float(rating)
ratings.append(rating)
predictions.append(predicted)
except Exception:
pass
end_time = time.time()
test_time = end_time - start_time
true_rat = np.array(ratings, dtype=np.uint8)
pred_rat = np.array(predictions, dtype=np.uint8)
#print true_rat < 3, true_rat
prec_rec = precision_recall_fscore_support(true_rat < 3,pred_rat < 3, average='binary')
print prec_rec
vabs = np.vectorize(abs)
distances = np.array(ratings) - np.array(predictions)
mae = vabs(distances).mean()
rmse = sqrt((distances ** 2).mean())
iteration_result = {
'iteration': j,
'k': k,
'momentum': momentum,
'mae': mae,
'rmse': rmse,
'lrate': current_l_w,
'train_time': train_time,
'test_time': test_time,
'prec_rec': prec_rec
}
config_result['results'].append(iteration_result)
print(iteration_str.format(j, k, current_l_w, momentum, mae, rmse))
with open('experiments/{}_{}.json'.format(config_name, name), 'wt') as res_output:
res_output.write(json.dumps(config_result, indent=4))
compare_with_2d_conv = 0
show_bench_results = 0
'''
if ((kernel_size % 2) == 0):
print("Error: the filter size is not odd !")
exit(1)
'''
'''
***** Pre-allocating *****
'''
l2=lena().astype("float32")
#l2=l2[:,:-1]
l2 = l2#[350:500,100:250]
if (expand == 1): l = utils.expand(l2, 4*std, mode=mode)
else: l=numpy.ascontiguousarray(l2,dtype="float32")
[imheight, imwidth] = l.shape
IMAGE_W = numpy.int32(imwidth)
IMAGE_H = numpy.int32(imheight)
FILTER_SIZE = numpy.int32(kernel_size)
output = numpy.zeros_like(l)
x=numpy.arange(kernel_size,dtype="float64")
shift = (kernel_size-1.0)/2.0
v = (x-shift)/std
g1 = numpy.exp(-v*v/2.0).astype("float32")
g1/=g1.sum()
print g1
elif mode == "reflect":
# 4 corners
output[s0 + k0:, s1 + k1:] = input_img[-1:-k0 - 1:-1, -1:-k1 - 1:-1]
output[:k0, :k1] = input_img[k0 - 1::-1, k1 - 1::-1]
output[:k0, s1 + k1:] = input_img[k0 - 1::-1, s1 - 1: s1 - k1 - 1:-1]
output[s0 + k0:, :k1] = input_img[s0 - 1: s0 - k0 - 1:-1, k1 - 1::-1]
# 4 sides
output[k0:k0 + s0, :k1] = input_img[:s0, k1 - 1::-1]
output[:k0, k1:k1 + s1] = input_img[k0 - 1::-1, :s1]
output[-k0:, k1:s1 + k1] = input_img[:s0 - k0 - 1:-1, :]
output[k0:s0 + k0, -k1:] = input_img[:, :s1 - k1 - 1:-1]
'''
output_mirror= utils.expand(input_img, sigma, mode="mirror")
output_reflect=utils.expand(input_img, sigma, mode="reflect")
output_nearest=utils.expand(input_img, sigma, mode="nearest")
output_wrap=utils.expand(input_img, sigma, mode="wrap")
#conv_ref = scipy.ndimage.gaussian_filter(l2,std,mode="mirror")
fig = plt.figure()
sp1 = fig.add_subplot(2,3,1,title="Input image")
sp2 = fig.add_subplot(2,3,2,title="Mirror")
sp3 = fig.add_subplot(2,3,3,title="Reflect")
sp4 = fig.add_subplot(2,3,4,title="Nearest")
sp5 = fig.add_subplot(2,3,5,title="Wrap")
#sp6 = fig.add_subplot(2,3,6,title="Error for: mirror")
def run(name, dataset, config, all_users, all_movies, tests, initial_v, sep):
config_name = config["name"]
number_hidden = config["number_hidden"]
epochs = config["epochs"]
ks = config["ks"]
momentums = config["momentums"]
l_w = config["l_w"]
l_v = config["l_v"]
l_h = config["l_h"]
decay = config["decay"]
config_result = config.copy()
config_result["results"] = []
vis = T.matrix()
vmasks = T.matrix()
rbm = CFRBM(len(all_users) * 5, number_hidden)
profiles = defaultdict(list)
with open(dataset, "rt") as data:
for i, line in enumerate(data):
