Enable PyPI installation

* Fix or toggle off warnings (mostly TF deprecation warnings)
* Update dependency to TF>=1.10.0
* Upgrade to version v0.2.2

Author: ZhitingHu

CHANGELOG.md

@@ -5,6 +5,16 @@
 
 ### Feature improvements
 
+### Fixes
+
+## [v0.2.2](https://github.com/asyml/texar/releases/tag/v0.2.2) (2019-08-05)
+
+### New features
+
+* Enable installation from [Pypi](https://pypi.org/project/texar/). ([#186](https://github.com/asyml/texar/pull/186))
+
+### 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

README.md

@@ -4,6 +4,8 @@
 
 -----------------
 
+
+[![pypi](https://img.shields.io/pypi/v/texar.svg)](https://pypi.python.org/pypi/texar)
 [![Build Status](https://travis-ci.org/asyml/texar.svg?branch=master)](https://travis-ci.org/asyml/texar)
 [![Documentation Status](https://readthedocs.org/projects/texar/badge/?version=latest)](https://texar.readthedocs.io/en/latest/?badge=latest)
 [![License](https://img.shields.io/badge/license-Apache%202.0-blue.svg)](https://github.com/asyml/texar/blob/master/LICENSE)
@@ -79,6 +81,13 @@ agent = tx.agents.SeqPGAgent(samples=outputs.sample_id,
 Many more examples are available [here](./examples)
 
