I'm currently using HParams to instigate a grid search hyperparameter optimisation session, which works fine, and is outputting logs to my tensorboard HParams plugin, and I can see the various different runs and the Parallel Co-Ordinates view. The code is structured like so, although it might not be necessary to review it for this question:
def hparam_wrap(args, n_classes, train_dataset, val_dataset, tokenizer):
log_date_subfolder = time.strftime("%Y%m%d-%H%M%S")
hparams_dict={
'HP_EMBEDDING_NODES': hp.HParam('embedding_nodes', hp.Discrete([200,300])),
'HP_LSTM_NODES': hp.HParam('lstm_nodes', hp.Discrete([200,300])),
'HP_TIMEDIST_NODES': hp.HParam('timedist_nodes', hp.Discrete([200,300])),
'HP_NUM_DENSE_LAYERS': hp.HParam('num_dense_layers', hp.Discrete([3,4, 5])),
'HP_DENSE_NODES': hp.HParam('dense_nodes', hp.Discrete([300,400, 500])),
'HP_LEARNING_RATE': hp.HParam('learning_rate', hp.Discrete([0.001, 0.0001, 0.00001])),
'HP_DROPOUT': hp.HParam('dropout', hp.Discrete([0.3, 0.4,0.5, 0.6])),
'HP_BATCH_SIZE': hp.HParam('batch_size', hp.Discrete([96]))
}
session_num = 0
for en in hparams_dict['HP_EMBEDDING_NODES'].domain.values:
for ln in hparams_dict['HP_LSTM_NODES'].domain.values:
for td in hparams_dict['HP_TIMEDIST_NODES'].domain.values:
for dl in hparams_dict['HP_NUM_DENSE_LAYERS'].domain.values:
for dn in hparams_dict['HP_DENSE_NODES'].domain.values:
for lr in hparams_dict['HP_LEARNING_RATE'].domain.values:
for do in hparams_dict['HP_DROPOUT'].domain.values:
for bs in hparams_dict['HP_BATCH_SIZE'].domain.values:
hparams ={
'HP_NUM_DENSE_LAYERS': dl,
'HP_LEARNING_RATE': lr,
'HP_DROPOUT': do,
'HP_DENSE_NODES': dn,
'HP_BATCH_SIZE': bs,
'HP_EMBEDDING_NODES': en,
'HP_LSTM_NODES': ln,
'HP_TIMEDIST_NODES': td
}
run_name = "run-%d" % session_num
print('--- Starting trial: %s' % run_name)
print({h: hparams[h] for h in hparams})
log_dir = os.path.join('s3://sn-classification', args.type, 'Logs', args.country,
args.subfolder, 'HParams', log_date_subfolder)
run_hparam(log_dir, hparams, hparams_dict, args, n_classes, train_dataset,
val_dataset, tokenizer)
session_num += 1
def run_hparam(log_dir, hparams, hparams_dict, args, n_classes, train_dataset, val_dataset, tokenizer):
with tf.summary.create_file_writer(log_dir).as_default():
hp.hparams_config(
hparams=list(hparams_dict.values()),
metrics=[hp.Metric('val_top_k_categorical_accuracy', display_name='TopK_Val_Accuracy'),hp.Metric('val_loss', display_name='val_loss')]
)
# hp.hparams(hparams) # record the values used in this trial
hp.hparams({hparams_dict[h]: hparams[h] for h in hparams_dict.keys()})
history = train(args,n_classes,hparams,train_dataset, val_dataset, tokenizer)
tf.summary.scalar('val_top_k_categorical_accuracy', history['val_top_k_categorical_accuracy'][-1], step=1)
tf.summary.scalar('val_loss', history['val_loss'][-1], step=1)
I've done a lot of googling, but I'm still unsure how to go about implementing a more efficient optimisation session, such as Bayesian Optimisation in order to find the optimum model in a faster way. All I want to know is - is it possible to do Bayesian Optimisation within HParams, or do I need to use a different package like Weights and Biases? If it's possible, any advice on where to find an example of such an implementation would be very helpful.
This is a long-time open feature request and is unfortunately still not currently implemented with the HPARAMS section but Keras-tuner will allow you to log the results of each run. Encoding the hyperparameter values into these directory names could be a quick and dirty workaround. For the benefit of future readers I have provided a guide for using TensorBoard for Bayesian optimisation at the end of this answer.
I might add that TensorBoard visualisation is useful for using grid or random search to inform a developer's manual tuning intuitions, but since Bayesian optimisation is a self-contained black-box optimiser, you should be able to let it run without the optimisation itself being affected by the lack of visualisations -- though I agree this would still be a nice feature to have.
To implement Bayesian optimisation in TensorFlow and log the losses for each run, I provide the following for future readers:
First define a HyperParameters object hp.
from kerastuner.engine.hyperparameters import HyperParameters
hp = HyperParameters()
Write a model_builder function with argument hp, incorporating the hyperparameters into the model using hp.get('name'). Define a Keras-tuner BayesianOptimization tuner.
import kerastuner as kt
tuner = kt.BayesianOptimization(model_builder,
hyperparameters = hp,
max_trials = 20,
objective = 'val_loss')
Include tf.keras.callbacks.TensorBoard(cb_dir) in your callbacks to log the loss plots of each run of the BaysianOptimiser in directory cb_dir. This includes the scalar plots against epoch but not the HPARAMS section. You may wish to name these run files such that they list the hyperparameters.
tuner.search(inputs, prices,
validation_split = 0.2,
batch_size = 32,
callbacks = [tf.keras.callbacks.TensorBoard(cb_dir)],
epochs = 30)
Access the dictionaries for the top n scoring hyperparameter combinations via
ith_best_hp_dict = tuner.get_best_hyperparameters(num_trials = n)[i]