def run_report_results(args, probe, dataset, model, loss, reporter, regimen):
"""
Reports results from a structural probe according to args.
By default, does so only for dev set.
Requires a simple code change to run on the test set.
"""
probe_params_path = os.path.join(args['reporting']['root'],
args['probe']['params_path'])
probe.load_state_dict(torch.load(probe_params_path))
probe.eval()
dev_dataloader = dataset.get_dev_dataloader()
dev_predictions = regimen.predict(probe, model, dev_dataloader)
dev_report = reporter(dev_predictions, dev_dataloader, 'dev', probe=probe)
#train_dataloader = dataset.get_train_dataloader(shuffle=False)
#train_predictions = regimen.predict(probe, model, train_dataloader)
#reporter(train_predictions, train_dataloader, 'train')
# Uncomment to run on the test set
test_dataloader = dataset.get_test_dataloader()
test_predictions = regimen.predict(probe, model, test_dataloader)
test_report = reporter(test_predictions,
test_dataloader,
'test',
probe=probe)
return dev_report, test_report
def run_report_results(args, probe, dataset, model, loss, reporter, regimen): """ Reports results from a structural probe according to args. By default, does so only for dev set. Requires a simple code change to run on the test set. """ probe_params_path = os.path.join(args['reporting']['root'],args['probe']['params_path']) probe.load_state_dict(torch.load(probe_params_path)) probe.eval() dev_dataloader = dataset.get_dev_dataloader() dev_predictions = regimen.predict(probe, model, dev_dataloader) reporter(dev_predictions, dev_dataloader, 'dev')
def run_report_results(args, probe, dataset, model, loss, reporter, regimen):
probe_params_path = os.path.join(args['reporting']['root'],
args['probe']['params_path'])
dev_dataloader = dataset.get_dev_dataloader()
try:
probe.load_state_dict(torch.load(probe_params_path))
probe.eval()
dev_predictions = regimen.predict(probe, model, dev_dataloader)
except FileNotFoundError:
print("No trained probe found.")
dev_predictions = None
reporter(dev_predictions, probe, model, dev_dataloader, 'dev')
def evaluate_vectors(args, probe, dataset, model, results_dir, output_name):
probe_params_path = os.path.join(results_dir, args['probe']['params_path'])
probe.load_state_dict(torch.load(probe_params_path))
probe.eval()
print(probe.proj)
dataloader = dataset.get_dev_dataloader()
projections = load_projected_representations(probe, model, dataloader)
relations_to_projections = defaultdict(list)
relations_to_sentences = defaultdict(list)
relations_to_idxs = defaultdict(list)
relations_to_words = defaultdict(list)
for projection_batch, (data_batch, label_batch, length_batch,
observation_batch) in zip(projections, dataloader):
for projection, label, length, (observation,
_) in zip(projection_batch,
label_batch, length_batch,
observation_batch):
for idx, word in enumerate(observation.sentence):
if observation.head_indices[idx] == '0':
pass # head word
else:
head_index = int(observation.head_indices[idx])
proj_diff = projection[idx] - projection[head_index - 1]
relation = observation.governance_relations[idx]
relations_to_projections[relation].append(proj_diff)
relations_to_sentences[relation].append(" ".join(
observation.sentence))
relations_to_idxs[relation].append(idx)
relations_to_words[relation].append(word)
relations_to_diffs = {}
all_relations = []
all_sentences = []
all_idxs = []
all_words = []
y_list = []
for relation in relations_to_projections:
diffs = torch.FloatTensor(relations_to_projections[relation])
# compute the SVD
u, s, v = diffs.svd()
average_diff = torch.mean(diffs, 0)
relations_to_diffs[relation] = average_diff
all_relations += relations_to_projections[relation]
all_sentences += relations_to_sentences[relation]
all_idxs += relations_to_idxs[relation]
all_words += relations_to_words[relation]
y_list += [relation] * len(relations_to_projections[relation])
allDiff = torch.FloatTensor(all_relations)
# print(y_list)
sentences_idxs_words = np.array([all_sentences, all_idxs, all_words])
if len(sys.argv) > 2:
np.save(
'/sailhome/ethanchi/structural-probes/relationOutputs/{}.npy'.
format(output_name), allDiff.numpy())
np.save(
'/sailhome/ethanchi/structural-probes/relationOutputs/{}Y.npy'.
format(output_name), np.array(y_list))
np.save(
'/sailhome/ethanchi/structural-probes/relationOutputs/{}-data.npy'.
format(output_name), sentences_idxs_words)
allDiff = torch.mean(allDiff, 0)
cos = torch.nn.CosineSimilarity(dim=0, eps=1e-10)