 ### Installation
+Besides the packages in requirements.txt, Texar additionally requires:
+
+* tensorflow >= 1.10.0 (but <= 2.0). Follow the tensorflow official instructions to install the appropriate version
+
+* tensorflow_probability >= 0.3.0. Follow the tensorflow_probability official instractions to install.
+
+After Tensorflow is installed, please run the following commands to install Texar:
 ```
 git clone https://github.com/asyml/texar.git
 cd texar

examples/bert/bert_classifier_main.py

@@ -107,7 +107,7 @@ def main(_):
     input_ids = batch["input_ids"]
     segment_ids = batch["segment_ids"]
     batch_size = tf.shape(input_ids)[0]
-    input_length = tf.reduce_sum(1 - tf.to_int32(tf.equal(input_ids, 0)),
+    input_length = tf.reduce_sum(1 - tf.cast(tf.equal(input_ids, 0), tf.int32),
                                  axis=1)
     # Builds BERT
     with tf.variable_scope('bert'):
@@ -198,7 +198,7 @@ def _train_epoch(sess):
         periodically.
         """
         iterator.restart_dataset(sess, 'train')
-        
+
         fetches = {
             'train_op': train_op,
             'loss': loss,

examples/bert/bert_classifier_main_v2.py

@@ -85,7 +85,7 @@ def main(_):
     input_ids = batch["input_ids"]
     segment_ids = batch["segment_ids"]
     batch_size = tf.shape(input_ids)[0]
-    input_length = tf.reduce_sum(1 - tf.to_int32(tf.equal(input_ids, 0)),
+    input_length = tf.reduce_sum(1 - tf.cast(tf.equal(input_ids, 0), tf.int32),
                                  axis=1)
     # Builds BERT
     hparams = {

examples/seq2seq_exposure_bias/interpolation_helper.py

@@ -22,6 +22,7 @@
 import tensorflow as tf
 import numpy as np
 
+from tensorflow_probability import distributions as tfpd
 from tensorflow.contrib.seq2seq import SampleEmbeddingHelper
 from texar.evals.bleu import sentence_bleu
 from rouge import Rouge
@@ -95,19 +96,19 @@ def sample(self, time, outputs, state, name=None):
         of 'state'([decoded_ids, rnn_state])
         """
         sample_method_sampler = \
-            tf.distributions.Categorical(probs=self._lambdas)
+            tfpd.Categorical(probs=self._lambdas)
         sample_method_id = sample_method_sampler.sample()
 
         truth_feeding = lambda: tf.cond(
             tf.less(time, tf.shape(self._ground_truth)[1]),
-            lambda: tf.to_int32(self._ground_truth[:, time]),
+            lambda: tf.cast(self._ground_truth[:, time], tf.int32),
             lambda: tf.ones_like(self._ground_truth[:, 0],
                                  dtype=tf.int32) * self._vocab.eos_token_id)
 
-        self_feeding = lambda : SampleEmbeddingHelper.sample(
+        self_feeding = lambda: SampleEmbeddingHelper.sample(
             self, time, outputs, state, name)
 
-        reward_feeding = lambda : self._sample_by_reward(time, state)
+        reward_feeding = lambda: self._sample_by_reward(time, state)
 
         sample_ids = tf.cond(
             tf.logical_or(tf.equal(time, 0), tf.equal(sample_method_id, 1)),
@@ -207,9 +208,9 @@ def _get_rewards(time, prefix_ids, target_ids, ground_truth_length):
 
             return result
 
-        sampler = tf.distributions.Categorical(
+        sampler = tfpd.Categorical(
             logits=tf.py_func(_get_rewards, [
                 time, state[0], self._ground_truth,
                 self._ground_truth_length], tf.float32))
         return tf.reshape(
-            sampler.sample(), (tf.shape(self._ground_truth)[0],))
+            sampler.sample(), (tf.shape(self._ground_truth)[0],))

examples/text_style_transfer/ctrl_gen_model.py

@@ -55,7 +55,7 @@ def _build_model(self, inputs, vocab, gamma, lambda_g):
         label_connector = MLPTransformConnector(self._hparams.dim_c)
 
         # Gets the sentence representation: h = (c, z)
-        labels = tf.to_float(tf.reshape(inputs['labels'], [-1, 1]))
+        labels = tf.cast(tf.reshape(inputs['labels'], [-1, 1]), tf.float32)
         c = label_connector(labels)
         c_ = label_connector(1 - labels)
         h = tf.concat([c, z], 1)
@@ -106,7 +106,7 @@ def _build_model(self, inputs, vocab, gamma, lambda_g):
             inputs=clas_embedder(ids=inputs['text_ids'][:, 1:]),
             sequence_length=inputs['length']-1)
         loss_d_clas = tf.nn.sigmoid_cross_entropy_with_logits(
-            labels=tf.to_float(inputs['labels']), logits=clas_logits)
+            labels=tf.cast(inputs['labels'], tf.float32), logits=clas_logits)
         loss_d_clas = tf.reduce_mean(loss_d_clas)
         accu_d = tx.evals.accuracy(labels=inputs['labels'], preds=clas_preds)
 
@@ -115,7 +115,7 @@ def _build_model(self, inputs, vocab, gamma, lambda_g):
