def _write_job_status(self, job, aborted=False):
"""
Write the current job status as set in this object to job status dir
and symlink them if needed.
* Job status goes to state/_all/<uuid>.
* Symlink goes from state/_all/<uuid> to state/<job_name>/uuid, unless
aborted == True, in which case it is removed.
* Symlink goes from FIXME
"""
jobdef_dir = os.path.join(self.status_dir, 'jobs', job.jobdef_name)
mkdir_p(jobdef_dir) # If it doesn't exist yet
job_status_path = os.path.join(self.all_dir, job.id)
job_status_link = os.path.join(jobdef_dir, job.id)
# Write the job status from the _all dir to the state/_all/<uuid> file
status = job.to_dict()
with open(job_status_path, 'w') as f:
json.dump(status, f)
# Link the job status from the _all dir to the state/<job_name>/<uuid>
# file, unless the build was aborted.
if not aborted:
if not os.path.islink(job_status_link):
os.symlink(os.path.join('..', '_all', job.id), job_status_link)
else:
logging.info("Removing status link")
if os.path.islink(job_status_link):
os.unlink(job_status_link)
def on_discard(self, _event):
name = self.names[self.index]
day = tools.get_day()
tools.mkdir_p(os.path.join('discard', day))
os.rename(os.path.join('infiles', name),
os.path.join('discard', day, name))
self.load_next_image()
def error_analysis(model, ids, notes, text_features, X, onehot_Y, label, task):
criteria, num_docs, clf = model
if task in ['sapsii', 'age', 'los']:
V = {}
labels_ = [0] + criteria
for i in range(len(labels_) - 1):
label = '[%d,%s)' % (labels_[i], labels_[i + 1])
V[i] = label
else:
V = {v: k for k, v in criteria.items()}
# for confidence
#P = clf.predict(list(X))
P = clf.predict(X)
pred = P.argmax(axis=1)
Y = onehot_Y.argmax(axis=1)
# convert predictions to right vs wrong
confidence = {}
for i, scores in enumerate(P.tolist()):
prediction = pred[i]
pid = ids[i]
ind = Y[i]
confidence[pid] = (scores[ind], scores, prediction, ind)
homedir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
methoddir = os.path.join(homedir, 'output', task, 'lstm')
mkdir_p(methoddir)
# order predictions by confidence
for pid, conf in sorted(confidence.items(), key=lambda t: t[1]):
if conf[2] == conf[3]:
success = ''
else:
success = '_'
filename = os.path.join(methoddir, '%s%s.pred' % (success, pid))
with open(filename, 'w') as f:
print >> f, ''
print >> f, '=' * 80
print >> f, ''
print >> f, pid
print >> f, 'scores:', conf[1]
print >> f, 'pred: ', V[conf[2]]
print >> f, 'ref: ', V[conf[3]]
print >> f, '#' * 20
print >> f, '#' * 20
print >> f, 'SCORES'
pind = conf[2]
rind = conf[3]
print >> f, '#' * 20
for dt, category, text in sorted(notes[pid]):
print >> f, dt
print >> f, category
print >> f, text
print >> f, '-' * 50
print >> f, ''
print >> f, '+' * 80
print >> f, ''
def process(infile, group_name, day, timeid, copies=1):
mask = Mask((A4_WIDTH, A4_HEIGHT))
page = mask.apply_mask(infile)
create_title(page, group_name, day, timeid)
tools.mkdir_p(os.path.join('png', day))
png_file = os.path.join(
'png', day, '{}_{}.jpg'.format(timeid,
tools.safe_filename(group_name)))
page.save(png_file, quality=75)
tools.print_image(png_file, copies)
def extract_templates(self):
template_dpath = os.path.join(var.CONFIG_DPATH, 'templates')
tools.mkdir_p(template_dpath)
if not os.listdir(template_dpath): # Is dir empty?
self.load_panel(self.panel_firstrun)
wx.Yield()
import bg_templates_tar_bz2
default_templates = bg_templates_tar_bz2.get_tarfile()
default_templates.extractall(template_dpath)
SetInitialDirectory(self.panel_main.file_picker_bg, template_dpath)
wx.CallAfter(self.load_panel, self.panel_main)
def on_restore(self, _event):
cwd = os.getcwd()
day = tools.get_day()
initial_dir = os.path.join(cwd, 'discard', day)
if not os.path.exists(initial_dir):
initial_dir = os.path.join(cwd, 'discard')
tools.mkdir_p(initial_dir)
dlg = wx.lib.imagebrowser.ImageDialog(self, initial_dir)
dlg.Centre()
if dlg.ShowModal() == wx.ID_OK:
self.filename = dlg.GetFile()
self.static_image.load_from_file(self.filename)
dlg.Destroy()
def process(infile, group_name, day, timeid):
page = Image.open(os.path.join('infiles', infile))
photo_size = (A4_WIDTH * 3 / 4, A4_HEIGHT * 3 / 4)
photo_left = (A4_WIDTH - photo_size[0]) / 2
photo_top = (A4_HEIGHT - photo_size[1]) / 2 - A4_HEIGHT / 64
photo_right = photo_left + photo_size[0]
photo_bottom = photo_top + photo_size[1]
photo_rect = (photo_left, photo_top, photo_right, photo_bottom)
create_title(page, (A4_WIDTH, A4_HEIGHT), photo_rect, group_name)
tools.mkdir_p(os.path.join('png', day))
png_file = os.path.join('png', day, '{}_{}.jpg'.format(timeid, tools.safe_filename(group_name)))
page.save(png_file, quality=75)
tools.print_image(png_file)
def on_get_photos(self, _event):
if not tools.mount_camera():
self.SetStatusText('Could not connect to camera. Try again.')
tools.get_camera_files()
if tools.umount_camera():
self.SetStatusText('You can disconnect the camera now.')
else:
self.SetStatusText('Could not disconnect from camera.')
names = os.listdir('infiles')
names.sort()
day = tools.get_day()
tools.mkdir_p(os.path.join('outfiles', day))
self.names = names
if len(self.names) > 0:
self.load_image(0)
else:
self.next_btn.Disable()
self.prev_btn.Disable()
self.load_blank()
def process(infile, day, timeid):
try:
page = Image.open(os.path.join('base', 'Calendar_2014.png'))
except IOError:
print "Cannot open calendar page base"
return
photo_size = (A4_WIDTH * 3 / 4, A4_HEIGHT * 3 / 8)
fade = Fade()
photo = fade.apply_mask(infile, photo_size)
photo_left = (A4_WIDTH - photo_size[0]) / 2
photo_top = 520
photo_right = photo_left + photo_size[0]
photo_bottom = photo_top + photo_size[1]
photo_rect = (photo_left, photo_top, photo_right, photo_bottom)
page.paste(photo, photo_rect)
create_title(page, (A4_WIDTH, A4_HEIGHT), photo_rect, day, timeid)
tools.mkdir_p(os.path.join('png', day))
png_file = os.path.join('png', day, '{}.jpg'.format(timeid))
page.save(png_file, quality=75)
tools.print_image(png_file)
def run():
if not tools.mount_camera():
print 'Could not connect to camera. Try again.'
tools.get_camera_files()
if tools.umount_camera():
print 'You can disconnect the camera now.'
else:
print 'Could not disconnect from camera.'
names = os.listdir('infiles')
names.sort()
day = tools.get_day()
tools.mkdir_p(os.path.join('outfiles', day))
for infile in names:
group_name = raw_input('Group name: ')
timeid = time.strftime('%H%M%S', time.localtime())
process(os.path.join('infiles', infile), group_name, day, timeid)
outfile_name = '{}_{}.jpg'.format(timeid,
tools.safe_filename(group_name))
os.rename(os.path.join('infiles', infile),
os.path.join('outfiles', day, outfile_name))
print 'Finished.'
