def get_predictions(config):
net_name = config[0]
hidden_activation, output_activation = activation_map[net_name]
# Pull datasets
b_predict_actions = "treatment_rnn" in net_name
b_use_actions_only = "rnn_action_inputs_only" in net_name
# Extract only relevant trajs and shift data
training_processed = core.get_processed_data(
training_data, b_predict_actions, b_use_actions_only,
b_use_predicted_confounders, b_use_oracle_confounders, b_remove_x1)
validation_processed = core.get_processed_data(
validation_data, b_predict_actions, b_use_actions_only,
b_use_predicted_confounders, b_use_oracle_confounders, b_remove_x1)
num_features = training_processed['scaled_inputs'].shape[
-1] # 4 if not b_use_actions_only else 3
num_outputs = training_processed['scaled_outputs'].shape[
-1] # 1 if not b_predict_actions else 3 # 5
# Unpack remaining variables
dropout_rate = config[1]
memory_multiplier = config[2] / num_features
num_epochs = config[3]
minibatch_size = config[4]
learning_rate = config[5]
max_norm = config[6]
model_folder = os.path.join(MODEL_ROOT, net_name)
means, outputs, _, _ = test(training_processed, validation_processed,
training_processed, tf_config, net_name,
expt_name, dropout_rate, num_features,
num_outputs, memory_multiplier, num_epochs,
minibatch_size, learning_rate, max_norm,
hidden_activation, output_activation,
model_folder)
return means, outputs
def rnn_fit(dataset_map,
networks_to_train,
MODEL_ROOT,
b_use_predicted_confounders,
b_use_oracle_confounders=False,
b_remove_x1=False):
logging.basicConfig(format='%(levelname)s:%(message)s', level=logging.INFO)
# Get the correct networks to train
if networks_to_train == "propensity_networks":
logging.info("Training propensity networks")
net_names = ['treatment_rnn_action_inputs_only', 'treatment_rnn']
elif networks_to_train == "encoder":
logging.info("Training R-MSN encoder")
net_names = ["rnn_propensity_weighted"]
elif networks_to_train == "user_defined":
logging.info("Training user defined network")
raise NotImplementedError("Specify network to use!")
else:
raise ValueError("Unrecognised network type")
logging.info("Running hyperparameter optimisation")
# Experiment name
expt_name = "treatment_effects"
# Possible networks to use along with their activation functions
activation_map = {
'rnn_propensity_weighted': ("elu", 'linear'),
'rnn_propensity_weighted_logistic': ("elu", 'linear'),
'rnn_model': ("elu", 'linear'),
'treatment_rnn': ("tanh", 'sigmoid'),
'treatment_rnn_action_inputs_only': ("tanh", 'sigmoid')
}
# Setup tensorflow
tf_device = 'gpu'
if tf_device == "cpu":
config = tf.ConfigProto(log_device_placement=False,
device_count={'GPU': 0})
else:
config = tf.ConfigProto(log_device_placement=False,
device_count={'GPU': 1})
config.gpu_options.allow_growth = True
training_data = dataset_map['training_data']
validation_data = dataset_map['validation_data']
test_data = dataset_map['test_data']
# Start Running hyperparam opt
opt_params = {}
for net_name in net_names:
# Re-run hyperparameter optimisation if parameters are not specified, otherwise train with defined params
max_hyperparam_runs = 3 if net_name not in specifications else 1
# Pull datasets
b_predict_actions = "treatment_rnn" in net_name
use_truncated_bptt = net_name != "rnn_model_bptt" # whether to train with truncated backpropagation through time
b_propensity_weight = "rnn_propensity_weighted" in net_name
b_use_actions_only = "rnn_action_inputs_only" in net_name
# Extract only relevant trajs and shift data
training_processed = core.get_processed_data(
training_data, b_predict_actions, b_use_actions_only,
b_use_predicted_confounders, b_use_oracle_confounders, b_remove_x1)
validation_processed = core.get_processed_data(
validation_data, b_predict_actions, b_use_actions_only,
b_use_predicted_confounders, b_use_oracle_confounders, b_remove_x1)
test_processed = core.get_processed_data(test_data, b_predict_actions,
b_use_actions_only,
b_use_predicted_confounders,
b_use_oracle_confounders,
b_remove_x1)
num_features = training_processed['scaled_inputs'].shape[-1]
num_outputs = training_processed['scaled_outputs'].shape[-1]
# Load propensity weights if they exist
if b_propensity_weight:
if net_name == 'rnn_propensity_weighted_den_only':
# use un-stabilised IPTWs generated by propensity networks
propensity_weights = np.load(
os.path.join(MODEL_ROOT, "propensity_scores_den_only.npy"))
elif net_name == "rnn_propensity_weighted_logistic":
# Use logistic regression weights
propensity_weights = np.load(
os.path.join(MODEL_ROOT, "propensity_scores.npy"))
tmp = np.load(
os.path.join(MODEL_ROOT, "propensity_scores_logistic.npy"))
propensity_weights = tmp[:propensity_weights.shape[0], :, :]
else:
# use stabilised IPTWs generated by propensity networks
propensity_weights = np.load(
os.path.join(MODEL_ROOT, "propensity_scores.npy"))
logging.info("Net name = {}. Mean-adjusting!".format(net_name))
propensity_weights /= propensity_weights.mean()
