def show():
data = Data()
d = data.train.make_one_shot_iterator().get_next()
d = utils.Object(**d)
from models import UTILS
sess = UTILS.get_session()
cnt = 0
hit = np.zeros(10)
pbar = tqdm()
while True:
try:
x = sess.run(d)
except tf.errors.OutOfRangeError:
break
cnt += 1
pbar.update(1)
# for i, v in enumerate(x.seq):
# neighbors = data.dp.neighbors[0][v]
# if x.ans in neighbors:
# hit[i] += 1
if cnt >= 10000: break
pbar.close()
print(hit, cnt, hit / cnt)
def augment_mask_probs(sample, state, probs, meta):
for prop, value in sample.items():
if prop not in meta.feat_labels:
print("skipping", prop)
continue
fid = meta.feat_labels.index(prop)
bid = meta.bag_idx[fid]
prob = probs.p[fid].item()
mask = state.mask[fid].item()
sample[prop] = {'value': value, 'prob': prob, 'mask': mask}
if bid is not None:
bag_prob = probs.bags[bid]
if bag_prob is None or len(bag_prob.p) == 0:
sample[prop]['value'] = None
continue
# compute softmax over all whole bag
bag_p = torch.softmax(bag_prob.p.flatten(),
dim=0).view_as(bag_prob.p)
# for all items go deeper
for idx, it in enumerate(value):
it_probs = utils.Object()
it_probs.p = bag_p[idx]
it_probs.bags = bag_prob.bags[idx] if len(
bag_prob.bags) > 0 else None
meta_bag = meta.bags[bid][1]
it_state = state.bags[bid][idx]
augment_mask_probs(it, it_state, it_probs, meta_bag)
def __init__(self,
name=None,
cmd=None,
cwd=None,
proc=None,
input=None,
output=None,
*args,
**kwargs):
if name is None: name = cmd
super().__init__(*args, name=name, readOnly=True, **kwargs)
self.input = None
self.cmd = cmd
self.cwd = cwd
self.proc = proc or utils.Object(stdout=input, stdin=output)
self.lines = collections.deque()
Thread(target=self.run, daemon=True).start()
class Base:
deep = True
args = utils.Object()
# feature: [type..], [tags..], mid
def __init__(self, data):
self.raw_adjs = data.adjs
# self.save_name = f'{utils.save_dir}/{args.run_name}/model.ckpt'
self.save_dir = f'{utils.save_dir}/{args.run_name}'
self.tb_name = f'{utils.save_dir}/{args.run_name}'
self.graph = tf.Graph()
with self.graph.as_default():
tf.set_random_seed(args.seed)
self.compile()
self.fit_step = 0
def compile(self):
self.make_io()
self.make_model()
if args.run_tb:
self.all_summary = tf.summary.merge_all()
self.tbfw = tf.summary.FileWriter(self.tb_name, self.sess.graph)
def placeholder(self, dtype, shape, name, to_list):
ph = tf.placeholder(dtype, shape, name)
self.placeholder_dict[name] = ph
to_list.append(ph)
return ph
def make_io(self):
self.placeholder_dict = {}
self.inputs = []
L = args.seq_length
self.placeholder(tf.int32, [None, L], 'share_seq', self.inputs)
self.placeholder(tf.int32, [None, L], 'click_seq', self.inputs)
self.placeholder(tf.int32, [None], 'pos', self.inputs)
self.placeholder(tf.int32, [None, None], 'neg', self.inputs)
self.adjs = [
tf.constant(adj, dtype=tf.int32) for adj in self.raw_adjs
] # [N, M] * 4, in_0, out_0, in_1, out_1
def get_data_map(self, data):
data_map = dict(zip(self.inputs, data))
return data_map
def make_model(self):
with tf.variable_scope(
'Graph', reuse=tf.AUTO_REUSE,
regularizer=self.l2_loss('all')) as self.graph_scope:
n = args.nb_nodes
k = args.dim_k
self.embedding_matrix = tf.get_variable(name='emb_w', shape=[n, k])
with tf.variable_scope(
'graph_agg', reuse=tf.AUTO_REUSE) as self.graph_agg_scope:
pass
with tf.variable_scope('Network',
reuse=tf.AUTO_REUSE,
regularizer=self.l2_loss('all')):
score, label = self.forward(*self.inputs)
seq_loss = tf.losses.softmax_cross_entropy(label, score)
tf.summary.scalar('seq_loss', seq_loss)
self.loss = seq_loss
self.loss += tf.losses.get_regularization_loss()
opt = tf.train.AdamOptimizer(learning_rate=args.lr)
self.minimizer = opt.minimize(self.loss)
tf.summary.scalar('loss', self.loss)
