Updates to Trainer/Tester/fastnlp

1. Tester has a parameter "print_every_step" to control printing. print_every_step == 0 means NO print.
2. Tester's evaluate return (list of) floats, rather than torch.cuda.tensor
3. Trainer also has a parameter "print_every_step". The same usage.
4. In training, validation steps are not shown.
5. Updates to code comments.
6. fastnlp.py is ready for CWS. test_fastNLP.py works.

fastNLP/core/preprocess.py

@@ -19,13 +19,13 @@ DEFAULT_WORD_TO_INDEX = {DEFAULT_PADDING_LABEL: 0, DEFAULT_UNKNOWN_LABEL: 1,
 def save_pickle(obj, pickle_path, file_name):
     with open(os.path.join(pickle_path, file_name), "wb") as f:
         _pickle.dump(obj, f)
-    print("{} saved in {}.".format(file_name, pickle_path))
+    print("{} saved in {}".format(file_name, pickle_path))
 
 
 def load_pickle(pickle_path, file_name):
     with open(os.path.join(pickle_path, file_name), "rb") as f:
         obj = _pickle.load(f)
-    print("{} loaded from {}.".format(file_name, pickle_path))
+    print("{} loaded from {}".format(file_name, pickle_path))
     return obj
 
 

fastNLP/core/tester.py

@@ -98,7 +98,7 @@ class BaseTester(object):
 
             print_output = "[test step {}] {}".format(step, eval_results)
             logger.info(print_output)
-            if step % self.print_every_step == 0:
+            if self.print_every_step > 0 and step % self.print_every_step == 0:
                 print(print_output)
             step += 1
 
@@ -187,7 +187,7 @@ class SeqLabelTester(BaseTester):
         # make sure "results" is in the same device as "truth"
         results = results.to(truth)
         accuracy = torch.sum(results == truth.view((-1,))).to(torch.float) / results.shape[0]
-        return [loss.data, accuracy.data]
+        return [float(loss), float(accuracy)]
 
     def metrics(self):
         batch_loss = np.mean([x[0] for x in self.eval_history])

fastNLP/core/trainer.py

@@ -4,7 +4,6 @@ import os
 import time
 from datetime import timedelta
 
-import numpy as np
 import torch
 
 from fastNLP.core.action import Action
@@ -47,7 +46,7 @@ class BaseTrainer(object):
             Otherwise, error will raise.
         """
         default_args = {"epochs": 3, "batch_size": 8, "validate": True, "use_cuda": True, "pickle_path": "./save/",
-                        "save_best_dev": True, "model_name": "default_model_name.pkl",
+                        "save_best_dev": True, "model_name": "default_model_name.pkl", "print_every_step": 1,
                         "loss": Loss(None),
                         "optimizer": Optimizer("Adam", lr=0.001, weight_decay=0)
                         }
@@ -86,6 +85,7 @@ class BaseTrainer(object):
         self.save_best_dev = default_args["save_best_dev"]
         self.use_cuda = default_args["use_cuda"]
         self.model_name = default_args["model_name"]
+        self.print_every_step = default_args["print_every_step"]
 
         self._model = None
         self._loss_func = default_args["loss"].get()  # return a pytorch loss function or None
@@ -93,48 +93,35 @@ class BaseTrainer(object):
         self._optimizer_proto = default_args["optimizer"]
 
     def train(self, network, train_data, dev_data=None):
-        """General Training Steps
+        """General Training Procedure
         :param network: a model
         :param train_data: three-level list, the training set.
         :param dev_data: three-level list, the validation data (optional)
-
-        The method is framework independent.
-        Work by calling the following methods:
-            - prepare_input
-            - mode
-            - define_optimizer
-            - data_forward
-            - get_loss
-            - grad_backward
-            - update
-        Subclasses must implement these methods with a specific framework.
         """
-        # prepare model and data, transfer model to gpu if available
+        # transfer model to gpu if available
         if torch.cuda.is_available() and self.use_cuda:
             self._model = network.cuda()
+            # self._model is used to access model-specific loss
         else:
             self._model = network
 
-        # define tester over dev data
+        # define Tester over dev data
         if self.validate:
             default_valid_args = {"save_output": True, "validate_in_training": True, "save_dev_input": True,
                                   "save_loss": True, "batch_size": self.batch_size, "pickle_path": self.pickle_path,
-                                  "use_cuda": self.use_cuda}
+                                  "use_cuda": self.use_cuda, "print_every_step": 0}
             validator = self._create_validator(default_valid_args)
             logger.info("validator defined as {}".format(str(validator)))
 
+        # optimizer and loss
         self.define_optimizer()
         logger.info("optimizer defined as {}".format(str(self._optimizer)))
         self.define_loss()
         logger.info("loss function defined as {}".format(str(self._loss_func)))
 
