FLAGS = flags.FLAGS
rng = np.random.RandomState(seed=FLAGS.synth_data_seed)
rnn_rngs = [
np.random.RandomState(seed=FLAGS.synth_data_seed + 1),
np.random.RandomState(seed=FLAGS.synth_data_seed + 2)
]
T = FLAGS.T
C = FLAGS.C
N = FLAGS.N
nreplications = FLAGS.nreplications
E = nreplications * C
train_percentage = FLAGS.train_percentage
ntimesteps = int(T / FLAGS.dt)
rnn_a = generate_rnn(rnn_rngs[0], N, FLAGS.g, FLAGS.tau, FLAGS.dt,
FLAGS.max_firing_rate)
rnn_b = generate_rnn(rnn_rngs[1], N, FLAGS.g, FLAGS.tau, FLAGS.dt,
FLAGS.max_firing_rate)
rnns = [rnn_a, rnn_b]
# pick which RNN is used on each trial
rnn_to_use = rng.randint(2, size=E)
ext_input = np.repeat(np.expand_dims(rnn_to_use, axis=1), ntimesteps, axis=1)
ext_input = np.expand_dims(ext_input, axis=2) # these are "a's" in the paper
x0s = []
condition_labels = []
condition_number = 0
for c in range(C):
x0 = FLAGS.x0_std * rng.randn(N, 1)
x0s.append(np.tile(x0, nreplications))
rng = np.random.RandomState(seed=FLAGS.synth_data_seed)
T = FLAGS.T
C = FLAGS.C
N = FLAGS.N
S = FLAGS.S
input_magnitude = FLAGS.input_magnitude
nreplications = FLAGS.nreplications
E = nreplications * C # total number of trials
# S is the number of measurements in each datasets, w/ each
# dataset having a different set of observations.
ndatasets = N/S # ok if rounded down
train_percentage = FLAGS.train_percentage
ntime_steps = int(T / FLAGS.dt)
# End of user parameters
rnn = generate_rnn(rng, N, FLAGS.g, FLAGS.tau, FLAGS.dt, FLAGS.max_firing_rate)
# Check to make sure the RNN is the one we used in the paper.
if N == 50:
assert abs(rnn['W'][0,0] - 0.06239899) < 1e-8, 'Error in random seed?'
rem_check = nreplications * train_percentage
assert abs(rem_check - int(rem_check)) < 1e-8, \
'Train percentage * nreplications should be integral number.'
# Initial condition generation, and condition label generation. This
# happens outside of the dataset loop, so that all datasets have the
# same conditions, which is similar to a neurophys setup.
condition_number = 0
x0s = []
condition_labels = []
flags.DEFINE_float("dt", 0.010, "Time bin")
flags.DEFINE_float("max_firing_rate", 30.0, "Map 1.0 of RNN to a spikes per second")
FLAGS = flags.FLAGS
rng = np.random.RandomState(seed=FLAGS.synth_data_seed)
rnn_rngs = [np.random.RandomState(seed=FLAGS.synth_data_seed+1),
np.random.RandomState(seed=FLAGS.synth_data_seed+2)]
T = FLAGS.T
C = FLAGS.C
N = FLAGS.N
nreplications = FLAGS.nreplications
E = nreplications * C
train_percentage = FLAGS.train_percentage
ntimesteps = int(T / FLAGS.dt)
rnn_a = generate_rnn(rnn_rngs[0], N, FLAGS.g, FLAGS.tau, FLAGS.dt,
FLAGS.max_firing_rate)
rnn_b = generate_rnn(rnn_rngs[1], N, FLAGS.g, FLAGS.tau, FLAGS.dt,
FLAGS.max_firing_rate)
rnns = [rnn_a, rnn_b]
# pick which RNN is used on each trial
rnn_to_use = rng.randint(2, size=E)
ext_input = np.repeat(np.expand_dims(rnn_to_use, axis=1), ntimesteps, axis=1)
ext_input = np.expand_dims(ext_input, axis=2) # these are "a's" in the paper
x0s = []
condition_labels = []
condition_number = 0
for c in range(C):
x0 = FLAGS.x0_std * rng.randn(N, 1)
x0s.append(np.tile(x0, nreplications))