def unpack(archive, targetdir = None, progress = None):
# Archive type?
atype = archive_type(archive)
if progress:
nfiles = float(len(get_file_list(archive)))
if targetdir:
if not os.path.isdir(targetdir):
mkdir_p(targetdir)
if atype == "rar":
p = pipe_unrar("x", "-y", archive, targetdir)
else:
p = pipe_7z("x", "-y", "-o%s" % targetdir, archive)
else:
if atype == "rar":
p = pipe_unrar("x", "-y", archive)
else:
p = pipe_7z("x", "-y", archive)
if progress:
gfiles = 0
name = ""
for l in p.stdout.readlines():
l = l.strip()
if l.startswith("Extracting"):
gfiles += 1
fs = l.split(None, 1)
if len(fs) > 0:
name = fs[-1]
else:
name = ""
progress(gfiles / nfiles, name)
p.wait()
if p.returncode != 0:
raise UnpackError("Caught error unpacking %s, return code %d" % (archive, p.returncode))
isfile = os.path.isfile
pj = os.path.join
sep = os.path.sep
import tools
def rec_glob(p, files):
for d in glob.glob(p):
if os.path.isfile(d):
files.append(d)
rec_glob("%s/*" % d, files)
outdir = sys.argv[1]
tools.mkdir_p(outdir)
print "Merging into ", outdir, ":",
sys.stdout.flush()
dfiles = {}
for d in sys.argv[2:]:
f = []
rec_glob("%s/*" % d, f)
print "%s (%d files)" % (d, len(f)),
dfiles[d] = f
print
for dir,files in dfiles.iteritems():
for f in files:
def error_analysis(model, ids, notes, text_features, X, Y, label, task):
criteria, num_docs, clf = model
if task in ['sapsii', 'age', 'los']:
V = {}
labels_ = [0] + criteria
for i in range(len(labels_)-1):
label = '[%d,%s)' % (labels_[i],labels_[i+1])
V[i] = label
else:
V = {v:k for k,v in criteria.items()}
V[len(V)] = '**wrong**'
# for confidence
P_ = clf.decision_function(X)
# sklearn has stupid changes in API when doing binary classification. make it conform to 3+
if len(criteria)==2:
m = X.shape[0]
P = np.zeros((m,2))
P[:,0] = -P_
P[:,1] = P_
else:
P = P_
pred = P.argmax(axis=1)
# convert predictions to right vs wrong
confidence = {}
for i,scores in enumerate(P.tolist()):
prediction = pred[i]
pid = ids[i]
ind = Y[i]
confidence[pid] = (scores[ind], scores, prediction, ind)
homedir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
methoddir = os.path.join(homedir, 'output', task, 'embeddings')
mkdir_p(methoddir)
# order predictions by confidence
for pid,conf in sorted(confidence.items(), key=lambda t:t[1]):
if conf[2] == conf[3]:
success = ''
else:
success = '_'
filename = os.path.join(methoddir, '%s%s.pred' % (success,pid))
with open(filename, 'w') as f:
print >>f, ''
print >>f, '=' * 80
print >>f, ''
print >>f, pid
print >>f, 'scores:', conf[1]
print >>f, 'pred: ', V[conf[2]]
print >>f, 'ref: ', V[conf[3]]
print >>f, '#'*20
print >>f, '#'*20
print >>f, 'SCORES'
pind = conf[2]
rind = conf[3]
print >>f, '#'*20
for dt,category,text in sorted(notes[pid]):
print >>f, dt
print >>f, category
print >>f, text
print >>f, '-'*50
print >>f, ''
print >>f, '+'*80
print >>f, ''
import re
import random
import psycopg2
import pandas as pd
import numpy as np
from tools import mkdir_p
# organization: data/$pid.txt (order: demographics, outcome, notes)
homedir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
pickled_dir = os.path.join(homedir, 'data', 'pickled')
readable_dir = os.path.join(homedir, 'data', 'readable')
# store the pickled and readable notes
mkdir_p(pickled_dir)
mkdir_p(readable_dir)
def main():
try:
size = sys.argv[1]
if size not in ['all', 'small']:
raise Exception('bad')
except Exception, e:
print '\n\tusage: python %s <all|small>\n' % sys.argv[0]
exit(1)
X, Y = gather_data(size)
assert sorted(X.keys()) == sorted(Y.keys())
def createDir(self, dirPath, prefix):
if dir and not os.path.isdir(dirPath) :
self.verbose(prefix + "- Creating directory '{0}'".format(dirPath))
tools.mkdir_p(dirPath)
self.okay(prefix + "> OK")
def _make_status_dir(self):
"""
Create required directories under the `status_dir`.
"""
mkdir_p(self.jobs_dir)
mkdir_p(self.all_dir)
line_index))
cv2.rectangle(block_img, (lx, ly+lh/2), (lx+lw, ly+lh/2), (250, 0, 0), 1)
#cv2.putText(block_img, str(line_index), (lx, ly+lh/2), cv2.FONT_HERSHEY_SIMPLEX, 0.8, 255, 2)
# Line Features:
# - mass/aria (ink_volume)
# - height
# - word_count/width (words_ratio)
return block_img
#cv2.rectangle(img, (x, y), (x+w, y+h), (255, 0, 255), 2)
if __name__ == '__main__':
mkdir_p('lines')
mkdir_p('blocks')
eprint('filename,ink_volume,fiber_ink_volume,height,width,area,words_ratio,line_index')
#find_text_block(sys.argv[1], 'marked.png', 'dilated.png')
for page_index, page_name in enumerate(sys.argv[1:]):
print('Page %d' % page_index)
dilated, img, contours = find_text_block(page_name)
#dilated, img = find_text_block(cv2.imread(sys.argv[1]),
# cv2.imread('marked.png'),
# cv2.imread('dilated.png'))
cv2.imwrite('dilated%d.png' % page_index, dilated)
cv2.imwrite('blocks%d.png' % page_index, img)
#lines = find_lines_in_block('blocks.png', contours)
for block_index, contour in enumerate(contours):
def write_jobfile(cmd, jobname, sbatchpath, scratchpath,
nodes=1, ppn=1, gpus=0, mem=16, nhours=18):
"""
Create a job file.
Args:
cmd : str, Command to execute.
jobname : str, Name of the job.
sbatchpath : str, Directory to store SBATCH file in.
scratchpath : str, Directory to store output files in.
nodes : int, optional, Number of compute nodes.
ppn : int, optional, Number of cores per node.
gpus : int, optional, Number of GPU cores.
mem : int, optional, Amount, in GB, of memory.
ndays : int, optional, Running time, in days.
queue : str, optional, Queue name.
Returns:
jobfile : str, Path to the job file.
"""
tools.mkdir_p(sbatchpath)
jobfile = os.path.join(sbatchpath, jobname + '.s')
logname = os.path.join('log', jobname)
if gpus == 0:
with open(jobfile, 'w') as f:
f.write(
'#! /bin/bash\n'
+ '\n'
+ '#SBATCH --nodes={}\n'.format(nodes)
#+ '#SBATCH --ntasks=1\n'
+ '#SBATCH --ntasks-per-node=1\n'
+ '#SBATCH --cpus-per-task={}\n'.format(ppn)
+ '#SBATCH --mem={}GB\n'.format(mem)
+ '#SBATCH --time={}:00:00\n'.format(nhours)
+ '#SBATCH --job-name={}\n'.format(jobname[0:16])
+ '#SBATCH --output={}log/{}.o\n'.format(scratchpath, jobname[0:16])
+ '\n'
+ 'cd {}\n'.format(scratchpath)
+ 'pwd > {}.log\n'.format(logname)
+ 'date >> {}.log\n'.format(logname)
+ 'which python >> {}.log\n'.format(logname)
+ '{} >> {}.log 2>&1\n'.format(cmd, logname)
+ '\n'
+ 'exit 0;\n'
)
else:
with open(jobfile, 'w') as f:
f.write(
'#! /bin/bash\n'
+ '\n'
+ '#SBATCH --nodes={}\n'.format(nodes)
+ '#SBATCH --ntasks-per-node=1\n'
+ '#SBATCH --cpus-per-task={}\n'.format(ppn)
+ '#SBATCH --mem={}GB\n'.format(mem)
+ '#SBATCH --partition=xwang_gpu\n'
+ '#SBATCH --gres=gpu:1\n'
+ '#SBATCH --time={}:00:00\n'.format(nhours)
+ '#SBATCH --job-name={}\n'.format(jobname[0:16])
+ '#SBATCH --output={}log/{}.o\n'.format(scratchpath, jobname[0:16])
+ '\n'
+ 'cd {}\n'.format(scratchpath)
+ 'pwd > {}.log\n'.format(logname)
+ 'date >> {}.log\n'.format(logname)
+ 'which python >> {}.log\n'.format(logname)
+ '{} >> {}.log 2>&1\n'.format(cmd, logname)
+ '\n'
+ 'exit 0;\n'
)
return jobfile
def train_sequential(
model_dir,
rule_trains,
hp=None,
max_steps=1e7,
display_step=500,
ruleset='mante',
seed=0,
):
'''Train the network sequentially.
