Fix infer bug on Release/1.6 (#3693)

* update downloads.py

* fix bug on ernie based inferring

PaddleNLP/lexical_analysis/creator.py

@@ -12,7 +12,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 """
 Define the function to create lexical analysis model and model's data reader
 """
@@ -37,15 +36,22 @@ def create_model(args, vocab_size, num_labels, mode='train'):
 
     # model's input data
     words = fluid.data(name='words', shape=[-1, 1], dtype='int64', lod_level=1)
-    targets = fluid.data(name='targets', shape=[-1, 1], dtype='int64', lod_level=1)
+    targets = fluid.data(
+        name='targets', shape=[-1, 1], dtype='int64', lod_level=1)
 
     # for inference process
     if mode == 'infer':
-        crf_decode = nets.lex_net(words, args, vocab_size, num_labels, for_infer=True, target=None)
-        return {"feed_list": [words], "words": words, "crf_decode": crf_decode, }
+        crf_decode = nets.lex_net(
+            words, args, vocab_size, num_labels, for_infer=True, target=None)
+        return {
+            "feed_list": [words],
+            "words": words,
+            "crf_decode": crf_decode,
+        }
 
     # for test or train process
-    avg_cost, crf_decode = nets.lex_net(words, args, vocab_size, num_labels, for_infer=False, target=targets)
+    avg_cost, crf_decode = nets.lex_net(
+        words, args, vocab_size, num_labels, for_infer=False, target=targets)
 
     (precision, recall, f1_score, num_infer_chunks, num_label_chunks,
      num_correct_chunks) = fluid.layers.chunk_eval(
@@ -73,7 +79,14 @@ def create_model(args, vocab_size, num_labels, mode='train'):
     return ret
 
 
-def create_pyreader(args, file_name, feed_list, place, model='lac', reader=None, return_reader=False, mode='train'):
+def create_pyreader(args,
+                    file_name,
+                    feed_list,
+                    place,
+                    model='lac',
+                    reader=None,
+                    return_reader=False,
+                    mode='train'):
     # init reader
 
     if model == 'lac':
@@ -81,8 +94,7 @@ def create_pyreader(args, file_name, feed_list, place, model='lac', reader=None,
             feed_list=feed_list,
             capacity=50,
             use_double_buffer=True,
-            iterable=True
-        )
+            iterable=True)
 
         if reader == None:
             reader = Dataset(args)
@@ -93,20 +105,16 @@ def create_pyreader(args, file_name, feed_list, place, model='lac', reader=None,
                 fluid.io.batch(
                     fluid.io.shuffle(
                         reader.file_reader(file_name),
-                        buf_size=args.traindata_shuffle_buffer
-                    ),
-                    batch_size=args.batch_size
-                ),
-                places=place
-            )
+                        buf_size=args.traindata_shuffle_buffer),
+                    batch_size=args.batch_size),
+                places=place)
         else:
             pyreader.decorate_sample_list_generator(
                 fluid.io.batch(
-                    reader.file_reader(file_name, mode=mode),
-                    batch_size=args.batch_size
-                ),
-                places=place
-            )
+                    reader.file_reader(
+                        file_name, mode=mode),
+                    batch_size=args.batch_size),
+                places=place)
 
