def train( i_exp = 0, mode = 'train', model_param = None, batch_size = None, sample_frac = None ):
# Memory before the training
mem = psutil.virtual_memory()
print( 'Memory:', mem[2] )
# Experiment parameters
e = exp.EXPERIMENTS[ i_exp ]
# Store path
overwrite = False
if sample_frac == None:
outpath = exp.OUTPATH + e['NAME'] + '/'
else:
outpath = exp.OUTPATH + e['NAME'] + '/sample_frac_best' + str(sample_frac) + '/'
create_dir(outpath, overwrite)
# Load the data
inpath = exp.INPATH + 'input_NOMINAL' + '.h5'
if sample_frac == None:
actionshist, codes, sites = fit_handler.load_data(inpath)
else:
actionshist, codes, sites = fit_handler.load_data(inpath, load_labels = False, sample = True, sample_frac = sample_frac)
# Setup the fit handler
handler = fit_handler.FitHandler( exp.MODEL, codes, sites, pruning_mode = e['PRUNE'],
callback_args = e['CALLBACK'], train_on_batch = False, path = outpath, verbose=2)
# Get the count matrix
X,y = handler.count_matrix(actionshist)
mem = psutil.virtual_memory()
print( 'Memory:', mem[2] )
if model_param is None:
model_param = e['HYPERPARAM']
if batch_size is None:
batch_size = exp.BATCH_SIZE
if mode == 'optimize':
score = handler.find_optimal_parameters( exp.SKOPT_DIM, model_param, X, y, cv=exp.CV, kfold_splits=exp.FOLDS,
num_calls=exp.SKOPTCALLS, max_epochs=exp.MAX_EPOCHS, batch_size=batch_size)
elif mode == 'cv':
score = handler.kfold_val( X, y, model_param = model_param, kfold_splits = exp.FOLDS,
max_epochs = exp.MAX_EPOCHS, batch_size = batch_size)
else:
score = handler.run_training(X, y, batch_size = batch_size, max_epochs = exp.MAX_EPOCHS,
model_param = model_param)
return score
def train( i_exp = 0, model_param = None, sample_frac = None ):
# Memory before the training
mem = psutil.virtual_memory()
print( 'Memory:', mem[2] )
# Experiment parameters
e = exp.experiments[ i_exp ]
# Load the data
if 'NOMINAL' in e.name:
path = e.inpath + 'input_' + 'NOMINAL' + '.h5'
e.nlp_param['embedding_matrix_path'] = e.inpath + 'embedding_matrix_' + 'NOMINAL' + '.npy'
elif 'AVG' in e.name:
path = e.inpath + 'input_' + 'NOMINAL' + '.h5'
e.nlp_param['embedding_matrix_path'] = e.inpath + 'embedding_matrix_' + 'NOMINAL' + '.npy'
else:
path = e.inpath + 'input_' + 'VAR_DIM' + '.h5'
e.nlp_param['embedding_matrix_path'] = e.inpath + 'embedding_matrix_' + 'VAR_DIM' + '.npy'
if sample_frac == None:
actionshist, codes, sites = fit_handler.load_data(path)
else:
actionshist, codes, sites = fit_handler.load_data(path, sample = True, sample_frac = sample_frac)
# Setup the fit handler
handler = fit_handler.FitHandler( e.model, codes, sites, e.max_words,
e.gen_param, pruning_mode = e.pruning,
model_args = e.nlp_param, callback_args = e.callback,
train_on_batch = e.train_on_batch, verbose=1 )
if model_param is None:
model_param = e.hyperparam
#score = handler.run_training(actionshist, batch_size = exp.BATCH_SIZE, max_epochs = exp.MAX_EPOCHS,
# model_param = e['HYPERPARAM'])
cvscores = handler.kfold_val( actionshist, model_param = model_param, kfold_splits = e.folds,