Args:
model_dir: str, training directory
rule_trains: a list of list of tasks to train sequentially
hp: dictionary of hyperparameters
max_steps: int, maximum number of training steps for each list of tasks
display_step: int, display steps
ruleset: the set of rules to train
seed: int, random seed to be used
Returns:
model is stored at model_dir/model.ckpt
training configuration is stored at model_dir/hp.json
'''
tools.mkdir_p(model_dir)
# Network parameters
default_hp = get_default_hp(ruleset)
if hp is not None:
default_hp.update(hp)
hp = default_hp
hp['seed'] = seed
hp['rng'] = np.random.RandomState(seed)
hp['rule_trains'] = rule_trains
# Get all rules by flattening the list of lists
hp['rules'] = [r for rs in rule_trains for r in rs]
# Number of training iterations for each rule
rule_train_iters = [len(r) * max_steps for r in rule_trains]
tools.save_hp(hp, model_dir)
# Display hp
for key, val in hp.items():
print('{:20s} = '.format(key) + str(val))
# Using continual learning or not
c, ksi = hp['c_intsyn'], hp['ksi_intsyn']
# Build the model
model = Model(model_dir, hp=hp)
grad_unreg = tf.gradients(model.cost_lsq, model.var_list)
# Store results
log = defaultdict(list)
log['model_dir'] = model_dir
# Record time
t_start = time.time()
# tensorboard summaries
placeholders = list()
for v_name in ['Omega0', 'omega0', 'vdelta']:
for v in model.var_list:
placeholder = tf.placeholder(tf.float32, shape=v.shape)
tf.summary.histogram(v_name + '/' + v.name, placeholder)
placeholders.append(placeholder)
merged = tf.summary.merge_all()
test_writer = tf.summary.FileWriter(model_dir + '/tb')
def relu(x):
return x * (x > 0.)
# Use customized session that launches the graph as well
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# penalty on deviation from initial weight
if hp['l2_weight_init'] > 0:
raise NotImplementedError()
# Looping
step_total = 0
for i_rule_train, rule_train in enumerate(hp['rule_trains']):
step = 0
# At the beginning of new tasks
# Only if using intelligent synapses
v_current = sess.run(model.var_list)
if i_rule_train == 0:
v_anc0 = v_current
Omega0 = [np.zeros(v.shape, dtype='float32') for v in v_anc0]
omega0 = [np.zeros(v.shape, dtype='float32') for v in v_anc0]
v_delta = [np.zeros(v.shape, dtype='float32') for v in v_anc0]
elif c > 0:
v_anc0_prev = v_anc0
v_anc0 = v_current
v_delta = [
v - v_prev for v, v_prev in zip(v_anc0, v_anc0_prev)
]
# Make sure all elements in omega0 are non-negative
# Penalty
Omega0 = [
relu(O + o / (v_d**2 + ksi))
for O, o, v_d in zip(Omega0, omega0, v_delta)
]
# Update cost
model.cost_reg = tf.constant(0.)
for v, w, v_val in zip(model.var_list, Omega0, v_current):
model.cost_reg += c * tf.reduce_sum(
tf.multiply(tf.constant(w),
tf.square(v - tf.constant(v_val))))
model.set_optimizer()
# Store Omega0 to tf summary
feed_dict = dict(zip(placeholders, Omega0 + omega0 + v_delta))
summary = sess.run(merged, feed_dict=feed_dict)
test_writer.add_summary(summary, i_rule_train)
# Reset
omega0 = [np.zeros(v.shape, dtype='float32') for v in v_anc0]
# Keep training until reach max iterations
while (step * hp['batch_size_train'] <=
rule_train_iters[i_rule_train]):
# Validation
if step % display_step == 0:
trial = step_total * hp['batch_size_train']
log['trials'].append(trial)
log['times'].append(time.time() - t_start)
log['rule_now'].append(rule_train)
log = do_eval(sess, model, log, rule_train)
if log['perf_avg'][-1] > model.hp['target_perf']:
print('Perf reached the target: {:0.2f}'.format(
hp['target_perf']))
break
# Training
rule_train_now = hp['rng'].choice(rule_train)
# Generate a random batch of trials.
# Each batch has the same trial length
trial = generate_trials(rule_train_now,
hp,
'random',
batch_size=hp['batch_size_train'])
# Generating feed_dict.
feed_dict = tools.gen_feed_dict(model, trial, hp)
# Continual learning with intelligent synapses
v_prev = v_current
# This will compute the gradient BEFORE train step
_, v_grad = sess.run([model.train_step, grad_unreg],
feed_dict=feed_dict)
# Get the weight after train step
v_current = sess.run(model.var_list)
# Update synaptic importance
omega0 = [
o - (v_c - v_p) * v_g for o, v_c, v_p, v_g in zip(
omega0, v_current, v_prev, v_grad)
]
step += 1
step_total += 1
print("Optimization Finished!")
def train_rule_only(
model_dir,
rule_trains,
max_steps,
hp=None,
ruleset='all',
seed=0,
):
'''Customized training function.
The network sequentially but only train rule for the second set.
First train the network to perform tasks in group 1, then train on group 2.
When training group 2, only rule connections are being trained.
Args:
model_dir: str, training directory
rule_trains: a list of list of tasks to train sequentially
hp: dictionary of hyperparameters
max_steps: int, maximum number of training steps for each list of tasks
display_step: int, display steps
ruleset: the set of rules to train
seed: int, random seed to be used
Returns:
model is stored at model_dir/model.ckpt
training configuration is stored at model_dir/hp.json
'''
tools.mkdir_p(model_dir)
# Network parameters
default_hp = get_default_hp(ruleset)
if hp is not None:
default_hp.update(hp)
hp = default_hp
hp['seed'] = seed
hp['rng'] = np.random.RandomState(seed)
hp['rule_trains'] = rule_trains
# Get all rules by flattening the list of lists
hp['rules'] = [r for rs in rule_trains for r in rs]
# Number of training iterations for each rule
if hasattr(max_steps, '__iter__'):
rule_train_iters = max_steps
else:
rule_train_iters = [len(r) * max_steps for r in rule_trains]
tools.save_hp(hp, model_dir)
# Display hp
for key, val in hp.items():
print('{:20s} = '.format(key) + str(val))
# Build the model
model = Model(model_dir, hp=hp)
# Store results
log = defaultdict(list)
log['model_dir'] = model_dir
# Record time
t_start = time.time()
# Use customized session that launches the graph as well
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# penalty on deviation from initial weight
if hp['l2_weight_init'] > 0:
raise NotImplementedError()
# Looping
step_total = 0
for i_rule_train, rule_train in enumerate(hp['rule_trains']):
step = 0
if i_rule_train == 0:
display_step = 200
else:
display_step = 50
if i_rule_train > 0:
# var_list = [v for v in model.var_list
# if ('input' in v.name) and ('rnn' not in v.name)]
var_list = [
v for v in model.var_list if 'rule_input' in v.name
]
model.set_optimizer(var_list=var_list)
# Keep training until reach max iterations
while (step * hp['batch_size_train'] <=
rule_train_iters[i_rule_train]):
# Validation
if step % display_step == 0:
trial = step_total * hp['batch_size_train']
log['trials'].append(trial)
log['times'].append(time.time() - t_start)
log['rule_now'].append(rule_train)
log = do_eval(sess, model, log, rule_train)
if log['perf_avg'][-1] > model.hp['target_perf']:
print('Perf reached the target: {:0.2f}'.format(
hp['target_perf']))
break
# Training
rule_train_now = hp['rng'].choice(rule_train)
# Generate a random batch of trials.
# Each batch has the same trial length
trial = generate_trials(rule_train_now,
hp,
'random',
batch_size=hp['batch_size_train'])
# Generating feed_dict.
feed_dict = tools.gen_feed_dict(model, trial, hp)
# This will compute the gradient BEFORE train step
_ = sess.run(model.train_step, feed_dict=feed_dict)
step += 1
step_total += 1
print("Optimization Finished!")
def createDir(self, dirPath, prefix):
if dir and not os.path.isdir(dirPath) :
self.verbose(prefix + "- Creating directory '{0}'".format(dirPath))
tools.mkdir_p(dirPath)
self.okay(prefix + "> OK")
def train(
model_dir,
hp=None,
max_steps=1e7,
display_step=500,
ruleset='mante',
rule_trains=None,
rule_prob_map=None,
seed=0,
rich_output=False,
load_dir=None,
trainables=None,
):
"""Train the network.
