additional_params.update(net.predict_end.collect_params())
elif args.bidaf_style_out:
additional_params.update(net.modeling_layer.collect_params())
additional_params.update(net.output_layer.collect_params())
if args.apply_self_attention:
net.multi_head_attention.collect_params().initialize(ctx=ctx)
additional_params.update(net.multi_head_attention.collect_params())
if args.apply_transformer:
net.transformer.collect_params().initialize(ctx=ctx)
additional_params.update(net.transformer.collect_params())
net.hybridize(static_alloc=True)
loss_function = net.loss(customize_loss=args.customize_loss)
loss_function.hybridize(static_alloc=True)
if version_2 and VERIFIER_ID is not None:
if VERIFIER_ID == 0:
verifier = AnswerVerifyThreshold(
tokenizer=nlp.data.BERTBasicTokenizer(lower=lower),
max_answer_length=max_answer_length,
n_best_size=n_best_size,
max_len=max_seq_length,
version_2=version_2,
ctx=verify_ctx)
elif VERIFIER_ID == 1:
verifier = AnswerVerify(
tokenizer=nlp.data.BERTBasicTokenizer(lower=lower),
max_answer_length=max_answer_length,
elif pretrained:
# only load BertModel parameters
net.span_classifier.initialize(init=mx.init.Normal(0.02), ctx=ctx)
else:
# no checkpoint is loaded
net.initialize(init=mx.init.Normal(0.02), ctx=ctx)
if args.apply_coattention:
net.co_attention.collect_params().initialize(ctx=ctx)
if args.apply_self_attention:
net.multi_head_attention.collect_params().initialize(ctx=ctx)
net.hybridize(static_alloc=True)
loss_function = net.loss()
loss_function.hybridize(static_alloc=True)
if args.verify:
if VERIFIER_ID == 1:
verifier = AnswerVerify(
tokenizer=nlp.data.BERTBasicTokenizer(lower=lower),
max_answer_length=max_answer_length,
null_score_diff_threshold=null_score_diff_threshold,
n_best_size=n_best_size,
max_len=max_seq_length,
version_2=version_2,
ctx=verify_ctx
) # debug: to be moved onto another GPU latter if space issue happens
elif VERIFIER_ID == 2:
verifier = AnswerVerify2(version_2=version_2,