     elif model == 'ernie':
         # create ernie pyreader
@@ -114,8 +122,7 @@ def create_pyreader(args, file_name, feed_list, place, model='lac', reader=None,
             feed_list=feed_list,
             capacity=50,
             use_double_buffer=True,
-            iterable=True
-        )
+            iterable=True)
         if reader == None:
             reader = SequenceLabelReader(
                 vocab_path=args.vocab_path,
@@ -127,17 +134,21 @@ def create_pyreader(args, file_name, feed_list, place, model='lac', reader=None,
         if mode == 'train':
             pyreader.set_batch_generator(
                 reader.data_generator(
-                    file_name, args.batch_size, args.epoch, shuffle=True, phase="train"
-                ),
-                places=place
-            )
+                    file_name,
+                    args.batch_size,
+                    args.epoch,
+                    shuffle=True,
+                    phase="train"),
+                places=place)
         else:
             pyreader.set_batch_generator(
                 reader.data_generator(
-                    file_name, args.batch_size, epoch=1, shuffle=False, phase=mode
-                ),
-                places=place
-            )
+                    file_name,
+                    args.batch_size,
+                    epoch=1,
+                    shuffle=False,
+                    phase=mode),
+                places=place)
     if return_reader:
         return pyreader, reader
     else:
@@ -150,14 +161,20 @@ def create_ernie_model(args, ernie_config):
     """
     # ERNIE's input data
 
-    src_ids = fluid.data(name='src_ids', shape=[-1, args.max_seq_len, 1], dtype='int64')
-    sent_ids = fluid.data(name='sent_ids', shape=[-1, args.max_seq_len, 1], dtype='int64')
-    pos_ids = fluid.data(name='pos_ids', shape=[-1, args.max_seq_len, 1], dtype='int64')
-    input_mask = fluid.data(name='input_mask', shape=[-1, args.max_seq_len, 1], dtype='float32')
+    src_ids = fluid.data(
+        name='src_ids', shape=[-1, args.max_seq_len, 1], dtype='int64')
+    sent_ids = fluid.data(
+        name='sent_ids', shape=[-1, args.max_seq_len, 1], dtype='int64')
+    pos_ids = fluid.data(
+        name='pos_ids', shape=[-1, args.max_seq_len, 1], dtype='int64')
+    input_mask = fluid.data(
+        name='input_mask', shape=[-1, args.max_seq_len, 1], dtype='float32')
 
-    padded_labels = fluid.data(name='padded_labels', shape=[-1, args.max_seq_len, 1], dtype='int64')
+    padded_labels = fluid.data(
+        name='padded_labels', shape=[-1, args.max_seq_len, 1], dtype='int64')
 
-    seq_lens = fluid.data(name='seq_lens', shape=[-1], dtype='int64', lod_level=0)
+    seq_lens = fluid.data(
+        name='seq_lens', shape=[-1], dtype='int64', lod_level=0)
 
     squeeze_labels = fluid.layers.squeeze(padded_labels, axes=[-1])
 
@@ -187,13 +204,14 @@ def create_ernie_model(args, ernie_config):
         input=emission,
         label=padded_labels,
         param_attr=fluid.ParamAttr(
-            name='crfw',
-            learning_rate=args.crf_learning_rate),
+            name='crfw', learning_rate=args.crf_learning_rate),
         length=seq_lens)
 
     avg_cost = fluid.layers.mean(x=crf_cost)
     crf_decode = fluid.layers.crf_decoding(
-        input=emission, param_attr=fluid.ParamAttr(name='crfw'), length=seq_lens)
+        input=emission,
+        param_attr=fluid.ParamAttr(name='crfw'),
+        length=seq_lens)
 
     (precision, recall, f1_score, num_infer_chunks, num_label_chunks,
      num_correct_chunks) = fluid.layers.chunk_eval(
@@ -206,9 +224,11 @@ def create_ernie_model(args, ernie_config):
     chunk_evaluator.reset()
 
     ret = {
-        "feed_list": [src_ids, sent_ids, pos_ids, input_mask, padded_labels, seq_lens],
+        "feed_list":
+        [src_ids, sent_ids, pos_ids, input_mask, padded_labels, seq_lens],
         "words": src_ids,
         "labels": padded_labels,
+        "seq_lens": seq_lens,
         "avg_cost": avg_cost,
         "crf_decode": crf_decode,
         "precision": precision,

PaddleNLP/lexical_analysis/run_ernie_sequence_labeling.py

@@ -39,6 +39,7 @@ from models.representation.ernie import ErnieConfig
 from models.model_check import check_cuda
 from models.model_check import check_version
 