training_processed['propensity_weights'] = propensity_weights
# Start hyperparamter optimisation
hyperparam_count = 0
while True:
if net_name not in specifications:
dropout_rate = np.random.choice([0.1, 0.2, 0.3, 0.4, 0.5])
memory_multiplier = np.random.choice([0.5, 1, 2, 3, 4])
num_epochs = 100
minibatch_size = np.random.choice([64, 128, 256])
learning_rate = np.random.choice([0.01, 0.005, 0.001
]) #([0.01, 0.001, 0.0001])
max_norm = np.random.choice([0.5, 1.0, 2.0, 4.0])
hidden_activation, output_activation = activation_map[net_name]
else:
spec = specifications[net_name]
logging.info("Using specifications for {}: {}".format(
net_name, spec))
dropout_rate = spec[0]
memory_multiplier = spec[1]
num_epochs = spec[2]
minibatch_size = spec[3]
learning_rate = spec[4]
max_norm = spec[5]
hidden_activation, output_activation = activation_map[net_name]
model_folder = os.path.join(MODEL_ROOT, net_name)
hyperparam_opt = train(
net_name, expt_name, training_processed, validation_processed,
test_processed, dropout_rate, memory_multiplier, num_epochs,
minibatch_size, learning_rate, max_norm, use_truncated_bptt,
num_features, num_outputs, model_folder, hidden_activation,
output_activation, config,
"hyperparam opt: {} of {}".format(hyperparam_count,
max_hyperparam_runs))
hyperparam_count = len(hyperparam_opt.columns)
if hyperparam_count >= max_hyperparam_runs:
opt_params[net_name] = hyperparam_opt.T
break
logging.info("Done")
logging.info(hyperparam_opt.T)
# Flag optimal params
logging.info(opt_params)
def rnn_test(dataset_map, MODEL_ROOT, b_use_predicted_confounders, b_use_oracle_confounders=False,
b_remove_x1=False):
logging.basicConfig(format='%(levelname)s:%(message)s', level=logging.INFO)
# Setup tensorflow
tf_device = 'gpu'
if tf_device == "cpu":
tf_config = tf.ConfigProto(log_device_placement=False, device_count={'GPU': 0})
else:
tf_config = tf.ConfigProto(log_device_placement=False, device_count={'GPU': 1})
tf_config.gpu_options.allow_growth = True
configs = [
load_optimal_parameters('rnn_propensity_weighted',
expt_name, MODEL_ROOT,
add_net_name=True)
]
# Config
activation_map = {'rnn_propensity_weighted': ("elu", 'linear'),
'rnn_propensity_weighted_binary': ("elu", 'linear'),
'rnn_propensity_weighted_logistic': ("elu", 'linear'),
'rnn_model': ("elu", 'linear'),
'treatment_rnn': ("tanh", 'sigmoid'),
'treatment_rnn_actions_only': ("tanh", 'sigmoid')}
projection_map = {}
mse_by_followup = {}
for config in configs:
net_name = config[0]
projection_map[net_name] = {}
training_data = dataset_map['training_data']
validation_data = dataset_map['validation_data']
test_data = dataset_map['test_data']
# scaling_data = pickle_map['scaling_data'] # use scaling data from above
# Setup some params
b_predict_actions = "treatment_rnn" in net_name
b_propensity_weight = "rnn_propensity_weighted" in net_name
b_use_actions_only = "treatment_rnn_action_inputs_only" in net_name
# In[*]: Compute base MSEs
# Extract only relevant trajs and shift data
training_processed = core.get_processed_data(training_data, b_predict_actions,
b_use_actions_only,
b_use_predicted_confounders, b_use_oracle_confounders, b_remove_x1)
validation_processed = core.get_processed_data(validation_data, b_predict_actions,
b_use_actions_only, b_use_predicted_confounders,
b_use_oracle_confounders, b_remove_x1)
test_processed = core.get_processed_data(test_data, b_predict_actions,
b_use_actions_only, b_use_predicted_confounders,
b_use_oracle_confounders, b_remove_x1)
num_features = training_processed['scaled_inputs'].shape[-1] # 4 if not b_use_actions_only else 3
num_outputs = training_processed['scaled_outputs'].shape[-1] # 1 if not b_predict_actions else 3 # 5
# Pull remaining params
dropout_rate = config[1]
memory_multiplier = config[2] / num_features
num_epochs = config[3]
minibatch_size = config[4]
learning_rate = config[5]
max_norm = config[6]
backprop_length = 60 # we've fixed this
hidden_activation = activation_map[net_name][0]
output_activation = activation_map[net_name][1]
# Run tests
model_folder = os.path.join(MODEL_ROOT, net_name)
means, output, mse, test_states \
= test(training_processed, validation_processed, test_processed, tf_config,
net_name, expt_name, dropout_rate, num_features, num_outputs,
memory_multiplier, num_epochs, minibatch_size, learning_rate, max_norm,
hidden_activation, output_activation, model_folder,
b_use_state_initialisation=False, b_dump_all_states=True)
active_entries = test_processed['active_entries']
def get_mse_at_follow_up_time(mean, output, active_entires):
mses = np.sum(np.sum((mean - output) ** 2 * active_entires, axis=-1), axis=0) \
/ active_entires.sum(axis=0).sum(axis=-1)
return pd.Series(mses, index=[idx for idx in range(len(mses))], name=net_name)
mse = get_mse_at_follow_up_time(means, output, active_entries)
projection_map[net_name] = mse
mse.to_csv(os.path.join(MODEL_ROOT, "results_mse.csv"))
return mse