def main():
args = get_args()
#model_name = 'nodepert_ae5_sgd_correctgeom'
model_name = 'nodepert_ae5_faauto'
#model_name = 'nodepert_ae5_bpauto'
#Model = AENPModel5
#Model = AENPModel5_ExactLsq
#Model = AENPModel5_ExactLsq_BPAuto
#Model = AENPModel5_ExactLsq_BPSelf
Model = AENPModel5_ExactLsq_FAAuto
#Model = AENPModel5_ExactLsq_FASelf
Data = MNISTDataGenerator
Trainer = AESFTrainer
config = process_config('./configs/np_optimized.json', model_name)
create_dirs([config.summary_dir, config.checkpoint_dir])
#var_vals = [1e-2]
N = 1
#M = 5
M = 10
T = config.num_epochs + 1
n_tags = 13
test_losses = np.zeros((N, M))
isnan = np.zeros((N, M))
metrics = np.zeros((N, M, T, n_tags))
for n in range(N):
tf.reset_default_graph()
model = Model(config)
data = Data(config)
for m in range(M):
with tf.Session() as sess:
logger = LoggerNumpy(sess, config, model)
model.load(sess)
trainer = Trainer(sess, model, data, config, logger)
try:
trainer.train()
except ValueError:
print("Method fails to converge for these parameters")
isnan[n, m] = 1
loss, acc = trainer.test()
metric = logger.get_data()
tags = logger.get_tags()
test_losses[n, m] = loss
metrics[n, m, :, :] = metric
#Save after each run
fn = os.path.join(
config.summary_dir) + "3_autoencoder_correctbatch.npz"
to_save = {
'test_losses': test_losses,
'metrics': metrics,
'isnan': isnan,
'tags': tags
}
pickle.dump(to_save, open(fn, "wb"))
return metrics
def main():
args = get_args()
model_name = 'nodepert4_fixedw_exact'
Model = NPModel4_ExactLsq
Data = MNISTDataGenerator
Trainer = SFTrainer
config = process_config('./configs/np.json', model_name)
create_dirs([config.summary_dir, config.checkpoint_dir])
#Param search parameters
attr = ['var_xi']
#var_vals = [1e-3, 1e-2, 1e-1, 1, 10]
var_vals = [1e-1]
N = len(var_vals)
M = 1
T = config.num_epochs + 1
n_tags = 8
test_losses = np.zeros((N, M))
isnan = np.zeros((N, M))
metrics = np.zeros((N, M, T, n_tags))
for n in range(N):
var_val = [var_vals[n]]
set_hyperparameters(config, attr, var_val)
model = Model(config)
data = Data(config)
print('Hyperparameters: ' + attr[0] + ' = %f' % var_vals[n])
for m in range(M):
with tf.Session() as sess:
logger = LoggerNumpy(sess, config, model)
model.load(sess)
trainer = Trainer(sess, model, data, config, logger)
try:
trainer.train()
except ValueError:
print("Method fails to converge for these parameters")
isnan[n, m] = 1
loss, acc = trainer.test()
metric = logger.get_data()
tags = logger.get_tags()
test_losses[n, m] = loss
metrics[n, m, :] = metric
fn = os.path.join(
config.summary_dir) + "2_establish_convergence_feedforward_output.npz"
to_save = {
'test_losses': test_losses,
'metrics': metrics,
'isnan': isnan,
'tags': tags
}
pickle.dump(to_save, open(fn, "wb"))
#np.savez(fn, test_losses=test_losses, metrics = metrics, isnan = isnan, tags = tags)
return metrics
def main():
args = get_args()
model_name = 'feedbackalignment4_small'
Model = FAModel4_Small
#Model = FAModel4
Data = MNISTDataGenerator
Trainer = SFTrainer
config = process_config('./configs/sf.json', model_name)
create_dirs([config.summary_dir, config.checkpoint_dir])
N = 1
M = 5
T = config.num_epochs + 1
n_tags = 10
test_losses = np.zeros((N, M))
isnan = np.zeros((N, M))
metrics = np.zeros((N, M, T, n_tags))
tfconfig = tf.ConfigProto()
tfconfig.gpu_options.allow_growth = True
n = 0
model = Model(config)
data = Data(config)
for m in range(M):
with tf.Session(config=tfconfig) as sess:
logger = LoggerNumpy(sess, config, model)
model.load(sess)
trainer = Trainer(sess, model, data, config, logger)
try:
trainer.train()
except ValueError:
print("Method fails to converge for these parameters")
isnan[n, m] = 1
loss, acc = trainer.test()
metric = logger.get_data()
tags = logger.get_tags()
test_losses[n, m] = loss
metrics[n, m, :, :] = metric
#Save after each run
