Make Texar-TF compatible with Texar-PyTorch (#183)

* Lazily load Texar-TF modules so these and TensorFlow will not be loaded when user only imports texar.torch.
* Fix cyclic dependencies between texar.hyperparams and texar.utils.
* Fix library imports that are not compatible with lazy loading.

CHANGELOG.md

@@ -3,6 +3,16 @@
 
 ### New features
 
+### Feature improvements
+
+* Use lazy import to be compatible with [texar-pytorch](https://github.com/asyml/texar-pytorch). ([#183](https://github.com/asyml/texar/pull/183)) 
+
+### Fixes
+
+## [v0.2.1](https://github.com/asyml/texar/releases/tag/v0.2.1) (2019-07-28)
+
+### New features
+
 * Add support for GPT-2 345M model in [examples/gpt-2](https://github.com/asyml/texar/tree/master/examples/gpt-2). ([#156](https://github.com/asyml/texar/pull/156)) 
 * Add BERT modules, including `texar.modules.BERTEncoder` ([doc](https://texar.readthedocs.io/en/latest/code/modules.html#texar.modules.BertEncoder)) and `texar.modules.BERTClassifier` ([doc](https://texar.readthedocs.io/en/latest/code/modules.html#bertclassifierv)). ([#167](https://github.com/asyml/texar/pull/167))
 

texar/__init__.py

@@ -21,17 +21,98 @@
 
 # pylint: disable=wildcard-import
 
-from texar.version import VERSION as __version__
-
-from texar.module_base import *
-from texar.hyperparams import *
-from texar.context import *
-from texar import modules
-from texar import core
-from texar import losses
-from texar import models
-from texar import data
-from texar import evals
-from texar import agents
-from texar import run
-from texar import utils
+import sys
+
+if sys.version_info.major < 3:
+    # PY 2.x, import as is because Texar-PyTorch cannot be installed.
+    from texar.version import VERSION as __version__
+
+    from texar.module_base import *
+    from texar.hyperparams import *
+    from texar.context import *
+    from texar import modules
+    from texar import core
+    from texar import losses
+    from texar import models
+    from texar import data
+    from texar import evals
+    from texar import agents
+    from texar import run
+    from texar import utils
+else:
+    # Lazily load Texar-TF modules upon usage. This is to ensure that Texar-TF
+    # and TensorFlow will not be imported if the user only requires
+    # Texar-PyTorch modules from `texar.torch`.
+    #
+    # Due to the lazy loading mechanism, it is now impossible to write
+    # `from texar import <module>` within library code (i.e., code that will be
+    # accessible from the `texar` module). Please use the following workarounds
+    # instead:
+    #
+    # 1. To import a class / function that is directly accessible from `texar`,
+    #    import them from their containing modules. For instance:
+    #
+    #    `from texar import HParams`
+    #      ->  `from texar.hyperparams import HParams`
+    #    `from texar import ModuleBase`
+    #      ->  `from texar.module_base import ModuleBase`
+    # 2. To import a module that is directly accessible from `texar`, use the
+    #    `import ... as` syntax. For instance:
+    #
+    #    `from texar import utils`  ->  `import texar.utils as utils`
+    #    `from texar import context`  ->  `import texar.context as context`
+
+    import importlib
+
+    __import_modules__ = [
+        "modules", "core", "losses", "models", "data", "evals",
+        "agents", "run", "utils",
+    ]
+    __import_star_modules__ = ["module_base", "hyperparams", "context"]
+
+
+    def _import_all():
+        from texar.version import VERSION
+        globals()["__version__"] = VERSION
+
+        for module_name in __import_star_modules__:
+            # from ... import *. Requires manually handling `__all__`.
+            module = importlib.import_module("." + module_name, package="texar")
+            try:
+                variables = module.__all__
+            except AttributeError:
+                variables = [name for name in module.__dict__
+                             if not name.startswith("_")]
+            globals().update({
+                name: module.__dict__[name] for name in variables})
+
+        for module_name in __import_modules__:
+            # from ... import module
+            module = importlib.import_module("." + module_name, package="texar")
+            globals()[module_name] = module
+
+
+    class _DummyTexarBaseModule:
+        # Credit: https://stackoverflow.com/a/7668273/4909228
+        def __getattr__(self, name):
+            if name in globals():
+                # Shortcut to global names.
+                return globals()[name]
+            if name == "torch":
+                # To use `texar.torch`, Texar-TF and TensorFlow should not be
+                # imported.
+                module = importlib.import_module(".torch", package="texar")
+                globals()["torch"] = module
+                return module
+
+            # The user tries to access Texar-TF modules, so we load all modules
+            # at this point, and restore the registered `texar` module.
+            sys.modules[__name__] = __module__
+            _import_all()
+            return globals()[name]
+
+
+    # Save `texar` module as `__module__`, ans replace the system-wide
+    # registered module with our dummy module.
+    __module__ = sys.modules[__name__]
+    sys.modules[__name__] = _DummyTexarBaseModule()

texar/hyperparams.py

@@ -22,7 +22,6 @@
 import copy
 import json
 
-from texar.utils.dtypes import is_callable
 
 __all__ = [
     "HParams"
@@ -212,6 +211,8 @@ def _parse(hparams, # pylint: disable=too-many-branches, too-many-statements
                 else:
                     parsed_hparams[name] = HParams(value, value)
 
+        from texar.utils.dtypes import is_callable
+
         # Parse hparams
         for name, value in hparams.items():
             if name not in default_hparams:

texar/models/model_base.py

@@ -19,7 +19,7 @@
 from __future__ import division
 from __future__ import print_function
 
-from texar import HParams
+from texar.hyperparams import HParams
 
 # pylint: disable=too-many-arguments
 

texar/models/seq2seq/seq2seq_base.py

@@ -25,7 +25,7 @@
 from texar.losses.mle_losses import sequence_sparse_softmax_cross_entropy
 from texar.data.data.paired_text_data import PairedTextData
 from texar.core.optimization import get_train_op
-from texar import HParams
+from texar.hyperparams import HParams
 from texar.utils import utils
 from texar.utils.variables import collect_trainable_variables
 

texar/modules/berts/berts.py

@@ -19,7 +19,7 @@
 from __future__ import division
 from __future__ import print_function
 
-from texar import ModuleBase
+from texar.module_base import ModuleBase
 from texar.modules.berts import bert_utils
 
 __all__ = [

texar/modules/encoders/bert_encoders.py

@@ -23,7 +23,7 @@
 from texar.core import layers
 from texar.modules.encoders.transformer_encoders import TransformerEncoder
 from texar.modules.embedders import WordEmbedder, PositionEmbedder
-from texar import HParams
+from texar.hyperparams import HParams
 from texar.modules.berts import BertBase, bert_utils
 from texar.modules.encoders import EncoderBase
 

texar/modules/encoders/transformer_encoders.py

@@ -26,7 +26,7 @@
 from texar.modules.encoders.encoder_base import EncoderBase
 from texar.modules.encoders.multihead_attention import MultiheadAttentionEncoder
 from texar.modules.networks.networks import FeedForwardNetwork
-from texar import utils
+import texar.utils as utils
 from texar.utils.shapes import shape_list
 from texar.utils.mode import is_train_mode
 

texar/utils/mode.py

@@ -21,7 +21,7 @@
 
 import tensorflow as tf
 
-from texar import context
+import texar.context as context
 
 __all__ = [
     "maybe_global_mode",

texar/utils/transformer_attentions.py

@@ -23,7 +23,6 @@
 
 import tensorflow as tf
 
-from texar import context
 
 # pylint: disable=too-many-arguments, invalid-name, no-member