def config(module_config, params):
c = dict()
c['embdim'] = 300
c['inp_e_dropout'] = 1 / 2
c['inp_w_dropout'] = 0
c['e_add_flags'] = True
c['ptscorer'] = B.mlp_ptscorer
c['mlpsum'] = 'sum'
c['Ddim'] = 1
c['opt'] = 'adam'
c['loss'] = pearsonobj
c['batch_size'] = 64
c['nb_epoch'] = 16
c['fix_layers'] = []
module_config(c)
for p in params:
k, v = p.split('=')
c[k] = eval(v)
ps, h = hash_params(c)
return c, ps, h
def config(module_config, params):
c = dict()
c['embdim'] = 300
c['inp_e_dropout'] = 1/2
c['inp_w_dropout'] = 0
c['e_add_flags'] = True
c['ptscorer'] = B.mlp_ptscorer
c['mlpsum'] = 'sum'
c['Ddim'] = 1
c['opt'] = 'adam'
c['loss'] = ranknet # XXX: binary_crossentropy back?
c['balance_class'] = True # seems essential
c['batch_size'] = 64
c['nb_epoch'] = 16
c['epoch_fract'] = 1/4 # XXX: or much smaller?
c['fix_layers'] = []
module_config(c)
for p in params:
k, v = p.split('=')
c[k] = eval(v)
ps, h = hash_params(c)
return c, ps, h
def config(module_config, params):
c = dict()
c['embdim'] = 300
c['inp_e_dropout'] = 1/2
c['inp_w_dropout'] = 0
c['e_add_flags'] = True
c['ptscorer'] = B.mlp_ptscorer
c['mlpsum'] = 'sum'
c['Ddim'] = 1
c['opt'] = 'adam'
c['loss'] = pearsonobj
c['batch_size'] = 64
c['nb_epoch'] = 16
c['fix_layers'] = []
module_config(c)
for p in params:
k, v = p.split('=')
c[k] = eval(v)
ps, h = hash_params(c)
return c, ps, h
def config(model_config, task_config, params):
c = default_config(model_config, task_config)
for p in params:
k, v = p.split('=')
c[k] = eval(v)
ps, h = hash_params(c)
return c, ps, h
def config(model_config, task_config, params):
c = default_config(model_config,
task_config) # by default, model_config > task_config
for p in params:
keyword, value = p.split('=')
c[keyword] = eval(value)
ps, h = hash_params(c)
return c, ps, h
def config(model_config, task_config, params):
c = default_config(model_config, task_config)
for p in params:
k, v = p.split('=')
c[k] = eval(v)
ps, h = hash_params(c)
# post-ps,h c-munging - only things that are redundant to whatever
# is user-visible
if c['prescoring'] is not None:
prescoring_setup(c, task_config)
return c, ps, h
def config(module_config, params):
c = dict()
c['embdim'] = 300
c['ptscorer'] = B.mlp_ptscorer
c['mlpsum'] = 'sum'
c['Ddim'] = 1
c['loss'] = ranknet
module_config(c)
for p in params:
k, v = p.split('=')
c[k] = eval(v)
ps, h = hash_params(c)
return c, ps, h
def config(model_config, task_config, params):
c = default_config(model_config, task_config)
for p in params:
k, v = p.split('=')
c[k] = eval(v)
ps, h = hash_params(c)
# post-ps,h c-munging - only things that are redundant to whatever
# is user-visible
if c['prescoring'] is not None:
prescoring_setup(c, task_config)
return c, ps, h
def config(module_config, params):
c = dict()
c['embdim'] = 300
c['ptscorer'] = B.mlp_ptscorer
c['mlpsum'] = 'sum'
c['Ddim'] = 1
c['loss'] = ranknet
module_config(c)
for p in params:
k, v = p.split('=')
c[k] = eval(v)
ps, h = hash_params(c)
return c, ps, h
def config(module_config, params):
c = dict()
c['embdim'] = 300
c['inp_e_dropout'] = 3/4
c['inp_w_dropout'] = 0
c['e_add_flags'] = True
c['ptscorer'] = B.dot_ptscorer
c['mlpsum'] = 'sum'
c['Ddim'] = 2
c['loss'] = pearsonobj # ...or 'categorical_crossentropy'
module_config(c)
for p in params:
k, v = p.split('=')
c[k] = eval(v)
ps, h = hash_params(c)
return c, ps, h
def config(module_config, params):
c = dict()
c['embdim'] = 50
