[Distributed] Upgrade the SOK version to v4.2. (#551)

Author: shijieliu

addons/sparse_operation_kit/README.md

@@ -77,14 +77,18 @@ There are 2 additional steps to use SOK embedding lookup with `Embedding Variabl
 git clone --recursive https://github.com/alibaba/DeepRec.git /DeepRec
 ```
 2. Build DeepRec from source code. You can follow the instruction in [DeepRec Build](https://deeprec.readthedocs.io/zh/latest/DeepRec-Compile-And-Install.html).
+3. Enter HugeCTR and do initialization
+```bash
+cd /DeepRec/addons/sparse_operation_kit/hugectr/ && git submodule update --init --recursive && cd sparse_operation_kit && mkdir build && cd build
+```
 3. Build and install SOK. You need to specify ${DeepRecBuild}, which is the directory that you store the DeepRec building intermidia results.
 ```
 DeepRecBuild=${DeepRecBuild} cmake -DENABLE_DEEPREC=ON -DSM=75 .. && make -j && make install;
 export PYTHONPATH=/DeepRec/addons/sparse_operation_kit/hugectr/sparse_operation_kit
 ```
 4. Run utest.
 ```
-cd /DeepRec/addons/sparse_operation_kit/adapter && horovodrun -np ${NUM_GPU} -H localhost:${NUM_GPU} python3 embedding_var_lookup_utest.py
+cd /DeepRec/addons/sparse_operation_kit/python && horovodrun -np ${NUM_GPU} -H localhost:${NUM_GPU} python3 embedding_var_lookup_utest.py
 ```
 ## Benchmark
 1. Download Kaggle Display Advertising Challenge Dataset (Criteo Dataset) from https://storage.googleapis.com/dataset-uploader/criteo-kaggle/large_version/train.csv

addons/sparse_operation_kit/example/demo.py

@@ -23,7 +23,7 @@
   values=tf.convert_to_tensor([1, 1, 3, 4, 5], dtype=tf.int64),
   dense_shape=[2, 3]
 )
-emb = sok.lookup_sparse([var], [indices], hotness=[3], combiners=['sum'])
+emb = sok.lookup_sparse([var], [indices], combiners=['sum'])
 fun = tf.multiply(emb, 2.0, name='multiply')
 loss = tf.reduce_sum(fun, name='reduce_sum')
 opt = tf.train.AdagradOptimizer(0.1)

addons/sparse_operation_kit/hugectr

@@ -81,7 +81,7 @@
     optimizer = tf.train.AdagradOptimizer(0.1)
 
     def step(params):
-        embeddings = sok.lookup_sparse(params, indices, hotness, combiners)
+        embeddings = sok.lookup_sparse(params, indices, combiners)
         loss = 0
         for i in range(len(embeddings)):
             loss = loss + tf.reduce_sum(embeddings[i])