uid, mid, rat, timstamp = line.strip().split(sep)
profiles[mid].append((uid, float(rat)))
print("Users and ratings loaded")
for j in range(epochs):
def get_index(col):
if j / (epochs / len(col)) < len(col):
return j / (epochs / len(col))
else:
return -1
index = get_index(ks)
mindex = get_index(momentums)
icurrent_l_w = get_index(l_w)
icurrent_l_v = get_index(l_v)
icurrent_l_h = get_index(l_h)
k = ks[index]
momentum = momentums[mindex]
current_l_w = l_w[icurrent_l_w]
current_l_v = l_v[icurrent_l_v]
current_l_h = l_h[icurrent_l_h]
train = rbm.cdk_fun(
vis, vmasks, k=k, w_lr=current_l_w, v_lr=current_l_v, h_lr=current_l_h, decay=decay, momentum=momentum
)
predict = rbm.predict(vis)
batch_size = 10
for batch_i, batch in enumerate(utils.chunker(profiles.keys(), batch_size)):
size = min(len(batch), batch_size)
# create needed binary vectors
bin_profiles = {}
masks = {}
for movieid in batch:
movie_profile = [0.0] * len(all_users)
mask = [0] * (len(all_users) * 5)
for user_id, rat in profiles[movieid]:
movie_profile[all_users.index(user_id)] = rat
for _i in range(5):
mask[5 * all_users.index(user_id) + _i] = 1
example = expand(np.array([movie_profile])).astype("float32")
bin_profiles[movieid] = example
masks[movieid] = mask
movies_batch = [bin_profiles[id] for id in batch]
masks_batch = [masks[id] for id in batch]
train_batch = np.array(movies_batch).reshape(size, len(all_users) * 5)
train_masks = np.array(masks_batch).reshape(size, len(all_users) * 5)
train_masks = train_masks.astype("float32")
train(train_batch, train_masks)
sys.stdout.write(".")
sys.stdout.flush()
batch_size = 10
ratings = []
predictions = []
for batch in utils.chunker(tests.keys(), batch_size):
size = min(len(batch), batch_size)
# create needed binary vectors
bin_profiles = {}
masks = {}
for movieid in batch:
movie_profile = [0.0] * len(all_users)
mask = [0] * (len(all_users) * 5)
for userid, rat in profiles[movieid]:
movie_profile[all_users.index(userid)] = rat
for _i in range(5):
mask[5 * all_users.index(userid) + _i] = 1
example = expand(np.array([movie_profile])).astype("float32")
bin_profiles[movieid] = example
masks[movieid] = mask
positions = {movie_id: pos for pos, movie_id in enumerate(batch)}
movies_batch = [bin_profiles[el] for el in batch]
test_batch = np.array(movies_batch).reshape(size, len(all_users) * 5)
movie_predictions = revert_expected_value(predict(test_batch))
for movie_id in batch:
test_users = tests[movie_id]
try:
for user, rating in test_users:
current_movie = movie_predictions[positions[movie_id]]
predicted = current_movie[all_users.index(user)]
rating = float(rating)
ratings.append(rating)
predictions.append(predicted)
except Exception:
pass
vabs = np.vectorize(abs)
distances = np.array(ratings) - np.array(predictions)
mae = vabs(distances).mean()
rmse = sqrt((distances ** 2).mean())
iteration_result = {
"iteration": j,
"k": k,
"momentum": momentum,
"mae": mae,
"rmse": rmse,
"lrate": current_l_w,
}
config_result["results"].append(iteration_result)
print(iteration_str.format(j, k, current_l_w, momentum, mae, rmse))
with open("{}_{}.json".format(config_name, name), "wt") as res_output:
res_output.write(json.dumps(config_result, indent=4))
def train(self, tasks, model_name, criterion, attention, device, n_epoch=30, lr=0.1, verbose=True):
"""
