mass config

mass-base-uncased-config.json

@@ -9,7 +9,7 @@
 	"decoder_embed_dim": 768,
 	"decoder_ffn_embed_dim": 3072,
 	"decoder_layers": 6,
-	"decoder_attention_heads": 6,
+	"decoder_attention_heads": 12,
 	"decoder_normalize_before": 0,
 	"attention_dropout": 0.1,
 	"activation_dropout": 0.1,
@@ -24,7 +24,7 @@
 	"decoder_output_dim": 768,
 	"decoder_input_dim": 768,
 	"no_scale_embedding": 0,
-	"adaptive_input": 1,
+	"adaptive_input": 0,
 	"tie_adaptive_weights": 0,
 	"layernorm_embedding": 1,
 	"encoder_layerdrop": 0,

mass_for_generation.ini

@@ -0,0 +1,50 @@
+[core/auto]
+task_type = core/auto/supervised_task
+cache_dir = ./cache
+
+# model
+[core/model/mass]
+pretrained_name = mass-base-uncased
+num_beams = 5
+no_repeat_ngram_size = 0
+max_gen_seq_length = 15
+repetition_penalty = 1.0
+
+# dataset
+[core/dataset]
+# data columns: id, num, query, doc, label, score
+data_name = fuliucansheng/data_for_test
+
+[core/dataset/train]
+preprocess_funcs = ['core/process/mass_for_generation(query, doc)']
+
+[core/dataset/dev]
+preprocess_funcs = ['core/process/mass_for_tokens(query)', 'core/process/mass_for_next_tokens(doc)']
+
+[core/dataset/test]
+preprocess_funcs = ['core/process/mass_for_tokens(query)']
+
+# process
+[core/process/mass]
+pretrained_name = mass-base-uncased
+max_seq_length = 24
+max_gen_seq_length = 15
+
+# task
+[core/auto/supervised_task]
+model = core/model/mass_for_generation
+optim = core/optim/adam
+dataset = core/dataset/auto
+loss_fn = core/loss/lm
+score_fn = core/score/bleu
+monitor_fns = ['core/score/bleu', 'core/score/rouge1', 'core/score/rouge2', 'core/score/rougel']
+output_header = query,doc
+post_process_fn = partial(core/process/mass_for_decode)
+
+opt_fp16 = O1
+from_ckpt_dir = ${core/auto:cache_dir}
+to_ckpt_dir = ${core/auto:cache_dir}
+output_path = ${core/auto:cache_dir}/output.txt
+train_batch_size = 128
+dev_batch_size = 128
+test_batch_size = 256