}
sherpa_algo = sherpa.algorithms.GPyOpt(max_num_trials=80)
study = sherpa.Study(parameters=params,
algorithm=sherpa_algo,
lower_is_better=True)
for trial in study:
trial_id = trial.id
savename = '{}/model/model.{}.h5'.format(algorithm, trial_id)
pars = deepcopy(trial.parameters)
for k, v in baseline_params.items():
pars[k] = v
pars['batches_per_epoch'] = set_batches_per_epoch(pars['batch_size'])
model = init_model(pars)
adam = init_adam(pars)
datagen = init_datagen(pars)
model.compile(optimizer=adam, loss='categorical_crossentropy')
# Train but reserve the last little bit of data for
# a final testing set.
t_loss, v_loss = train(model,
pars,
X_train,
Y_train,
X_test[:6000],
Y_test[:6000],
patience=10,
savename=savename,
datagen=datagen)
'beta2':0.998234
}
baseline_params['batches_per_epoch'] = set_batches_per_epoch(baseline_params['batch_size'])
sherpa_algo = sherpa.algorithms.GPyOpt(max_num_trials=200)
study = sherpa.Study(parameters=params, algorithm=sherpa_algo, lower_is_better=True)
for trial in study:
trial_id = trial.id
savename = '{}/model/model.{}.h5'.format(algorithm, trial_id)
pars = deepcopy(trial.parameters)
for k,v in baseline_params.items():
pars[k] = v
model = init_model(baseline_params)
adam = init_adam(baseline_params)
datagen = init_datagen(pars)
model.compile(optimizer=adam, loss='categorical_crossentropy')
# Train but reserve the last little bit of data for
# a final testing set.
t_loss, v_loss = train(model, baseline_params, X_train, Y_train, X_test[:6000], Y_test[:6000],
patience=10, savename=savename, datagen=datagen)
# Log
study.add_observation(trial=trial, iteration=0, objective=min(v_loss))
# Save if it is not saved
if not os.path.exists(savename):
model.save(savename)
'depth': 3,
'dense_neurons': 512,
'init_filters': 20,
'use_batchnorm': True,
'dropout': 0.4,
'batch_size': 128,
'max_epochs': 5,
'learning_rate': 0.008,
'beta1': 0.5,
'beta2': 0.999
}
params['batches_per_epoch'] = set_batches_per_epoch(params['batch_size'])
print(params)
model = init_model(params)
adam = init_adam(params)
model.compile(optimizer=adam, loss='categorical_crossentropy')
print(model.summary())
# Train but reserve the last little bit of data for
# a final testing set.
t_loss, v_loss = train(model,
params,
X_train,
Y_train,
X_test[:6000],
Y_test[:6000],
patience=10,
savename=savename)
# Save if it is not saved