c['inp_e_dropout'] = 1 / 2
c['e_add_flags'] = True
c['ptscorer'] = B.mlp_ptscorer
c['mlpsum'] = 'sum'
c['Ddim'] = 1
c['loss'] = 'binary_crossentropy'
c['nb_epoch'] = 2
module_config(c)
for p in params:
k, v = p.split('=')
c[k] = eval(v)
ps, h = hash_params(c)
return c, ps, h
def config(module_config, params):
c = dict()
c['embdim'] = 300
c['inp_e_dropout'] = 1/2
c['e_add_flags'] = True
c['ptscorer'] = B.mlp_ptscorer
c['mlpsum'] = 'sum'
c['Ddim'] = 1
c['batch_size'] = 192
c['loss'] = 'binary_crossentropy'
module_config(c)
for p in params:
k, v = p.split('=')
c[k] = eval(v)
ps, h = hash_params(c)
return c, ps, h
def config(module_config, params):
c = dict()
c['embdim'] = 300
c['inp_e_dropout'] = 3 / 4
c['inp_w_dropout'] = 0
c['e_add_flags'] = True
c['ptscorer'] = B.dot_ptscorer
c['mlpsum'] = 'sum'
c['Ddim'] = 2
c['loss'] = pearsonobj # ...or 'categorical_crossentropy'
c['batch_size'] = 160
c['nb_epoch'] = 32
module_config(c)
for p in params:
k, v = p.split('=')
c[k] = eval(v)
ps, h = hash_params(c)
return c, ps, h
def config(module_config, params):
c = dict()
c['embdim'] = 300
c['inp_e_dropout'] = 1/2
c['inp_w_dropout'] = 0
c['e_add_flags'] = True
c['ptscorer'] = B.mlp_ptscorer
c['mlpsum'] = 'sum'
c['Ddim'] = 1
c['loss'] = ranknet
c['balance_class'] = False
c['batch_size'] = 160
c['nb_epoch'] = 16
c['epoch_fract'] = 1/4
module_config(c)
for p in params:
k, v = p.split('=')
c[k] = eval(v)
ps, h = hash_params(c)
return c, ps, h
def config(module_config, params):
c = dict()
c['embdim'] = 300
c['inp_e_dropout'] = 1/2
c['inp_w_dropout'] = 0
c['e_add_flags'] = True
c['ptscorer'] = B.mlp_ptscorer
c['mlpsum'] = 'sum'
c['Ddim'] = 2
c['loss'] = 'categorical_crossentropy'
c['balance_class'] = False
c['batch_size'] = 160
c['nb_epoch'] = 32
module_config(c)
for p in params:
print(p)
k, v = p.split('=')
c[k] = eval(v)
ps, h = hash_params(c)
return c, ps, h
task.load_vocab(conf['vocabf'])
task.load_data(trainf, valf)
tuneargs = dict()
for p in params:
k, v = p.split('=')
v = eval(v)
if isinstance(v, list) or isinstance(v, dict):
tuneargs[k] = v
rs = RandomSearch(modelname+'_'+taskname+'_log.txt', **tuneargs)
for ps, h, pardict in rs():
# final config for this run
conf, ps, h = config(model_module.config, task.config, [])
for k, v in pardict.items():
conf[k] = v
ps, h = hash_params(conf)
task.set_conf(conf)
runid = '%s-%s-%x' % (taskname, modelname, h)
print()
print(' ...... %s .................... %s' % (runid, ps))
try:
model, res = train_and_eval(runid, model_module.prep_model, task, conf)
rs.report(ps, h, res[1])
except Exception as e:
print(e)
time.sleep(1)
k, v = p.split('=')
v = eval(v)
if isinstance(v, list) or isinstance(v, dict):
tuneargs[k] = v
elif isinstance(v, int) or isinstance(v, string):
tuneargs[k] = [v]
# rs = RandomSearch(modelname+'_'+taskname+'_log.txt', **tuneargs)
# Permutation does not support dict argument like "cdim"
rs = PermutationSearch(modelname + '_' + taskname + '_log.txt', **tuneargs)
for ps, h, pardict in rs():
# final config for this run
conf, ps, h = config(model_module.config, task.config, [])
for k, v in pardict.items():
conf[k] = v
ps, h = hash_params(conf)
task.set_conf(conf)
runid = '%s-%s-%x' % (taskname, modelname, h)
print()
print(' ...... %s .................... %s' % (runid, ps))
try:
model, res = train_and_eval(runid, model_module.prep_model, task,
conf)
rs.report(ps, h, res[1])
except Exception as e:
print(e)
time.sleep(1)