graph_var_list = tf.trainable_variables(scope='^Graph/')
network_var_list = tf.trainable_variables(scope='^Network/')
for v in graph_var_list:
print('graph', v)
for v in network_var_list:
print('network', v)
self.saver = tf.train.Saver()
self.sess = self.get_session()
self.sess.run(tf.global_variables_initializer())
def get_session(self):
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=1,
visible_device_list=args.gpu,
allow_growth=True,
)
config = tf.ConfigProto(gpu_options=gpu_options)
session = tf.Session(config=config)
return session
def fit(self, data):
data = dict(zip(self.inputs, data))
if args.run_tb:
_, loss, summary = self.sess.run(
[self.minimizer, self.loss, self.all_summary], data)
self.tbfw.add_summary(summary, self.fit_step)
else:
_, loss = self.sess.run([self.minimizer, self.loss], data)
self.fit_step += 1
return loss
def topk(self, data):
data = self.get_data_map(data)
return self.sess.run([self.topkV, self.topkI], data)
def save(self):
name = f'{self.save_dir}/model.ckpt'
self.saver.save(self.sess, name)
def restore(self):
name = f'{self.save_dir}/model.ckpt'
try:
self.saver.restore(self.sess, name)
except Exception as e:
print(f'can not restore model: {name}')
raise e
def l2_loss(self, name):
alpha = args.get(f'l2_{name}', 0)
if alpha < 1e-7:
return None
return lambda x: alpha * tf.nn.l2_loss(x)
def Mean(self, seq, seq_length=None, mask=None, name=None):
# seq: (None, L, k), seq_length: (None, ), mask: (None, L)
# ret: (None, k)
if seq_length is None and mask is None:
with tf.variable_scope('Mean'):
return tf.reduce_sum(seq, -2)
with tf.variable_scope('MaskMean'):
if mask is None:
mask = tf.sequence_mask(seq_length,
maxlen=tf.shape(seq)[1],
dtype=tf.float32)
mask = tf.expand_dims(mask, -1) # (None, L, 1)
seq = seq * mask
seq = tf.reduce_sum(seq, -2) # (None, k)
seq = seq / (tf.reduce_sum(mask, -2) + eps)
return seq
def MLP(self, x, fc, activation, name):
with tf.variable_scope(f'MLP_{name}'):
for i in range(len(fc)):
x = tf.layers.dense(x,
fc[i],
activation=activation,
name=f'dense_{i}')
return x
def gate(self, a, b, name):
with tf.variable_scope(name):
alpha = tf.layers.dense(tf.concat([a, b], -1),
1,
activation=tf.nn.sigmoid,
name='gateW')
ret = alpha * a + (1 - alpha) * b
return ret
def Embedding(self, node, name='node', mask_zero=False):
# node: [BS]
with tf.variable_scope(f'Emb_{name}'):
emb_w = self.embedding_matrix
t = tf.gather(emb_w, node)
if mask_zero:
mask = tf.not_equal(node, 0)
mask = tf.cast(mask, tf.float32)
else:
mask = None
return t, mask
def forward(self, share_seq, click_seq, pos, neg):
pos2 = tf.expand_dims(pos, -1)
nxt = tf.concat([pos2, neg], -1) # [BS, M + 1]
label = tf.concat([
tf.ones_like(pos2, dtype=tf.int32),
tf.zeros_like(neg, dtype=tf.int32)
], -1) # [BS, M + 1]
seq_emb = self.merge_seq(share_seq, click_seq)
seq_emb = tf.layers.dense(seq_emb,
args.dim_k,
name='dense_W',
use_bias=False)
score = tf.matmul(seq_emb, self.embedding_matrix, transpose_b=True)
topk = tf.math.top_k(score, k=500)
self.topkV = topk.values
self.topkI = topk.indices
nxt_embs, _ = self.Embedding(nxt) # [BS, M + 1, k]
nxt_score = tf.reduce_sum(tf.expand_dims(seq_emb, 1) * nxt_embs, -1)
return nxt_score, label
def node_embedding(self, node):
# node: [BS, L]
embs, mask = self.Embedding(node, mask_zero=True)
return embs, mask
def merge_seq(self, share_seq, click_seq):
with tf.variable_scope(f'merge_seq', reuse=tf.AUTO_REUSE):
share_seq_embs, share_mask = self.node_embedding(share_seq)
share_emb = self.seq_embedding(share_seq_embs, share_mask, 'share')
click_seq_embs, click_mask = self.node_embedding(click_seq)
click_emb = self.seq_embedding(click_seq_embs, click_mask, 'click')
emb = self.gate(share_emb, click_emb, 'merge_share_and_click_seq')
