Update requirements versions, change tf.compat.v1. usages to tf.

PiperOrigin-RevId: 288659456

alphafold_casp13/contacts.py

@@ -71,10 +71,10 @@ def evaluate(crop_size_x, crop_size_y, feature_normalization, checkpoint_path,
       normalization_exclusion=normalization_exclusion)
 
   checkpoint = snt.get_saver(experiment.model, collections=[
-      tf.compat.v1.GraphKeys.GLOBAL_VARIABLES,
+      tf.GraphKeys.GLOBAL_VARIABLES,
-      tf.compat.v1.GraphKeys.MOVING_AVERAGE_VARIABLES])
+      tf.GraphKeys.MOVING_AVERAGE_VARIABLES])
 
-  with tf.compat.v1.train.SingularMonitoredSession(hooks=[]) as sess:
+  with tf.train.SingularMonitoredSession(hooks=[]) as sess:
     logging.info('Restoring from checkpoint %s', checkpoint_path)
     checkpoint.restore(sess, checkpoint_path)
 

alphafold_casp13/contacts_dataset.py

@@ -167,7 +167,7 @@ def parse_tfexample(raw_data, features):
   for k, v in parsed_features.items():
     new_shape = shape(feature_name=k, num_residues=num_residues)
     # Make sure the feature we are reshaping is not empty.
-    assert_non_empty = tf.compat.v1.assert_greater(
+    assert_non_empty = tf.assert_greater(
         tf.size(v), 0, name='assert_%s_non_empty' % k,
         message='The feature %s is not set in the tf.Example. Either do not '
         'request the feature or use a tf.Example that has the feature set.' % k)
@@ -245,7 +245,7 @@ def normalize_from_stats_file(
       train_mean = tf.cast(norm_stats['mean'][key], dtype=tf.float32)
       train_range = tf.sqrt(tf.cast(norm_stats['var'][key], dtype=tf.float32))
       value -= train_mean
-      value = tf.compat.v1.where(
+      value = tf.where(
           train_range > range_epsilon, value / train_range, value)
       features[key] = value
     else:

alphafold_casp13/contacts_experiment.py

@@ -22,12 +22,12 @@ from alphafold_casp13 import contacts_network
 
 
 def _int_ph(shape, name):
-  return tf.compat.v1.placeholder(
+  return tf.placeholder(
       dtype=tf.int32, shape=shape, name=('%s_placeholder' % name))
 
 
 def _float_ph(shape, name):
-  return tf.compat.v1.placeholder(
+  return tf.placeholder(
       dtype=tf.float32, shape=shape, name=('%s_placeholder' % name))
 
 
@@ -102,7 +102,7 @@ class Contacts(object):
       dataset = dataset.batch(1)
 
       # Get a batch of tensors in the legacy ProteinsDataset format.
-      iterator = tf.compat.v1.data.make_one_shot_iterator(dataset)
+      iterator = tf.data.make_one_shot_iterator(dataset)
       self._input_batch = iterator.get_next()
 
       self.num_eval_examples = sum(

alphafold_casp13/contacts_network.py

@@ -41,7 +41,7 @@ def call_on_tuple(f):
 
 
 class ContactsNet(sonnet.AbstractModule):
-  """A network to go from sequence to secondary structure."""
+  """A network to go from sequence to distance histograms."""
 