Args:
model_dir: str, training directory
hp: dictionary of hyperparameters
max_steps: int, maximum number of training steps
display_step: int, display steps
ruleset: the set of rules to train
rule_trains: list of rules to train, if None then all rules possible
rule_prob_map: None or dictionary of relative rule probability
seed: int, random seed to be used
Returns:
model is stored at model_dir/model.ckpt
training configuration is stored at model_dir/hp.json
"""
tools.mkdir_p(model_dir)
# Network parameters
default_hp = get_default_hp(ruleset)
if hp is not None:
default_hp.update(hp)
hp = default_hp
hp['seed'] = seed
hp['rng'] = np.random.RandomState(seed)
# Rules to train and test. Rules in a set are trained together
if rule_trains is None:
# By default, training all rules available to this ruleset
hp['rule_trains'] = task.rules_dict[ruleset]
else:
hp['rule_trains'] = rule_trains
hp['rules'] = hp['rule_trains']
# Assign probabilities for rule_trains.
if rule_prob_map is None:
rule_prob_map = dict()
# Turn into rule_trains format
hp['rule_probs'] = None
if hasattr(hp['rule_trains'], '__iter__'):
# Set default as 1.
rule_prob = np.array(
[rule_prob_map.get(r, 1.) for r in hp['rule_trains']])
hp['rule_probs'] = list(rule_prob / np.sum(rule_prob))
tools.save_hp(hp, model_dir)
# Build the model
model = Model(model_dir, hp=hp)
# Display hp
for key, val in hp.items():
print('{:20s} = '.format(key) + str(val))
# Store results
log = defaultdict(list)
log['model_dir'] = model_dir
# Record time
t_start = time.time()
with tf.Session() as sess:
if load_dir is not None:
model.restore(load_dir) # complete restore
else:
# Assume everything is restored
sess.run(tf.global_variables_initializer())
# Set trainable parameters
if trainables is None or trainables == 'all':
var_list = model.var_list # train everything
elif trainables == 'input':
# train all nputs
var_list = [
v for v in model.var_list
if ('input' in v.name) and ('rnn' not in v.name)
]
elif trainables == 'rule':
# train rule inputs only
var_list = [v for v in model.var_list if 'rule_input' in v.name]
else:
raise ValueError('Unknown trainables')
model.set_optimizer(var_list=var_list)
# penalty on deviation from initial weight
if hp['l2_weight_init'] > 0:
anchor_ws = sess.run(model.weight_list)
for w, w_val in zip(model.weight_list, anchor_ws):
model.cost_reg += (hp['l2_weight_init'] *
tf.nn.l2_loss(w - w_val))
model.set_optimizer(var_list=var_list)
# partial weight training
if ('p_weight_train' in hp and (hp['p_weight_train'] is not None)
and hp['p_weight_train'] < 1.0):
for w in model.weight_list:
w_val = sess.run(w)
w_size = sess.run(tf.size(w))
w_mask_tmp = np.linspace(0, 1, w_size)
hp['rng'].shuffle(w_mask_tmp)
ind_fix = w_mask_tmp > hp['p_weight_train']
w_mask = np.zeros(w_size, dtype=np.float32)
w_mask[ind_fix] = 1e-1 # will be squared in l2_loss
w_mask = tf.constant(w_mask)
w_mask = tf.reshape(w_mask, w.shape)
model.cost_reg += tf.nn.l2_loss((w - w_val) * w_mask)
model.set_optimizer(var_list=var_list)
step = 0
while step * hp['batch_size_train'] <= max_steps:
try:
# Validation
if step % display_step == 0:
log['trials'].append(step * hp['batch_size_train'])
log['times'].append(time.time() - t_start)
log = do_eval(sess, model, log, hp['rule_trains'])
#if log['perf_avg'][-1] > model.hp['target_perf']:
#check if minimum performance is above target
if log['perf_min'][-1] > model.hp['target_perf']:
print('Perf reached the target: {:0.2f}'.format(
hp['target_perf']))
break
if rich_output:
display_rich_output(model, sess, step, log, model_dir)
# Training
rule_train_now = hp['rng'].choice(hp['rule_trains'],
p=hp['rule_probs'])
# Generate a random batch of trials.
# Each batch has the same trial length
trial = generate_trials(rule_train_now,
hp,
'random',
batch_size=hp['batch_size_train'])
# Generating feed_dict.
feed_dict = tools.gen_feed_dict(model, trial, hp)
sess.run(model.train_step, feed_dict=feed_dict)
step += 1
except KeyboardInterrupt:
print("Optimization interrupted by user")
break
print("Optimization finished!")
def train_sequential_orthogonalized(model_dir,
rule_trains,
hp=None,
max_steps=1e7,
display_step=500,
rich_output=False,
ruleset='mante',
applyProj='both',
seed=0,
nEpisodeBatches=100,
projGrad=True,
alpha=0.001,
fixReadout=False):
'''Train the network sequentially.
Args:
model_dir: str, training directory
rule_trains: a list of list of tasks to train sequentially
hp: dictionary of hyperparameters
max_steps: int, maximum number of training steps for each list of tasks
display_step: int, display steps
ruleset: the set of rules to train
seed: int, random seed to be used
Returns:
model is stored at model_dir/model.ckpt
training configuration is stored at model_dir/hp.json
'''
tools.mkdir_p(model_dir)
# Network parameters
default_hp = get_default_hp(ruleset)
if hp is not None:
default_hp.update(hp)
hp = default_hp
hp['seed'] = seed
hp['rng'] = np.random.RandomState(seed)
hp['rule_trains'] = rule_trains
# Get all rules by flattening the list of lists
# hp['rules'] = [r for rs in rule_trains for r in rs]
hp['rules'] = rule_trains
# save some other parameters
hp['alpha_projection'] = alpha
hp['max_steps'] = max_steps
# Number of training iterations for each rule
rule_train_iters = [max_steps for _ in rule_trains]
tools.save_hp(hp, model_dir)
# Display hp
for key, val in hp.items():
print('{:20s} = '.format(key) + str(val))
# Build the model
model = Sequential_Model(model_dir,
projGrad=projGrad,
applyProj=applyProj,
hp=hp)
# Store results
log = defaultdict(list)
log['model_dir'] = model_dir
# Record time
t_start = time.time()
def relu(x):
return x * (x > 0.)
# -------------------------------------------------------
# Use customized session that launches the graph as well
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# penalty on deviation from initial weight
if hp['l2_weight_init'] > 0:
raise NotImplementedError()
# Looping
step_total = 0
taskNumber = 0
if fixReadout is True:
my_var_list = [
var for var in model.var_list
if 'rnn/leaky_rnn_cell/kernel:0' in var.name
]
else:
my_var_list = [
var for var in model.var_list
if 'rnn/leaky_rnn_cell/kernel:0' in var.name
or 'output/weights:0' in var.name
]
# initialise projection matrices
input_proj = tf.zeros(
(hp['n_rnn'] + hp['n_input'], hp['n_rnn'] + hp['n_input']))
activity_proj = tf.zeros((hp['n_rnn'], hp['n_rnn']))
output_proj = tf.zeros((hp['n_output'], hp['n_output']))
recurrent_proj = tf.zeros((hp['n_rnn'], hp['n_rnn']))
for i_rule_train, rule_train in enumerate(hp['rule_trains']):
step = 0
model.set_optimizer(activity_proj=activity_proj,
input_proj=input_proj,
output_proj=output_proj,
recurrent_proj=recurrent_proj,
taskNumber=taskNumber,
var_list=my_var_list,
alpha=alpha)
# Keep training until reach max iterations
while (step * hp['batch_size_train'] <=
rule_train_iters[i_rule_train]):
# Validation
if step % display_step == 0:
trial = step_total * hp['batch_size_train']
log['trials'].append(trial)
log['times'].append(time.time() - t_start)
log['rule_now'].append(rule_train)
log = do_eval(sess, model, log, rule_train)
if log['perf_avg'][-1] > model.hp['target_perf']:
print('Perf reached the target: {:0.2f}'.format(
hp['target_perf']))
break
# Training
# rule_train_now = hp['rng'].choice(rule_train)
# Generate a random batch of trials.
# Each batch has the same trial length
trial = generate_trials(rule_train,
hp,
'random',
batch_size=hp['batch_size_train'],
delay_fac=hp['delay_fac'])
# Generating feed_dict.
feed_dict = tools.gen_feed_dict(model, trial, hp)
# update model
sess.run(model.train_step, feed_dict=feed_dict)
# # Get the weight after train step
# v_current = sess.run(model.var_list)
step += 1
step_total += 1
if step % display_step == 0:
model.save_ckpt(step_total)
# ---------- save model after its completed training the current task ----------
model.save_after_task(taskNumber)
# ---------- generate task activity for continual learning -------
trial = generate_trials(rule_train,
hp,
'random',
batch_size=hp['batch_size_test'],
delay_fac=hp['delay_fac'])
# Generating feed_dict.
feed_dict = tools.gen_feed_dict(model, trial, hp)
eval_h, eval_x, eval_y, Wrec, Win = sess.run(
[model.h, model.x, model.y, model.w_rec, model.w_in],
feed_dict=feed_dict)
full_state = np.concatenate([eval_x, eval_h], -1)
# get weight matrix after current task
Wfull = np.concatenate([Win, Wrec], 0)
# joint covariance matrix of input and activity
Shx_task = compute_covariance(
np.reshape(full_state, (-1, hp['n_rnn'] + hp['n_input'])).T)
# covariance matrix of output
Sy_task = compute_covariance(
np.reshape(eval_y, (-1, hp['n_output'])).T)
# get block matrices from Shx_task
# Sh_task = Shx_task[-hp['n_rnn']:, -hp['n_rnn']:]
Sh_task = np.matmul(np.matmul(Wfull.T, Shx_task), Wfull)
# ---------- update stored covariance matrices for continual learning -------
if taskNumber == 0:
input_cov = Shx_task
activity_cov = Sh_task
output_cov = Sy_task
else:
input_cov = taskNumber / (
taskNumber + 1) * input_cov + Shx_task / (taskNumber + 1)
activity_cov = taskNumber / (
taskNumber + 1) * activity_cov + Sh_task / (taskNumber + 1)
output_cov = taskNumber / (
taskNumber + 1) * output_cov + Sy_task / (taskNumber + 1)
# ---------- update projection matrices for continual learning ----------
activity_proj, input_proj, output_proj, recurrent_proj = compute_projection_matrices(
activity_cov, input_cov, output_cov,
input_cov[-hp['n_rnn']:, -hp['n_rnn']:], alpha)
# update task number
taskNumber += 1
print("Optimization Finished!")