trains the given model
args:
task: task to train the model on
criterion: loss for train
attention: whether to use attention mechanism
model_name: name of the model to train
n_epoch: number of epochs for training
lr: learning rate for the optimization
device: whether to use CPU or GPU
verbose: if set to True prints additional info
returns: the obtained metrics, the final predictions and the wrong ones
"""
total_loss_train = []
total_loss_val = []
total_accuracy_val = []
total_accuracy_train = []
total_accuracy_val_pix = []
total_accuracy_train_pix = []
total_predictions = []
total_wrong_predictions = []
criterion = criterion()
for task in tasks:
sh1_big = 0
sh2_big = 0
for i in range(len(task['train'])):
sh1 = task['train'][i]['input'].shape[0]
sh2 = task['train'][i]['input'].shape[1]
if sh1 > sh1_big:
sh1_big = sh1
if sh2 > sh2_big:
sh2_big = sh2
for i in range(len(task['test'])):
sh1 = task['test'][i]['input'].shape[0]
sh2 = task['test'][i]['input'].shape[1]
if sh1 > sh1_big:
sh1_big = sh1
if sh2 > sh2_big:
sh2_big = sh2
net = model_name(task['train'], sh1_big, sh2_big, attention).to(device)
optimizer = Adam(net.parameters(), lr = lr)
loss_train = []
loss_val = []
accuracy_val = []
accuracy_train = []
accuracy_val_pix = []
accuracy_train_pix = []
for epoch in tqdm(range(n_epoch)):
net.train()
loss_iter = 0
for sample in task['train']:
img = FloatTensor(expand(sample['input'])).to(device)
labels = LongTensor(sample['output']).unsqueeze(dim=0).to(device)
optimizer.zero_grad()
outputs = net(img)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
net.eval()
with torch.no_grad():
correct_val = 0
correct_val_pix = 0
total_val = 0
loss_iter_val = 0
predictions = []
wrong_pred = []
n_pixels_val = 0
for sample in task['test']:
img = FloatTensor(expand(sample['input'])).to(device)
labels = LongTensor(sample['output']).unsqueeze(dim=0).to(device)
outputs = net(img)
_, pred = torch.max(outputs.data, 1)
predictions.append((img, pred))
n_pixels_val += pred.shape[1]*pred.shape[2]
total_val += labels.size(0)
flag = (torch.all(torch.eq(pred, labels))).sum().item()
correct_val += flag
if flag == 0:
wrong_pred.append((img, pred))
correct_val_pix += (torch.eq(pred, labels)).sum().item()
loss = criterion(outputs, labels)
loss_iter_val += loss.item()
correct_train = 0
correct_train_pix = 0
total_train = 0
loss_iter_train = 0
n_pixels_train = 0
for sample in task['train']:
img = FloatTensor(expand(sample['input'])).to(device)
labels = LongTensor(sample['output']).unsqueeze(dim=0).to(device)
outputs = net(img)
_, pred = torch.max(outputs.data, 1)
n_pixels_train += pred.shape[1]*pred.shape[2]
total_train += labels.size(0)
correct_train += (torch.all(torch.eq(pred, labels))).sum().item()
correct_train_pix += (torch.eq(pred, labels)).sum().item()
loss = criterion(outputs, labels)
loss_iter_train += loss.item()
loss_train.append(loss_iter_train/len(task['train']))
loss_val.append(loss_iter_val/len(task['test']))
val_accuracy = 100 * correct_val / total_val
val_accuracy_pix = 100 * correct_val_pix/(n_pixels_val)
accuracy_val.append(val_accuracy)
accuracy_val_pix.append(val_accuracy_pix)
train_accuracy = 100 * correct_train / total_train
train_accuracy_pix = 100 * correct_train_pix/(n_pixels_train)
accuracy_train.append(train_accuracy)
accuracy_train_pix.append(train_accuracy_pix)
