def initialize_folder(self, save_location, folder_name):
parent_folder = os.path.join(save_location, folder_name)
# make folder if it does not exist
if not gfile.IsDirectory(parent_folder):
gfile.MkDir(parent_folder)
self.parent_folder = parent_folder
save_location = os.path.join(parent_folder, self.short_filename)
if not gfile.IsDirectory(save_location):
gfile.MkDir(save_location)
self.save_location = save_location
self.save_filename = os.path.join(self.save_location, self.short_filename)
def initialize_folder(self, save_location, folder_name):
"""Intialize saving location of the model."""
parent_folder = os.path.join(save_location, folder_name)
# Make folder if it does not exist.
if not gfile.IsDirectory(parent_folder):
gfile.MkDir(parent_folder)
self.parent_folder = parent_folder
save_location = os.path.join(parent_folder, self.short_filename)
if not gfile.IsDirectory(save_location):
gfile.MkDir(save_location)
self.save_location = save_location
self.save_filename = os.path.join(self.save_location, self.short_filename)
def convert_to_TFRecords_chunks(self,
prefix,
save_location='~/tmp',
examples_per_file=1000):
# converts the stimulus-response data into a TFRecords file.
tf.logging.info('making TFRecords chunks')
stimulus = self.stimulus.astype(np.float32)
response = self.response.astype(np.float32)
num_examples = stimulus.shape[0]
num_files = np.ceil(num_examples / examples_per_file).astype(np.int)
tf.logging.info('Number of files: %d, examples: %d' %
(num_files, num_examples))
def _bytes_feature(value):
# value: bytes to convert to tf.train.Feature
return tf.train.Feature(bytes_list=tf.train.BytesList(
value=[value]))
# make folder for storing .tfrecords files
folder_tfrec = os.path.join(save_location, prefix)
# Make folder if it does not exist.
if not gfile.IsDirectory(folder_tfrec):
tf.logging.info('making folder to store tfrecords')
gfile.MkDir(folder_tfrec)
else:
tf.logging.info('folder exists, will overwrite results')
index = -1
for ifile in range(num_files):
filename = os.path.join(folder_tfrec,
'chunk_' + str(ifile) + '.tfrecords')
tf.logging.info('Writing %s , starting index %d' %
(filename, index + 1))
writer = tf.python_io.TFRecordWriter(filename)
for iexample in range(examples_per_file):
index += 1
stimulus_raw = stimulus[index, :].tostring()
response_raw = response[index, :].tostring()
#print(index, stimulus[index,:].shape, response[index,:].shape)
example = tf.train.Example(features=tf.train.Features(
feature={
'stimulus': _bytes_feature(stimulus_raw),
'response': _bytes_feature(response_raw)
}))
writer.write(example.SerializeToString())
writer.close()
def initialize_model(save_folder, file_name, sess):
"""Setup model variables and saving information.
Args:
save_folder (string) : Folder to store model.
Makes one if it does not exist.
filename (string) : Prefix of model/checkpoint files.
sess : Tensorflow session.
"""
# Make folder.
if not gfile.IsDirectory(save_folder):
gfile.MkDir(save_folder)
# Initialize variables.
saver_var, start_iter = initialize_variables(sess, save_folder, file_name)
return saver_var, start_iter
def RunComputation():
# filename for saving file
if FLAGS.architecture == '2 layer_stimulus':
architecture_string = ('_architecture=' + str(FLAGS.architecture) +
'_stim_downsample_window=' +
str(FLAGS.stim_downsample_window) +
'_stim_downsample_stride=' +
str(FLAGS.stim_downsample_stride))
else:
architecture_string = ('_architecture=' + str(FLAGS.architecture))
short_filename = ('model=' + str(FLAGS.model_id) + '_loss=' +
str(FLAGS.loss) + '_batch_sz=' + str(FLAGS.batchsz) +
'_lam_w=' + str(FLAGS.lam_w) + '_step_sz' +
str(FLAGS.step_sz) + '_tlen=' + str(FLAGS.train_len) +
'_window=' + str(FLAGS.window) + '_stride=' +
str(FLAGS.stride) + str(architecture_string) + '_jitter')
# make a folder with name derived from parameters of the algorithm - it saves checkpoint files and summaries used in tensorboard
parent_folder = FLAGS.save_location + FLAGS.folder_name + '/'
# make folder if it does not exist
if not gfile.IsDirectory(parent_folder):
gfile.MkDir(parent_folder)
FLAGS.save_location = parent_folder + short_filename + '/'
print('Does the file exist?', gfile.IsDirectory(FLAGS.save_location))
if not gfile.IsDirectory(FLAGS.save_location):
gfile.MkDir(FLAGS.save_location)
save_filename = FLAGS.save_location + short_filename
"""Main function which runs all TensorFlow computations."""
with tf.Graph().as_default() as gra:
with tf.device(tf.ReplicaDeviceSetter(FLAGS.ps_tasks)):
print(FLAGS.config_params)
tf.logging.info(FLAGS.config_params)
# set up training dataset
tc_mean = get_data_mat.init_chunks(FLAGS.n_chunks)
'''
# plot histogram of a training dataset
stim_train, resp_train, train_len = get_data_mat.get_stim_resp('train',
num_chunks=FLAGS.num_chunks_to_load)
plt.hist(np.ndarray.flatten(stim_train[:,:,0:]))
plt.show()
plt.draw()
'''
# Create computation graph.
#
# Graph should be fully constructed before you create supervisor.
# Attempt to modify graph after supervisor is created will cause an error.
with tf.name_scope('model'):
if FLAGS.architecture == '1 layer':
# single GPU model
if False:
global_step = tf.contrib.framework.create_global_step()
model, stim, resp = jitter_model.approximate_conv_jitter(
FLAGS.n_cells, FLAGS.lam_w, FLAGS.window,
FLAGS.stride, FLAGS.step_sz, tc_mean,
FLAGS.su_channels)
# multiGPU model
if True:
model, stim, resp, global_step = jitter_model.approximate_conv_jitter_multigpu(
FLAGS.n_cells, FLAGS.lam_w, FLAGS.window,
FLAGS.stride, FLAGS.step_sz, tc_mean,
FLAGS.su_channels, FLAGS.config_params)
if FLAGS.architecture == '2 layer_stimulus':
# stimulus is first smoothened to lower dimensions, then same model is applied
print(' put stimulus to lower dimenstions!')
model, stim, resp, global_step, stim_tuple = jitter_model.approximate_conv_jitter_multigpu_stim_lr(
FLAGS.n_cells, FLAGS.lam_w, FLAGS.window, FLAGS.stride,
FLAGS.step_sz, tc_mean, FLAGS.su_channels,
FLAGS.config_params, FLAGS.stim_downsample_window,
FLAGS.stim_downsample_stride)
# Print the number of variables in graph
print('Calculating model size') # Hope we do not exceed memory
PrintModelAnalysis(gra, max_depth=10)
#import pdb; pdb.set_trace()
# Builds our summary op.
summary_op = model.merged_summary
# Create a Supervisor. It will take care of initialization, summaries,
# checkpoints, and recovery.
#
# When multiple replicas of this program are running, the first one,
# identified by --task=0 is the 'chief' supervisor. It is the only one
# that takes case of initialization, etc.
is_chief = (FLAGS.task == 0) # & (FLAGS.learn==1)
print(save_filename)
if FLAGS.learn == 1:
# use supervisor only for learning,
# otherwise it messes up data as it tries to store variables while you are doing analysis
sv = tf.train.Supervisor(logdir=save_filename,
is_chief=is_chief,
saver=tf.train.Saver(),
summary_op=None,
save_model_secs=100,
global_step=global_step,
recovery_wait_secs=5)
if (is_chief and FLAGS.learn == 1):
tf.train.write_graph(tf.get_default_graph().as_graph_def(),
save_filename, 'graph.pbtxt')
# Get an initialized, and possibly recovered session. Launch the
# services: Checkpointing, Summaries, step counting.
#
# When multiple replicas of this program are running the services are
# only launched by the 'chief' replica.
session_config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
#import pdb; pdb.set_trace()
sess = sv.PrepareSession(FLAGS.master, config=session_config)
FitComputation(sv, sess, model, stim, resp, global_step,
summary_op, stim_tuple)
sv.Stop()
else:
# if not learn, then analyse
session_config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
with tf.Session(config=session_config) as sess:
saver_var = tf.train.Saver(
tf.all_variables(),
keep_checkpoint_every_n_hours=float('inf'))
restore_file = tf.train.latest_checkpoint(save_filename)
print(restore_file)
start_iter = int(
restore_file.split('/')[-1].split('-')[-1])
saver_var.restore(sess, restore_file)
if FLAGS.architecture == '2 layer_stimulus':
AnalyseModel_lr(sess, model)
else:
AnalyseModel(sv, sess, model)
def main(argv):