fn = os.path.join(
config.summary_dir
) + "2b_establish_convergence_feedforward_feedbackalignment_output.npz"
to_save = {
'test_losses': test_losses,
'metrics': metrics,
'isnan': isnan,
'tags': tags
}
pickle.dump(to_save, open(fn, "wb"))
return metrics
def main():
args = get_args()
#model_name = 'nodepert_ae5_sgd'
#model_name = 'nodepert_ae5_bpauto'
model_name = 'nodepert_ae5_adam'
#model_name = 'nodepert_ae5_bpself'
#model_name = 'nodepert_ae5_faauto'
#model_name = 'nodepert_ae5_faself'
#Model = AENPModel5
Model = AENPModel5_ADAM
#Model = AENPModel5_ExactLsq
#Model = AENPModel5_ExactLsq_BPAuto
#Model = AENPModel5_ExactLsq_BPSelf
#Model = AENPModel5_ExactLsq_FAAuto
#Model = AENPModel5_ExactLsq_FASelf
Data = MNISTDataGenerator
Trainer = AESFTrainer
config = process_config('./configs/np.json', model_name)
create_dirs([config.summary_dir, config.checkpoint_dir])
#Param search parameters
attr = ['var_xi', 'learning_rate', 'lmda_learning_rate']
attr_ranges = [[-4, -1], [-6, -3], [-6, -3]]
log_scale = [True, True, True]
N = 20
#M = 5
M = 1
T = config.num_epochs + 1
n_tags = 13
test_losses = np.zeros((N, M))
isnan = np.zeros((N, M))
metrics = np.zeros((N, M, T, n_tags))
params = []
for n in range(N):
param = set_random_hyperparameters(config, attr, attr_ranges,
log_scale)
params.append(param)
tf.reset_default_graph()
model = Model(config)
data = Data(config)
print('Hyperparameters: ' + ' '.join(
[attr[ii] + ' = %f' % param[ii] for ii in range(len(attr))]))
for m in range(M):
with tf.Session() as sess:
logger = LoggerNumpy(sess, config, model)
model.load(sess)
trainer = Trainer(sess, model, data, config, logger)
try:
trainer.train()
except ValueError:
print("Method fails to converge for these parameters")
isnan[n, m] = 1
loss, acc = trainer.test()
metric = logger.get_data()
tags = logger.get_tags()
test_losses[n, m] = loss
metrics[n, m, :, :] = metric
#Save after each run
fn = os.path.join(
config.summary_dir) + "3_autoencoder_correctbatch_hyperparam.npz"
to_save = {
'attr': attr,
'params': params,
'test_losses': test_losses,
'metrics': metrics,
'isnan': isnan,
'tags': tags
}
pickle.dump(to_save, open(fn, "wb"))
return metrics
def main():
args = get_args()
model_name = 'WM_4layer'
Model = WM_4layer
Data = MNISTDataGenerator
Trainer = SFTrainer
config = process_config('/home/prashanth/synthfeedback/configs/wm.json',
model_name)
create_dirs([config.summary_dir, config.checkpoint_dir])
#Param search parameters
#attr = ['']
#var_vals = [1e-3, 1e-2, 1e-1, 1, 10]
#var_vals = [30,100]
N = 1
M = 3
#config.num_epoch=5
#print("\n\n",config.num_epoch)
T = config.num_epochs + 1
#config.gamma=250000
n_tags = 10
test_losses = np.zeros((N, M))
isnan = np.zeros((N, M))
metrics = np.zeros((N, M, T, n_tags))
tfconfig = tf.ConfigProto()
tfconfig.gpu_options.allow_growth = True
for n in range(N):
#var_val = [var_vals[n]]
#set_hyperparameters(config, attr, var_val)
tf.reset_default_graph()
model = Model(config)
#print("\n\n",config.num_epoch)
data = Data(config)
#print('Hyperparameters: ' + attr[0] + ' = %f'%var_vals[n])
for m in range(M):
with tf.Session(config=tfconfig) as sess:
#options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
#run_metadata = tf.RunMetadata()
logger = LoggerNumpy(sess, config, model)
model.load(sess)
#trainer = Trainer(sess, model, data, config, logger, options, run_metadata)
trainer = Trainer(sess, model, data, config, logger)
try:
trainer.train()
except ValueError:
print("Method fails to converge for these parameters")
isnan[n, m] = 1
loss, acc = trainer.test()
metric = logger.get_data()
tags = logger.get_tags()
test_losses[n, m] = loss
metrics[n, m, :, :] = metric
#fetched_timeline = timeline.Timeline(run_metadata.step_stats)
#chrome_trace = fetched_timeline.generate_chrome_trace_format()
#with open('./timeline_02_n_%d_m_%d.json'%(n,m), 'w') as f:
# f.write(chrome_trace)
#Save after each run
#fn = os.path.join(config.summary_dir) + "WM_4layer(condnum).npz"
to_save = {
'test_losses': test_losses,
'metrics': metrics,
'isnan': isnan,
'tags': tags
}
#pickle.dump(to_save, open(fn, "wb"))
#np.savez(fn, test_losses=test_losses, metrics = metrics, isnan = isnan, tags = tags)
return metrics
def main():
args = get_args()
model_name = 'nodepert4_exact'
Model = NPModel4_ExactLsq_CorrectBatch
Data = MNISTDataGenerator
Trainer = SFTrainer
config = process_config('./configs/np.json', model_name)
create_dirs([config.summary_dir, config.checkpoint_dir])
#Param search parameters
attr = ['var_xi']
var_vals = [1e-4, 1e-3, 1e-2, 1e-1]
#var_vals = [1e-3]
N = len(var_vals)
M = 5
#M = 1
T = config.num_epochs + 1
n_tags = 10
test_losses = np.zeros((N, M))
isnan = np.zeros((N, M))
metrics = np.zeros((N, M, T, n_tags))
tfconfig = tf.ConfigProto()
tfconfig.gpu_options.allow_growth = True
for n in range(N):
var_val = [var_vals[n]]
set_hyperparameters(config, attr, var_val)
tf.reset_default_graph()
model = Model(config)
data = Data(config)
print('Hyperparameters: ' + attr[0] + ' = %f' % var_vals[n])
for m in range(M):
with tf.Session(config=tfconfig) as sess:
logger = LoggerNumpy(sess, config, model)
model.load(sess)
trainer = Trainer(sess, model, data, config, logger)
try:
trainer.train()
except ValueError:
print("Method fails to converge for these parameters")
isnan[n, m] = 1
loss, acc = trainer.test()
metric = logger.get_data()
tags = logger.get_tags()
test_losses[n, m] = loss
metrics[n, m, :, :] = metric
#Save after each run
fn = os.path.join(
config.summary_dir
) + "2b_establish_convergence_feedforward_output_correctbatch.npz"
to_save = {
'test_losses': test_losses,
'metrics': metrics,
'isnan': isnan,
'tags': tags
}
pickle.dump(to_save, open(fn, "wb"))
return metrics
def main():
args = get_args()
model_name = 'nodepert4_fixedw_exact'
Model = NPModel4_ExactLsq_CorrectBatch
Data = MNISTDataGenerator
Trainer = SFTrainer
config = process_config('./configs/np.json', model_name)
create_dirs([config.summary_dir, config.checkpoint_dir])
#Param search parameters
attr = ['var_xi']
var_vals = [1e-3, 1e-2, 1e-1]
#var_vals = [1e-1]
N = len(var_vals)
M = 3
T = config.num_epochs+1
n_tags = 10
test_losses = np.zeros((N, M))
isnan = np.zeros((N, M))
metrics = np.zeros((N, M, T, n_tags))
tfconfig = tf.ConfigProto()
tfconfig.gpu_options.allow_growth = True
for n in range(N):
var_val = [var_vals[n]]
set_hyperparameters(config, attr, var_val)
tf.reset_default_graph()
model = Model(config)
data = Data(config)
print('Hyperparameters: ' + attr[0] + ' = %f'%var_vals[n])
for m in range(M):
with tf.Session(config=tfconfig) as sess:
#options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
#run_metadata = tf.RunMetadata()
logger = LoggerNumpy(sess, config, model)
model.load(sess)
#trainer = Trainer(sess, model, data, config, logger, options, run_metadata)
trainer = Trainer(sess, model, data, config, logger)
try:
trainer.train()
except ValueError:
print("Method fails to converge for these parameters")
isnan[n,m] = 1
loss, acc = trainer.test()
metric = logger.get_data()
tags = logger.get_tags()
test_losses[n,m] = loss
metrics[n,m,:,:] = metric
#See here: https://towardsdatascience.com/howto-profile-tensorflow-1a49fb18073d
#fetched_timeline = timeline.Timeline(run_metadata.step_stats)
#chrome_trace = fetched_timeline.generate_chrome_trace_format()
#with open('./timeline_02_n_%d_m_%d.json'%(n,m), 'w') as f:
# f.write(chrome_trace)
#Save after each run
fn = os.path.join(config.summary_dir) + "2_establish_convergence_feedforward_output_correctbatches.npz"
to_save = {
'test_losses': test_losses,
'metrics': metrics,
'isnan': isnan,
'tags': tags
}
pickle.dump(to_save, open(fn, "wb"))
return metrics