def train():
"""Training loop."""
tf.reset_default_graph()
# Create the motion models for training and evaluation
data_reader = dataset_reader.DataReader(
FLAGS.task_dataset_info, root=FLAGS.task_root, num_threads=4)
train_traj = data_reader.read(batch_size=FLAGS.training_minibatch_size)
# Create the ensembles that provide targets during training
place_cell_ensembles = utils.get_place_cell_ensembles(
env_size=FLAGS.task_env_size,
neurons_seed=FLAGS.task_neurons_seed,
targets_type=FLAGS.task_targets_type,
lstm_init_type=FLAGS.task_lstm_init_type,
n_pc=FLAGS.task_n_pc,
pc_scale=FLAGS.task_pc_scale)
head_direction_ensembles = utils.get_head_direction_ensembles(
neurons_seed=FLAGS.task_neurons_seed,
targets_type=FLAGS.task_targets_type,
lstm_init_type=FLAGS.task_lstm_init_type,
n_hdc=FLAGS.task_n_hdc,
hdc_concentration=FLAGS.task_hdc_concentration)
target_ensembles = place_cell_ensembles + head_direction_ensembles
# Model creation
rnn_core = model.GridCellsRNNCell(
target_ensembles=target_ensembles,
nh_lstm=FLAGS.model_nh_lstm,
nh_bottleneck=FLAGS.model_nh_bottleneck,
dropoutrates_bottleneck=np.array(FLAGS.model_dropout_rates),
bottleneck_weight_decay=FLAGS.model_weight_decay,
bottleneck_has_bias=FLAGS.model_bottleneck_has_bias,
init_weight_disp=FLAGS.model_init_weight_disp)
rnn = model.GridCellsRNN(rnn_core, FLAGS.model_nh_lstm)
# Get a trajectory batch
input_tensors = []
init_pos, init_hd, ego_vel, target_pos, target_hd = train_traj
if FLAGS.task_velocity_inputs:
# Add the required amount of noise to the velocities
vel_noise = tf.distributions.Normal(0.0, 1.0).sample(
sample_shape=ego_vel.get_shape()) * FLAGS.task_velocity_noise
input_tensors = [ego_vel + vel_noise] + input_tensors
# Concatenate all inputs
inputs = tf.concat(input_tensors, axis=2)
# Replace euclidean positions and angles by encoding of place and hd ensembles
# Note that the initial_conds will be zeros if the ensembles were configured
# to provide that type of initialization
initial_conds = utils.encode_initial_conditions(
init_pos, init_hd, place_cell_ensembles, head_direction_ensembles)
# Encode targets as well
ensembles_targets = utils.encode_targets(
target_pos, target_hd, place_cell_ensembles, head_direction_ensembles)
# Estimate future encoding of place and hd ensembles inputing egocentric vels
outputs, _ = rnn(initial_conds, inputs, training=True)
ensembles_logits, bottleneck, lstm_output = outputs
# Training loss
pc_loss = tf.nn.softmax_cross_entropy_with_logits_v2(
labels=ensembles_targets[0], logits=ensembles_logits[0], name='pc_loss')
hd_loss = tf.nn.softmax_cross_entropy_with_logits_v2(
labels=ensembles_targets[1], logits=ensembles_logits[1], name='hd_loss')
total_loss = pc_loss + hd_loss
train_loss = tf.reduce_mean(total_loss, name='train_loss')
# Optimisation ops
optimizer_class = eval(FLAGS.training_optimizer_class) # pylint: disable=eval-used
optimizer = optimizer_class(**eval(FLAGS.training_optimizer_options)) # pylint: disable=eval-used
grad = optimizer.compute_gradients(train_loss)
clip_gradient = eval(FLAGS.training_clipping_function) # pylint: disable=eval-used
clipped_grad = [
clip_gradient(g, var, FLAGS.training_clipping) for g, var in grad
]
train_op = optimizer.apply_gradients(clipped_grad)
# Store the grid scores
grid_scores = dict()
grid_scores['btln_60'] = np.zeros((FLAGS.model_nh_bottleneck,))
grid_scores['btln_90'] = np.zeros((FLAGS.model_nh_bottleneck,))
grid_scores['btln_60_separation'] = np.zeros((FLAGS.model_nh_bottleneck,))
grid_scores['btln_90_separation'] = np.zeros((FLAGS.model_nh_bottleneck,))
grid_scores['lstm_60'] = np.zeros((FLAGS.model_nh_lstm,))
grid_scores['lstm_90'] = np.zeros((FLAGS.model_nh_lstm,))
# Create scorer objects
starts = [0.2] * 10
ends = np.linspace(0.4, 1.0, num=10)
masks_parameters = zip(starts, ends.tolist())
latest_epoch_scorer = scores.GridScorer(20, data_reader.get_coord_range(),
masks_parameters)
with tf.train.SingularMonitoredSession() as sess:
for epoch in range(FLAGS.training_epochs):
loss_acc = list()
for _ in range(FLAGS.training_steps_per_epoch):
res = sess.run({'train_op': train_op, 'total_loss': train_loss})
loss_acc.append(res['total_loss'])