# global variables will be used for getting training data
global cells_choose
global chosen_mask
global chunk_order
# set random seeds: when same algorithm run with different FLAGS,
# the sequence of random data is same.
np.random.seed(FLAGS.np_randseed)
random.seed(FLAGS.randseed)
# initial chunk order (will be re-shuffled everytime we go over a chunk)
chunk_order = np.random.permutation(np.arange(FLAGS.n_chunks - 1))
# Load data summary
data_filename = FLAGS.data_location + 'data_details.mat'
summary_file = gfile.Open(data_filename, 'r')
data_summary = sio.loadmat(summary_file)
cells = np.squeeze(data_summary['cells'])
# which cells to train subunits for
if FLAGS.all_cells == 'True':
cells_choose = np.array(np.ones(np.shape(cells)), dtype='bool')
else:
cells_choose = (cells == 3287) | (cells == 3318) | (cells == 3155) | (
cells == 3066)
n_cells = np.sum(cells_choose) # number of cells
# load spikes and relevant stimulus pixels for chosen cells
tot_spks = np.squeeze(data_summary['tot_spks'])
tot_spks_chosen_cells = np.array(tot_spks[cells_choose], dtype='float32')
total_mask = np.squeeze(data_summary['totalMaskAccept_log']).T
# chosen_mask = which pixels to learn subunits over
if FLAGS.masked_stimulus == 'True':
chosen_mask = np.array(np.sum(total_mask[cells_choose, :], 0) > 0,
dtype='bool')
else:
chosen_mask = np.array(np.ones(3200).astype('bool'))
stim_dim = np.sum(chosen_mask) # stimulus dimensions
print('\ndataset summary loaded')
# print parameters
print('Save folder name: ' + str(FLAGS.folder_name) + '\nmodel:' +
str(FLAGS.model_id) + '\nLoss:' + str(FLAGS.loss) +
'\nmasked stimulus:' + str(FLAGS.masked_stimulus) + '\nall_cells?' +
str(FLAGS.all_cells) + '\nbatch size' + str(FLAGS.batchsz) +
'\nstep size' + str(FLAGS.step_sz) + '\ntraining length: ' +
str(FLAGS.train_len) + '\nn_cells: ' + str(n_cells))
# decide the number of subunits to fit
n_su = FLAGS.ratio_SU * n_cells
# filename for saving file
short_filename = ('_masked_stim=' + str(FLAGS.masked_stimulus) +
'_all_cells=' + str(FLAGS.all_cells) + '_loss=' +
str(FLAGS.loss) + '_batch_sz=' + str(FLAGS.batchsz) +
'_step_sz' + str(FLAGS.step_sz) + '_tlen=' +
str(FLAGS.train_len) + '_bg')
with tf.Session() as sess:
# set up stimulus and response palceholders
stim = tf.placeholder(tf.float32, shape=[None, stim_dim], name='stim')
resp = tf.placeholder(tf.float32, name='resp')
data_len = tf.placeholder(tf.float32, name='data_len')
if FLAGS.loss == 'poisson':
b_init = np.array(
0.000001 * np.ones(n_cells)
) # a very small positive bias needed to avoid log(0) in poisson loss
else:
b_init = np.log(
(tot_spks_chosen_cells) / (216000. - tot_spks_chosen_cells)
) # log-odds, a good initialization for some losses (like logistic)
# different firing rate models
if FLAGS.model_id == 'exp_additive':
# This model was implemented for earlier work.
# firing rate for cell c: lam_c = sum_s exp(w_s.x + a_sc)
# filename
short_filename = ('model=' + str(FLAGS.model_id) + short_filename)
# variables
w = tf.Variable(np.array(0.01 * np.random.randn(stim_dim, n_su),
dtype='float32'),
name='w')
a = tf.Variable(np.array(0.01 * np.random.rand(n_cells, 1, n_su),
dtype='float32'),
name='a')
# firing rate model
lam = tf.transpose(tf.reduce_sum(tf.exp(tf.matmul(stim, w) + a),
2))
regularization = 0
vars_fit = [w, a]
def proj(
): # called after every training step - to project to parameter constraints
pass
if FLAGS.model_id == 'relu':
# firing rate for cell c: lam_c = a_c'.relu(w.x) + b
# we know a>0 and for poisson loss, b>0
# for poisson loss: small b added to prevent lam_c going to 0
# filename
short_filename = ('model=' + str(FLAGS.model_id) + '_lam_w=' +
str(FLAGS.lam_w) + '_lam_a=' + str(FLAGS.lam_a) +
'_nsu=' + str(n_su) + short_filename)
# variables
w = tf.Variable(np.array(0.01 * np.random.randn(stim_dim, n_su),
dtype='float32'),
name='w')
a = tf.Variable(np.array(0.01 * np.random.rand(n_su, n_cells),
dtype='float32'),
name='a')
b = tf.Variable(np.array(b_init, dtype='float32'), name='b')
# firing rate model
lam = tf.matmul(tf.nn.relu(tf.matmul(stim, w)), a) + b
vars_fit = [w, a] # which variables are learnt
if not FLAGS.loss == 'poisson': # don't learn b for poisson loss
vars_fit = vars_fit + [b]
# regularization of parameters
regularization = (FLAGS.lam_w * tf.reduce_sum(tf.abs(w)) +
FLAGS.lam_a * tf.reduce_sum(tf.abs(a)))
# projection to satisfy constraints
a_pos = tf.assign(a, (a + tf.abs(a)) / 2)
b_pos = tf.assign(b, (b + tf.abs(b)) / 2)
def proj():
sess.run(a_pos)
if FLAGS.loss == 'poisson':
sess.run(b_pos)
if FLAGS.model_id == 'relu_window':
# firing rate for cell c: lam_c = a_c'.relu(w.x) + b,
# where w_i are over a small window which are convolutionally related with each other.
# we know a>0 and for poisson loss, b>0
# for poisson loss: small b added to prevent lam_c going to 0
# filename
short_filename = ('model=' + str(FLAGS.model_id) + '_window=' +
str(FLAGS.window) + '_stride=' +
str(FLAGS.stride) + '_lam_w=' +
str(FLAGS.lam_w) + short_filename)
mask_tf, dimx, dimy, n_pix = get_windows(
) # get convolutional windows
# variables
w = tf.Variable(np.array(0.1 +
0.05 * np.random.rand(dimx, dimy, n_pix),
dtype='float32'),
name='w')
a = tf.Variable(np.array(np.random.rand(dimx * dimy, n_cells),
dtype='float32'),
name='a')
b = tf.Variable(np.array(b_init, dtype='float32'), name='b')
vars_fit = [w, a] # which variables are learnt
if not FLAGS.loss == 'poisson': # don't learn b for poisson loss
vars_fit = vars_fit + [b]
# stimulus filtered with convolutional windows
stim4D = tf.expand_dims(tf.reshape(stim, (-1, 40, 80)), 3)
stim_masked = tf.nn.conv2d(
stim4D,
mask_tf,
strides=[1, FLAGS.stride, FLAGS.stride, 1],
padding="VALID")
stim_wts = tf.nn.relu(tf.reduce_sum(tf.mul(stim_masked, w), 3))
# get firing rate
lam = tf.matmul(tf.reshape(stim_wts, [-1, dimx * dimy]), a) + b
# regularization
regularization = FLAGS.lam_w * tf.reduce_sum(tf.nn.l2_loss(w))
# projection to satisfy hard variable constraints
a_pos = tf.assign(a, (a + tf.abs(a)) / 2)
b_pos = tf.assign(b, (b + tf.abs(b)) / 2)
def proj():
sess.run(a_pos)
if FLAGS.loss == 'poisson':
sess.run(b_pos)
if FLAGS.model_id == 'relu_window_mother':
# firing rate for cell c: lam_c = a_c'.relu(w.x) + b,
# where w_i are over a small window which are convolutionally related with each other.
# w_i = w_mother + w_del_i,
# where w_mother is common accross all 'windows' and w_del is different for different windows.
# we know a>0 and for poisson loss, b>0
# for poisson loss: small b added to prevent lam_c going to 0
# filename
short_filename = ('model=' + str(FLAGS.model_id) + '_window=' +
str(FLAGS.window) + '_stride=' +
str(FLAGS.stride) + '_lam_w=' +
str(FLAGS.lam_w) + short_filename)
mask_tf, dimx, dimy, n_pix = get_windows()
# variables
w_del = tf.Variable(np.array(0.05 *
np.random.randn(dimx, dimy, n_pix),
dtype='float32'),
name='w_del')
w_mother = tf.Variable(np.array(np.ones(
(2 * FLAGS.window + 1, 2 * FLAGS.window + 1, 1, 1)),
dtype='float32'),
name='w_mother')
a = tf.Variable(np.array(np.random.rand(dimx * dimy, n_cells),
dtype='float32'),
name='a')
b = tf.Variable(np.array(b_init, dtype='float32'), name='b')
vars_fit = [w_mother, w_del, a] # which variables to learn
if not FLAGS.loss == 'poisson':
vars_fit = vars_fit + [b]
# stimulus filtered with convolutional windows
stim4D = tf.expand_dims(tf.reshape(stim, (-1, 40, 80)), 3)
stim_convolved = tf.reduce_sum(
tf.nn.conv2d(stim4D,
w_mother,
strides=[1, FLAGS.stride, FLAGS.stride, 1],
padding="VALID"), 3)
stim_masked = tf.nn.conv2d(
stim4D,
mask_tf,
strides=[1, FLAGS.stride, FLAGS.stride, 1],
padding="VALID")
stim_del = tf.reduce_sum(tf.mul(stim_masked, w_del), 3)
# activation of differnet subunits
su_act = tf.nn.relu(stim_del + stim_convolved)
# get firing rate
lam = tf.matmul(tf.reshape(su_act, [-1, dimx * dimy]), a) + b
# regularization
regularization = FLAGS.lam_w * tf.reduce_sum(tf.nn.l2_loss(w_del))
# projection to satisfy hard variable constraints
a_pos = tf.assign(a, (a + tf.abs(a)) / 2)
b_pos = tf.assign(b, (b + tf.abs(b)) / 2)
def proj():
sess.run(a_pos)
if FLAGS.loss == 'poisson':
sess.run(b_pos)
if FLAGS.model_id == 'relu_window_mother_sfm':