+
 def evaluate(exe, test_program, test_pyreader, test_ret):
     """
     Evaluation Function
@@ -55,8 +56,7 @@ def evaluate(exe, test_program, test_pyreader, test_ret):
                 test_ret["num_label_chunks"],
                 test_ret["num_correct_chunks"],
             ],
-            feed=data[0]
-        )
+            feed=data[0])
         total_loss.append(loss)
 
         test_ret["chunk_evaluator"].update(nums_infer, nums_label, nums_correct)
@@ -64,9 +64,11 @@ def evaluate(exe, test_program, test_pyreader, test_ret):
     precision, recall, f1 = test_ret["chunk_evaluator"].eval()
     end_time = time.time()
 
-    print("\t[test] loss: %.5f, P: %.5f, R: %.5f, F1: %.5f, elapsed time: %.3f s"
+    print(
+        "\t[test] loss: %.5f, P: %.5f, R: %.5f, F1: %.5f, elapsed time: %.3f s"
         % (np.mean(total_loss), precision, recall, f1, end_time - start_time))
 
+
 def do_train(args):
     """
     Main Function
@@ -80,14 +82,15 @@ def do_train(args):
     else:
         dev_count = min(multiprocessing.cpu_count(), args.cpu_num)
         if (dev_count < args.cpu_num):
-            print("WARNING: The total CPU NUM in this machine is %d, which is less than cpu_num parameter you set. "
-                  "Change the cpu_num from %d to %d"%(dev_count, args.cpu_num, dev_count))
+            print(
+                "WARNING: The total CPU NUM in this machine is %d, which is less than cpu_num parameter you set. "
+                "Change the cpu_num from %d to %d" %
+                (dev_count, args.cpu_num, dev_count))
         os.environ['CPU_NUM'] = str(dev_count)
         place = fluid.CPUPlace()
 
     exe = fluid.Executor(place)
 
-
     startup_prog = fluid.Program()
     if args.random_seed is not None:
         startup_prog.random_seed = args.random_seed
@@ -99,22 +102,27 @@ def do_train(args):
             train_ret = creator.create_ernie_model(args, ernie_config)
 
             # ernie pyreader
-            train_pyreader = creator.create_pyreader(args, file_name=args.train_data,
+            train_pyreader = creator.create_pyreader(
+                args,
+                file_name=args.train_data,
                 feed_list=train_ret['feed_list'],
                 model="ernie",
                 place=place)
 
             test_program = train_program.clone(for_test=True)
-            test_pyreader = creator.create_pyreader(args, file_name=args.test_data,
+            test_pyreader = creator.create_pyreader(
+                args,
+                file_name=args.test_data,
                 feed_list=train_ret['feed_list'],
                 model="ernie",
                 place=place)
 
-            optimizer = fluid.optimizer.Adam(learning_rate=args.base_learning_rate)
-            fluid.clip.set_gradient_clip(clip=fluid.clip.GradientClipByGlobalNorm(clip_norm=1.0))
+            optimizer = fluid.optimizer.Adam(
+                learning_rate=args.base_learning_rate)
+            fluid.clip.set_gradient_clip(
+                clip=fluid.clip.GradientClipByGlobalNorm(clip_norm=1.0))
             optimizer.minimize(train_ret["avg_cost"])
 
-
     lower_mem, upper_mem, unit = fluid.contrib.memory_usage(
         program=train_program, batch_size=args.batch_size)
     print("Theoretical memory usage in training: %.3f - %.3f %s" %
@@ -129,16 +137,18 @@ def do_train(args):
     if args.init_checkpoint:
         utils.init_checkpoint(exe, args.init_checkpoint, startup_prog)
     elif args.init_pretraining_params:
-        utils.init_pretraining_params(exe, args.init_pretraining_params, startup_prog)
+        utils.init_pretraining_params(exe, args.init_pretraining_params,
+                                      startup_prog)
 