tf.logging.info('Epoch %i, mean loss %.5f, std loss %.5f', epoch,
np.mean(loss_acc), np.std(loss_acc))
if epoch % FLAGS.saver_eval_time == 0:
res = dict()
for _ in xrange(FLAGS.training_evaluation_minibatch_size //
FLAGS.training_minibatch_size):
mb_res = sess.run({
'bottleneck': bottleneck,
'lstm': lstm_output,
'pos_xy': target_pos
})
res = utils.concat_dict(res, mb_res)
# Store at the end of validation
filename = 'rates_and_sac_latest_hd.pdf'
grid_scores['btln_60'], grid_scores['btln_90'], grid_scores[
'btln_60_separation'], grid_scores[
'btln_90_separation'] = utils.get_scores_and_plot(
latest_epoch_scorer, res['pos_xy'], res['bottleneck'],
FLAGS.saver_results_directory, filename)
def train():
"""Training loop."""
# Create the ensembles that provide targets during training
place_cell_ensembles = utils.get_place_cell_ensembles(
env_size=FLAGS['task_env_size'],
neurons_seed=FLAGS['task_neurons_seed'],
targets_type=FLAGS['task_targets_type'],
lstm_init_type=FLAGS['task_lstm_init_type'],
n_pc=FLAGS['task_n_pc'],
pc_scale=FLAGS['task_pc_scale'])
head_direction_ensembles = utils.get_head_direction_ensembles(
neurons_seed=FLAGS['task_neurons_seed'],
targets_type=FLAGS['task_targets_type'],
lstm_init_type=FLAGS['task_lstm_init_type'],
n_hdc=FLAGS['task_n_hdc'],
hdc_concentration=FLAGS['task_hdc_concentration'])
target_ensembles = place_cell_ensembles + head_direction_ensembles
# Store the grid scores
'''grid_scores = dict()
grid_scores['btln_60'] = np.zeros((FLAGS['model_nh_bottleneck'],))
grid_scores['btln_90'] = np.zeros((FLAGS['model_nh_bottleneck'],))
grid_scores['btln_60_separation'] = np.zeros((FLAGS['model_nh_bottleneck'],))
grid_scores['btln_90_separation'] = np.zeros((FLAGS['model_nh_bottleneck'],))
grid_scores['lstm_60'] = np.zeros((FLAGS['model_nh_lstm'],))
grid_scores['lstm_90'] = np.zeros((FLAGS['model_nh_lstm'],))
# Create scorer objects
starts = [0.2] * 10
ends = np.linspace(0.4, 1.0, num=10)
masks_parameters = zip(starts, ends.tolist())
latest_epoch_scorer = scores.GridScorer(20, data_reader.get_coord_range(),
masks_parameters)'''
#tf.compat.v1.reset_default_graph()
data_reader = dataset_reader.DataReader(
FLAGS['task_dataset_info'],
root=FLAGS['task_root'],
num_threads=4,
batch_size=FLAGS['training_minibatch_size'])
# Model creation
rnn = model.GridCellNetwork(
target_ensembles=target_ensembles,
nh_lstm=FLAGS['model_nh_lstm'],
nh_bottleneck=FLAGS['model_nh_bottleneck'],
dropoutrates_bottleneck=np.array(FLAGS['model_dropout_rates']),
bottleneck_weight_decay=FLAGS['model_weight_decay'],
bottleneck_has_bias=FLAGS['model_bottleneck_has_bias'],
init_weight_disp=FLAGS['model_init_weight_disp'])
optimizer = tf.keras.optimizers.RMSprop(learning_rate=1e-5,
momentum=0.9,
clipvalue=1e-5)
for epoch in range(1000):
loss_metric = tf.keras.metrics.Mean()
for batch in range(1000):
train_traj = data_reader.read()
init_pos, init_hd, ego_vel, target_pos, target_hd = train_traj
input_tensors = []
if FLAGS['task_velocity_inputs']:
vel_noise = tfp.distributions.Normal(
0.0, 1.0).sample(sample_shape=tf.shape(
ego_vel)) * FLAGS['task_velocity_noise']
input_tensors = [ego_vel + vel_noise] + input_tensors
inputs = tf.concat(input_tensors, axis=2)
initial_conds = utils.encode_initial_conditions(
init_pos, init_hd, place_cell_ensembles,
head_direction_ensembles)
ensembles_targets = utils.encode_targets(target_pos, target_hd,
place_cell_ensembles,
head_direction_ensembles)
loss, gradients = train_step(inputs, initial_conds,
ensembles_targets, rnn)
back_pass(rnn, optimizer, gradients)
loss_metric(loss)
print("epoch {}, loss {}".format(epoch, loss_metric.result()))
loss_metric.reset_states()
dataset = Dataset(batch_size=data_params['batch_size'])
data_generator = data.DataLoader(dataset, **data_params)
test_generator = data.DataLoader(dataset, **test_params)
# Create the ensembles that provide targets during training
place_cell_ensembles = utils.get_place_cell_ensembles(
env_size=ENV_SIZE,
neurons_seed=SEED,
targets_type='softmax',
lstm_init_type='softmax',
n_pc=N_PC,
pc_scale=[0.01])
head_direction_ensembles = utils.get_head_direction_ensembles(
neurons_seed=SEED,
targets_type='softmax',
lstm_init_type='softmax',
n_hdc=N_HDC,
hdc_concentration=[20.])