# firing rate for cell c: lam_c = a_sfm_c'.relu(w.x) + b,
# a_sfm_c = softmax(a) : so a cell cannot be connected to all subunits equally well.
# where w_i are over a small window which are convolutionally related with each other.
# w_i = w_mother + w_del_i,
# where w_mother is common accross all 'windows' and w_del is different for different windows.
# we know a>0 and for poisson loss, b>0
# for poisson loss: small b added to prevent lam_c going to 0
# filename
short_filename = ('model=' + str(FLAGS.model_id) + '_window=' +
str(FLAGS.window) + '_stride=' +
str(FLAGS.stride) + '_lam_w=' +
str(FLAGS.lam_w) + short_filename)
mask_tf, dimx, dimy, n_pix = get_windows()
# variables
w_del = tf.Variable(np.array(0.05 *
np.random.randn(dimx, dimy, n_pix),
dtype='float32'),
name='w_del')
w_mother = tf.Variable(np.array(np.ones(
(2 * FLAGS.window + 1, 2 * FLAGS.window + 1, 1, 1)),
dtype='float32'),
name='w_mother')
a = tf.Variable(np.array(np.random.randn(dimx * dimy, n_cells),
dtype='float32'),
name='a')
a_sfm = tf.transpose(tf.nn.softmax(tf.transpose(a)))
b = tf.Variable(np.array(b_init, dtype='float32'), name='b')
vars_fit = [w_mother, w_del, a] # which variables to fit
if not FLAGS.loss == 'poisson':
vars_fit = vars_fit + [b]
# stimulus filtered with convolutional windows
stim4D = tf.expand_dims(tf.reshape(stim, (-1, 40, 80)), 3)
stim_convolved = tf.reduce_sum(
tf.nn.conv2d(stim4D,
w_mother,
strides=[1, FLAGS.stride, FLAGS.stride, 1],
padding="VALID"), 3)
stim_masked = tf.nn.conv2d(
stim4D,
mask_tf,
strides=[1, FLAGS.stride, FLAGS.stride, 1],
padding="VALID")
stim_del = tf.reduce_sum(tf.mul(stim_masked, w_del), 3)
# activation of differnet subunits
su_act = tf.nn.relu(stim_del + stim_convolved)
# get firing rate
lam = tf.matmul(tf.reshape(su_act, [-1, dimx * dimy]), a_sfm) + b
# regularization
regularization = FLAGS.lam_w * tf.reduce_sum(tf.nn.l2_loss(w_del))
# projection to satisfy hard variable constraints
b_pos = tf.assign(b, (b + tf.abs(b)) / 2)
def proj():
if FLAGS.loss == 'poisson':
sess.run(b_pos)
if FLAGS.model_id == 'relu_window_mother_sfm_exp':
# firing rate for cell c: lam_c = exp(a_sfm_c'.relu(w.x)) + b,
# a_sfm_c = softmax(a) : so a cell cannot be connected to all subunits equally well.
# exponential output NL would cancel the log() in poisson and might get better estimation properties.
# where w_i are over a small window which are convolutionally related with each other.
# w_i = w_mother + w_del_i,
# where w_mother is common accross all 'windows' and w_del is different for different windows.
# we know a>0 and for poisson loss, b>0
# for poisson loss: small b added to prevent lam_c going to 0
# filename
short_filename = ('model=' + str(FLAGS.model_id) + '_window=' +
str(FLAGS.window) + '_stride=' +
str(FLAGS.stride) + '_lam_w=' +
str(FLAGS.lam_w) + short_filename)
# get windows
mask_tf, dimx, dimy, n_pix = get_windows()
# declare variables
w_del = tf.Variable(np.array(0.05 *
np.random.randn(dimx, dimy, n_pix),
dtype='float32'),
name='w_del')
w_mother = tf.Variable(np.array(np.ones(
(2 * FLAGS.window + 1, 2 * FLAGS.window + 1, 1, 1)),
dtype='float32'),
name='w_mother')
a = tf.Variable(np.array(np.random.randn(dimx * dimy, n_cells),
dtype='float32'),
name='a')
a_sfm = tf.transpose(tf.nn.softmax(tf.transpose(a)))
b = tf.Variable(np.array(b_init, dtype='float32'), name='b')
vars_fit = [w_mother, w_del, a]
if not FLAGS.loss == 'poisson':
vars_fit = vars_fit + [b]
# filter stimulus
stim4D = tf.expand_dims(tf.reshape(stim, (-1, 40, 80)), 3)
stim_convolved = tf.reduce_sum(
tf.nn.conv2d(stim4D,
w_mother,
strides=[1, FLAGS.stride, FLAGS.stride, 1],
padding="VALID"), 3)
stim_masked = tf.nn.conv2d(
stim4D,
mask_tf,
strides=[1, FLAGS.stride, FLAGS.stride, 1],
padding="VALID")
stim_del = tf.reduce_sum(tf.mul(stim_masked, w_del), 3)
# get subunit activation
su_act = tf.nn.relu(stim_del + stim_convolved)
# get cell firing rates
lam = tf.exp(
tf.matmul(tf.reshape(su_act, [-1, dimx * dimy]), a_sfm)) + b
# regularization
regularization = FLAGS.lam_w * tf.reduce_sum(tf.nn.l2_loss(w_del))
# projection to satisfy hard variable constraints
b_pos = tf.assign(b, (b + tf.abs(b)) / 2)
def proj():
if FLAGS.loss == 'poisson':
sess.run(b_pos)
# different loss functions
if FLAGS.loss == 'poisson':
loss_inter = (tf.reduce_sum(lam) / 120. -
tf.reduce_sum(resp * tf.log(lam))) / data_len
if FLAGS.loss == 'logistic':
loss_inter = tf.reduce_sum(tf.nn.softplus(
-2 * (resp - 0.5) * lam)) / data_len
if FLAGS.loss == 'hinge':
loss_inter = tf.reduce_sum(
tf.nn.relu(1 - 2 * (resp - 0.5) * lam)) / data_len
loss = loss_inter + regularization # add regularization to get final loss function
# training consists of calling training()
# which performs a train step and
# project parameters to model specific constraints using proj()
train_step = tf.train.AdagradOptimizer(FLAGS.step_sz).minimize(
loss, var_list=vars_fit)
def training(inp_dict):
sess.run(train_step,
feed_dict=inp_dict) # one step of gradient descent
proj() # model specific projection operations
# evaluate loss on given data.
def get_loss(inp_dict):
ls = sess.run(loss, feed_dict=inp_dict)
return ls
# saving details
# make a folder with name derived from parameters of the algorithm
# - it saves checkpoint files and summaries used in tensorboard
parent_folder = FLAGS.save_location + FLAGS.folder_name + '/'
# make folder if it does not exist
if not gfile.IsDirectory(parent_folder):
gfile.MkDir(parent_folder)
FLAGS.save_location = parent_folder + short_filename + '/'
if not gfile.IsDirectory(FLAGS.save_location):
gfile.MkDir(FLAGS.save_location)
save_filename = FLAGS.save_location + short_filename
# create summary writers
# create histogram summary for all parameters which are learnt
for ivar in vars_fit:
tf.histogram_summary(ivar.name, ivar)
# loss summary
l_summary = tf.scalar_summary('loss', loss)
# loss without regularization summary
l_inter_summary = tf.scalar_summary('loss_inter', loss_inter)
# Merge all the summary writer ops into one op (this way,
# calling one op stores all summaries)
merged = tf.merge_all_summaries()
# training and testing has separate summary writers
train_writer = tf.train.SummaryWriter(FLAGS.save_location + 'train',
sess.graph)
test_writer = tf.train.SummaryWriter(FLAGS.save_location + 'test')
## Fitting procedure
print('Start fitting')
sess.run(tf.initialize_all_variables())
saver_var = tf.train.Saver(tf.all_variables(),
keep_checkpoint_every_n_hours=0.05)
load_prev = False
start_iter = 0
try:
# restore previous fits if they are available
# - useful when programs are preempted frequently on .
latest_filename = short_filename + '_latest_fn'
restore_file = tf.train.latest_checkpoint(FLAGS.save_location,
latest_filename)