def createDir(self, dirPath, prefix):
if dir and not os.path.isdir(dirPath) :
self.verbose(prefix + "- Creating directory '{0}'".format(dirPath))
tools.mkdir_p(dirPath)
print (colored(prefix + "> OK", 'green'))
cv2.rectangle(block_img, (lx, ly + lh / 2), (lx + lw, ly + lh / 2),
(250, 0, 0), 1)
#cv2.putText(block_img, str(line_index), (lx, ly+lh/2), cv2.FONT_HERSHEY_SIMPLEX, 0.8, 255, 2)
# Line Features:
# - mass/aria (ink_volume)
# - height
# - word_count/width (words_ratio)
return block_img
#cv2.rectangle(img, (x, y), (x+w, y+h), (255, 0, 255), 2)
if __name__ == '__main__':
mkdir_p('lines')
mkdir_p('blocks')
eprint(
'filename,ink_volume,fiber_ink_volume,height,width,area,words_ratio,line_index'
)
#find_text_block(sys.argv[1], 'marked.png', 'dilated.png')
for page_index, page_name in enumerate(sys.argv[1:]):
print('Page %d' % page_index)
dilated, img, contours = find_text_block(page_name)
#dilated, img = find_text_block(cv2.imread(sys.argv[1]),
# cv2.imread('marked.png'),
# cv2.imread('dilated.png'))
cv2.imwrite('dilated%d.png' % page_index, dilated)
cv2.imwrite('blocks%d.png' % page_index, img)
def error_analysis(model, ids, notes, text_features, X, Y, label, task):
criteria, vect, clf = model
if task in ['sapsii', 'age', 'los']:
V = {}
labels_ = [0] + criteria
for i in range(len(labels_) - 1):
label = '[%d,%s)' % (labels_[i], labels_[i + 1])
V[i] = label
else:
V = {v: k for k, v in criteria.items()}
V[len(V)] = '**wrong**'
# for confidence
P_ = clf.decision_function(X)
# sklearn has stupid changes in API when doing binary classification. make it conform to 3+
if len(criteria) == 2:
m = X.shape[0]
P = np.zeros((m, 2))
P[:, 0] = -P_
P[:, 1] = P_
n = clf.coef_.shape[1]
coef_ = np.zeros((2, n))
coef_[0, :] = -clf.coef_[0, :]
coef_[1, :] = clf.coef_[0, :]
else:
P = P_
coef_ = clf.coef_
# hard predictions
pred = clf.predict(X)
# convert predictions to right vs wrong
confidence = {}
for i, scores in enumerate(P.tolist()):
prediction = pred[i]
pid = ids[i]
ind = Y[i]
confidence[pid] = (scores[ind], scores, prediction, ind)
homedir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
methoddir = os.path.join(homedir, 'output', task, 'bow')
mkdir_p(methoddir)
def importance(featname, p):
if featname in vect.vocabulary_:
ind = vect.vocabulary_[featname]
return coef_[p, ind]
else:
return float('-inf')
# order predictions by confidence
for pid, conf in sorted(confidence.items(), key=lambda t: t[1]):
if conf[2] == conf[3]:
success = ''
else:
success = '_'
#if success == '_':
if True:
filename = os.path.join(methoddir, '%s%s.pred' % (success, pid))
with open(filename, 'w') as f:
print >> f, ''
print >> f, '=' * 80
print >> f, ''
print >> f, pid
print >> f, 'scores:', conf[1]
print >> f, 'pred: ', V[conf[2]]
print >> f, 'ref: ', V[conf[3]]
print >> f, '#' * 20
for feat, v in sorted(text_features[pid].items(),
key=lambda t: importance(t[0], conf[3])):
#for feat,v in sorted(text_features[pid].items(), key=lambda t:t[1]):
imp = importance(feat, conf[2])
#imp = v
#print >>f, feat,
print >> f, '%.3f %s' % (imp, feat)
print >> f, '#' * 20
print >> f, 'SCORES'
pind = conf[2]
rind = conf[3]
print >> f, 'predicted:', sum([
val * importance(feat, pind)
for feat, val in text_features[pid].items()
if float('-inf') < importance(feat, pind) < float('inf')
])
print >> f, 'true: ', sum([
val * importance(feat, rind)
for feat, val in text_features[pid].items()
if float('-inf') < importance(feat, rind) < float('inf')
])
print >> f, '#' * 20
#'''
for dt, category, text in sorted(notes[pid]):
print >> f, dt
print >> f, category
print >> f, text
print >> f, '-' * 50
#'''
print >> f, ''
print >> f, '+' * 80
print >> f, ''
def main():
global best_acc
start_epoch = args.start_epoch # start from epoch 0 or last checkpoint epoch
if not os.path.isdir(args.checkpoint):
mkdir_p(args.checkpoint)
print('==> Preparing dataset %s' % args.dataset)
if args.dataset == 'cifar100':
training_loader = get_training_dataloader(settings.CIFAR100_TRAIN_MEAN,
settings.CIFAR100_TRAIN_STD,
num_workers=4,
batch_size=args.train_batch,
shuffle=True)
test_loader = get_test_dataloader(settings.CIFAR100_TRAIN_MEAN,
settings.CIFAR100_TRAIN_STD,
num_workers=4,
batch_size=args.test_batch,
shuffle=False)
num_classes = 100
else:
training_loader = get_training_dataloader_10(
settings.CIFAR10_TRAIN_MEAN,
settings.CIFAR10_TRAIN_STD,
num_workers=4,
batch_size=args.train_batch,
shuffle=True)
test_loader = get_test_dataloader_10(settings.CIFAR10_TRAIN_MEAN,
settings.CIFAR10_TRAIN_STD,
num_workers=4,
batch_size=args.test_batch,
shuffle=False)
num_classes = 10
#data preprocessing:
print("==> creating model '{}'".format(args.arch))
model = get_network(args, num_classes=num_classes)
model = torch.nn.DataParallel(model).cuda()
cudnn.benchmark = True
print(' Total params: %.2fM' %
(sum(p.numel() for p in model.parameters()) / 1000000.0))
criterion1 = am_softmax.AMSoftmax()
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=5e-4)
title = 'cifar-10-' + args.arch
if args.resume:
# Load checkpoint.
print('==> Resuming from checkpoint..')
assert os.path.isfile(
args.resume), 'Error: no checkpoint directory found!'
args.checkpoint = os.path.dirname(args.resume)
checkpoint = torch.load(args.resume)
best_acc = checkpoint['best_acc']
start_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
logger = Logger(os.path.join(args.checkpoint, 'log.txt'),
title=title,
resume=True)
else:
logger = Logger(os.path.join(args.checkpoint, 'log.txt'), title=title)
logger.set_names([
'Learning Rate', 'Train Loss', 'Valid Loss', 'Train Acc.',
'Valid Acc.'
])
train_scheduler = optim.lr_scheduler.MultiStepLR(
optimizer, milestones=args.schedule, gamma=0.2) #learning rate decay
iter_per_epoch = len(training_loader)
warmup_scheduler = WarmUpLR(optimizer, iter_per_epoch * args.warm)
for epoch in range(start_epoch, args.epochs):
if epoch > args.warm:
train_scheduler.step(epoch)
train_loss, train_acc = train(training_loader, model, warmup_scheduler,
criterion, criterion1, optimizer, epoch,
use_cuda)
test_loss, test_acc = eval_training(test_loader, model, criterion,
epoch, use_cuda)
logger.append([
optimizer.param_groups[0]['lr'], train_loss, test_loss, train_acc,
test_acc
])
# save model
is_best = test_acc > best_acc
best_acc = max(test_acc, best_acc)
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'acc': test_acc,
'best_acc': best_acc,
'optimizer': optimizer.state_dict(),
},
is_best,
checkpoint=args.checkpoint)
logger.close()
# logger.plot()
# savefig(os.path.join(args.checkpoint, 'log.eps'))
print('Best acc:')
print(best_acc)
def run(self):
thumbnail_dpath = os.path.join(self.clip_dpath, 'media')
try:
tools.mkdir_p(thumbnail_dpath)
except OSError as err:
wx.CallAfter(
self.wx_parent.error_message,
_('Permission error creating clip directory.')