             inputs=clas_embedder(soft_ids=soft_outputs_.sample_id),
             sequence_length=soft_length_)
         loss_g_clas = tf.nn.sigmoid_cross_entropy_with_logits(
-            labels=tf.to_float(1-inputs['labels']), logits=soft_logits)
+            labels=tf.cast(1-inputs['labels'], tf.float32), logits=soft_logits)
         loss_g_clas = tf.reduce_mean(loss_g_clas)
 
         # Accuracy on soft samples, for training progress monitoring

examples/transformer/transformer_main.py

@@ -22,8 +22,8 @@
 import random
 import os
 import importlib
-from torchtext import data
 import tensorflow as tf
+from torchtext import data
 import texar as tx
 from texar.modules import TransformerEncoder, TransformerDecoder
 from texar.utils import transformer_utils
@@ -74,10 +74,10 @@ def main():
     batch_size = tf.shape(encoder_input)[0]
     # (text sequence length excluding padding)
     encoder_input_length = tf.reduce_sum(
-        1 - tf.to_int32(tf.equal(encoder_input, 0)), axis=1)
+        1 - tf.cast(tf.equal(encoder_input, 0), tf.int32), axis=1)
 
     labels = tf.placeholder(tf.int64, shape=(None, None))
-    is_target = tf.to_float(tf.not_equal(labels, 0))
+    is_target = tf.cast(tf.not_equal(labels, 0), tf.float32)
 
     global_step = tf.Variable(0, dtype=tf.int64, trainable=False)
     learning_rate = tf.placeholder(tf.float64, shape=(), name='lr')

requirements.txt

@@ -1,5 +1,6 @@
-tensorflow >= 1.7.0
-tensorflow-gpu >= 1.7.0
+tensorflow >= 1.10.0
+tensorflow-gpu >= 1.10.0
 tensorflow-probability >= 0.3.0
 tensorflow-probability-gpu >= 0.3.0
-funcsigs >= 1.0.2
+funcsigs >= 1.0.2
+packaging

setup.py

@@ -15,7 +15,7 @@
 
 setuptools.setup(
     name="texar",
-    version="0.2.1",
+    version="0.2.2",
     url="https://github.com/asyml/texar",
 
     description="Toolkit for Machine Learning and Text Generation",
@@ -26,19 +26,20 @@
     platforms='any',
 
     install_requires=[
-        'numpy',
+        'numpy<1.17.0',
         'pyyaml',
         'requests',
-        'funcsigs',
+        'funcsigs>=1.0.2',
+        'packaging'
     ],
     extras_require={
         'tensorflow-cpu': [
-            'tensorflow>=1.7.0',
-            'tensorflow-probability >= 0.3.0'
+            'tensorflow>=1.10.0,<2.0',
+            'tensorflow-probability>=0.3.0'
         ],
         'tensorflow-gpu': [
-            'tensorflow-gpu>=1.7.0',
-            'tensorflow-probability-gpu >= 0.3.0'
+            'tensorflow-gpu>=1.10.0,<2.0',
+            'tensorflow-probability-gpu>=0.3.0'
         ]
     },
     package_data={

texar/__init__.py

@@ -19,9 +19,20 @@
 from __future__ import division
 from __future__ import print_function
 
+
 # pylint: disable=wildcard-import
 
 import sys
+from packaging import version
+import tensorflow as tf
+
+
+VERSION_WARNING = "1.13.2"
+
+if version.parse(tf.__version__) <= version.parse(VERSION_WARNING):
+    tf.logging.set_verbosity(tf.logging.ERROR)
+else:
+    tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
 
 if sys.version_info.major < 3:
     # PY 2.x, import as is because Texar-PyTorch cannot be installed.

texar/agents/dqn_agent.py

@@ -292,7 +292,7 @@ def _get_target_outputs(self, state_inputs):
         return self._target(inputs=state_inputs, **self._qnet_caller_kwargs)
 
     def _get_td_error(self, qnet_qvalues, actions, y):
-        return y - tf.reduce_sum(qnet_qvalues * tf.to_float(actions), axis=1)
+        return y - tf.reduce_sum(qnet_qvalues * tf.cast(actions, tf.float), axis=1)
 
     def _get_train_op(self):
         train_op = opt.get_train_op(

texar/core/layers.py

@@ -573,7 +573,7 @@ def get_layer(hparams):
     if not is_str(layer_type) and not isinstance(layer_type, type):
         layer = layer_type
     else:
-        layer_modules = ["tensorflow.layers", "texar.core", "texar.costum"]
+        layer_modules = ["tensorflow.layers", "texar.core", "texar.custom"]
         layer_class = utils.check_or_get_class(layer_type, layer_modules)
         if isinstance(hparams, dict):
             default_kwargs = _layer_class_to_default_kwargs_map.get(layer_class,

texar/core/optimization.py

@@ -259,8 +259,8 @@ def get_learning_rate_decay_fn(hparams=None):
     if fn_kwargs is HParams:
         fn_kwargs = fn_kwargs.todict()
 
-    start_step = tf.to_int32(hparams["start_decay_step"])
-    end_step = tf.to_int32(hparams["end_decay_step"])
+    start_step = tf.cast(hparams["start_decay_step"], tf.int32)
+    end_step = tf.cast(hparams["end_decay_step"], tf.int32)
 
     def lr_decay_fn(learning_rate, global_step):
         """Learning rate decay function.
@@ -273,7 +273,8 @@ def lr_decay_fn(learning_rate, global_step):
             scalar float Tensor: decayed learning rate.
