class experiment:
'''experiment variables'''
# None at the moment
def __init__(self, name="Uninitialized Experiment"):
self.name = name
self.initialized = False
self.trained = False
self.num_trains = 0
self.trained = False
self.exp = {} # this is our experiment meta-data. Dictionary with lots of parameters
self.filepath = None
self.W = None
def read(self, filepath, loadW=False):
self.filepath = filepath
if self.filepath == None:
raise Exception("No filepath initialized for this experiment")
expfile = open(filepath, 'rb')
#############################################
# Read in the experiment file
#############################################
# swallow everything in memory
block = expfile.read()
lines = block.split("\n")
expfile.close()
# free our in-core file copy asap
del block
# First, set a line index for reading in the input
i = 0
print "Reading in experiment: {}".format(filepath)
# Now read in the experiment line by line and fill in data
while i < len(lines):
if not lines[i].startswith("#") and lines[i] != '\n' and lines[i] != '':
line = lines[i].split(None, 1)
# used to be line = lines[i].split('\t')
try:
self.exp[line[0]] = int(line[1])
except:
self.exp[line[0]] = line[1]
i += 1
try:
if self.exp['train_pct'] > 100 or self.exp['train_pct'] <= 0:
raise Exception("train_pct must be > 0 and <= 100")
except:
self.exp['train_pct'] = 100
# get the dictionary of phase times
self.exp['phase_times'] = json.loads(self.exp['phase_times'])
# get the dictionary of delay times
self.exp['phase_var'] = json.loads(self.exp['phase_var'])
if "gauss_inputs" in self.exp:
self.exp['gauss_inputs'] = util.isTrue(self.exp['gauss_inputs'])
else:
self.exp['gauss_inputs'] = False
for phase in self.exp['phase_times']:
self.exp['phase_times'][phase] = int(self.exp['phase_times'][phase])
for phase in self.exp['phase_var']:
self.exp['phase_var'][phase] = util.isTrue(self.exp['phase_var'][phase])
if not 'name' in self.exp:
if not 'directory' in self.exp:
self.exp['name'] = "default_name"
else:
self.exp['name'] = self.exp['directory'].split('/')[-1]
if not 'directory' in self.exp:
self.exp['directory'] = None
elif not self.exp['directory'].endswith('/'):
self.exp['directory'] += '/'
if not 'train_file' in self.exp:
self.exp['train_file'] = "{}.train".format(self.exp['name'])
if "version" in self.exp:
self.exp['input_layer'] = self.exp["num_locs"] + 4
else:
self.exp["version"] = 1
self.exp['input_layer'] = self.exp['input_side']*self.exp['input_side'] + 4
if loadW == True:
self.exp["W"] = "W"
else:
self.exp["W"] = None
self.rnn = HessianRNN(layers=[self.exp['input_layer'], self.exp['hidden_layer'], self.exp['out_layer']], struc_damping=0.5,
use_GPU=False, debug=False, loadW=loadW, dataPath=self.exp['directory'])
np.set_printoptions(edgeitems = 10)
self.exp['epath'] = self.exp['directory'] + self.exp['name'] + '.exp'
self.initialized = True
def createTrainSet(self):
trainFilepath = "{}{}".format(self.exp['directory'], self.exp['train_file'])
if self.exp['version'] == 1:
if self.exp['type'] == "simple":
pass # ig.simpleGen()
elif self.exp['type'] == "attention":
ig.attenGen(self.exp['input_side'], self.exp['num_locs'], self.exp['phase_times'],
self.exp['phase_var'], self.exp['train_pct'], filepath=self.exp['directory'],
ifile=self.exp['train_file'])
elif self.exp['type'] == "selection":
ig.selGen(self.exp['input_side'], self.exp['num_locs'], self.exp['phase_times'],
self.exp['phase_var'], self.exp['train_pct'], filepath=self.exp['directory'],
ifile=self.exp['train_file'])
elif self.exp['type'] == "combined":
ig.combGen(self.exp['input_side'], self.exp['num_locs'], self.exp['phase_times'],
self.exp['phase_var'], self.exp['train_pct'], filepath=self.exp['directory'],
ifile=self.exp['train_file'])
else:
raise Exception("Experiment type not in list: {}".format(self.exp['type']))
elif self.exp['version'] == 2:
igV2.inputGen(self)
elif self.exp['version'] == 3:
igV2.inputGen(self)
elif self.exp['version'] == 4:
igV2.inputGen(self)
elif self.exp['version'] == 5:
igV2.inputGen(self)
else:
raise Exception("Unkown Exp Version")
def getTrainInputs(self):
return util.readTrials(self.exp['directory'] + self.exp['train_file'])
def getTestInputs(self):
return util.readTrials(self.exp['directory'] + "Unused_Locs.train")
def train(self, loadW=False):
if loadW is True:
if self.exp['W'] is None:
loadW = "{}W".format(self.exp['directory'])
else:
loadW = "{}{}".format(self.exp['directory'], self.exp['W'])
if self.exp['directory'] == None or self.exp['directory'] == '':
trainFilepath = self.exp['train_file']
else:
trainFilepath = "{}{}".format(self.exp['directory'], self.exp['train_file'])
ret = util.readTrials(trainFilepath)
inputs = ret['inputs']
targets = ret['targets']
inputNames = ret["inputNames"]
targetNames = ret["targetNames"]
