def main():
READER_TRAIN = DataReader(DATA_FILE_TRAIN, 0, True, 256)
qrels, scores = load_file_to_rerank(QRELS, 1)
_, to_rerank = load_file_to_rerank(DATA_FILE_PREDICT,0)
torch.manual_seed(1)
print_message('Starting')
try:
net = torch.load('tensors.duet').to(DEVICE)
print_message("Previous Model Found and Loaded")
except:
print_message("No Previous Model Found. Creating new")
net = Duet().to(DEVICE)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(), lr=LEARNING_RATE)
print_message('Number of learnable parameters: {}'.format(net.parameter_count()))
print_message('Training for {} epochs'.format(NUM_EPOCHS))
for ep_idx in range(NUM_EPOCHS):
net, train_loss = train(READER_TRAIN,net, optimizer,criterion)
print_message('epoch:{}, loss: {}'.format(ep_idx + 1, train_loss / EPOCH_SIZE))
mrr = evaluate(net,qrels,scores, to_rerank)
print_message('MRR @1000:{}'.format(mrr))
torch.cuda.empty_cache()
torch.save(net,'tensors.duet')
print_message('Done Training')
print_message('Evaluating')
metrics = msmarco_eval.compute_metrics_from_files(QRELS,DATA_FILE_PREDICTOUT)
print_message('#####################')
for metric in sorted(metrics):
print_message('{}: {}'.format(metric, metrics[metric]))
print_message('#####################')
def main():
args = get_args()
if args.sp_folder is not None:
eval_specific_folder(args)
exit(0)
if args.full_partition:
cache_fname = 'full_collection.npy'
if os.path.exists(cache_fname):
collection = np.load(cache_fname, allow_pickle=True).item()
else:
collection = load_collection(args)
np.save(cache_fname, collection)
else:
collection = load_collection(args)
if args.each_layer:
eval_each_layer(args, collection)
exit(0)
if args.detail:
print_relevant_pair_details(args, collection)
final_scores, avg_layer = generate_final_scores_confidence(
args, collection)
run_fname = generate_run_file(args, final_scores)
metrics = compute_metrics_from_files(args.reference, run_fname)
print('Avg layer: {}'.format(avg_layer))
for metric in sorted(metrics):
print('{}: {}'.format(metric, metrics[metric]))
break
if not args.no_save_np:
save_to_np(args, metrics, avg_layer)
def eval_each_layer(args, collection):
with open('each_layer.txt', 'w') as fout:
for i in range(args.n_layers):
final_scores = defaultdict(
list) # key: qid; value: list of [pid, score]
for qid, qid_value in collection.items():
for pid, score_list in qid_value.items():
final_scores[qid].append(
[pid, score_list[i], i + 1, score_list])
run_fname = generate_run_file(args, final_scores)
metrics = compute_metrics_from_files(args.reference, run_fname)
print(f'Layer {i}:\t{metrics["MRR @10"]}', file=fout)
def eval_specific_folder(args):
final_scores = defaultdict(list) # key: qid; value: list of [pid, score]
used_layers = 0
pairs = 0
for par in args.partition:
with open(
os.path.join(args.sp_folder,
f'{args.split}.partition{par}.score')) as fin:
for line in fin:
qid, pid, score, layer = line.strip().split('\t')
pairs += 1
final_scores[qid].append([pid, float(score), int(layer), []])
used_layers += int(layer)
run_fname = generate_run_file(args, final_scores)
print('Avg-layer:', used_layers / pairs)
metrics = compute_metrics_from_files(args.reference, run_fname)
for metric in sorted(metrics):
print('{}: {}'.format(metric, metrics[metric]))
class FakeScorer():
def __init__(self):
pass
def score_query_passages(self, query, passages, batch_size):
return [1] * len(passages)
bm25_path = "../data/evaluation/bm25/run.dev.small.tsv"
queries_path = "../data/queries/queries.dev.small.tsv"
passages_path = "../data/passages/passages.bm25.small.json"
candidate_path = "../data/evaluation/test/run.tsv"
reference_path = "../data/evaluation/gold/qrels.dev.small.tsv"
n_top = 50
n_queries_to_evaluate = None
mrr_ref = 0.18741227770955546
model = FakeScorer()
mrr = EvaluationQueries(bm25_path, queries_path, passages_path, n_top)
mrr.score(model, candidate_path, n_queries_to_evaluate)
mrr_metrics = compute_metrics_from_files(reference_path, candidate_path)
assert int(1000*mrr_ref) == int(1000*mrr_metrics['MRR @10']), "Test failed"
print(mrr_metrics)
print('Test ok')
def evaluate(path_to_reference,path_to_candidate):
metrics = msmarco_eval.compute_metrics_from_files(path_to_reference, path_to_candidate)
print('#####################')
for metric in sorted(metrics):
print('{}: {}'.format(metric, metrics[metric]))
print('#####################')