if verbose:
print('\nEpoch: ['+str(epoch+1)+'/'+str(n_epoch)+']')
print('Train loss is: {}'.format(loss_train[-1]))
print('Validation loss is: {}'.format(loss_val[-1]))
print('Train accuracy is: {} %'.format(accuracy_train[-1]))
print('Train accuracy for pixels is: {} %'.format(accuracy_train_pix[-1]))
print('Validation accuracy is: {} %'.format(accuracy_val[-1]))
print('Validation accuracy for pixels is: {} %'.format(accuracy_val_pix[-1]))
total_loss_train += loss_train
total_loss_val += loss_val
total_accuracy_train += accuracy_train
total_accuracy_train_pix += accuracy_train_pix
total_accuracy_val += accuracy_val
total_accuracy_val_pix += accuracy_val_pix
total_predictions += predictions
total_wrong_predictions += wrong_pred
metrics = {'loss_train': total_loss_train, 'loss_val': total_loss_val, 'accuracy_train':total_accuracy_train,
'accuracy_train_pix': total_accuracy_train_pix, 'accuracy_val':total_accuracy_val,
'accuracy_val_pix': total_accuracy_val_pix}
final_pred = total_predictions
return metrics, final_pred, total_wrong_predictions
class RNN:
h_size = 100
learning_rate = 1e-1
batch_size = 32
regul = .5
initial_h = expand(np.zeros(h_size))
filename = 'input8.txt'
def __init__(self):
self.h = self.initial_h
self.text = open(self.filename, 'r').read()[:-1]
chars = set(self.text)
self.x_size = len(chars)
self.i_to_char = {i: char for i, char in enumerate(chars)}
self.char_to_i = {char: i for i, char in enumerate(chars)}
self.W_xh = np.random.randn(self.h_size, self.x_size) * .01
self.W_hh = np.random.randn(self.h_size, self.h_size) * .01
self.W_hy = np.random.randn(self.x_size, self.h_size) * .01
self.B_h = np.zeros((self.h_size, 1))
self.B_y = np.zeros((self.x_size, 1))
def forward_pass(self):
self.prev_h = self.h #np.copy(self.h)
self.h = np.tanh(
np.dot(self.W_xh, self.x) + np.dot(self.W_hh, self.h) + self.B_h)
return np.tanh(np.dot(self.W_hy, self.h) + self.B_y)
def backward_pass(self):
dCdY = self.y - self.target
dYdZ = np.diag(
tanh_prime(np.dot(self.W_hy, self.h)).reshape(self.x_size, ))
dZdWhy = np.repeat(self.h.T, self.x_size, axis=0)
dYdWhy = np.dot(dYdZ, dZdWhy)
dCdWhy = dCdY * dYdWhy
dCdBy = np.dot(dCdY.T, dYdZ)
dZdHt = self.W_hy
dYdHt = np.dot(dYdZ, dZdHt)
dHtdZ2 = tanh_prime(
np.dot(self.W_xh, self.x) + np.dot(self.W_hh, self.h))
dHtdZ2 = np.diag(dHtdZ2.reshape(self.h_size, ))
dZ2dWxh = np.repeat(self.x.T, self.h_size, axis=0)
dZ2dWhh = np.repeat(self.prev_h.T, self.h_size, axis=0)
dYdZ2 = np.dot(dYdHt, dHtdZ2)
dYdWxh = np.dot(dYdZ2, dZ2dWxh)
dYdWhh = np.dot(dYdZ2, dZ2dWhh)
dCdWxh = np.repeat(np.dot(dCdY.T, dYdWxh), self.h_size, axis=0)
dCdWhh = np.repeat(np.dot(dCdY.T, dYdWhh), self.h_size, axis=0)
dCdBh = np.dot(dCdY.T, dYdZ2)
return [dCdWxh, dCdWhh, dCdWhy, dCdBh, dCdBy]
def diagnose(self, i):
if i % 100 == 0:
print('Iteration: {}. Cost: {}'.format(
i, np.average(self.costs[-10:])))
x = np.random.randint(self.x_size)
for j in range(200):
print(self.i_to_char[x], end='')
self.x = expand(np.eye(self.x_size)[x])
y = self.forward_pass()
y = normalize(y.reshape(self.x_size, ))
x = np.argmax(y)
#x = np.random.choice(np.arange(self.x_size), p=y)
print('\n')
def train(self):
self.costs = []
gradients = []
for i, (char, target) in enumerate(zip(self.text, self.text[1:])):
self.x = expand(np.eye(self.x_size)[self.char_to_i[char]])