max_epochs = e.max_epochs, batch_size = e.batch_size)
def evaluate(o, fold=None):
hash_value = o.pop('hash')
i_exp = int(o.pop('i_exp'))
# Memory before the training
mem = psutil.virtual_memory()
print('Memory:', mem[2])
# Experiment parameters
e = exp.experiments[i_exp]
out_path = e.outpath + e.name + '/'
# Load the data
if 'NOMINAL' in e.name:
path = e.inpath + 'input_' + 'NOMINAL' + '.h5'
e.nlp_param[
'embedding_matrix_path'] = e.inpath + 'embedding_matrix_' + 'NOMINAL' + '.npy'
elif 'AVG' in e.name:
path = e.inpath + 'input_' + 'NOMINAL' + '.h5'
e.nlp_param[
'embedding_matrix_path'] = e.inpath + 'embedding_matrix_' + 'NOMINAL' + '.npy'
else:
path = e.inpath + 'input_' + 'VAR_DIM' + '.h5'
e.nlp_param[
'embedding_matrix_path'] = e.inpath + 'embedding_matrix_' + 'VAR_DIM' + '.npy'
actionshist, codes, sites = fit_handler.load_data(path)
# Setup the fit handler
handler = fit_handler.FitHandler(e.model,
codes,
sites,
e.max_words,
e.gen_param,
pruning_mode=e.pruning,
model_args=e.nlp_param,
callback_args=e.callback,
train_on_batch=e.train_on_batch,
verbose=2)
# Initial hyper parameters
model_param = e.hyperparam
# Overwrite with bayesian suggestion
for name, value in o.items():
model_param[name] = value
"""
score = handler.run_training(actionshist, batch_size = exp.BATCH_SIZE, max_epochs = exp.MAX_EPOCHS,
model_param = model_param)
#value = -1 * score
#print( value )
"""
cvscores = handler.kfold_val(actionshist,
model_param=model_param,
kfold_splits=e.folds,
max_epochs=e.max_epochs,
batch_size=e.batch_size)
value = -1 * np.mean(cvscores)
std_dv = np.std(cvscores)
res = {
'result': value,
'params': o,
'annotate': 'a free comment',
'std_dv': std_dv
}
print(res)
if fold is not None:
res['fold'] = fold
out = out_path + hash_value
dest = '%s.json' % out if fold is None else '%s_f%d.json' % (out, fold)
open(dest, 'w').write(json.dumps(res))
def evaluate(o, fold=None):
hash_value = o.pop('hash')
i_exp = int(o.pop('i_exp'))
# Memory before the training
mem = psutil.virtual_memory()
print('Memory:', mem[2])
# Experiment parameters
e = exp.EXPERIMENTS[i_exp]
out_path = exp.OUTPATH + e['NAME'] + '/'
# Load the data
if 'VAR_LOW' in e['NAME']:
path = exp.INPATH + 'input_' + 'VAR_LOW' + '.h5'
e['NLP_PARAM'][
'embedding_matrix_path'] = exp.INPATH + 'embedding_matrix_' + 'VAR_LOW' + '.npy'
elif 'VAR_DIM' in e['NAME']:
path = exp.INPATH + 'input_' + 'VAR_DIM' + '.h5'
e['NLP_PARAM'][
'embedding_matrix_path'] = exp.INPATH + 'embedding_matrix_' + 'VAR_DIM' + '.npy'
elif 'AVG' in e['NAME']:
path = exp.INPATH + 'input_' + 'VAR_DIM' + '.h5'
e['NLP_PARAM'][
'embedding_matrix_path'] = exp.INPATH + 'embedding_matrix_' + 'VAR_DIM' + '.npy'
else:
path = exp.INPATH + 'input_' + 'NOMINAL' + '.h5'
e['NLP_PARAM'][
'embedding_matrix_path'] = exp.INPATH + 'embedding_matrix_' + 'NOMINAL' + '.npy'
actionshist, codes, sites = fit_handler.load_data(
path,
msg_only=exp.MSG_ONLY,
sample=exp.SAMPLE,
sample_fact=exp.SAMPLE_FACT)
