def run_morse_nets(stim_file, net_file, input_gain=1.0, noise_std=0.0, num_trials=1, exp_desc='default', net_keys=None, fixed_seed=None):
stimset = MorseStimSet()
stimset.from_hdf5(stim_file)
f = h5py.File(net_file, 'a')
burn_time = 100
post_stim_time = 30
pre_stim_time = 30
if net_keys is None:
net_keys = f.keys()
for net_key in net_keys:
if fixed_seed is not None:
#use a fixed random seed so each network sees the same set of random stimuli
print 'Using fixed random seed..'
np.random.seed(fixed_seed)
print 'Running network %s' % net_key
net_grp = f[net_key]
#read network weights from hdf5
net = EchoStateNetwork()
net.from_hdf5(net_grp)
net.noise_std = noise_std
#set the input gain
net.connect_input(0, 0, input_gain)
#rebuild network and run
net.compile()
net_sims = run_sim(net, stimset.all_stims,
burn_time=burn_time,
pre_stim_time=pre_stim_time, post_stim_time=post_stim_time,
num_trials=num_trials)
#save experiment to network
exp_grp = net_grp.create_group(exp_desc)
exp_grp['noise_std'] = noise_std
exp_grp['num_stims'] = len(net_sims)
exp_grp['input_gain'] = input_gain
resp_grp = exp_grp.create_group('responses')
for stim_key,sim in net_sims.iteritems():
stim_grp = resp_grp.create_group(stim_key)
sim.to_hdf5(stim_grp)
f.close()
def run_morse_nets_online(stim_file, num_trials=20, stack_size=5,
num_nodes=3, input_gain=1.0, noise_std=0.0, fixed_seed=None,
stim_class='standard', num_nets=10, output_file=None,
symmetric=False, feedforward=False, connection_rate=1.0,
output_nl_type=None, output_nl_args=None):
stimset = MorseStimSet()
stimset.from_hdf5(stim_file)
burn_time = 100
post_stim_time = 1
pre_stim_time = -1
if fixed_seed is not None:
#use a fixed random seed so each network sees the same set of random stimuli
print 'Using fixed random seed..'
np.random.seed(fixed_seed)
#get stims
stim_md5s = stimset.class_to_md5[stim_class]
all_stims = {}
for md5 in stim_md5s:
all_stims[md5] = stimset.all_stims[md5]
net_stack = [(None, None, None)]*stack_size
net_performances = np.zeros([stack_size])
for k in range(num_nets):
#read network weights from hdf5
net = create_fullyconnected_net(num_nodes=num_nodes, symmetric=symmetric,
feedforward=feedforward)
net.noise_std = noise_std
if output_nl_type is not None:
output_nl = output_nl_type(net, **output_nl_args)
net.output_nl = output_nl
#try out each input node, see which one works best
best_performance = -np.Inf
best_input_node = 0
best_sims = None
for input_node in range(num_nodes):
if feedforward and input_node != 0:
continue
net_copy = copy.deepcopy(net)
net_copy.create_input(0)
net_copy.connect_input(0, input_node, input_gain)
net_copy.compile()
net_sims = run_sim(net_copy, all_stims,
burn_time=burn_time,
pre_stim_time=pre_stim_time, post_stim_time=post_stim_time,
num_trials=num_trials)
all_samps = {}
for md5,sim in net_sims.iteritems():
all_samps[md5] = sim.responses[:, 0, :].squeeze()
percent_correct = train_readout_mnlogit(stimset, all_samps)
if percent_correct > best_performance:
best_performance = percent_correct
best_input_node = input_node
best_net = net_copy
best_sims = net_sims
print 'Net %d (input=%d) got %0.2f correct' % (k, best_input_node, best_performance)
"""
for md5,sim in best_sims.iteritems():
sp = (sim.responses[:, 0, :].squeeze()).sum(axis=0) / float(len(all_samps[md5]))
spstr = ['%d' % f for f in sp]
print '%s,%s' % (''.join(spstr), stimset.md5_to_symbol[md5])
"""
min_perf = net_performances.min()
if min_perf < best_performance:
min_perf_index = net_performances.argmin()
net_stack[min_perf_index] = (net, best_input_node, best_net)
net_performances[min_perf_index] = best_performance
if output_file is not None:
f = h5py.File(output_file, 'w')
for k,(sim, input_node, net) in enumerate(net_stack):
perf = net_performances[k]
net_name = 'net_%d' % k
net_grp = f.create_group(net_name)
net.to_hdf5(net_grp)
perf_grp = net_grp.create_group('performance')
perf_grp[stim_class] = perf
f.close()
return (net_stack, net_performances)
