[v0.8.x][BUGFIX] avoid using dict for attention cell parameter creation (#1051)

* avoid using dict for attention cell parameter creation

* fix

* Update test_sequence_sampler.py

* Update cache.py

Author: eric-haibin-lin

src/gluonnlp/model/attention_cell.py

@@ -209,8 +209,8 @@ def __init__(self, base_cell, query_units, key_units, value_units, num_heads, us
         self._base_cell = base_cell
         self._num_heads = num_heads
         self._use_bias = use_bias
-        units = {'query': query_units, 'key': key_units, 'value': value_units}
-        for name, unit in units.items():
+        units = [('query', query_units), ('key', key_units), ('value', value_units)]
+        for name, unit in units:
             if unit % self._num_heads != 0:
                 raise ValueError(
                     'In MultiHeadAttetion, the {name}_units should be divided exactly'

src/gluonnlp/model/train/cache.py

@@ -73,17 +73,13 @@ def __init__(self, lm_model, vocab_size, window, theta, lambdas, **kwargs):
         with self.name_scope():
             self.lm_model = lm_model
 
-    def save_parameters(self, filename, deduplicate=False):  # pylint: disable=arguments-differ
+    def save_parameters(self, filename):  # pylint: disable=arguments-differ
         """Save parameters to file.
 
         filename : str
             Path to file.
-        deduplicate : bool, default False
-            If True, save shared parameters only once. Otherwise, if a Block
-            contains multiple sub-blocks that share parameters, each of the
-            shared parameters will be separately saved for every sub-block.
         """
-        self.lm_model.save_parameters(filename, deduplicate=deduplicate)
+        self.lm_model.save_parameters(filename)
 
     def load_parameters(self, filename, ctx=mx.cpu()):  # pylint: disable=arguments-differ
         """Load parameters from file.

tests/unittest/test_sequence_sampler.py

@@ -53,8 +53,8 @@ def context_free_distribution(step_input, states):
         true_dist = dist.softmax().asnumpy()
     assert_allclose(true_dist, np.array(emp_dist), atol=0.01, rtol=0.1)
 
-# temporarily disabled model.HybridBeamSearchSampler test
-# due to https://github.com/dmlc/gluon-nlp/issues/706
+
+@pytest.mark.skip(reason='https://github.com/dmlc/gluon-nlp/issues/1020')
 @pytest.mark.seed(1)
 @pytest.mark.parametrize('hybridize', [False, True])
 @pytest.mark.parametrize('sampler_cls', [model.BeamSearchSampler, model.HybridBeamSearchSampler])