# restore previous iteration count and start from there.
start_iter = int(restore_file.split('/')[-1].split('-')[-1])
saver_var.restore(sess, restore_file) # restore variables
load_prev = True
except:
print('No previous dataset')
if load_prev:
print('Previous results loaded')
else:
print('Variables initialized')
# Finally, do fitting
icnt = 0
# get test data and make test dictionary
stim_test, resp_test, test_length = get_test_data()
fd_test = {stim: stim_test, resp: resp_test, data_len: test_length}
for istep in np.arange(start_iter, 400000):
print(istep)
# get training data and make test dictionary
stim_train, resp_train, train_len = get_next_training_batch(istep)
fd_train = {
stim: stim_train,
resp: resp_train,
data_len: train_len
}
# take training step
training(fd_train)
if istep % 10 == 0:
# compute training and testing losses
ls_train = get_loss(fd_train)
ls_test = get_loss(fd_test)
latest_filename = short_filename + '_latest_fn'
saver_var.save(sess,
save_filename,
global_step=istep,
latest_filename=latest_filename)
# add training summary
summary = sess.run(merged, feed_dict=fd_train)
train_writer.add_summary(summary, istep)
# add testing summary
summary = sess.run(merged, feed_dict=fd_test)
test_writer.add_summary(summary, istep)
print(istep, ls_train, ls_test)
icnt += FLAGS.batchsz
if icnt > 216000 - 1000:
icnt = 0
tms = np.random.permutation(np.arange(216000 - 1000))
def main(argv):
print('\nCode started')
global cells_choose
global chosen_mask
np.random.seed(FLAGS.np_randseed)
random.seed(FLAGS.randseed)
global chunk_order
chunk_order = np.random.permutation(np.arange(FLAGS.n_chunks - 1))
## Load data summary
filename = FLAGS.data_location + 'data_details.mat'
summary_file = gfile.Open(filename, 'r')
data_summary = sio.loadmat(summary_file)
cells = np.squeeze(data_summary['cells'])
if FLAGS.model_id == 'poisson' or FLAGS.model_id == 'logistic' or FLAGS.model_id == 'hinge' or FLAGS.model_id == 'poisson_relu':
cells_choose = (cells == 3287) | (cells == 3318) | (cells == 3155) | (
cells == 3066)
if FLAGS.model_id == 'poisson_full':
cells_choose = np.array(np.ones(np.shape(cells)), dtype='bool')
n_cells = np.sum(cells_choose)
tot_spks = np.squeeze(data_summary['tot_spks'])
total_mask = np.squeeze(data_summary['totalMaskAccept_log']).T
tot_spks_chosen_cells = np.array(tot_spks[cells_choose], dtype='float32')
chosen_mask = np.array(np.sum(total_mask[cells_choose, :], 0) > 0,
dtype='bool')
print(np.shape(chosen_mask))
print(np.sum(chosen_mask))
stim_dim = np.sum(chosen_mask)
print(FLAGS.model_id)
print('\ndataset summary loaded')
# use stim_dim, chosen_mask, cells_choose, tot_spks_chosen_cells, n_cells
# decide the number of subunits to fit
n_su = FLAGS.ratio_SU * n_cells
# saving details
if FLAGS.model_id == 'poisson':
short_filename = ('data_model=ASM_pop_batch_sz=' + str(FLAGS.batchsz) +
'_n_b_in_c' + str(FLAGS.n_b_in_c) + '_step_sz' +
str(FLAGS.step_sz) + '_tlen=' +
str(FLAGS.train_len) + '_bg')
else:
short_filename = ('data_model=' + str(FLAGS.model_id) + '_batch_sz=' +
str(FLAGS.batchsz) + '_n_b_in_c' +
str(FLAGS.n_b_in_c) + '_step_sz' +
str(FLAGS.step_sz) + '_tlen=' +
str(FLAGS.train_len) + '_bg')
parent_folder = FLAGS.save_location + FLAGS.folder_name + '/'
if not gfile.IsDirectory(parent_folder):
gfile.MkDir(parent_folder)
FLAGS.save_location = parent_folder + short_filename + '/'
print(gfile.IsDirectory(FLAGS.save_location))
if not gfile.IsDirectory(FLAGS.save_location):
gfile.MkDir(FLAGS.save_location)
print(FLAGS.save_location)
save_filename = FLAGS.save_location + short_filename
with tf.Session() as sess:
# Learn population model!
stim = tf.placeholder(tf.float32, shape=[None, stim_dim], name='stim')
resp = tf.placeholder(tf.float32, name='resp')
data_len = tf.placeholder(tf.float32, name='data_len')
if FLAGS.model_id == 'poisson' or FLAGS.model_id == 'poisson_full':
# variables
w = tf.Variable(
np.array(0.01 * np.random.randn(stim_dim, n_su),
dtype='float32'))
a = tf.Variable(
np.array(0.01 * np.random.rand(n_cells, 1, n_su),
dtype='float32'))
lam = tf.transpose(tf.reduce_sum(tf.exp(tf.matmul(stim, w) + a),
2))
#loss_inter = (tf.reduce_sum(lam/tot_spks_chosen_cells)/120. - tf.reduce_sum(resp*tf.log(lam)/tot_spks_chosen_cells)) / data_len
loss_inter = (tf.reduce_sum(lam) / 120. -
tf.reduce_sum(resp * tf.log(lam))) / data_len
loss = loss_inter
train_step = tf.train.AdagradOptimizer(FLAGS.step_sz).minimize(
loss, var_list=[w, a])
def training(inp_dict):
sess.run(train_step, feed_dict=inp_dict)
def get_loss(inp_dict):
ls = sess.run(loss, feed_dict=inp_dict)
return ls
if FLAGS.model_id == 'poisson_relu':
# variables
w = tf.Variable(
np.array(0.01 * np.random.randn(stim_dim, n_su),
dtype='float32'))
a = tf.Variable(
np.array(0.01 * np.random.rand(n_su, n_cells),
dtype='float32'))
b_init = np.log(
(tot_spks_chosen_cells) / (216000. - tot_spks_chosen_cells))
b = tf.Variable(b_init, dtype='float32')
f = tf.matmul(
tf.exp(tf.nn.relu(stim, w)), a
) + b #loss_inter = (tf.reduce_sum(lam/tot_spks_chosen_cells)/120. - tf.reduce_sum(resp*tf.log(lam)/tot_spks_chosen_cells)) / data_len
loss_inter = (tf.reduce_sum(f) / 120. -
tf.reduce_sum(resp * tf.log(f))) / data_len
loss = loss_inter
train_step = tf.train.AdagradOptimizer(FLAGS.step_sz).minimize(
loss, var_list=[w, a])
def training(inp_dict):
sess.run(train_step, feed_dict=inp_dict)
def get_loss(inp_dict):
ls = sess.run(loss, feed_dict=inp_dict)
return ls
if FLAGS.model_id == 'logistic':
w = tf.Variable(
np.array(0.01 * np.random.randn(stim_dim, n_su),
dtype='float32'))
a = tf.Variable(
np.array(0.01 * np.random.rand(n_su, n_cells),
dtype='float32'))
b_init = np.log(
(tot_spks_chosen_cells) / (216000. - tot_spks_chosen_cells))
b = tf.Variable(b_init, dtype='float32')
f = tf.matmul(tf.nn.relu(tf.matmul(stim, w)), a) + b
loss_inter = tf.reduce_sum(tf.nn.softplus(
-2 * (resp - 0.5) * f)) / data_len
loss = loss_inter
train_step = tf.train.AdagradOptimizer(FLAGS.step_sz).minimize(
loss, var_list=[w, a, b])
a_pos = tf.assign(a, (a + tf.abs(a)) / 2)
def training(inp_dict):
sess.run(train_step, feed_dict=inp_dict)
sess.run(a_pos)
def get_loss(inp_dict):
ls = sess.run(loss, feed_dict=inp_dict)
return ls
if FLAGS.model_id == 'hinge':
w = tf.Variable(
np.array(0.01 * np.random.randn(stim_dim, n_su),
dtype='float32'))
a = tf.Variable(
np.array(0.01 * np.random.rand(n_su, n_cells),
dtype='float32'))
b_init = np.log(
(tot_spks_chosen_cells) / (216000. - tot_spks_chosen_cells))
b = tf.Variable(b_init, dtype='float32')
f = tf.matmul(tf.nn.relu(tf.matmul(stim, w)), a) + b
loss_inter = tf.reduce_sum(tf.nn.relu(1 - 2 *
(resp - 0.5) * f)) / data_len
loss = loss_inter
train_step = tf.train.AdagradOptimizer(FLAGS.step_sz).minimize(
loss, var_list=[w, a, b])
a_pos = tf.assign(a, (a + tf.abs(a)) / 2)
def training(inp_dict):
sess.run(train_step, feed_dict=inp_dict)
sess.run(a_pos)
def get_loss(inp_dict):
ls = sess.run(loss, feed_dict=inp_dict)
return ls
# summaries
l_summary = tf.scalar_summary('loss', loss)
l_inter_summary = tf.scalar_summary('loss_inter', loss_inter)
# Merge all the summaries and write them out to /tmp/mnist_logs (by default)
merged = tf.merge_all_summaries()
train_writer = tf.train.SummaryWriter(FLAGS.save_location + 'train',
sess.graph)
test_writer = tf.train.SummaryWriter(FLAGS.save_location + 'test')
print('\nStarting new code')
sess.run(tf.initialize_all_variables())
saver_var = tf.train.Saver(tf.all_variables(),
keep_checkpoint_every_n_hours=0.05)
load_prev = False
start_iter = 0
try:
latest_filename = short_filename + '_latest_fn'
restore_file = tf.train.latest_checkpoint(FLAGS.save_location,
latest_filename)
start_iter = int(restore_file.split('/')[-1].split('-')[-1])
saver_var.restore(sess, restore_file)
load_prev = True
except:
print('No previous dataset')
if load_prev:
print('\nPrevious results loaded')
else:
print('\nVariables initialized')
# Do the fitting
icnt = 0
stim_test, resp_test, test_length = get_test_data()
fd_test = {stim: stim_test, resp: resp_test, data_len: test_length}
#logfile.close()
for istep in np.arange(start_iter, 400000):
print(istep)
# get training data
stim_train, resp_train, train_len = get_next_training_batch(istep)
fd_train = {
stim: stim_train,
resp: resp_train,
data_len: train_len
}
# take training step
training(fd_train)
if istep % 10 == 0:
# compute training and testing losses
ls_train = get_loss(fd_train)
ls_test = get_loss(fd_test)
latest_filename = short_filename + '_latest_fn'
saver_var.save(sess,
save_filename,
global_step=istep,
latest_filename=latest_filename)
# add training summary
summary = sess.run(merged, feed_dict=fd_train)
train_writer.add_summary(summary, istep)
# add testing summary
summary = sess.run(merged, feed_dict=fd_test)
test_writer.add_summary(summary, istep)
print(istep, ls_train, ls_test)
icnt += FLAGS.batchsz
if icnt > 216000 - 1000:
icnt = 0
tms = np.random.permutation(np.arange(216000 - 1000))
def main(argv):
# initialize training and testing chunks
init_chunks()
# setup dataset
_,cids_select, n_cells, tc_select, tc_mean = setup_dataset()
# print parameters
print('Save folder name: ' + str(FLAGS.folder_name) +
'\nmodel:' + str(FLAGS.model_id) +
'\nLoss:' + str(FLAGS.loss) +
'\nbatch size' + str(FLAGS.batchsz) +
'\nstep size' + str(FLAGS.step_sz) +
'\ntraining length: ' + str(FLAGS.train_len) +
'\nn_cells: '+str(n_cells))
# filename for saving file
short_filename = ('_loss='+
str(FLAGS.loss) + '_batch_sz='+ str(FLAGS.batchsz) +
'_step_sz'+ str(FLAGS.step_sz) +
'_tlen=' + str(FLAGS.train_len) + '_jitter')
# setup model
with tf.Session() as sess:
# initialize stuff
if FLAGS.loss == 'poisson':
b_init = np.array(0.000001*np.ones(n_cells)) # a very small positive bias needed to avoid log(0) in poisson loss
else:
b_init = np.log((tot_spks_chosen_cells)/(216000. - tot_spks_chosen_cells)) # log-odds, a good initialization for some
# RGB time filter
tm4D = np.zeros((30,1,3,3))
for ichannel in range(3):
tm4D[:,0,ichannel,ichannel] = tc_mean[:,ichannel]
tc = tf.Variable((tm4D).astype('float32'),name = 'tc')
d1=640
d2=320
colors=3
# make data placeholders
stim = tf.placeholder(tf.float32,shape=[None,d1,d2,colors],name='stim')
resp = tf.placeholder(tf.float32,shape=[None,n_cells],name='resp')
data_len = tf.placeholder(tf.float32,name='data_len')
# time convolution
# time course should be time,d1,color,color
# original stimulus is (time, d1,d2,color). Permute it to (d2,time,d1,color) so that 1D convolution could be mimicked using conv_2d.