-        # main training epochs
-        n_samples = len(train_data)
-        n_batches = n_samples // self.batch_size
-        n_print = 1
+        # main training procedure
         start = time.time()
         logger.info("training epochs started")
-
         for epoch in range(1, self.n_epochs + 1):
             logger.info("training epoch {}".format(epoch))
 
@@ -144,23 +131,30 @@ class BaseTrainer(object):
             data_iterator = iter(Batchifier(RandomSampler(train_data), self.batch_size, drop_last=False))
             logger.info("prepared data iterator")
 
-            self._train_step(data_iterator, network, start=start, n_print=n_print, epoch=epoch)
+            # one forward and backward pass
+            self._train_step(data_iterator, network, start=start, n_print=self.print_every_step, epoch=epoch)
 
+            # validation
             if self.validate:
                 logger.info("validation started")
                 validator.test(network, dev_data)
 
                 if self.save_best_dev and self.best_eval_result(validator):
                     self.save_model(network, self.model_name)
-                    print("saved better model selected by dev")
-                    logger.info("saved better model selected by dev")
+                    print("Saved better model selected by validation.")
+                    logger.info("Saved better model selected by validation.")
 
                 valid_results = validator.show_matrices()
                 print("[epoch {}] {}".format(epoch, valid_results))
                 logger.info("[epoch {}] {}".format(epoch, valid_results))
 
     def _train_step(self, data_iterator, network, **kwargs):
-        """Training process in one epoch."""
+        """Training process in one epoch.
+            kwargs should contain:
+                - n_print: int, print training information every n steps.
+                - start: time.time(), the starting time of this step.
+                - epoch: int,
+        """
         step = 0
         for batch_x, batch_y in self.make_batch(data_iterator):
 
@@ -287,10 +281,11 @@ class BaseTrainer(object):
         raise NotImplementedError
 
     def save_model(self, network, model_name):
-        """
+        """Save this model with such a name.
+        This method may be called multiple times by Trainer to overwritten a better model.
+
         :param network: the PyTorch model
         :param model_name: str
-        model_best_dev.pkl may be overwritten by a better model in future epochs.
         """
         if model_name[-4:] != ".pkl":
             model_name += ".pkl"
@@ -300,33 +295,9 @@ class BaseTrainer(object):
         raise NotImplementedError
 
 
-class ToyTrainer(BaseTrainer):
-    """
-        An example to show the definition of Trainer.
-    """
-
-    def __init__(self, training_args):
-        super(ToyTrainer, self).__init__(training_args)
-
-    def load_train_data(self, data_path):
-        data_train = _pickle.load(open(data_path + "/data_train.pkl", "rb"))
-        data_dev = _pickle.load(open(data_path + "/data_train.pkl", "rb"))
-        return data_train, data_dev, 0, 1
-
-    def data_forward(self, network, x):
-        return network(x)
-
-    def grad_backward(self, loss):
-        self._model.zero_grad()
-        loss.backward()
-
-    def get_loss(self, pred, truth):
-        return np.mean(np.square(pred - truth))
-
-
 class SeqLabelTrainer(BaseTrainer):
     """
-    Trainer for Sequence Modeling
+    Trainer for Sequence Labeling
 
     """
 
@@ -384,7 +355,7 @@ class SeqLabelTrainer(BaseTrainer):
 
 
 class ClassificationTrainer(BaseTrainer):
-    """Trainer for classification."""
+    """Trainer for text classification."""
 
     def __init__(self, **train_args):
         super(ClassificationTrainer, self).__init__(**train_args)

fastNLP/fastnlp.py

@@ -1,4 +1,4 @@
-# from fastNLP.core.predictor import SeqLabelInfer, ClassificationInfer
+from fastNLP.core.predictor import SeqLabelInfer, ClassificationInfer
 from fastNLP.core.preprocess import load_pickle
 from fastNLP.loader.config_loader import ConfigLoader, ConfigSection
 from fastNLP.loader.model_loader import ModelLoader
@@ -74,9 +74,11 @@ class FastNLP(object):
             self._download(model_name, FastNLP_MODEL_COLLECTION[model_name]["url"])
 
         model_class = self._get_model_class(FastNLP_MODEL_COLLECTION[model_name]["class"])
+        print("Restore model class {}".format(str(model_class)))
 
         model_args = ConfigSection()
         ConfigLoader.load_config(self.model_dir + config_file, {section_name: model_args})
+        print("Restore model hyper-parameters {}".format(str(model_args.data)))
 
         # fetch dictionary size and number of labels from pickle files
         word2index = load_pickle(self.model_dir, "word2id.pkl")
@@ -86,14 +88,16 @@ class FastNLP(object):
 
         # Construct the model
         model = model_class(model_args)
+        print("Model constructed.")
 
         # To do: framework independent
         ModelLoader.load_pytorch(model, self.model_dir + FastNLP_MODEL_COLLECTION[model_name]["pickle"])
+        print("Model weights loaded.")
 
         self.model = model
         self.infer_type = FastNLP_MODEL_COLLECTION[model_name]["type"]
 
-        print("Model loaded. ")
+        print("Inference ready.")
 