self.target = expand(np.eye(self.x_size)[self.char_to_i[target]])
self.y = self.forward_pass()
gradients.append(self.backward_pass())
if (i + 1) % self.batch_size == 0:
dCdWxh, dCdWhh, dCdWhy, dCdBh, dCdBy = np.average(gradients,
axis=0)
self.W_xh -= self.learning_rate * dCdWxh
self.W_hh -= self.learning_rate * dCdWhh
self.W_hy -= self.learning_rate * dCdWhy
self.B_h -= self.learning_rate * dCdBh.T
self.B_y -= self.learning_rate * dCdBy.T
self.costs.append(cost(self.target, self.y))
self.diagnose(i)
plt.plot(self.costs)
plt.show(block=False)
time.sleep(1000)
def train(self, tasks, model_name, criterion, n_epoch=30, lr=0.1, device = "cpu", verbose=True, inner_lr = 0.1, inner_iter = 10, meta_size = 50):
"""
trains the given model
args:
task: task to train the model on
criterion: loss for train
model_name: name of the model to train
n_epoch: number of epochs for training
lr: learning rate for the optimization
device: whether to use CPU or GPU
verbose: if set to True prints additional info
inner_lr: if using a meta-learning algorithm, the learning rate of the inner loop
inner_iter: if using a meta-learning algorithm, the iterations of the inner loop
meta_size: if using a meta-learning algorithm, how big to set the meta samples size
returns: the obtained metrics, the final predictions and the wrong ones
"""
total_loss_train = []
total_loss_val = []
total_accuracy_val = []
total_accuracy_train = []
total_accuracy_val_pix = []
total_accuracy_train_pix = []
total_predictions = []
total_wrong_predictions = []
criterion = criterion()
sh1_big = 0
sh2_big = 0
for task in tasks:
for i in range(len(task['train'])):
sh1 = task['train'][i]['input'].shape[0]
sh2 = task['train'][i]['input'].shape[1]
if sh1 > sh1_big:
sh1_big = sh1
if sh2 > sh2_big:
sh2_big = sh2
for i in range(len(task['test'])):
sh1 = task['test'][i]['input'].shape[0]
sh2 = task['test'][i]['input'].shape[1]
if sh1 > sh1_big:
sh1_big = sh1
if sh2 > sh2_big:
sh2_big = sh2
net = model_name(device, sh1_big, sh2_big).to(device)
optimizer = Adam(net.parameters(), lr = lr)
for epoch in tqdm(range(n_epoch)):
losses = []
for task in tasks:
loss_train = []
loss_val = []
accuracy_val = []
accuracy_train = []
accuracy_val_pix = []
accuracy_train_pix = []
inputs = []
outputs = []
for sample in task["train"]:
x = Tensor(sample['input'])
x = pad_crop(x, sh1_big, sh2_big, x.shape[0], x.shape[1], goal = "pad")
inputs.append(FloatTensor(expand(x).float()).to(device))
y = Tensor(sample['output'])
y = pad_crop(y, sh1_big, sh2_big, y.shape[0], y.shape[1], goal = "pad")
outputs.append(LongTensor(y.long()).unsqueeze(0).to(device))
inputs_train = inputs[:meta_size]
inputs_val = inputs[meta_size:]
outputs_train = outputs[:meta_size]
outputs_val = outputs[meta_size:]
fast_weights = OrderedDict(net.named_parameters())
for _ in range(inner_iter):
grads = []
loss = 0
for x,y in zip(inputs_train, outputs_train):
logits = net._forward(x.to(device), fast_weights)
loss += criterion(logits.to(device), y.to(device))
loss /= len(inputs_train)
gradients = torch.autograd.grad(loss, fast_weights.values(), create_graph=True)
fast_weights = OrderedDict((name, param - inner_lr * grad)
for ((name, param), grad) in zip(fast_weights.items(), gradients))
loss = 0
for x,y in zip(inputs_val, outputs_val):
logits = net._forward(x.to(device), fast_weights)