stim_time_filtered = tf.transpose(tf.nn.conv2d(tf.transpose(stim,(2,0,1,3)),tc, strides=[1,1,1,1], padding='VALID'), (1,2,0,3))
# learn almost convolutional model
short_filename = ('model=' + str(FLAGS.model_id) + short_filename)
mask_tf, dimx, dimy, n_pix = get_windows()
w_del = tf.Variable(np.array( 0.05*np.random.randn(dimx, dimy, n_pix),dtype='float32'), name='w_del')
w_mother = tf.Variable(np.array( np.ones((2 * FLAGS.window + 1, 2 * FLAGS.window + 1, FLAGS.su_channels, 1)),dtype='float32'), name='w_mother')
a = tf.Variable(np.array(np.random.randn(dimx*dimy, n_cells),dtype='float32'), name='a')
a_sfm = tf.transpose(tf.nn.softmax(tf.transpose(a)))
b = tf.Variable(np.array(b_init,dtype='float32'), name='b')
vars_fit = [w_mother, w_del, a] # which variables to fit
if not FLAGS.loss == 'poisson':
vars_fit = vars_fit + [b]
# stimulus filtered with convolutional windows
stim4D = stim_time_filtered#tf.expand_dims(tf.reshape(stim, (-1,40,80)), 3)
stim_convolved = tf.reduce_sum(tf.nn.conv2d(stim4D, w_mother, strides=[1, FLAGS.stride, FLAGS.stride, 1], padding="VALID"),3)
stim_masked = tf.nn.conv2d(stim4D, mask_tf, strides=[1, FLAGS.stride, FLAGS.stride, 1], padding="VALID" )
stim_del = tf.reduce_sum(tf.mul(stim_masked, w_del), 3)
# activation of different subunits
su_act = tf.nn.relu(stim_del + stim_convolved)
# get firing rate
lam = tf.matmul(tf.reshape(su_act, [-1, dimx*dimy]), a_sfm) + b
# regularization
regularization = FLAGS.lam_w * tf.reduce_sum(tf.nn.l2_loss(w_del))
# projection to satisfy hard variable constraints
b_pos = tf.assign(b, (b + tf.abs(b))/2)
def proj():
if FLAGS.loss == 'poisson':
sess.run(b_pos)
if FLAGS.loss == 'poisson':
loss_inter = (tf.reduce_sum(lam)/120. - tf.reduce_sum(resp*tf.log(lam))) / data_len
loss = loss_inter + regularization # add regularization to get final loss function
# training consists of calling training()
# which performs a train step and project parameters to model specific constraints using proj()
train_step = tf.train.AdagradOptimizer(FLAGS.step_sz).minimize(loss, var_list=vars_fit)
def training(inp_dict):
sess.run(train_step, feed_dict=inp_dict) # one step of gradient descent
proj() # model specific projection operations
# evaluate loss on given data.
def get_loss(inp_dict):
ls = sess.run(loss,feed_dict = inp_dict)
return ls
# saving details
# make a folder with name derived from parameters of the algorithm - it saves checkpoint files and summaries used in tensorboard
parent_folder = FLAGS.save_location + FLAGS.folder_name + '/'
# make folder if it does not exist
if not gfile.IsDirectory(parent_folder):
gfile.MkDir(parent_folder)
FLAGS.save_location = parent_folder + short_filename + '/'
if not gfile.IsDirectory(FLAGS.save_location):
gfile.MkDir(FLAGS.save_location)
save_filename = FLAGS.save_location + short_filename
# create summary writers
# create histogram summary for all parameters which are learnt
for ivar in vars_fit:
tf.histogram_summary(ivar.name, ivar)
# loss summary
l_summary = tf.scalar_summary('loss',loss)
# loss without regularization summary
l_inter_summary = tf.scalar_summary('loss_inter',loss_inter)
# Merge all the summary writer ops into one op (this way, calling one op stores all summaries)
merged = tf.merge_all_summaries()
# training and testing has separate summary writers
train_writer = tf.train.SummaryWriter(FLAGS.save_location + 'train',
sess.graph)
test_writer = tf.train.SummaryWriter(FLAGS.save_location + 'test')
## load previous results
sess.run(tf.initialize_all_variables())
saver_var = tf.train.Saver(tf.all_variables(), keep_checkpoint_every_n_hours=0.05)
load_prev = False
start_iter=0
try:
# restore previous fits if they are available - useful when programs are preempted frequently on .
latest_filename = short_filename + '_latest_fn'
restore_file = tf.train.latest_checkpoint(FLAGS.save_location, latest_filename)
start_iter = int(restore_file.split('/')[-1].split('-')[-1]) # restore previous iteration count and start from there.
saver_var.restore(sess, restore_file) # restore variables
load_prev = True
except:
print('No previous dataset')
if load_prev:
print('\nPrevious results loaded')
else:
print('\nVariables initialized')
# Finally, do fitting
# get test data and make test dictionary
stim_test,resp_test,test_length = get_stim_resp('test')
fd_test = {stim: stim_test,
resp: resp_test,
data_len: test_length}
for istep in np.arange(start_iter,400000):
print(istep)
# get training data and make test dictionary
stim_train, resp_train, train_len = get_stim_resp('train')
fd_train = {stim: stim_train,
resp: resp_train,
data_len: train_len}
print('loaded a training set')
# take training step
training(fd_train)
print('performed a training step')
if istep%10 == 0:
# save variables
latest_filename = short_filename + '_latest_fn'
saver_var.save(sess, save_filename, global_step=istep, latest_filename = latest_filename)
# add training summary
summary = sess.run(merged, feed_dict=fd_train)
train_writer.add_summary(summary,istep)
print('training loss calculated')
# add testing summary
summary = sess.run(merged, feed_dict=fd_test)
test_writer.add_summary(summary,istep)
print('testing loss calculated')
def main(argv):
np.random.seed(23)
# Figure out dictionary path.
dict_list = gfile.ListDirectory(FLAGS.src_dict)
dict_path = os.path.join(FLAGS.src_dict, dict_list[FLAGS.taskid])
# Load the dictionary
if dict_path[-3:] == 'pkl':
data = pickle.load(gfile.Open(dict_path, 'r'))
if dict_path[-3:] == 'mat':
data = sio.loadmat(gfile.Open(dict_path, 'r'))
#FLAGS.save_dir = '/home/bhaishahster/stimulation_algos/dictionaries/' + dict_list[FLAGS.taskid][:-4]
FLAGS.save_dir = FLAGS.save_dir + dict_list[FLAGS.taskid][:-4]
if not gfile.Exists(FLAGS.save_dir):
gfile.MkDir(FLAGS.save_dir)
# S_collection = data['S'] # Target
A = data['A'] # Decoder
D = data['D'].T # Dictionary
# clean dictionary
thr_list = np.arange(0, 1, 0.01)
dict_val = []
for thr in thr_list:
dict_val += [np.sum(np.sum(D.T > thr, 1) != 0)]
plt.ion()
plt.figure()
plt.plot(thr_list, dict_val)
plt.xlabel('Threshold')
plt.ylabel(
'Number of dictionary elements with \n atleast one element above threshold'
)
plt.title('Please choose threshold')
thr_use = float(input('What threshold to use?'))