if not loadW:
isSuccess = train_saccade(self.num_trains, inputNames, targetNames,
inputs, targets, self.rnn)
self.num_trains += 1
if isSuccess != True:
print "Error: Overfitting in batch run"
return False
print "Training Successful \(0_0)/"
# Indicate that we finished at least one training cycle and may now load old weights
self.trained = True
else:
print "Loading weights from previous training"
print "\n\n\n(0_0)\n\n\n"
isSuccess = train_saccade(self.num_trains, inputNames, targetNames,
inputs, targets, self.rnn, load_weights=loadW)
self.num_trains += 1
if isSuccess != True:
print "Error: Overfitting in batch run"
return False
print "Training Successful \(0_0)/"
return True
def test(self, W=None):
raise Exception("Function not implemented yet")
def testError(self):
"""Compute RMS error of the testing set"""
testFilepath = "{}{}".format(self.exp['directory'], "Unused_Locs.train")
ret = util.readTrials(testFilepath)
return self.rnn.error(inputs=ret['inputs'], targets=ret['targets'])
def trainError(self):
"""Compute RMS error of the training set"""
trainFilepath = "{}{}".format(self.exp['directory'], self.exp['train_file'])
ret = util.readTrials(trainFilepath)
return self.rnn.error(inputs=ret['inputs'], targets=ret['targets'])
def __repr__(self):
return "Experiment Object, Name: {}".format(self.name)
def __str__(self):
return "\nExperiment Data: \n\n{}\n".format(self.exp)
def read(self, filepath, loadW=False):
self.filepath = filepath
if self.filepath == None:
raise Exception("No filepath initialized for this experiment")
expfile = open(filepath, 'rb')
#############################################
# Read in the experiment file
#############################################
# swallow everything in memory
block = expfile.read()
lines = block.split("\n")
expfile.close()
# free our in-core file copy asap
del block
# First, set a line index for reading in the input
i = 0
print "Reading in experiment: {}".format(filepath)
# Now read in the experiment line by line and fill in data
while i < len(lines):
if not lines[i].startswith("#") and lines[i] != '\n' and lines[i] != '':
line = lines[i].split(None, 1)
# used to be line = lines[i].split('\t')
try:
self.exp[line[0]] = int(line[1])
except:
self.exp[line[0]] = line[1]
i += 1
try:
if self.exp['train_pct'] > 100 or self.exp['train_pct'] <= 0:
raise Exception("train_pct must be > 0 and <= 100")
except:
self.exp['train_pct'] = 100
# get the dictionary of phase times
self.exp['phase_times'] = json.loads(self.exp['phase_times'])
# get the dictionary of delay times
self.exp['phase_var'] = json.loads(self.exp['phase_var'])
if "gauss_inputs" in self.exp:
self.exp['gauss_inputs'] = util.isTrue(self.exp['gauss_inputs'])
else:
self.exp['gauss_inputs'] = False
for phase in self.exp['phase_times']:
self.exp['phase_times'][phase] = int(self.exp['phase_times'][phase])
for phase in self.exp['phase_var']:
self.exp['phase_var'][phase] = util.isTrue(self.exp['phase_var'][phase])
if not 'name' in self.exp:
if not 'directory' in self.exp:
self.exp['name'] = "default_name"
else:
self.exp['name'] = self.exp['directory'].split('/')[-1]
if not 'directory' in self.exp:
self.exp['directory'] = None
elif not self.exp['directory'].endswith('/'):
self.exp['directory'] += '/'
if not 'train_file' in self.exp:
self.exp['train_file'] = "{}.train".format(self.exp['name'])
if "version" in self.exp:
self.exp['input_layer'] = self.exp["num_locs"] + 4
else:
self.exp["version"] = 1
self.exp['input_layer'] = self.exp['input_side']*self.exp['input_side'] + 4
if loadW == True:
self.exp["W"] = "W"
else:
self.exp["W"] = None
self.rnn = HessianRNN(layers=[self.exp['input_layer'], self.exp['hidden_layer'], self.exp['out_layer']], struc_damping=0.5,
use_GPU=False, debug=False, loadW=loadW, dataPath=self.exp['directory'])
np.set_printoptions(edgeitems = 10)
self.exp['epath'] = self.exp['directory'] + self.exp['name'] + '.exp'
self.initialized = True
for temp in range(trial_time):
row = np.array(lines[index].split(' '))
try:
data = row[0:num_targets].astype(np.float32)
except:
print row
exit()
index += 1
targets[trial][temp] = data
assert(lines[index] == 'end') # verify where we are
index += 1
########################################################
# create a recurrent NN
rnn = HessianRNN(layers=[lay[0], lay[1], lay[3]], struc_damping=0.5,
use_GPU=False, debug=False, loadW = wFileString)
#create an output file for the activations. I don't actually use this yet.
# out2 = open("activations", "wb+")
# shows activations for each time step for each input
index = 1
for i, t in zip(inputs, targets):
j = np.asarray(i)
# print >> out2, "\nTrail Num: {}".format(index)
index += 1
hidden = rnn.forward(j, rnn.W)[-2]
out = rnn.forward(j, rnn.W)[-1]
# for x in range(len(output)):
# for y in range(len(output[x][0])):