-    if dev_count>1 and not args.use_cuda:
+    if dev_count > 1 and not args.use_cuda:
         device = "GPU" if args.use_cuda else "CPU"
-        print("%d %s are used to train model"%(dev_count, device))
+        print("%d %s are used to train model" % (dev_count, device))
 
         # multi cpu/gpu config
         exec_strategy = fluid.ExecutionStrategy()
         build_strategy = fluid.BuildStrategy()
-        compiled_prog = fluid.compiler.CompiledProgram(train_program).with_data_parallel(
+        compiled_prog = fluid.compiler.CompiledProgram(
+            train_program).with_data_parallel(
                 loss_name=train_ret['avg_cost'].name,
                 build_strategy=build_strategy,
                 exec_strategy=exec_strategy)
@@ -162,16 +172,23 @@ def do_train(args):
 
             start_time = time.time()
 
-            outputs = exe.run(program=compiled_prog, feed=data[0], fetch_list=fetch_list)
+            outputs = exe.run(program=compiled_prog,
+                              feed=data[0],
+                              fetch_list=fetch_list)
             end_time = time.time()
             if steps % args.print_steps == 0:
-                loss, precision, recall, f1_score = [np.mean(x) for x in outputs]
-                print("[train] batch_id = %d, loss = %.5f, P: %.5f, R: %.5f, F1: %.5f, elapsed time %.5f, "
-                        "pyreader queue_size: %d " % (steps, loss, precision, recall, f1_score,
+                loss, precision, recall, f1_score = [
+                    np.mean(x) for x in outputs
+                ]
+                print(
+                    "[train] batch_id = %d, loss = %.5f, P: %.5f, R: %.5f, F1: %.5f, elapsed time %.5f, "
+                    "pyreader queue_size: %d " %
+                    (steps, loss, precision, recall, f1_score,
                      end_time - start_time, train_pyreader.queue.size()))
 
             if steps % args.save_steps == 0:
-                save_path = os.path.join(args.model_save_dir, "step_" + str(steps))
+                save_path = os.path.join(args.model_save_dir,
+                                         "step_" + str(steps))
                 print("\tsaving model as %s" % (save_path))
                 fluid.io.save_persistables(exe, save_path, train_program)
 
@@ -182,7 +199,6 @@ def do_train(args):
     fluid.io.save_persistables(exe, save_path, train_program)
 
 
-
 def do_eval(args):
     # init executor
     if args.use_cuda:
@@ -198,11 +214,13 @@ def do_eval(args):
             test_ret = creator.create_ernie_model(args, ernie_config)
     test_program = test_program.clone(for_test=True)
 
-    pyreader = creator.create_pyreader(args, file_name=args.test_data,
+    pyreader = creator.create_pyreader(
+        args,
+        file_name=args.test_data,
         feed_list=test_ret['feed_list'],
         model="ernie",
         place=place,
-                                          mode='test',)
+        mode='test', )
 
     print('program startup')
 
@@ -212,11 +230,13 @@ def do_eval(args):
     print('program loading')
     # load model
     if not args.init_checkpoint:
-        raise ValueError("args 'init_checkpoint' should be set if only doing test or infer!")
+        raise ValueError(
+            "args 'init_checkpoint' should be set if only doing test or infer!")
     utils.init_checkpoint(exe, args.init_checkpoint, test_program)
 
     evaluate(exe, test_program, pyreader, test_ret)
 
+
 def do_infer(args):
     # init executor
     if args.use_cuda:
@@ -233,7 +253,9 @@ def do_infer(args):
             infer_ret = creator.create_ernie_model(args, ernie_config)
     infer_program = infer_program.clone(for_test=True)
     print(args.test_data)
-    pyreader, reader = creator.create_pyreader(args, file_name=args.test_data,
+    pyreader, reader = creator.create_pyreader(
+        args,
+        file_name=args.test_data,
         feed_list=infer_ret['feed_list'],
         model="ernie",
         place=place,
@@ -245,29 +267,38 @@ def do_infer(args):
 