plt.axvline(thr_use)
plt.title('Using threshold: %.5f' % thr_use)
dict_valid = np.sum(D.T > thr_use, 1) > 0
D = D[:, dict_valid]
D = np.append(D, np.zeros((D.shape[0], 1)), 1)
print(
'Appending a "dummy" dictionary element that does not activate any cell'
)
# Vary stimulus resolution
for itarget in range(20):
n_targets = 1
for stix_resolution in [32, 64, 16, 8]:
# Get the target
x_dim = int(640 / stix_resolution)
y_dim = int(320 / stix_resolution)
targets = (np.random.rand(y_dim, x_dim, n_targets) < 0.5) - 0.5
upscale = stix_resolution / 8
targets = np.repeat(np.repeat(targets, upscale, axis=0),
upscale,
axis=1)
targets = np.reshape(targets, [-1, targets.shape[-1]])
S_actual = targets[:, 0]
# Remove null component of A from S
S = A.dot(np.linalg.pinv(A).dot(S_actual))
# Run Greedy first to initialize
x_greedy = greedy_stimulation(S,
A,
D,
max_stims=FLAGS.t_max * FLAGS.delta,
file_suffix='%d_%d' %
(stix_resolution, itarget),
save=True,
save_dir=FLAGS.save_dir)
# Load greedy output from previous run
#data_greedy = pickle.load(gfile.Open('/home/bhaishahster/greedy_2000_32_0.pkl', 'r'))
#x_greedy = data_greedy['x_chosen']
# Plan for multiple time points
x_init = np.zeros((x_greedy.shape[0], FLAGS.t_max))
for it in range(FLAGS.t_max):
print((it + 1) * FLAGS.delta - 1)
x_init[:, it] = x_greedy[:, (it + 1) * FLAGS.delta - 1]
#simultaneous_planning(S, A, D, t_max=FLAGS.t_max, lr=FLAGS.learning_rate,
# delta=FLAGS.delta, normalization=FLAGS.normalization,
# file_suffix='%d_%d_normal' % (stix_resolution, itarget), x_init=x_init, save_dir=FLAGS.save_dir)
from IPython import embed
embed()
simultaneous_planning_interleaved_discretization(
S,
A,
D,
t_max=FLAGS.t_max,
lr=FLAGS.learning_rate,
delta=FLAGS.delta,
normalization=FLAGS.normalization,
file_suffix='%d_%d_pgd' % (stix_resolution, itarget),
x_init=x_init,
save_dir=FLAGS.save_dir,
freeze_freq=np.inf,
steps_max=500 * 20 - 1)
# Interleaved discretization.
simultaneous_planning_interleaved_discretization(
S,
A,
D,
t_max=FLAGS.t_max,
lr=FLAGS.learning_rate,
delta=FLAGS.delta,
normalization=FLAGS.normalization,
file_suffix='%d_%d_pgd_od' % (stix_resolution, itarget),
x_init=x_init,
save_dir=FLAGS.save_dir,
freeze_freq=500,
steps_max=500 * 20 - 1)
# Exponential weighing.
simultaneous_planning_interleaved_discretization_exp_gradient(
S,
A,
D,
t_max=FLAGS.t_max,
lr=FLAGS.learning_rate,
delta=FLAGS.delta,
normalization=FLAGS.normalization,
file_suffix='%d_%d_ew' % (stix_resolution, itarget),
x_init=x_init,
save_dir=FLAGS.save_dir,
freeze_freq=np.inf,
steps_max=500 * 20 - 1)
# Exponential weighing with interleaved discretization.
simultaneous_planning_interleaved_discretization_exp_gradient(
S,
A,
D,
t_max=FLAGS.t_max,
lr=FLAGS.learning_rate,
delta=FLAGS.delta,
normalization=FLAGS.normalization,
file_suffix='%d_%d_ew_od' % (stix_resolution, itarget),
x_init=x_init,
save_dir=FLAGS.save_dir,
freeze_freq=500,
steps_max=500 * 20 - 1)
'''
# Plot results
data_fractional = pickle.load(gfile.Open('/home/bhaishahster/2012-09-24-0_SAD_fr1/pgd_20_5.000000_C_100_32_0_normal.pkl', 'r'))
plt.ion()
start = 0
end = 20
error_target_frac = np.linalg.norm(S - data_fractional['S'])
#error_target_greedy = np.linalg.norm(S - data_greedy['S'])
print('Did target change? \n Err from fractional: %.3f ' % error_target_frac)
#'\n Err from greedy %.3f' % (error_target_frac,
# error_target_greedy))
# from IPython import embed; embed()
normalize = np.sum(data_fractional['S'] ** 2)
plt.ion()
plt.plot(np.arange(120, len(data_fractional['f_log'])), data_fractional['f_log'][120:] / normalize, 'b')
plt.axhline(data_fractional['errors'][5:].mean() / normalize, color='m')
plt.axhline(data_fractional['errors_ht_discrete'][5:].mean() / normalize, color='y')
plt.axhline(data_fractional['errors_rr_discrete'][5:].mean() / normalize, color='r')
plt.axhline(data_greedy['error_curve'][120:].mean() / normalize, color='k')
plt.legend(['fraction_curve', 'fractional error', 'HT', 'RR', 'Greedy'])
plt.pause(1.0)
'''
from IPython import embed
embed()
def main(unused_argv=()):
## copy data locally
dst = FLAGS.tmp_dir
print('Starting Copy')
if not gfile.IsDirectory(dst):
gfile.MkDir(dst)
files = gfile.ListDirectory(FLAGS.src_dir)
for ifile in files:
ffile = os.path.join(dst, ifile)
if not gfile.Exists(ffile):
gfile.Copy(os.path.join(FLAGS.src_dir, ifile), ffile)
print('Copied %s' % os.path.join(FLAGS.src_dir, ifile))
else:
print('File exists %s' % ffile)
print('File copied to destination')
## load data
# load stimulus
data = h5py.File(os.path.join(dst, 'stimulus.mat'))
stimulus = np.array(data.get('stimulus')) - 0.5
# load responses from multiple retina
datasets_list = os.path.join(dst, 'datasets.txt')
datasets = open(datasets_list, "r").read()
training_datasets = [line for line in datasets.splitlines()]
responses = []
for idata in training_datasets:
print(idata)
data_file = os.path.join(dst, idata)
data = sio.loadmat(data_file)
responses += [data]
print(np.max(data['centers'], 0))
# generate additional features for responses
num_cell_types = 2
dimx = 80
dimy = 40
for iresp in responses:
# remove cells which are outside 80x40 window.
process_dataset(iresp, dimx, dimy, num_cell_types)
## generate graph -
if FLAGS.is_test == 0:
is_training = True
if FLAGS.is_test == 1:
is_training = True # False
with tf.Session() as sess:
## Make graph
# embed stimulus.
time_window = 30
stimx = stimulus.shape[1]
stimy = stimulus.shape[2]
stim_tf = tf.placeholder(tf.float32,
shape=[None, stimx,
stimy, time_window]) # batch x X x Y x time_window
batch_norm = FLAGS.batch_norm
stim_embed = embed_stimulus(FLAGS.stim_layers.split(','),
batch_norm, stim_tf, is_training,
reuse_variables=False)
'''
ttf_tf = tf.Variable(np.ones(time_window).astype(np.float32)/10, name='stim_ttf')
filt = tf.expand_dims(tf.expand_dims(tf.expand_dims(ttf_tf, 0), 0), 3)
stim_time_filt = tf.nn.conv2d(stim_tf, filt,
strides=[1, 1, 1, 1], padding='SAME') # batch x X x Y x 1
ilayer = 0
stim_time_filt = slim.conv2d(stim_time_filt, 1, [3, 3],
stride=1,
scope='stim_layer_wt_%d' % ilayer,
reuse=False,
normalizer_fn=slim.batch_norm,
activation_fn=tf.nn.softplus,
normalizer_params={'is_training': is_training},
padding='SAME')
'''