)
return
tools.save_safe_image(self.clip_dpath, self.bg_input_fpath)
tools.save_safe_image(self.clip_dpath, self.pip_input_fpath)
tools.save_thumbnails(thumbnail_dpath, self.pip_input_fpath)
pip_width, pip_height = tools.get_pip_size(
self.pip_input_fpath,
self.pip_scale
)
bg_fname = tools.sanitize_path(os.path.split(self.bg_input_fpath)[1])
pip_fname = tools.sanitize_path(os.path.split(self.pip_input_fpath)[1])
clip_config = tools.default_clip_config()
clip_config['name'] = self.clip_name
clip_config['data'] = bg_fname
clip_config['pip'] = pip_fname
clip_config['pipwidth'] = pip_width
clip_config['pipheight'] = pip_height
clip_config['category'] = self.clip_category
clip_config_buf = ''
# name, data, pip must be first 3 items in order
ordered_config_items = ['name', 'data', 'pip']
for line in ordered_config_items:
clip_config_buf += line + ' = ' + clip_config[line] + os.linesep
# Add the rest of the items whose order probably doesn't matter
for key, value in clip_config.iteritems():
key = unicode(key)
value = unicode(value)
if value and (key not in ordered_config_items):
clip_config_buf += key + ' = ' + value + os.linesep
clip_config_fpath = os.path.join(self.clip_dpath, 'playlist.sos')
with open(clip_config_fpath, 'wb') as clip_config_fobj:
clip_config_fobj.write(tools.encode(clip_config_buf))
# # Update the SOS library
# # Not used because updating the library doesn't automatically update
# # the playlist panel in SOS Stream GUI, and updating the playlist
# # panel in SOS Stream GUI the right way would be tedious.
# # Rescanning the library automatically without refreshing the
# # playlist panel could confuse users.
# child = subprocess.Popen(
# ['/shared/sos/default/bin/scan_library'], stdout=subprocess.PIPE
# )
# while child.poll() is None: # Read command output as it comes
# out = child.stdout.readline()
# print(out.strip()) # Put this in a status window later
# # The menu can be updated, at least:
# wish = Tkinter.Tk()
# wish.withdraw() # Hide Tk window. Possible with constructor?
# try:
# wish.send('sos_stream_gui',
# 'destroy',
# '.menubar.library')
# wish.send('sos_stream_gui',
# 'mk_library_menu',
# '.menubar.library')
# except Tkinter.TclError as err:
# pass # SOS Stream GUI not open
# Done
success_message = '\n'.join(
[
_('Clip "{}" successfully added.'),
_('To activate the changes, '
'perform the following actions in SOS Stream GUI:'),
_('Library -> Update Library...\n'
'Library -> (all/{})')
]
).format(
clip_config['name'],
clip_config['category']
)
wx.CallAfter(self.wx_parent.information_message, success_message)
def plot_tuning_feature(self,
epochs,
rules = None,
trial_list = None,
neuron_types = [('exh_neurons','mix_neurons'),('inh_neurons',)]):
from scipy.optimize import curve_fit
import math
if trial_list is None:
trial_list = self.trial_list
if rules is None:
rules = self.rules
neuron_cate = OrderedDict()
tuning_by_trial = OrderedDict()
for rule in rules:
neuron_cate[rule] = OrderedDict()
for epoch in epochs:
neuron_cate[rule][epoch] = OrderedDict()
tools.mkdir_p('figure/figure_'+self.model_dir+'/'+rule+'/'+epoch+'/')
if epoch not in self.epoch_info[rule].keys():
raise KeyError('Rule ',rule,' dose not have epoch ',epoch,'!')
for trial_num in trial_list:
tuning_by_trial[trial_num] = OrderedDict()
for rule in rules:
tuning_by_trial[trial_num][rule] = OrderedDict()
for epoch in epochs:
tuning_by_trial[trial_num][rule][epoch] = dict()
for key, value in self.neuron_info[trial_num][rule][epoch].items():
if 'neurons' in key:
print(key,"number:",len(value))
if key not in neuron_cate[rule][epoch].keys():
neuron_cate[rule][epoch][key] = [len(value)]
else:
neuron_cate[rule][epoch][key].append(len(value))
for type_pair in neuron_types:
type_pair_folder = 'figure/figure_'+self.model_dir+'/'+rule+'/'+epoch+'/'+'_'.join(type_pair)+'/'
tools.mkdir_p(type_pair_folder)
fig_by_epoch = plt.figure()
tuning_by_neuron = list()
for n_type in type_pair:
for neuron in self.neuron_info[trial_num][rule][epoch][n_type]:
tuning_by_neuron.append(self.neuron_info[trial_num][rule][epoch]['firerate_max_central'][neuron])
tuning_by_neuron = np.array(tuning_by_neuron)
plt.plot(np.arange(len(self.neuron_info[trial_num][rule][epoch]['firerate_max_central'][neuron])),
tuning_by_neuron.T)
plt.title("Rule:"+rule+" Epoch:"+epoch+" Trial:"+str(trial_num)+\
" Perf:"+str(self.log['perf_'+rule][trial_num//self.log['trials'][1]])[:4])
save_name_neuron = type_pair_folder+str(trial_num)+'.png'
plt.tight_layout()
plt.savefig(save_name_neuron, transparent=False, bbox_inches='tight')
plt.close(fig_by_epoch)
tuning_by_trial[trial_num][rule][epoch][type_pair] = tuning_by_neuron.mean(axis=0)
for type_pair in neuron_types:
for rule in rules:
for epoch in epochs:
fig_tuning = plt.figure()
for trial_num in trial_list:
plt.plot(np.arange(len(tuning_by_trial[trial_num][rule][epoch][type_pair])),
tuning_by_trial[trial_num][rule][epoch][type_pair],
label = str(trial_num)+' Perf:'+str(self.log['perf_'+rule][trial_num//self.log['trials'][1]])[:4])
type_pair_folder = 'figure/figure_'+self.model_dir+'/'+rule+'/'+epoch+'/'+'_'.join(type_pair)+'/'
save_name_trial = type_pair_folder+'tuning_all_'+str(trial_list[0])+'to'+str(trial_list[-1])+'.png'
plt.legend(bbox_to_anchor=(1.05, 0), loc=3, borderaxespad=0)
plt.tight_layout()
plt.savefig(save_name_trial, transparent=False,bbox_inches='tight')
plt.close(fig_tuning)
#tunning by growth and gaussian curve fit
tuning_by_growth = OrderedDict()
def gaussian(x, a,u, sig):
return a*np.exp(-(x - u) ** 2 / (2 * sig ** 2)) / (sig * math.sqrt(2 * math.pi))
for type_pair in neuron_types:
tuning_by_growth[type_pair] = OrderedDict()
for rule in rules:
tuning_by_growth[type_pair][rule] = OrderedDict()
for epoch in epochs:
tuning_by_growth[type_pair][rule][epoch] = OrderedDict()
for growth_key in ['less_than_I','I_to_Y','elder_than_Y']:
tuning_by_growth[type_pair][rule][epoch][growth_key]=dict()
for type_key in ['growth','tuning']:
tuning_by_growth[type_pair][rule][epoch][growth_key][type_key]=list()
for trial_num in trial_list:
growth = self.log['perf_'+rule][trial_num//self.log['trials'][1]]
if growth <= self.hp['infancy_target_perf']:
tuning_by_growth[type_pair][rule][epoch]['less_than_I']['growth'].append(growth)
tuning_by_growth[type_pair][rule][epoch]['less_than_I']['tuning'].\
append(tuning_by_trial[trial_num][rule][epoch][type_pair])
elif growth <= self.hp['young_target_perf']:
tuning_by_growth[type_pair][rule][epoch]['I_to_Y']['growth'].append(growth)
tuning_by_growth[type_pair][rule][epoch]['I_to_Y']['tuning'].\
append(tuning_by_trial[trial_num][rule][epoch][type_pair])
else:
tuning_by_growth[type_pair][rule][epoch]['elder_than_Y']['growth'].append(growth)
tuning_by_growth[type_pair][rule][epoch]['elder_than_Y']['tuning'].\
append(tuning_by_trial[trial_num][rule][epoch][type_pair])
for growth_key in ['less_than_I','I_to_Y','elder_than_Y']:
for type_key in ['growth','tuning']:
try:
tuning_by_growth[type_pair][rule][epoch][growth_key][type_key]=\
np.array(tuning_by_growth[type_pair][rule][epoch][growth_key][type_key]).mean(axis=0)
except:
pass
for type_pair in neuron_types:
for rule in rules:
for epoch in epochs:
fig_tuning_growth = plt.figure()
for growth_key in ['less_than_I','I_to_Y','elder_than_Y']:
tuning_temp = tuning_by_growth[type_pair][rule][epoch][growth_key]['tuning']
temp_x = np.arange(len(tuning_temp)) ###############################################TODO:something wrong for anti saccade
gaussian_x = np.arange(-0.1,len(tuning_temp)-0.9,0.1)
#gaussian fit
paras , _ = curve_fit(gaussian,temp_x,tuning_temp+(-1)*np.min(tuning_temp),\
p0=[np.max(tuning_temp)+1,len(tuning_temp)//2,1])
gaussian_y = gaussian(gaussian_x,paras[0],paras[1],paras[2])-np.min(tuning_temp)*(-1)
tuning_by_growth[type_pair][rule][epoch][growth_key]['gaussian_fit']=gaussian_y
if growth_key == 'less_than_I':
color = 'green'
elif growth_key == 'I_to_Y':
color = 'blue'
else:
color = 'red'
plt.scatter(temp_x, tuning_temp, marker = '+',color = color, s = 70 ,\
label = growth_key+'_'+str(tuning_by_growth[type_pair][rule][epoch][growth_key]['growth'])[:4])
plt.plot(gaussian_x, gaussian_y, color=color,linestyle = '--',\
label = growth_key+'_'+str(tuning_by_growth[type_pair][rule][epoch][growth_key]['growth'])[:4])
plt.legend()
type_pair_folder = 'figure/figure_'+self.model_dir+'/'+rule+'/'+epoch+'/'+'_'.join(type_pair)+'/'
save_name_tun_growth = type_pair_folder+'tuning_growth_'+str(trial_list[0])+'to'+str(trial_list[-1])+'.png'
plt.legend(bbox_to_anchor=(1.05, 0), loc=3, borderaxespad=0)
plt.tight_layout()
plt.savefig(save_name_tun_growth, transparent=False,bbox_inches='tight')
plt.close(fig_tuning_growth)
#plt.show()
#plot neuron category changes
for rule in rules:
for epoch in epochs:
fig_cate = plt.figure()
for key,value in neuron_cate[rule][epoch].items():
plt.plot(trial_list, value, label = key)
plt.legend(bbox_to_anchor=(1.05, 0), loc=3, borderaxespad=0)
plt.tight_layout()
plt.savefig('figure/figure_'+self.model_dir+'/'+rule+'/'+epoch+'/'+'neuron_cate_'+str(trial_list[0])+'to'+str(trial_list[-1])+'.png',transparent=False,bbox_inches='tight')
plt.close(fig_cate)
def plot_PSTH(self,
epochs,
rules=None,
trial_list=None,
neuron_types=[('exh_neurons','mix_neurons')],
norm = True,
separate_plot = False,
fuse_rules = False):
if trial_list is None:
trial_list = self.trial_list
if rules is None:
rules = self.rules
psth_to_plot = OrderedDict()
for trial_num in trial_list:
psth_to_plot[trial_num] = OrderedDict()
for rule in rules:
H = tools.load_pickle(self.model_dir+'/'+str(trial_num)+'/'+'H_'+rule+'.pkl')
psth_to_plot[trial_num][rule] = OrderedDict()
for epoch in epochs:
psth_to_plot[trial_num][rule][epoch] = OrderedDict()
for type_pair in neuron_types:
psth_to_plot[trial_num][rule][epoch][type_pair] = OrderedDict()
psth_neuron = list()
anti_dir_psth = list()
for n_type in type_pair:
for neuron in self.neuron_info[trial_num][rule][epoch][n_type]:
sel_loc = np.argmax(self.neuron_info[trial_num][rule][epoch]['firerate_loc_order'][neuron])
anti_loc = (sel_loc+len(self.in_loc_set[rule])//2)%len(self.in_loc_set[rule])
psth_temp = H[:,self.in_loc[rule] == sel_loc, neuron].mean(axis=1)
fix_level = H[self.epoch_info[rule]['fix1'][0]:self.epoch_info[rule]['fix1'][1], \
self.in_loc[rule] == sel_loc, neuron].mean(axis=1).mean(axis=0)
if len(self.in_loc_set[rule])%2:
anti_dir_psth_temp = (H[:,self.in_loc[rule] == anti_loc, neuron].mean(axis=1)+\
H[:,self.in_loc[rule] == (anti_loc+1), neuron].mean(axis=1))/2.0
else:
anti_dir_psth_temp = H[:,self.in_loc[rule] == anti_loc, neuron].mean(axis=1)
anti_dir_psth_norm = anti_dir_psth_temp/fix_level-1
psth_norm = psth_temp/fix_level-1
if norm:
psth_neuron.append(psth_norm)
anti_dir_psth.append(anti_dir_psth_norm)
else:
psth_neuron.append(psth_temp)
anti_dir_psth.append(anti_dir_psth_temp)
try:
psth_to_plot[trial_num][rule][epoch][type_pair]['sel_dir'] = np.array(psth_neuron).mean(axis=0)
except:
pass
try:
psth_to_plot[trial_num][rule][epoch][type_pair]['anti_sel_dir'] = np.array(anti_dir_psth).mean(axis=0)
except:
pass
for rule in rules:
for epoch in epochs:
for type_pair in neuron_types:
if not separate_plot:
fig_psth = plt.figure()
for trial_num in trial_list:
if separate_plot:
fig_psth = plt.figure()
color = None
else:
color = kelly_colors[(trial_list.index(trial_num)+1)%len(kelly_colors)]
try:
plt.plot(np.arange(len(psth_to_plot[trial_num][rule][epoch][type_pair]['sel_dir']))*self.hp['dt']/1000,
psth_to_plot[trial_num][rule][epoch][type_pair]['sel_dir'],label=str(trial_num)+'sel',color=color)
except:
pass
try:
plt.plot(np.arange(len(psth_to_plot[trial_num][rule][epoch][type_pair]['anti_sel_dir']))*self.hp['dt']/1000,
psth_to_plot[trial_num][rule][epoch][type_pair]['anti_sel_dir'],linestyle = '--',label=str(trial_num)+'anti',color=color)
except:
pass
if separate_plot:
plt.title("Rule:"+rule+" Epoch:"+epoch+" Neuron_type:"+"_".join(type_pair))
plt.legend(bbox_to_anchor=(1.05, 0), loc=3, borderaxespad=0)
type_pair_folder = 'figure/figure_'+self.model_dir+'/'+rule+'/'+epoch+'/'+'_'.join(type_pair)+'/'
tools.mkdir_p(type_pair_folder)
plt.tight_layout()
plt.savefig(type_pair_folder+'PSTH-'+str(trial_num)+'.png',transparent=False,bbox_inches='tight')
plt.close(fig_psth)
if not separate_plot:
plt.title("Rule:"+rule+" Epoch:"+epoch+" Neuron_type:"+"_".join(type_pair))
plt.legend(bbox_to_anchor=(1.05, 0), loc=3, borderaxespad=0)
type_pair_folder = 'figure/figure_'+self.model_dir+'/'+rule+'/'+epoch+'/'+'_'.join(type_pair)+'/'
tools.mkdir_p(type_pair_folder)
plt.tight_layout()
plt.savefig(type_pair_folder+'PSTH_all_'+str(trial_list[0])+'to'+str(trial_list[-1])+'.png',
transparent=False,bbox_inches='tight')
plt.close(fig_psth)
plot_by_growth = dict()
for rule in rules:
plot_by_growth[rule] = dict()
for epoch in epochs:
plot_by_growth[rule][epoch] = dict()
for type_pair in neuron_types:
plot_by_growth[rule][epoch][type_pair] = dict()
plot_by_growth[rule][epoch][type_pair]['less_than_I'] = dict()
plot_by_growth[rule][epoch][type_pair]['I_to_Y'] = dict()
plot_by_growth[rule][epoch][type_pair]['elder_than_Y'] = dict()
for key in plot_by_growth[rule][epoch][type_pair].keys():
plot_by_growth[rule][epoch][type_pair][key]['sel'] = list()
plot_by_growth[rule][epoch][type_pair][key]['anti'] = list()
plot_by_growth[rule][epoch][type_pair][key]['growth'] = list()
for trial_num in trial_list:
growth = self.log['perf_'+rule][trial_num//self.log['trials'][1]]
if growth <= self.hp['infancy_target_perf']:
plot_by_growth[rule][epoch][type_pair]['less_than_I']['sel'].append(\
psth_to_plot[trial_num][rule][epoch][type_pair]['sel_dir'])
plot_by_growth[rule][epoch][type_pair]['less_than_I']['anti'].append(\
psth_to_plot[trial_num][rule][epoch][type_pair]['anti_sel_dir'])
plot_by_growth[rule][epoch][type_pair]['less_than_I']['growth'].append(growth)
elif growth <= self.hp['young_target_perf']:
plot_by_growth[rule][epoch][type_pair]['I_to_Y']['sel'].append(\
psth_to_plot[trial_num][rule][epoch][type_pair]['sel_dir'])
plot_by_growth[rule][epoch][type_pair]['I_to_Y']['anti'].append(\
psth_to_plot[trial_num][rule][epoch][type_pair]['anti_sel_dir'])
plot_by_growth[rule][epoch][type_pair]['I_to_Y']['growth'].append(growth)