         """
         offset_global_step = tf.maximum(
-            tf.minimum(tf.to_int32(global_step), end_step) - start_step, 0)
+            tf.minimum(tf.cast(global_step, tf.int32), end_step) - start_step,
+            0)
         if decay_fn == tf.train.piecewise_constant:
             decayed_lr = decay_fn(x=offset_global_step, **fn_kwargs)
         else:

texar/custom/__init__.py

@@ -14,10 +14,14 @@
 import copy
 import shutil
 import tempfile
+import ssl
 import tensorflow as tf
 import texar as tx
 
 
+ssl._create_default_https_context = ssl._create_unverified_context
+
+
 class TFRecordDataTest(tf.test.TestCase):
     """Tests tfrecord data class.
     """

texar/data/data/tfrecord_data_test.py

@@ -630,7 +630,7 @@ def decode(self, data, items):
             if from_type is to_type:
                 continue
             elif to_type is tf.string:
-                decoded_data[key] = tf.dtypes.as_string(decoded_data[key])
+                decoded_data[key] = tf.as_string(decoded_data[key])
             elif from_type is tf.string:
                 decoded_data[key] = tf.string_to_number(
                     decoded_data[key], to_type)

texar/data/data_decoders.py

@@ -182,7 +182,7 @@ def map_ids_to_tokens(self, ids):
         Returns:
             A tensor of text tokens of the same shape.
         """
-        return self.id_to_token_map.lookup(tf.to_int64(ids))
+        return self.id_to_token_map.lookup(tf.cast(ids, tf.int64))
 
     def map_tokens_to_ids(self, tokens):
         """Maps text tokens into ids.

texar/data/vocabulary.py

@@ -26,7 +26,7 @@ def accuracy(labels, preds):
         A float scalar Tensor containing the accuracy.
     """
     labels = tf.cast(labels, preds.dtype)
-    return tf.reduce_mean(tf.to_float(tf.equal(preds, labels)))
+    return tf.reduce_mean(tf.cast(tf.equal(preds, labels), tf.float32))
 
 def binary_clas_accuracy(pos_preds=None, neg_preds=None):
     """Calculates the accuracy of binary predictions.
@@ -44,7 +44,7 @@ def binary_clas_accuracy(pos_preds=None, neg_preds=None):
     """
     pos_accu = accuracy(tf.ones_like(pos_preds), pos_preds)
     neg_accu = accuracy(tf.zeros_like(neg_preds), neg_preds)
-    psize = tf.to_float(tf.size(pos_preds))
-    nsize = tf.to_float(tf.size(neg_preds))
+    psize = tf.cast(tf.size(pos_preds), tf.float32)
+    nsize = tf.cast(tf.size(neg_preds), tf.float32)
     accu = (pos_accu * psize + neg_accu * nsize) / (psize + nsize)
     return accu

texar/evals/metrics.py

@@ -102,7 +102,7 @@ def test_sequence_sigmoid_cross_entropy(self):
         labels = tf.placeholder(dtype=tf.int32, shape=None)
         loss = tx.losses.sequence_sigmoid_cross_entropy(
             logits=self._logits[:, :, 0],
-            labels=tf.to_float(labels),
+            labels=tf.cast(labels, tf.float32),
             sequence_length=self._sequence_length)
         with self.test_session() as sess:
             rank = sess.run(

texar/losses/mle_losses_test.py

@@ -30,4 +30,4 @@
 from texar.modules.policies import *
 from texar.modules.qnets import *
 from texar.modules.memory import *
-from texar.modules.berts import *
+from texar.modules.berts import *

texar/modules/__init__.py

@@ -273,7 +273,7 @@ def _build(self,
             else:
                 pred = tf.argmax(logits, axis=-1)
             pred = tf.reshape(pred, [-1])
-        pred = tf.to_int64(pred)
+        pred = tf.cast(pred, tf.int64)
 
         if not self._built:
             self._add_internal_trainable_variables()

texar/modules/classifiers/bert_classifiers.py

@@ -189,7 +189,7 @@ def _build(self,    # pylint: disable=arguments-differ
             logits = tf.reshape(logits, [-1])
         else:
             pred = tf.argmax(logits, 1)
-        pred = tf.to_int64(tf.reshape(pred, [-1]))
+        pred = tf.cast(tf.reshape(pred, [-1]), tf.int64)
 
         self._built = True
 

texar/modules/classifiers/conv_classifiers.py

@@ -328,7 +328,7 @@ def _build(self,
             else:
                 pred = tf.argmax(logits, axis=-1)
             pred = tf.reshape(pred, [-1])
-        pred = tf.to_int64(pred)
+        pred = tf.cast(pred, tf.int64)
 
         if not self._built:
             self._add_internal_trainable_variables()

texar/modules/classifiers/rnn_classifiers.py

@@ -664,7 +664,7 @@ class instance.
         dstr = check_or_get_instance(
             distribution, distribution_kwargs,
             ["tensorflow.distributions", "tensorflow_probability.distributions",
-             "tensorflow.contrib.distributions", "texar.custom"])
+             "texar.custom"])
 
         if num_samples:
             sample = dstr.sample(num_samples)