# embed responses.
num_cell_types = 2
layers = FLAGS.resp_layers # format: window x filters x stride .. NOTE: final filters=1, stride =1 throughout
batch_norm = FLAGS.batch_norm
time_window = 1
anchor_model = conv.ConvolutionalProsthesisScore(sess, time_window=1,
layers=layers,
batch_norm=batch_norm,
is_training=is_training,
reuse_variables=False,
num_cell_types=2,
dimx=dimx, dimy=dimy)
neg_model = conv.ConvolutionalProsthesisScore(sess, time_window=1,
layers=layers,
batch_norm=batch_norm,
is_training=is_training,
reuse_variables=True,
num_cell_types=2,
dimx=dimx, dimy=dimy)
d_s_r_pos = tf.reduce_sum((stim_embed - anchor_model.responses_embed)**2, [1, 2, 3]) # batch
d_pairwise_s_rneg = tf.reduce_sum((tf.expand_dims(stim_embed, 1) -
tf.expand_dims(neg_model.responses_embed, 0))**2, [2, 3, 4]) # batch x batch_neg
beta = 10
# if FLAGS.save_suffix == 'lr=0.001':
loss = tf.reduce_sum(beta * tf.reduce_logsumexp(tf.expand_dims(d_s_r_pos / beta, 1) -
d_pairwise_s_rneg / beta, 1), 0)
# else :
# loss = tf.reduce_sum(tf.nn.softplus(1 + tf.expand_dims(d_s_r_pos, 1) - d_pairwise_s_rneg))
accuracy_tf = tf.reduce_mean(tf.sign(-tf.expand_dims(d_s_r_pos, 1) + d_pairwise_s_rneg))
lr = 0.001
train_op = tf.train.AdagradOptimizer(lr).minimize(loss)
# set up training and testing data
training_datasets_all = [1, 2, 3, 5, 6, 7, 9, 10, 11, 13, 14, 15]
testing_datasets = [0, 4, 8, 12]
print('Testing datasets', testing_datasets)
n_training_datasets_log = [1, 3, 5, 7, 9, 11, 12]
if (np.floor(FLAGS.taskid / 4)).astype(np.int) < len(n_training_datasets_log):
# do randomly sampled training data for 0<= FLAGS.taskid < 28
prng = RandomState(23)
n_training_datasets = n_training_datasets_log[(np.floor(FLAGS.taskid / 4)).astype(np.int)]
for _ in range(10*FLAGS.taskid):
print(prng.choice(training_datasets_all,
n_training_datasets, replace=False))
training_datasets = prng.choice(training_datasets_all,
n_training_datasets, replace=False)
# training_datasets = [i for i in range(7) if i< 7-FLAGS.taskid] #[0, 1, 2, 3, 4, 5]
else:
# do 1 training data, chosen in order for FLAGS.taskid >= 28
datasets_all = np.arange(16)
training_datasets = [datasets_all[FLAGS.taskid % (4 * len(n_training_datasets_log))]]
print('Task ID %d' % FLAGS.taskid)
print('Training datasets', training_datasets)
# Initialize stuff.
file_name = ('end_to_end_stim_%s_resp_%s_beta_%d_taskid_%d'
'_training_%s_testing_%s_%s' % (FLAGS.stim_layers,
FLAGS.resp_layers, beta, FLAGS.taskid,
str(training_datasets)[1: -1],
str(testing_datasets)[1: -1],
FLAGS.save_suffix))
saver_var, start_iter = initialize_model(FLAGS.save_folder, file_name, sess)
# print model graph
PrintModelAnalysis(tf.get_default_graph())
# Add summary ops
retina_number = tf.placeholder(tf.int16, name='input_retina');
summary_ops = []
for iret in np.arange(len(responses)):
print(iret)
r1 = tf.summary.scalar('loss_%d' % iret , loss)
r2 = tf.summary.scalar('accuracy_%d' % iret , accuracy_tf)
summary_ops += [tf.summary.merge([r1, r2])]
# tf.summary.scalar('loss', loss)
# tf.summary.scalar('accuracy', accuracy_tf)
# summary_op = tf.summary.merge_all()
# Setup summary writers
summary_writers = []
for loc in ['train', 'test']:
summary_location = os.path.join(FLAGS.save_folder, file_name,
'summary_' + loc )
summary_writer = tf.summary.FileWriter(summary_location, sess.graph)
summary_writers += [summary_writer]
# start training
batch_neg_train_sz = 100
batch_train_sz = 100
def batch(dataset_id):
batch_train = get_batch(stimulus, responses[dataset_id]['responses'],
batch_size=batch_train_sz,
batch_neg_resp=batch_neg_train_sz,
stim_history=30, min_window=10)
stim_batch, resp_batch, resp_batch_neg = batch_train
feed_dict = {stim_tf: stim_batch,
anchor_model.responses_tf: np.expand_dims(resp_batch, 2),
neg_model.responses_tf: np.expand_dims(resp_batch_neg, 2),
anchor_model.map_cell_grid_tf: responses[dataset_id]['map_cell_grid'],
anchor_model.cell_types_tf: responses[dataset_id]['ctype_1hot'],
anchor_model.mean_fr_tf: responses[dataset_id]['mean_firing_rate'],
neg_model.map_cell_grid_tf: responses[dataset_id]['map_cell_grid'],
neg_model.cell_types_tf: responses[dataset_id]['ctype_1hot'],
neg_model.mean_fr_tf: responses[dataset_id]['mean_firing_rate'],
retina_number : dataset_id}
return feed_dict
def batch_few_cells(responses):
batch_train = get_batch(stimulus, responses['responses'],
batch_size=batch_train_sz,
batch_neg_resp=batch_neg_train_sz,
stim_history=30, min_window=10)
stim_batch, resp_batch, resp_batch_neg = batch_train
feed_dict = {stim_tf: stim_batch,
anchor_model.responses_tf: np.expand_dims(resp_batch, 2),
neg_model.responses_tf: np.expand_dims(resp_batch_neg, 2),
anchor_model.map_cell_grid_tf: responses['map_cell_grid'],
anchor_model.cell_types_tf: responses['ctype_1hot'],
anchor_model.mean_fr_tf: responses['mean_firing_rate'],
neg_model.map_cell_grid_tf: responses['map_cell_grid'],
neg_model.cell_types_tf: responses['ctype_1hot'],
neg_model.mean_fr_tf: responses['mean_firing_rate'],
}
return feed_dict
if FLAGS.is_test == 1:
print('Testing')
save_dict = {}
from IPython import embed; embed()
## Estimate one, fix others
'''
grad_resp = tf.gradients(d_s_r_pos, anchor_model.responses_tf)
t_start = 1000
t_len = 100
stim_history = 30
stim_batch = np.zeros((t_len, stimulus.shape[1],
stimulus.shape[2], stim_history))
for isample, itime in enumerate(np.arange(t_start, t_start + t_len)):
stim_batch[isample, :, :, :] = np.transpose(stimulus[itime: itime-stim_history:-1, :, :], [1, 2, 0])
iretina = testing_datasets[0]
resp_batch = np.expand_dims(np.random.rand(t_len, responses[iretina]['responses'].shape[1]), 2)
step_sz = 0.01
eps = 1e-2
dist_prev = np.inf
for iiter in range(10000):
feed_dict = {stim_tf: stim_batch,
anchor_model.map_cell_grid_tf: responses[iretina]['map_cell_grid'],
anchor_model.cell_types_tf: responses[iretina]['ctype_1hot'],
anchor_model.mean_fr_tf: responses[iretina]['mean_firing_rate'],
anchor_model.responses_tf: resp_batch}
dist_np, resp_grad_np = sess.run([d_s_r_pos, grad_resp], feed_dict=feed_dict)
if np.sum(np.abs(dist_prev - dist_np)) < eps:
break
print(np.sum(dist_np), np.sum(np.abs(dist_prev - dist_np)))
dist_prev = dist_np
resp_batch = resp_batch - step_sz * resp_grad_np[0]
resp_batch = resp_batch.squeeze()
'''
# from IPython import embed; embed()
## compute distances between s-r pairs for small number of cells
test_retina = []
for iretina in range(len(testing_datasets)):
dataset_id = testing_datasets[iretina]
num_cells_total = responses[dataset_id]['responses'].shape[1]
dataset_center = responses[dataset_id]['centers'].mean(0)
dataset_cell_distances = np.sqrt(np.sum((responses[dataset_id]['centers'] -
dataset_center), 1))
order_cells = np.argsort(dataset_cell_distances)
test_sr_few_cells = {}
for num_cells_prc in [5, 10, 20, 30, 50, 100]:
num_cells = np.percentile(np.arange(num_cells_total),
num_cells_prc).astype(np.int)
choose_cells = order_cells[:num_cells]
resposnes_few_cells = {'responses': responses[dataset_id]['responses'][:, choose_cells],
'map_cell_grid': responses[dataset_id]['map_cell_grid'][:, :, choose_cells],
'ctype_1hot': responses[dataset_id]['ctype_1hot'][choose_cells, :],
'mean_firing_rate': responses[dataset_id]['mean_firing_rate'][choose_cells]}
# get a batch
d_pos_log = np.array([])
d_neg_log = np.array([])
for test_iter in range(1000):
print(iretina, num_cells_prc, test_iter)
feed_dict = batch_few_cells(resposnes_few_cells)
d_pos, d_neg = sess.run([d_s_r_pos, d_pairwise_s_rneg], feed_dict=feed_dict)
d_neg = np.diag(d_neg) # np.mean(d_neg, 1) #
d_pos_log = np.append(d_pos_log, d_pos)
d_neg_log = np.append(d_neg_log, d_neg)
precision_log, recall_log, F1_log, FPR_log, TPR_log = ROC(d_pos_log, d_neg_log)
print(np.sum(d_pos_log > d_neg_log))
print(np.sum(d_pos_log < d_neg_log))
test_sr= {'precision': precision_log, 'recall': recall_log,
'F1': F1_log, 'FPR': FPR_log, 'TPR': TPR_log,
'd_pos_log': d_pos_log, 'd_neg_log': d_neg_log,
'num_cells': num_cells}
test_sr_few_cells.update({'num_cells_prc_%d' % num_cells_prc : test_sr})
test_retina += [test_sr_few_cells]
save_dict.update({'few_cell_analysis': test_retina})
