def run(filter_len):
np.random.seed(3)
tf.set_random_seed(3)
D_init = np.random.randn(filter_len, input_size, layer_size)
# D_init = generate_dct_dictionary(filter_len, layer_size).reshape((filter_len, input_size, layer_size))*0.1
D = tf.Variable(D_init.reshape((filter_len, 1, input_size, layer_size)),
dtype=tf.float32)
x = tf.placeholder(tf.float32,
shape=(batch_size, seq_size, 1, input_size),
name="x")
xc = tf.nn.conv2d(x, D, strides=[1, 1, 1, 1], padding='VALID', name="xc")
x_v = np.zeros((seq_size, batch_size, input_size))
for bi in xrange(batch_size):
for ni in xrange(input_size):
x_v[:, bi, ni] = generate_ts(seq_size)
x_v = x_v.transpose((1, 0, 2)).reshape(
(batch_size, seq_size, input_size, 1))
sess = tf.Session()
sess.run(tf.global_variables_initializer())
xc_v = sess.run(xc, {x: x_v})
# dst_pic = "/home/alexeyche/tmp/xc_{0}_{1:.2f}.png".format(int(filter_len), factor)
dst_pic = "/home/alexeyche/tmp/xc_{0}.png".format(int(filter_len))
shl(xc_v, file=dst_pic)
print filter_len
def gen_ts(seq_size):
x = np.zeros((seq_size, batch_size, input_size))
for bi in xrange(batch_size):
for ni in xrange(input_size):
x[:, bi, ni] = np.diff(generate_ts(seq_size + 1))
x[:, bi, ni] /= np.std(x[:, bi, ni])
return x.transpose((1, 0, 2)).reshape(
(batch_size, seq_size, 1, input_size))
grads_and_vars = []
for li, s in enumerate(finstate):
dF = s[-1]
grads_and_vars += [
(-tf.reduce_mean(dF, 0), net._cells[li].F_flat),
]
sess = tf.Session()
env = Env("lca_simple")
x_v = np.zeros((seq_size, batch_size, input_size))
for bi in xrange(batch_size):
for ni in xrange(input_size):
x_v[:, bi, ni] = generate_ts(seq_size)
x_v = x_v.reshape((seq_size, batch_size, input_size))
state_v = get_zero_state()
sample_size = 20
E = np.zeros((sample_size, sample_size))
a_v_res = np.zeros((sample_size, sample_size, seq_size, layer_size))
x_v_res = np.zeros((sample_size, sample_size, seq_size))
W0 = np.linspace(-2.0, 2.0, sample_size)
W1 = np.linspace(-2.0, 2.0, sample_size)
# optimizer = tf.train.GradientDescentOptimizer(lrate)
apply_grads_step = tf.group(
optimizer.apply_gradients([(D_grad, D)]),
tf.nn.l2_normalize(D, 0)
)
##
sess = tf.Session()
sess.run(tf.global_variables_initializer())
x_v = np.zeros((seq_size, batch_size, input_size))
for bi in xrange(batch_size):
for ni in xrange(input_size):
x_v[:,bi,ni] = np.diff(generate_ts(seq_size+1))
x_v[:,bi,ni] /= np.std(x_v[:,bi,ni])
# x_v[:,bi,ni] = generate_ts(seq_size)
# x_v[:,bi,ni] = np.random.randn(seq_size)
x_v = x_v.transpose((1, 0, 2)).reshape((batch_size, seq_size, input_size, 1))
h_v = np.random.random(h.get_shape().as_list())
h_v[h_v < 0.999] = 0
e_m_arr, l_m_arr = [], []
lookback, tol = 10, 1e-05