return emb
def seq_embedding(self, seq, mask, name):
# seq: [BS, L, k]
with tf.variable_scope(f'seq_embedding_{name}', reuse=tf.AUTO_REUSE):
seq_emb = self.Mean(seq, mask=mask)
return seq_emb
def get_input_seq(self, user, pred_pos, name):
if name in ('train', 'vali') and (user, pred_pos) in self.tv_cache:
return self.tv_cache[(user, pred_pos)]
item_seq = self.dp.user2item_seq[user]
ts_seq = self.dp.user2ts_seq[user]
if name == 'train':
ans = item_seq[pred_pos]
if ans == 1:
vali_phase, vali_pos, vali_ans = self.dp.vali_user2ppa[user]
if args.use_unused_vali and args.mode_pred_phase != 'all' and str(
vali_phase) not in args.mode_pred_phase:
ans = vali_ans
if ans < 3:
return None
elif name == 'vali':
phase, vali_pos, ans = self.dp.vali_user2ppa[user]
assert vali_pos == pred_pos
assert item_seq[vali_pos] == 1
elif name == 'test':
phase, test_pos, ans = self.dp.test_user2ppa[user]
assert test_pos == pred_pos
assert item_seq[test_pos] == 2
if user in self.dp.vali_user2ppa:
vali_phase, vali_pos, vali_ans = self.dp.vali_user2ppa[user]
item_seq = list(item_seq)
item_seq[vali_pos] = vali_ans
else:
raise Exception
q_ts = ts_seq[pred_pos]
pre_ts = ts_seq[pred_pos - 1] if pred_pos else -100
assert args.data_dt_less_than < 0 or args.data_dt_greater_than < 0
# noinspection PyChainedComparisons
if args.data_dt_less_than > 0 and name != 'train' and not q_ts - pre_ts < args.data_dt_less_than:
return None
if args.data_dt_greater_than > 0 and name != 'train' and not q_ts - pre_ts >= args.data_dt_greater_than:
return None
_item_seq, _ts_seq = [], []
# _item_seq = []
for i in range(pred_pos)[::-1]:
item, ts = item_seq[i], ts_seq[i]
# item = item_seq[i]
# if _item_seq and ts_seq[i + 1] - ts > args.seq_max_dt:
if q_ts - ts > args.seq_max_dt:
break
if item == 1:
vali_phase, vali_pos, vali_ans = self.dp.vali_user2ppa[user]
if args.use_unused_vali and args.mode_pred_phase != 'all' and str(
vali_phase) not in args.mode_pred_phase:
item = vali_ans
if item >= 3 and item != ans:
_item_seq.append(item)
_ts_seq.append(ts)
if len(_item_seq) >= args.seq_length:
break
if not _item_seq:
_item_seq, _ts_seq = [0], [-100]
out = utils.Object(
seq=_item_seq,
ts=_ts_seq,
)
self.tv_cache[(user, pred_pos)] = out
return out
def load_data(self):
self.raw_ids = utils.Object(**self.dp.raw_id_list)
self.tv_cache = {}
def __init__(self, input_param, para, **kwargs):
input_param_new = {}
for k in input_param.keys():
if '__' not in k:
input_param_new[k] = input_param[k]
del input_param
input_param = input_param_new
del input_param_new
from imports import *
for k in para:
globals()[k] = para[k]
setattr(self, k, para[k])
for k in input_param.keys():
if '__' not in k:
setattr(self, k, input_param[k])
for key, value in kwargs.items():
input_param[key] = value
setattr(self, key, value)
if self.scale == 'Default':
print('Getting scale by filename:')
a = self.filename
a1 = a.split('.')[0]
a1 = a1.split('_')
for k in range(0, len(a1)):
if 'scale' == a1[k]:
self.scale = float(a1[k + 1])
print(self.scale)
else:
print(self.scale)
input_param['scale'] = self.scale
print('')
if self.timeunit == 'Default':
print('Getting timestep by filename:')
a = self.filename
a1 = a.split('.')[0]
a1 = a1.split('_')
for k in range(0, len(a1)):
if 'timestep' == a1[k]:
self.timeunit = a1[k + 1]
print(self.timeunit)
if self.timeunit != 'days' and self.timeunit != 'seconds':
print('Could not obtain proper timestep')
else:
print(self.timeunit)
print('')
input_param['timeunit'] = self.timeunit
self.input_param = input_param
stat = utils.Object()
for k in input_param.keys():
if '__' not in k:
try:
setattr(stat, k, input_param[k])
except:
print('Not copy:')
print(k, input_param[k])
print('')
pass
self.stat = stat
import os
import imports