## compute distances between s-r pairs - pos and neg.
test_retina = []
for iretina in range(len(testing_datasets)):
# stim-resp log
d_pos_log = np.array([])
d_neg_log = np.array([])
for test_iter in range(1000):
print(test_iter)
feed_dict = batch(testing_datasets[iretina])
d_pos, d_neg = sess.run([d_s_r_pos, d_pairwise_s_rneg], feed_dict=feed_dict)
d_neg = np.diag(d_neg) # np.mean(d_neg, 1) #
d_pos_log = np.append(d_pos_log, d_pos)
d_neg_log = np.append(d_neg_log, d_neg)
precision_log, recall_log, F1_log, FPR_log, TPR_log = ROC(d_pos_log, d_neg_log)
print(np.sum(d_pos_log > d_neg_log))
print(np.sum(d_pos_log < d_neg_log))
test_sr = {'precision': precision_log, 'recall': recall_log,
'F1': F1_log, 'FPR': FPR_log, 'TPR': TPR_log,
'd_pos_log': d_pos_log, 'd_neg_log': d_neg_log}
test_retina += [test_sr]
save_dict.update({'test_sr': test_retina})
## ROC curves of responses from repeats - dataset 1
repeats_datafile = '/home/bhaishahster/metric_learning/datasets/2015-09-23-7.mat'
repeats_data = sio.loadmat(gfile.Open(repeats_datafile, 'r'));
repeats_data['cell_type'] = repeats_data['cell_type'].T
# process repeats data
process_dataset(repeats_data, dimx, dimy, num_cell_types)
# analyse and store the result
test_reps = analyse_response_repeats(repeats_data, anchor_model, neg_model, sess)
save_dict.update({'test_reps_2015-09-23-7': test_reps})
## ROC curves of responses from repeats - dataset 2
repeats_datafile = '/home/bhaishahster/metric_learning/examples_pc2005_08_03_0/data005_test.mat'
repeats_data = sio.loadmat(gfile.Open(repeats_datafile, 'r'));
process_dataset(repeats_data, dimx, dimy, num_cell_types)
# analyse and store the result
'''
test_clustering = analyse_response_repeats_all_trials(repeats_data, anchor_model, neg_model, sess)
save_dict.update({'test_reps_2005_08_03_0': test_clustering})
'''
#
# get model params
save_dict.update({'model_pars': sess.run(tf.trainable_variables())})
save_analysis_filename = os.path.join(FLAGS.save_folder, file_name + '_analysis.pkl')
pickle.dump(save_dict, gfile.Open(save_analysis_filename, 'w'))
print(save_analysis_filename)
return
test_iiter = 0
for iiter in range(start_iter, FLAGS.max_iter): # TODO(bhaishahster) :add FLAGS.max_iter
# get a new batch
# stim_tf, anchor_model.responses_tf, neg_model.responses_tf
# training step
train_dataset = training_datasets[iiter % len(training_datasets)]
feed_dict_train = batch(train_dataset)
_, loss_np_train = sess.run([train_op, loss], feed_dict=feed_dict_train)
print(train_dataset, loss_np_train)
# write summary
if iiter % 10 == 0:
# write train summary
test_iiter = test_iiter + 1
train_dataset = training_datasets[test_iiter % len(training_datasets)]
feed_dict_train = batch(train_dataset)
summary_train = sess.run(summary_ops[train_dataset], feed_dict=feed_dict_train)
summary_writers[0].add_summary(summary_train, iiter)
# write test summary
test_dataset = testing_datasets[test_iiter % len(testing_datasets)]
feed_dict_test = batch(test_dataset)
l_test, summary_test = sess.run([loss, summary_ops[test_dataset]], feed_dict=feed_dict_test)
summary_writers[1].add_summary(summary_test, iiter)
print('Test retina: %d, loss: %.3f' % (test_dataset, l_test))
# save model
if iiter % 10 == 0:
save_model(saver_var, FLAGS.save_folder, file_name, sess, iiter)
def RunComputation():
# filename for saving files, derived from FLAGS.
short_filename = get_filename()
# make a folder with name derived from parameters of the algorithm
# it saves checkpoint files and summaries used in tensorboard
parent_folder = FLAGS.save_location + FLAGS.folder_name + '/'
# make folder if it does not exist
if not gfile.IsDirectory(parent_folder):
gfile.MkDir(parent_folder)
FLAGS.save_location = parent_folder + short_filename + '/'
print('Does the file exist?', gfile.IsDirectory(FLAGS.save_location))
if not gfile.IsDirectory(FLAGS.save_location):
gfile.MkDir(FLAGS.save_location)
save_filename = FLAGS.save_location + short_filename
if FLAGS.learn == 0:
# for analysis, use smaller batch sizes, so that we can work with single GPU.
FLAGS.batchsz = 600
#Set up tensorflow
with tf.Graph().as_default() as gra:
with tf.device(tf.ReplicaDeviceSetter(FLAGS.ps_tasks)):
print(FLAGS.config_params)
tf.logging.info(FLAGS.config_params)
# set up training dataset
# tc_mean = get_data_mat.init_chunks(FLAGS.n_chunks) <- use this with old get_data_mat
tc_mean = get_data_mat.init_chunks(FLAGS.batchsz)
#plt.plot(tc_mean)
#plt.show()
#plt.draw()
# Create computation graph.
#
# Graph should be fully constructed before you create supervisor.
# Attempt to modify graph after supervisor is created will cause an error.
with tf.name_scope('model'):
if FLAGS.architecture == '1 layer':
# single GPU model
if False:
global_step = tf.contrib.framework.create_global_step()
model, stim, resp = jitter_model.approximate_conv_jitter(
FLAGS.n_cells, FLAGS.lam_w, FLAGS.window,
FLAGS.stride, FLAGS.step_sz, tc_mean,
FLAGS.su_channels)
# multiGPU model
if True:
model, stim, resp, global_step = jitter_model.approximate_conv_jitter_multigpu(
FLAGS.n_cells, FLAGS.lam_w, FLAGS.window,
FLAGS.stride, FLAGS.step_sz, tc_mean,
FLAGS.su_channels, FLAGS.config_params)
if FLAGS.architecture == '2 layer_stimulus':
# stimulus is first smoothened to lower dimensions, then same model is applied
print('First take a low resolution version of stimulus')
model, stim, resp, global_step, stim_tuple = (
jitter_model.approximate_conv_jitter_multigpu_stim_lr(
FLAGS.n_cells, FLAGS.lam_w, FLAGS.window,
FLAGS.stride, FLAGS.step_sz, tc_mean,
FLAGS.su_channels, FLAGS.config_params,
FLAGS.stim_downsample_window,
FLAGS.stim_downsample_stride))
if FLAGS.architecture == 'complex':
print(' Multiple modifications over 2 layered model above')
model, stim, resp, global_step = (
jitter_model_2.
approximate_conv_jitter_multigpu_complex(
FLAGS.n_cells, FLAGS.lam_w, FLAGS.window,
FLAGS.stride, FLAGS.step_sz, tc_mean,
FLAGS.su_channels, FLAGS.config_params,
FLAGS.stim_downsample_window,
FLAGS.stim_downsample_stride))
# Print the number of variables in graph
print('Calculating model size') # Hope we do not exceed memory
PrintModelAnalysis(gra, max_depth=10)
# Builds our summary op.
summary_op = model.merged_summary
# Create a Supervisor. It will take care of initialization, summaries,
# checkpoints, and recovery.
#
# When multiple replicas of this program are running, the first one,
# identified by --task=0 is the 'chief' supervisor. It is the only one
# that takes case of initialization, etc.
is_chief = (FLAGS.task == 0) # & (FLAGS.learn==1)
print(save_filename)
if FLAGS.learn == 1:
# use supervisor only for learning,
# otherwise it messes up data as it tries to store variables while you are doing analysis
sv = tf.train.Supervisor(logdir=save_filename,
is_chief=is_chief,
saver=tf.train.Saver(),
summary_op=None,
save_model_secs=100,
global_step=global_step,
recovery_wait_secs=5)
if (is_chief and FLAGS.learn == 1):
# save graph only if task id =0 (is_chief) and learning the model
tf.train.write_graph(tf.get_default_graph().as_graph_def(),
save_filename, 'graph.pbtxt')
# Get an initialized, and possibly recovered session. Launch the
# services: Checkpointing, Summaries, step counting.
#
# When multiple replicas of this program are running the services are
# only launched by the 'chief' replica.
session_config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
sess = sv.PrepareSession(FLAGS.master, config=session_config)
# Finally, learn the parameters of the model
FitComputation(sv, sess, model, stim, resp, global_step,
summary_op)
sv.Stop()
else:
# Analyse the model
session_config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
with tf.Session(config=session_config) as sess:
# First, recover the model
saver_var = tf.train.Saver(
tf.all_variables(),
keep_checkpoint_every_n_hours=float('inf'))
restore_file = tf.train.latest_checkpoint(save_filename)
print(restore_file)
start_iter = int(
restore_file.split('/')[-1].split('-')[-1])
saver_var.restore(sess, restore_file)
# model specific analysis
if FLAGS.architecture == '2 layer_stimulus':
AnalyseModel_lr(sess, model)
elif FLAGS.architecture == 'complex':
AnalyseModel_complex(sess, model, stim, resp,
save_filename)
else:
AnalyseModel(sv, sess, model)