# Setup the fit handler
handler = fit_handler.FitHandler(exp.MODEL,
codes,
sites,
exp.MAX_WORDS,
exp.GEN_PARAM,
pruning_mode=exp.PRUNING,
model_args=e['NLP_PARAM'],
callback_args=e['CALLBACK'],
train_on_batch=exp.TRAIN_ON_BATCH,
verbose=2)
# Initial hyper parameters
model_param = e['HYPERPARAM']
# Overwrite with bayesian suggestion
for name, value in o.items():
model_param[name] = value
"""
score = handler.run_training(actionshist, batch_size = exp.BATCH_SIZE, max_epochs = exp.MAX_EPOCHS,
model_param = model_param)
"""
cvscores = handler.kfold_val(actionshist,
model_param=model_param,
kfold_splits=exp.FOLDS,
max_epochs=exp.MAX_EPOCHS,
batch_size=exp.BATCH_SIZE)
#value = -1 * score
#print( value )
value = -1 * np.mean(cvscores)
std_dv = np.std(cvscores)
#X = (o['learning_rate'], o['learning_rate']*2)
#value = dummy_func( X , fold = fold)
#dummy_func( X , fold = fold)
res = {
'result': value,
'params': o,
'annotate': 'a free comment',
'std_dv': std_dv
}
print(res)
if fold is not None:
res['fold'] = fold
out = out_path + hash_value
dest = '%s.json' % out if fold is None else '%s_f%d.json' % (out, fold)
open(dest, 'w').write(json.dumps(res))
def train(i_exp=0, model_param=None):
# Memory before the training
mem = psutil.virtual_memory()
print('Memory:', mem[2])
# Experiment parameters
e = exp.EXPERIMENTS[i_exp]
# Load the data
if 'VAR_LOW' in e['NAME']:
path = exp.INPATH + 'input_' + 'VAR_LOW' + '.h5'
e['NLP_PARAM'][
'embedding_matrix_path'] = exp.INPATH + 'embedding_matrix_' + 'VAR_LOW' + '.npy'
elif 'VAR_DIM' in e['NAME']:
path = exp.INPATH + 'input_' + 'VAR_DIM' + '.h5'
e['NLP_PARAM'][
'embedding_matrix_path'] = exp.INPATH + 'embedding_matrix_' + 'VAR_DIM' + '.npy'
elif 'AVG' in e['NAME']:
path = exp.INPATH + 'input_' + 'VAR_DIM' + '.h5'
e['NLP_PARAM'][
'embedding_matrix_path'] = exp.INPATH + 'embedding_matrix_' + 'VAR_DIM' + '.npy'
else:
path = exp.INPATH + 'input_' + 'NOMINAL' + '.h5'
e['NLP_PARAM'][
'embedding_matrix_path'] = exp.INPATH + 'embedding_matrix_' + 'NOMINAL' + '.npy'
actionshist, codes, sites = fit_handler.load_data(
path,
msg_only=exp.MSG_ONLY,
sample=exp.SAMPLE,
sample_fact=exp.SAMPLE_FACT)
# Setup the fit handler
handler = fit_handler.FitHandler(exp.MODEL,
codes,
sites,
exp.MAX_WORDS,
exp.GEN_PARAM,
pruning_mode=exp.PRUNING,
model_args=e['NLP_PARAM'],
callback_args=e['CALLBACK'],
train_on_batch=exp.TRAIN_ON_BATCH)
if model_param is None:
#score = handler.run_training(actionshist, batch_size = exp.BATCH_SIZE, max_epochs = exp.MAX_EPOCHS,
# model_param = e['HYPERPARAM'])
cvscores = handler.kfold_val(actionshist,
model_param=e['HYPERPARAM'],
kfold_splits=exp.FOLDS,
max_epochs=exp.MAX_EPOCHS,
batch_size=exp.BATCH_SIZE)
else:
#score = handler.run_training(actionshist, batch_size = exp.BATCH_SIZE, max_epochs = exp.MAX_EPOCHS,
# model_param = model_param)
cvscores = handler.kfold_val(actionshist,
model_param=model_param,
kfold_splits=exp.FOLDS,
max_epochs=exp.MAX_EPOCHS,
batch_size=exp.BATCH_SIZE)
return score