     # load model
     if not args.init_checkpoint:
-        raise ValueError("args 'init_checkpoint' should be set if only doing test or infer!")
+        raise ValueError(
+            "args 'init_checkpoint' should be set if only doing test or infer!")
     utils.init_checkpoint(exe, args.init_checkpoint, infer_program)
 
     # create dict
-    id2word_dict = dict([(str(word_id), word) for word, word_id in reader.vocab.items()])
-    id2label_dict = dict([(str(label_id), label) for label, label_id in reader.label_map.items()])
+    id2word_dict = dict(
+        [(str(word_id), word) for word, word_id in reader.vocab.items()])
+    id2label_dict = dict([(str(label_id), label)
+                          for label, label_id in reader.label_map.items()])
     Dataset = namedtuple("Dataset", ["id2word_dict", "id2label_dict"])
     dataset = Dataset(id2word_dict, id2label_dict)
 
     # make prediction
     for data in pyreader():
-        (words, crf_decode) = exe.run(infer_program,
-                                      fetch_list=[infer_ret["words"], infer_ret["crf_decode"]],
+        (words, crf_decode, seq_lens) = exe.run(infer_program,
+                                                fetch_list=[
+                                                    infer_ret["words"],
+                                                    infer_ret["crf_decode"],
+                                                    infer_ret["seq_lens"]
+                                                ],
                                                 feed=data[0],
-                                      return_numpy=False)
+                                                return_numpy=True)
         # User should notice that words had been clipped if long than args.max_seq_len
-        results = utils.parse_result(words, crf_decode, dataset)
+        results = utils.parse_padding_result(words, crf_decode, seq_lens,
+                                             dataset)
         for sent, tags in results:
-            result_list = ['(%s, %s)' % (ch, tag) for ch, tag in zip(sent, tags)]
+            result_list = [
+                '(%s, %s)' % (ch, tag) for ch, tag in zip(sent, tags)
+            ]
             print(''.join(result_list))
 
 
-
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(__doc__)
     utils.load_yaml(parser, './conf/ernie_args.yaml')
@@ -284,4 +315,3 @@ if __name__ == "__main__":
         do_infer(args)
     else:
         print("Usage: %s --mode train|eval|infer " % sys.argv[0])
-

PaddleNLP/lexical_analysis/utils.py

@@ -148,6 +148,50 @@ def parse_result(words, crf_decode, dataset):
     return batch_out
 
 
+def parse_padding_result(words, crf_decode, seq_lens, dataset):
+    """ parse padding result """
+    words = np.squeeze(words)
+    batch_size = len(seq_lens)
+
+    batch_out = []
+    for sent_index in range(batch_size):
+
+        sent = [
+            dataset.id2word_dict[str(id)]
+            for id in words[sent_index][1:seq_lens[sent_index] - 1]
+        ]
+        tags = [
+            dataset.id2label_dict[str(id)]
+            for id in crf_decode[sent_index][1:seq_lens[sent_index] - 1]
+        ]
+
+        sent_out = []
+        tags_out = []
+        parital_word = ""
+        for ind, tag in enumerate(tags):
+            # for the first word
+            if parital_word == "":
+                parital_word = sent[ind]
+                tags_out.append(tag.split('-')[0])
+                continue
+
+            # for the beginning of word
+            if tag.endswith("-B") or (tag == "O" and tags[ind - 1] != "O"):
+                sent_out.append(parital_word)
+                tags_out.append(tag.split('-')[0])
+                parital_word = sent[ind]
+                continue
+
+            parital_word += sent[ind]
+
+        # append the last word, except for len(tags)=0
+        if len(sent_out) < len(tags_out):
+            sent_out.append(parital_word)
+
+        batch_out.append([sent_out, tags_out])
+    return batch_out
+
+
 def init_checkpoint(exe, init_checkpoint_path, main_program):
     """
     Init CheckPoint