loss += criterion(logits.to(device), y.to(device))
loss /= len(inputs_val)
loss.backward(retain_graph=True)
losses.append(loss.float())
gradients = torch.autograd.grad(loss, fast_weights.values(), create_graph=True)
net.train()
optimizer.zero_grad()
meta_loss = torch.stack(losses).mean()
meta_loss.backward()
optimizer.step()
net.eval()
with torch.no_grad():
correct_val = 0
correct_val_pix = 0
total_val = 0
loss_iter_val = 0
predictions = []
wrong_pred = []
n_pixels_val = 0
for sample in task['test']:
img = FloatTensor(expand(sample['input'])).to(device)
y = LongTensor(sample['output'])
labels = pad_crop(y, sh1_big, sh2_big, y.shape[0], y.shape[1], "pad").unsqueeze(0).to(device)
outputs = net(img)
_, pred = torch.max(outputs.data, 1)
predictions.append((img, pred))
n_pixels_val += pred.shape[1]*pred.shape[2]
total_val += labels.size(0)
flag = (torch.all(torch.eq(pred, labels))).sum().item()
correct_val += flag
if flag == 0:
wrong_pred.append((img, pred))
correct_val_pix += (torch.eq(pred, labels)).sum().item()
loss = criterion(outputs, labels)
loss_iter_val += loss.item()
correct_train = 0
correct_train_pix = 0
total_train = 0
loss_iter_train = 0
n_pixels_train = 0
for sample in task['train']:
img = FloatTensor(expand(sample['input'])).to(device)
y = LongTensor(sample['output'])
labels = pad_crop(y, sh1_big, sh2_big, y.shape[0], y.shape[1], "pad").unsqueeze(0).to(device)
outputs = net(img)
_, pred = torch.max(outputs.data, 1)
n_pixels_train += pred.shape[1]*pred.shape[2]
total_train += labels.size(0)
correct_train += (torch.all(torch.eq(pred, labels))).sum().item()
correct_train_pix += (torch.eq(pred, labels)).sum().item()
loss = criterion(outputs, labels)
loss_iter_train += loss.item()
loss_train.append(loss_iter_train/len(task['train']))
loss_val.append(loss_iter_val/len(task['test']))
val_accuracy = 100 * correct_val / total_val
val_accuracy_pix = 100 * correct_val_pix/(n_pixels_val)
accuracy_val.append(val_accuracy)
accuracy_val_pix.append(val_accuracy_pix)
train_accuracy = 100 * correct_train / total_train
train_accuracy_pix = 100 * correct_train_pix/(n_pixels_train)
accuracy_train.append(train_accuracy)
accuracy_train_pix.append(train_accuracy_pix)
if verbose:
print('\nEpoch: ['+str(epoch+1)+'/'+str(n_epoch)+']')
print('Train loss is: {}'.format(loss_train[-1]))
print('Validation loss is: {}'.format(loss_val[-1]))
print('Train accuracy is: {} %'.format(accuracy_train[-1]))
print('Train accuracy for pixels is: {} %'.format(accuracy_train_pix[-1]))
print('Validation accuracy is: {} %'.format(accuracy_val[-1]))
print('Validation accuracy for pixels is: {} %'.format(accuracy_val_pix[-1]))
total_loss_train.append(loss_train)
total_loss_val.append(loss_val)
total_accuracy_train.append(accuracy_train)
total_accuracy_train_pix.append(accuracy_train_pix)
total_accuracy_val.append(accuracy_val)
total_accuracy_val_pix.append(accuracy_val_pix)
total_predictions.append(total_predictions)
total_wrong_predictions.append(wrong_pred)
metrics = {'loss_train': total_loss_train, 'loss_val': total_loss_val, 'accuracy_train':total_accuracy_train,
'accuracy_train_pix': total_accuracy_train_pix, 'accuracy_val':total_accuracy_val,
'accuracy_val_pix': total_accuracy_val_pix}
final_pred = total_predictions
return metrics, final_pred, total_wrong_predictions