     def run(self, raw_input):
         """
@@ -168,10 +172,15 @@ class FastNLP(object):
         :param language: str, one of ('zh', 'en'), Chinese or English.
         :return data: list of list of string, each string is a token.
         """
+        assert language in ("zh", "en")
         data = []
-        delimiter = " " if language is "en" else ""
         for sent in text:
-            tokens = sent.strip().split(delimiter)
+            if language == "en":
+                tokens = sent.strip().split()
+            elif language == "zh":
+                tokens = [char for char in sent]
+            else:
+                raise RuntimeError("Unknown language {}".format(language))
             data.append(tokens)
         return data
 

reproduction/chinese_word_segment/run.py

@@ -6,21 +6,21 @@ sys.path.append(os.path.join(os.path.dirname(__file__), '../..'))
 from fastNLP.loader.config_loader import ConfigLoader, ConfigSection
 from fastNLP.core.trainer import SeqLabelTrainer
 from fastNLP.loader.dataset_loader import TokenizeDatasetLoader, BaseLoader
-from fastNLP.loader.preprocess import POSPreprocess, load_pickle
+from fastNLP.core.preprocess import SeqLabelPreprocess, load_pickle
 from fastNLP.saver.model_saver import ModelSaver
 from fastNLP.loader.model_loader import ModelLoader
 from fastNLP.core.tester import SeqLabelTester
 from fastNLP.models.sequence_modeling import AdvSeqLabel
-from fastNLP.core.inference import SeqLabelInfer
+from fastNLP.core.predictor import SeqLabelInfer
 
 # not in the file's dir
 if len(os.path.dirname(__file__)) != 0:
     os.chdir(os.path.dirname(__file__))
-datadir = 'icwb2-data'
-cfgfile = 'cws.cfg'
+datadir = "/home/zyfeng/data/"
+cfgfile = './cws.cfg'
 data_name = "pku_training.utf8"
 
-cws_data_path = os.path.join(datadir, "training/pku_training.utf8")
+cws_data_path = os.path.join(datadir, "pku_training.utf8")
 pickle_path = "save"
 data_infer_path = os.path.join(datadir, "infer.utf8")
 
@@ -70,9 +70,10 @@ def train():
     train_data = loader.load_pku()
 
     # Preprocessor
-    p = POSPreprocess(train_data, pickle_path, train_dev_split=0.3)
-    train_args["vocab_size"] = p.vocab_size
-    train_args["num_classes"] = p.num_classes
+    preprocessor = SeqLabelPreprocess()
+    data_train, data_dev = preprocessor.run(train_data, pickle_path=pickle_path, train_dev_split=0.3)
+    train_args["vocab_size"] = preprocessor.vocab_size
+    train_args["num_classes"] = preprocessor.num_classes
 
     # Trainer
     trainer = SeqLabelTrainer(**train_args.data)
@@ -83,10 +84,11 @@ def train():
         ModelLoader.load_pytorch(model, "./save/saved_model.pkl")
         print('model parameter loaded!')
     except Exception as e:
+        print("No saved model. Continue.")
         pass
         
     # Start training
-    trainer.train(model)
+    trainer.train(model, data_train, data_dev)
     print("Training finished!")
 
     # Saver
@@ -106,6 +108,9 @@ def test():
     index2label = load_pickle(pickle_path, "id2class.pkl")
     test_args["num_classes"] = len(index2label)
 
+    # load dev data
+    dev_data = load_pickle(pickle_path, "data_dev.pkl")
+
     # Define the same model
     model = AdvSeqLabel(test_args)
 
@@ -114,10 +119,10 @@ def test():
     print("model loaded!")
 
     # Tester
-    tester = SeqLabelTester(test_args)
+    tester = SeqLabelTester(**test_args.data)
 
     # Start testing
-    tester.test(model)
+    tester.test(model, dev_data)
 
     # print test results
     print(tester.show_matrices())

test/seq_labeling.py

@@ -123,7 +123,7 @@ def train_and_test():
     tester = SeqLabelTester(save_output=False,
                             save_loss=False,
                             save_best_dev=False,
-                            batch_size=8,
+                            batch_size=4,
                             use_cuda=False,
                             pickle_path=pickle_path,
                             model_name="seq_label_in_test.pkl",
@@ -140,4 +140,4 @@ def train_and_test():
 
 if __name__ == "__main__":
     train_and_test()
-    infer()
+    # infer()

test/test_fastNLP.py

@@ -3,7 +3,7 @@ import sys
 sys.path.append("..")
 from fastNLP.fastnlp import FastNLP
 
-PATH_TO_CWS_PICKLE_FILES = "/home/zyfeng/data/save/"
+PATH_TO_CWS_PICKLE_FILES = "/home/zyfeng/fastNLP/reproduction/chinese_word_segment/save/"
 
 def word_seg():
     nlp = FastNLP(model_dir=PATH_TO_CWS_PICKLE_FILES)