target_ensembles = place_cell_ensembles + head_direction_ensembles
model = model.GridTorch(target_ensembles, (BATCH_SIZE, 100, 3)).cuda()
params = model.parameters()
def train():
"""Training loop."""
#tf.reset_default_graph()
# Create the motion models for training and evaluation
data_reader = dataset_reader.DataReader(
FLAGS['task_dataset_info'],
root=FLAGS['task_root'],
batch_size=FLAGS['training_minibatch_size'])
dataset = data_reader.read()
#train_traj = data_reader.read()
# Create the ensembles that provide targets during training
place_cell_ensembles = utils.get_place_cell_ensembles(
env_size=FLAGS['task_env_size'],
neurons_seed=FLAGS['task_neurons_seed'],
targets_type=FLAGS['task_targets_type'],
lstm_init_type=FLAGS['task_lstm_init_type'],
n_pc=FLAGS['task_n_pc'],
pc_scale=FLAGS['task_pc_scale'])
head_direction_ensembles = utils.get_head_direction_ensembles(
neurons_seed=FLAGS['task_neurons_seed'],
targets_type=FLAGS['task_targets_type'],
lstm_init_type=FLAGS['task_lstm_init_type'],
n_hdc=FLAGS['task_n_hdc'],
hdc_concentration=FLAGS['task_hdc_concentration'])
target_ensembles = place_cell_ensembles + head_direction_ensembles
'''
# Get a trajectory batch
input_tensors = []
init_pos, init_hd, ego_vel, target_pos, target_hd = train_traj
if FLAGS['task_velocity_inputs']:
# Add the required amount of noise to the velocities
vel_noise = tfb.distributions.Normal(0.0, 1.0).sample(
sample_shape=ego_vel.get_shape()) * FLAGS['task_velocity_noise']
input_tensors = [ego_vel + vel_noise] + input_tensors
# Concatenate all inputs
inputs = tf.concat(input_tensors, axis=2)
# Replace euclidean positions and angles by encoding of place and hd ensembles
# Note that the initial_conds will be zeros if the ensembles were configured
# to provide that type of initialization
initial_conds = utils.encode_initial_conditions(
init_pos, init_hd, place_cell_ensembles, head_direction_ensembles)
# Encode targets as well
ensembles_targets = utils.encode_targets(
target_pos, target_hd, place_cell_ensembles, head_direction_ensembles)
# Estimate future encoding of place and hd ensembles inputing egocentric vels
'''
#Defining model
grid_cell_model = GridCellNetwork(
target_ensembles=target_ensembles,
nh_lstm=FLAGS['model_nh_lstm'],
nh_bottleneck=FLAGS['model_nh_bottleneck'],
dropoutrates_bottleneck=FLAGS['model_dropout_rates'],
bottleneck_weight_decay=FLAGS['model_weight_decay'],
bottleneck_has_bias=FLAGS['model_bottleneck_has_bias'],
init_weight_disp=FLAGS['model_init_weight_disp'],
)
# Store the grid scores
grid_scores = dict()
grid_scores['btln_60'] = np.zeros((FLAGS['model_nh_bottleneck'], ))
grid_scores['btln_90'] = np.zeros((FLAGS['model_nh_bottleneck'], ))
grid_scores['btln_60_separation'] = np.zeros(
(FLAGS['model_nh_bottleneck'], ))
grid_scores['btln_90_separation'] = np.zeros(
(FLAGS['model_nh_bottleneck'], ))
grid_scores['lstm_60'] = np.zeros((FLAGS['model_nh_lstm'], ))
grid_scores['lstm_90'] = np.zeros((FLAGS['model_nh_lstm'], ))
# Create scorer objects
starts = [0.2] * 10
ends = np.linspace(0.4, 1.0, num=10)
masks_parameters = zip(starts, ends.tolist())
latest_epoch_scorer = scores.GridScorer(20, data_reader.get_coord_range(),
masks_parameters)
# we can run without feed_dict either because we are in the singular monitored session
# so we use fetches on sess.run
# pos_xy is simply the input data. In tf1 it's the computed graph up until target_pos, which is very early and basically just fetched the data.
