def reconstruct_changed_activation(device, tokenizer, emb_map, model):
"""Reconstruct the activation for a given sentence after they have been shifted.
Args:
device: The device to use for training the model.
tokenizer: Used to convert between sentences, tokens, and ids.
emb_map: Map containing all the pretrained embeddings of the model.
model: BERT model used for the dreaming process.
"""
data = []
results = {}
params = {}
# Create a folder for this experiment
layer_dir = os.path.join(FLAGS.output_dir, str(FLAGS.layer_id))
folder_helper.make_folder_if_not_exists(layer_dir)
# Actually do the optimization
deep_dream(data, results, params, device, tokenizer, emb_map, model)
# If the top k file is to be written, write it
if FLAGS.write_top_k:
for i in range(len(data)):
top_k_path = os.path.join(layer_dir, 'top_k' + str(i) + '.json')
top_k_file = open(top_k_path, 'w')
json.dump(data[i], top_k_file)
top_k_file.close()
output_helper.write_results(layer_dir, results, params,
'reconstruct_changed')
def get_dream(device, tokenizer, embedding_map, model, base_path):
"""Obtain a dream from the modle given the parameters passed by the user.
Args:
device: The device to use for training the model.
tokenizer: Used to convert between sentences, tokens, and ids.
embedding_map: Map containing all the pretrained embeddings of the model.
model: BERT model used for the dreaming process.
base_path: Location of where to write the results.
"""
results = {}
params = {}
# Actually do the optimization
deep_dream(results, params, device, tokenizer, embedding_map, model)
output_helper.write_results(base_path, results, params, 'dream_mlm')
def reconstruct_activation(device, tokenizer, embedding_map, model, base_path):
"""Reconstruct the activation for a given sentence.
Args:
device: The device to use for training the model.
tokenizer: Used to convert between sentences, tokens, and ids.
embedding_map: Map containing all the pretrained embeddings of the model.
model: BERT model used for the dreaming process.
base_path: Location of where to write the results.
"""
data = []
results = {}
params = {}
# Actually do the optimization
deep_dream(data, results, params, device, tokenizer, embedding_map, model)
# If the top k file is to be written, write it
if FLAGS.write_top_k:
output_helper.write_top_ks(base_path, data, FLAGS.dream_start, params)
output_helper.write_results(base_path, results, params, 'reconstruct')