def test_siamese_model():
network_params_siam['load_saved_model'] = True
test_data_x = None; test_data_y = None; batch_creator = None
if network_params_siam['batch_params'] is not None:
network_params_siam['batch_params']['data_file_indices'] = network_params_siam['batch_params']['test_data_file_indices']
batch_creator = BatchCreator(network_params_siam['batch_params'])
else:
test_data_x, test_data_y = get_data('test')
siamese_model = SiamesePushModel(sess=get_session(), network_params=network_params_siam)
siamese_model.init_model(epoch = 560)
siamese_model.configure_data(data_x=test_data_x, data_y=test_data_y, batch_creator=batch_creator)
# h = siamese_model.run_op('last_hidden',data_x)
# plt.figure(figsize=(8, 8))
# plt.plot(h, data_y,'ro', alpha=0.3)
# plt.show()
print "Got the data, gonna train the model..."
epochs = network_params_siam['epochs']
siamese_model.test(iterations=epochs)
def train_siamese_model():
network_params_siam['load_saved_model'] = False
train_data_x = None; train_data_y = None; batch_creator = None
if network_params_siam['batch_params'] is not None:
batch_creator = BatchCreator(network_params_siam['batch_params'])
else:
train_data_x, train_data_y = get_data('train')
siamese_model = SiamesePushModel(sess=get_session(), network_params=network_params_siam)
siamese_model.init_model(epoch = 50)
siamese_model.configure_data(data_x=train_data_x, data_y=train_data_y, batch_creator=batch_creator)
# h = siamese_model.run_op('last_hidden',data_x)
# plt.figure(figsize=(8, 8))
# plt.plot(h, data_y,'ro', alpha=0.3)
# plt.show()
print "Got the data, gonna train the model..."
epochs = network_params_siam['epochs']
siamese_model.train2(iterations=epochs, chk_pnt_save_invl=network_params_siam['check_point_save_interval'])
siamese_model.save_model('final') #providing subscript for the final model saved, so the model name is final_blah_lah
if network_params_siam['write_summary']:
logdir = siamese_model._tf_sumry_wrtr._summary_dir
instruction = 'tensorboard --logdir=' + logdir
os.system(instruction)
def main():
batch_params = {
'buffer_size':45,
'batch_size': 20,
'data_file_indices': range(1,10),
'model_type':'siam',
'use_random_batches':False,
'files_per_read':10,
'load_pre_processd_data':True}
batch_creator = BatchCreator(batch_params)
x_batch1 = None; y_batch1 = None
x_batch2 = None; y_batch2 = None
#while loops are given since if we make batch_creator multi threaded, this will be
#useful
while x_batch1 is None and y_batch1 is None:
x_batch1, y_batch1, _ = batch_creator.get_batch()
while x_batch2 is None and y_batch2 is None:
x_batch2, y_batch2, _ = batch_creator.get_batch()
x_check_list = []
y_check_list = []
for itm1, itm2, itm3, itm4 in zip(x_batch1, x_batch2, y_batch1, y_batch2):
x_check_list.append((itm1==itm2).all())
y_check_list.append((itm3==itm4).all())
if any(x_check_list):
print "Found a match in x_batch1 and x_batch2"
else:
print "No matches found in x_batch1 and x_batch2"
if any(y_check_list):
print "Found a match in y_batch1 and y_batch2"
else:
print "No matches found in y_batch1 and y_batch2"
def train_fwd_model():
if network_params_fwd['write_summary']:
sess = tf.InteractiveSession()
else:
sess = tf.Session()
train_data_x = None
train_data_y = None
batch_creator = None
if network_params_fwd['batch_params'] is not None:
batch_creator = BatchCreator(network_params_fwd['batch_params'])
else:
train_data_x, train_data_y = get_data('train')
network_params_fwd['load_saved_model'] = False
forward_model = MDNPushFwdModel(sess=sess,
network_params=network_params_fwd)
forward_model.init_model()
forward_model.configure_data(data_x=train_data_x,
data_y=train_data_y,
batch_creator=batch_creator)
# h = forward_model.run_op('last_hidden',data_x)
# plt.figure(figsize=(8, 8))
# plt.plot(h, data_y,'ro', alpha=0.3)
# plt.show()
print "Got the data, gonna train the model..."
epochs = 100 #10000
loss = forward_model.train(epochs=epochs)
forward_model.save_model()
if network_params_fwd['write_summary']:
logdir = forward_model._tf_sumry_wrtr._summary_dir
instruction = 'tensorboard --logdir=' + logdir
os.system(instruction)
else:
plt.figure(figsize=(8, 8))
plt.plot(np.arange(100, epochs, 1), loss[100:], 'r-')
plt.show()