optimizer = tf.keras.optimizers.RMSprop(
learning_rate=1e-5, momentum=0.9, clipvalue=FLAGS['training_clipping'])
train_loss = tf.keras.metrics.Mean(name='train_loss')
train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(
name='train_accuracy')
pos_xy = []
bottleneck = []
lstm_output = []
@tf.function
def loss_function(targets, logits):
pc_loss = tf.nn.softmax_cross_entropy_with_logits(labels=targets[0],
logits=logits[0],
name='pc_loss')
hd_loss = tf.nn.softmax_cross_entropy_with_logits(labels=targets[1],
logits=logits[1],
name='hd_loss')
total_loss = pc_loss + hd_loss
return tf.reduce_mean(total_loss, name='train_loss')
@tf.function
def train_step(velocities, targets, logits):
global preprocess_time
with tf.GradientTape() as tape:
predictions, _ = grid_cell_model(velocities,
initial_conds,
trainable=True)
ensembles_logits, bottleneck, lstm_output = predictions
loss = loss_function(targets, ensembles_logits)
gradients = tape.gradient(loss, grid_cell_model.trainable_weights)
optimizer.apply_gradients(
zip(gradients, grid_cell_model.trainable_weights))
train_loss(loss)
return {
"bottleneck": bottleneck,
"lstm_output": lstm_output,
"pos_xy": target_pos
}
for epoch in range(FLAGS['training_epochs']):
loss_acc = list()
res = dict()
#for _ in range(FLAGS['training_steps_per_epoch']):
train_loss.reset_states()
for batch, train_trajectory in enumerate(dataset):
print(batch)
'''some preprocessing that maybe should be done in the data_pipeline'''
start_time = time.time()
init_pos = train_trajectory['init_pos']
init_hd = train_trajectory['init_hd']
ego_vel = train_trajectory['ego_vel']
target_pos = train_trajectory['target_pos']
target_hd = train_trajectory['target_hd']
input_tensors = []
if FLAGS['task_velocity_inputs']:
# Add the required amount of noise to the velocities
vel_noise = tfb.distributions.Normal(
0.0, 1.0).sample(sample_shape=tf.shape(
ego_vel)) * FLAGS['task_velocity_noise']
input_tensors = [ego_vel + vel_noise] + input_tensors
velocities = tf.concat(input_tensors, axis=2)
initial_conds = utils.encode_initial_conditions(
init_pos, init_hd, place_cell_ensembles,
head_direction_ensembles)
ensembles_targets = utils.encode_targets(target_pos, target_hd,
place_cell_ensembles,
head_direction_ensembles)
mb_res = train_step(velocities, ensembles_targets, initial_conds)
#res = utils.concat_dict(res, mb_res)
if batch % 1000 > 600:
pos_xy.append(mb_res['pos_xy'])
bottleneck.append(mb_res['bottleneck'])
lstm_output.append(mb_res['lstm_output'])
if batch % 1000 == 0 and batch != 0:
print(preprocess_time)
print('Epoch {}, batch {}, loss {}'.format(
epoch, batch, train_loss.result()))
for i in range(len(pos_xy)):
mb_res = {
"bottleneck": bottleneck[i],
"lstm_out": lstm_output[i],
"pos_xy": pos_xy[i]
}
utils.concat_dict(res, mb_res)
pos_xy = []
bottleneck = []
lstm_output = []
mb_res = dict()
# Store at the end of validation
filename = 'rates_and_sac_latest_hd.pdf'
grid_scores['btln_60'], grid_scores['btln_90'], grid_scores[
'btln_60_separation'], grid_scores[
'btln_90_separation'] = utils.get_scores_and_plot(
latest_epoch_scorer, res['pos_xy'],
res['bottleneck'],
FLAGS['saver_results_directory'], filename)
res = dict()