import utils
import numpy as np
# Static settings
stat = utils.Object()
stat.filename = ''
stat.read_max = 'all'
stat.scale = 'Default'
stat.method = 'fsolve'
stat.new_folder = True
stat.calc_method = 'Abram'
stat.dir_savefig = os.getcwd(
) + '/' # The directory where the figures will be saved. If False, it will be in the current working directory
stat.noise_check = False
stat.home = '/home/ester/git/synthlisa/' # Home directory
stat.directory_imp = False
stat.num_back = 0
stat.dir_orbits = '/home/ester/git/synthlisa/orbits/' # Folder with orbit files
stat.length_calc = 'all' # Length of number of imported datapoints of orbit files. 'all' is also possible
stat.dir_extr = 'zzzAbram_no_abberation' # This will be added to the folder name of the figures
stat.timeunit = 'Default' # The timeunit of the plots (['minutes'],['days']['years'])
stat.LISA_opt = 'cache' # If a LISA object from syntheticLISA will be used for further calculations (not sure if it works properly if this False)
stat.arm_influence = True # Set True to consider the travel time of the photons when calculating the nominal armlengths
stat.tstep = False
stat.delay = True #'Not ahead' or False
stat.method = 'fsolve' # Method used to solve the equation for the photon traveling time
stat.valorfunc = 'Function' #
stat.select = 'Hallion' # Select which orbit files will be imported ('all' is all)
stat.aberration = True
def fit(self, args, args_i, args_n):
self.args_i = args_i
self.args_n = args_n
self.args = utils.Object(**args)
self.pre_fit()
self.prt_info()
self.make_model()
unique_fn = '{}-{}'.format(logger.unique_fn, self.args_i)
tensorboard_dir = 'tensorboard/{}'.format(unique_fn)
self.tb_dirs.append(tensorboard_dir)
train_writer = tf.summary.FileWriter(tensorboard_dir, self.sess.graph)
saver = tf.train.Saver()
summ_loss = tf.Summary()
summ_loss_v = summ_loss.value.add()
summ_loss_v.tag = 'loss_per_batch'
summaries = tf.summary.merge_all()
batch_cnt = 0
best_vali = None
brk = 0
# ret, _ = self.data.evaluate('vali', self.predict)
# print(ret)
has_ckpt = False
try:
for epochs in range(self.max_epochs):
loss = []
progress_bar = utils.ProgressBar(self.args.batch_steps, msg='training')
for step in range(self.args.batch_steps):
batch = next(self.data_generator)
data = dict(zip(self.train_inputs, batch))
# print(len(self.train_inputs), len(batch)); input()
# print(data); input()
if step == 0 and summaries is not None:
summ = self.sess.run(summaries, data)
train_writer.add_summary(summ, global_step=batch_cnt)
if self.minimize2 is None:
_, _loss = self.sess.run([self.minimize, self.loss], data)
else:
_, _, _loss = self.sess.run([self.minimize, self.minimize2, self.loss],
data)
batch_cnt += 1
loss.append(_loss)
summ_loss_v.simple_value = _loss
train_writer.add_summary(summ_loss, global_step=batch_cnt)
progress_bar.make_a_step()
self.good_log = True
train_time = progress_bar.stop()
vali, vali_time = self.data.evaluate('vali', self.predict)
if vali.is_better_than(best_vali):
brk = 0
best_vali = vali
saver.save(self.sess, 'save/{}_model.ckpt'.format(unique_fn))
has_ckpt = True
else:
brk += 1
if self.run_test:
test, test_time = self.data.evaluate('test', self.predict)
vali = '{} {}'.format(vali, test)
vali_time += test_time
msg = '#{}/{}, loss: {:.5f}, vali: {}, brk: {}, time: {:.1f}s {:.1f}s'.format(
epochs + 1, self.max_epochs, np.mean(loss), vali, brk, train_time, vali_time)
log(msg, i=-1, red=(brk == 0))
if self.early_stop > 0 and brk >= self.early_stop:
break
except KeyboardInterrupt:
utils.timer.stop()
log('KeyboardInterrupt')
except Exception as e:
utils.timer.stop()
log('Exception: {}'.format(e), red=True)
if has_ckpt:
saver.restore(self.sess, 'save/{}_model.ckpt'.format(unique_fn))
return self.after_fit()