   def __init__(self,
                binary_code_bits,
@@ -102,7 +102,7 @@ class ContactsNet(sonnet.AbstractModule):
       if self.asa_multiplier > 0:
         self._asa = asa_output.ASAOutputLayer()
       if self._position_specific_bias_size:
-        self._position_specific_bias = tf.compat.v1.get_variable(
+        self._position_specific_bias = tf.get_variable(
             'position_specific_bias',
             [self._position_specific_bias_size, self._num_bins or 1],
             initializer=tf.zeros_initializer())
@@ -338,7 +338,7 @@ class ContactsNet(sonnet.AbstractModule):
     layers_forward = None
     if config_2d_deep.extra_blocks:
       # Optionally put some extra double-size blocks at the beginning.
-      with tf.compat.v1.variable_scope('Deep2DExtra'):
+      with tf.variable_scope('Deep2DExtra'):
         hidden_2d = two_dim_resnet.make_two_dim_resnet(
             input_node=hidden_2d,
             num_residues=None,  # Unused
@@ -362,7 +362,7 @@ class ContactsNet(sonnet.AbstractModule):
         if features_forward is not None:
           hidden_2d = tf.concat([hidden_2d, features_forward], 1
                                 if data_format == 'NCHW' else 3)
-    with tf.compat.v1.variable_scope('Deep2D'):
+    with tf.variable_scope('Deep2D'):
       logging.info('2d hidden shape is %s', str(hidden_2d.shape.as_list()))
       contact_pre_logits = two_dim_resnet.make_two_dim_resnet(
           input_node=hidden_2d,

alphafold_casp13/requirements.txt

@@ -1,6 +1,7 @@
-absl-py>=0.8.0
+setuptools==41.0.0
-numpy>=1.16
+absl-py==0.8.1
-six>=1.12
+numpy==1.16
-dm-sonnet>=1.35
+six==1.12
+dm-sonnet==1.35
 tensorflow==1.14
 tensorflow-probability==0.7.0

alphafold_casp13/secstruct.py

@@ -55,7 +55,7 @@ class Secstruct(object):
 
   def make_layer_new(self, activations):
     """Make the layer."""
-    with tf.compat.v1.variable_scope(self.name, reuse=tf.AUTO_REUSE):
+    with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
       logging.info('Creating secstruct %s', activations)
       self.logits = contrib_layers.linear(activations, self._dimension)
       self.ss_q8_probs = tf.nn.softmax(self.logits)

alphafold_casp13/two_dim_convnet.py

@@ -22,12 +22,12 @@ from tensorflow.contrib import layers as contrib_layers
 def weight_variable(shape, stddev=0.01):
   """Returns the weight variable."""
   logging.vlog(1, 'weight init for shape %s', str(shape))
-  return tf.compat.v1.get_variable(
+  return tf.get_variable(
       'w', shape, initializer=tf.random_normal_initializer(stddev=stddev))
 
 
 def bias_variable(shape):
-  return tf.compat.v1.get_variable(
+  return tf.get_variable(
       'b', shape, initializer=tf.zeros_initializer())
 
 
@@ -61,12 +61,12 @@ def make_conv_sep2d_layer(input_node,
     filter_size_2 = filter_size
   logging.vlog(1, 'layer %s in %d out %d chan mult %d', layer_name, in_channels,
                out_channels, channel_multiplier)
-  with tf.compat.v1.variable_scope(layer_name):
+  with tf.variable_scope(layer_name):
-    with tf.compat.v1.variable_scope('depthwise'):
+    with tf.variable_scope('depthwise'):
       w_depthwise = weight_variable(
           [filter_size, filter_size_2, in_channels, channel_multiplier],
           stddev=stddev)
-    with tf.compat.v1.variable_scope('pointwise'):
+    with tf.variable_scope('pointwise'):
       w_pointwise = weight_variable(
           [1, 1, in_channels * channel_multiplier, out_channels], stddev=stddev)
     h_conv = tf.nn.separable_conv2d(
@@ -119,7 +119,7 @@ def make_conv_layer(input_node,
     filter_size_2 = filter_size
   logging.vlog(
       1, 'layer %s in %d out %d', layer_name, in_channels, out_channels)
-  with tf.compat.v1.variable_scope(layer_name):
+  with tf.variable_scope(layer_name):
     w_conv = weight_variable(
         [filter_size, filter_size_2, in_channels, out_channels], stddev=stddev)
     h_conv = conv2d(