else:
plot_by_growth[rule][epoch][type_pair]['elder_than_Y']['sel'].append(\
psth_to_plot[trial_num][rule][epoch][type_pair]['sel_dir'])
plot_by_growth[rule][epoch][type_pair]['elder_than_Y']['anti'].append(\
psth_to_plot[trial_num][rule][epoch][type_pair]['anti_sel_dir'])
plot_by_growth[rule][epoch][type_pair]['elder_than_Y']['growth'].append(growth)
for growth_key in plot_by_growth[rule][epoch][type_pair].keys():
for type_key, value in plot_by_growth[rule][epoch][type_pair][growth_key].items():
try:
plot_by_growth[rule][epoch][type_pair][growth_key][type_key] = np.array(value).mean(axis=0)
except:
pass
for rule in rules:
for epoch in epochs:
for type_pair in neuron_types:
fig_psth = plt.figure()
for growth_key,value in plot_by_growth[rule][epoch][type_pair].items():
if growth_key == 'less_than_I':
color = 'green'
elif growth_key == 'I_to_Y':
color = 'blue'
else:
color = 'red'
try:
plt.plot(np.arange(len(value['sel']))*self.hp['dt']/1000,value['sel'],
label = growth_key+'_'+str(value['growth'])[:4], color = color)
except:
pass
try:
plt.plot(np.arange(len(value['anti']))*self.hp['dt']/1000,value['anti'],
linestyle = '--',label = growth_key+'_'+str(value['growth'])[:4], color = color)
except:
pass
plt.title("Rule:"+rule+" Epoch:"+epoch+" Neuron_type:"+"_".join(type_pair))
plt.legend(bbox_to_anchor=(1.05, 0), loc=3, borderaxespad=0)
type_pair_folder = 'figure/figure_'+self.model_dir+'/'+rule+'/'+epoch+'/'+'_'.join(type_pair)+'/'
tools.mkdir_p(type_pair_folder)
plt.tight_layout()
plt.savefig(type_pair_folder+'PSTH_bygrowth_'+str(trial_list[0])+'to'+str(trial_list[-1])+'.png',
transparent=False,bbox_inches='tight')
plt.close(fig_psth)
if fuse_rules:
ls_list = ['-','--','-.',':']
for type_pair in neuron_types:
for epoch in epochs:
fig_psth_fr = plt.figure()
for rule, linestyle in zip(rules,ls_list[0:len(rules)]):
for key,value in plot_by_growth[rule][epoch][type_pair].items():
if key == 'less_than_I':
color = 'green'
elif key == 'I_to_Y':
color = 'blue'
else:
color = 'red'
try:
plt.plot(np.arange(len(value['sel']))*self.hp['dt']/1000,value['sel'],
label = rule+'_'+key+'_'+str(value['growth'])[:4], color = color, linestyle=linestyle)
except:
pass
plt.legend()
plt.title("Rule:"+rule+" Epoch:"+epoch+" Neuron_type:"+"_".join(type_pair))
plt.legend(bbox_to_anchor=(1.05, 0), loc=3, borderaxespad=0)
plt.tight_layout()
plt.savefig('figure/figure_'+self.model_dir+'/'+'-'.join(rules)+'_'+epoch+'_'+'-'.join(type_pair)\
+str(trial_list[0])+'to'+str(trial_list[-1])+'.png',
transparent=False,bbox_inches='tight')
plt.close(fig_psth_fr)
def train(
model_dir,
hp=None,
max_steps=1e7,
display_step=500,
ruleset='mante',
rule_trains=None,
rule_prob_map=None,
seed=0,
rich_output=True,
load_dir=None,
trainables=None,
fixReadoutandBias=False,
fixBias=False,
):
"""Train the network.
Args:
model_dir: str, training directory
hp: dictionary of hyperparameters
max_steps: int, maximum number of training steps
display_step: int, display steps
ruleset: the set of rules to train
rule_trains: list of rules to train, if None then all rules possible
rule_prob_map: None or dictionary of relative rule probability
seed: int, random seed to be used
Returns:
model is stored at model_dir/model.ckpt
training configuration is stored at model_dir/hp.json
"""
tools.mkdir_p(model_dir)
# Network parameters
default_hp = get_default_hp(ruleset)
if hp is not None:
default_hp.update(hp)
hp = default_hp
hp['seed'] = seed
hp['rng'] = np.random.RandomState(seed)
# Rules to train and test. Rules in a set are trained together
if rule_trains is None:
# By default, training all rules available to this ruleset
hp['rule_trains'] = task.rules_dict[ruleset]
else:
hp['rule_trains'] = rule_trains
hp['rules'] = hp['rule_trains']
# Assign probabilities for rule_trains.
if rule_prob_map is None:
rule_prob_map = dict()
# Turn into rule_trains format
hp['rule_probs'] = None
if hasattr(hp['rule_trains'], '__iter__'):
# Set default as 1.
rule_prob = np.array(
[rule_prob_map.get(r, 1.) for r in hp['rule_trains']])
hp['rule_probs'] = list(rule_prob / np.sum(rule_prob))
tools.save_hp(hp, model_dir)
# Build the model
with tf.device('gpu:0'):
model = Model(model_dir, hp=hp)
# Display hp
for key, val in hp.items():
print('{:20s} = '.format(key) + str(val))
if fixReadoutandBias is True:
my_var_list = [
var for var in model.var_list
if 'rnn/leaky_rnn_cell/kernel:0' in var.name
]
print(my_var_list)
elif fixBias is True:
my_var_list = [
var for var in model.var_list
if 'rnn/leaky_rnn_cell/kernel:0' in var.name
or 'output/weights:0' in var.name
]
else:
my_var_list = model.var_list
model.set_optimizer(var_list=my_var_list)
# Store results
log = defaultdict(list)
log['model_dir'] = model_dir
# Record time
t_start = time.time()
# Use customized session that launches the graph as well
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# penalty on deviation from initial weight
if hp['l2_weight_init'] > 0:
anchor_ws = sess.run(model.weight_list)
for w, w_val in zip(model.weight_list, anchor_ws):
model.cost_reg += (hp['l2_weight_init'] *
tf.nn.l2_loss(w - w_val))
model.set_optimizer(var_list=my_var_list)
# partial weight training
if ('p_weight_train' in hp and (hp['p_weight_train'] is not None)
and hp['p_weight_train'] < 1.0):
for w in model.weight_list:
w_val = sess.run(w)
w_size = sess.run(tf.size(w))
w_mask_tmp = np.linspace(0, 1, w_size)
hp['rng'].shuffle(w_mask_tmp)
ind_fix = w_mask_tmp > hp['p_weight_train']
w_mask = np.zeros(w_size, dtype=np.float32)
w_mask[ind_fix] = 1e-1 # will be squared in l2_loss
w_mask = tf.constant(w_mask)
w_mask = tf.reshape(w_mask, w.shape)
model.cost_reg += tf.nn.l2_loss((w - w_val) * w_mask)
model.set_optimizer(var_list=my_var_list)
step = 0
run_ave_time = []
while step * hp['batch_size_train'] <= max_steps:
try:
# Validation
if step % display_step == 0:
grad_norm = tf.global_norm(model.clipped_gs)
grad_norm_np = sess.run(grad_norm)
# import pdb
# pdb.set_trace()
log['grad_norm'].append(grad_norm_np.item())
log['trials'].append(step * hp['batch_size_train'])
log['times'].append(time.time() - t_start)
log = do_eval(sess, model, log, hp['rule_trains'])
# if log['perf_avg'][-1] > model.hp['target_perf']:
# check if minimum performance is above target
if log['perf_min'][-1] > model.hp['target_perf']:
print('Perf reached the target: {:0.2f}'.format(
hp['target_perf']))
break
if rich_output:
display_rich_output(model, sess, step, log,
model_dir)
# Training
dtStart = datetime.now()
sess.run(model.train_step)
dtEnd = datetime.now()
if len(run_ave_time) is 0:
run_ave_time = np.expand_dims(
(dtEnd - dtStart).total_seconds(), axis=0)
else:
run_ave_time = np.concatenate(
(run_ave_time,
np.expand_dims((dtEnd - dtStart).total_seconds(),
axis=0)))
# print(np.mean(run_ave_time))
# print((dtEnd-dtStart).total_seconds())
step += 1
if step < 10:
model.save_ckpt(step)
if step < 1000:
if step % display_step / 10 == 0:
model.save_ckpt(step)
if step % display_step == 0:
model.save_ckpt(step)
except KeyboardInterrupt:
print("Optimization interrupted by user")
break
print("Optimization finished!")