def odd_expansion(n):
i = len(expand(n)) - 1
return 0 < i and i % 2 == 1
def scan(self): #login items files loginItems = [] #dbg msg utils.logMessage(utils.MODE_INFO, 'running scan') #init results dictionary results = self.initResults(LOGIN_ITEM_NAME, LOGIN_ITEM_DESCRIPTION) #process # ->open file and read each line for userLoginItems in utils.expand(LOGIN_ITEM_FILE): #wrap try: #dbg msg utils.logMessage(utils.MODE_INFO, 'scanning %s' % userLoginItems) #load plist and check plistData = utils.loadPlist(userLoginItems) #extract sessions items sesssionItems = plistData['SessionItems'] #extract custom list items customListItems = sesssionItems['CustomListItems'] #iterate over all login items for customListItem in customListItems: #wrap it try: #extact alias data aliasData = list((customListItem['Alias']).bytes()) #parse alias data loginItem = self.parseAliasData(aliasData) #save extracted login item if loginItem: #save results['items'].append(file.File(loginItem)) #ignore exceptions except Exception, e: #skip continue #ignore exceptions except: #skip continue return results
def _build_model(self):
if args.mann == 'none':
def single_cell(num_units):
return tf.contrib.rnn.BasicLSTMCell(num_units, forget_bias=1.0)
cell = tf.contrib.rnn.OutputProjectionWrapper(
tf.contrib.rnn.MultiRNNCell([
single_cell(args.num_units) for _ in range(args.num_layers)
]),
args.num_bits_per_vector,
activation=None)
initial_state = tuple(
tf.contrib.rnn.LSTMStateTuple(
c=expand(tf.tanh(learned_init(args.num_units)),
dim=0,
N=args.batch_size),
h=expand(tf.tanh(learned_init(args.num_units)),
dim=0,
N=args.batch_size))
for _ in range(args.num_layers))
elif args.mann == 'ntm':
cell = NTMCell(args.num_layers,
args.num_units,
args.num_memory_locations,
args.memory_size,
args.num_read_heads,
args.num_write_heads,
addressing_mode='content_and_location',
shift_range=args.conv_shift_range,
reuse=False,
output_dim=args.num_bits_per_vector,
clip_value=args.clip_value,
init_mode=args.init_mode)
initial_state = cell.zero_state(args.batch_size, tf.float32)
elif args.mann == 'dnc':
access_config = {
'memory_size': args.num_memory_locations,
'word_size': args.memory_size,
'num_reads': args.num_read_heads,
'num_writes': args.num_write_heads,
}
controller_config = {
'hidden_size': args.num_units,
}
cell = DNC(access_config, controller_config,
args.num_bits_per_vector, args.clip_value)
initial_state = cell.initial_state(args.batch_size)
output_sequence, _ = tf.nn.dynamic_rnn(cell=cell,
inputs=self.inputs,
time_major=False,
initial_state=initial_state)
if args.task == 'copy' or args.task == 'repeat_copy':
self.output_logits = output_sequence[:, self.max_seq_len + 1:, :]
elif args.task == 'associative_recall':
self.output_logits = output_sequence[:,
3 * (self.max_seq_len + 1) +
2:, :]
elif args.task in ('traversal', 'shortest_path'):
self.output_logits = output_sequence[:, -self.max_seq_len:, :]
if args.task in ('copy', 'repeat_copy', 'associative_recall'):
self.outputs = tf.sigmoid(self.output_logits)
if args.task in ('traversal', 'shortest_path'):
output_logits_split = tf.split(self.output_logits, 9, axis=2)
self.outputs = tf.concat(
[tf.nn.softmax(logits) for logits in output_logits_split],
axis=2)
def expand(self, w):
return expand(w, { 'name': self.name, 'number': str(self.hits.current) })