Resolve conflicts

Author: Lifann

.github/workflows/ci_test.yml

@@ -10,7 +10,7 @@ on:
       - master
       - r*
 env:
-  USE_BAZEL_VERSION: "3.7.2"
+  USE_BAZEL_VERSION: "5.1.1"
 
 jobs:
   yapf-test:

.github/workflows/release.yml

@@ -17,7 +17,7 @@ on:
 env:
   MIN_PY_VERSION: '3.7'
   MAX_PY_VERSION: '3.9'
-  USE_BAZEL_VERSION: "3.7.2"
+  USE_BAZEL_VERSION: "5.1.1"
   HOROVOD_VERSION: '0.23.0'
 
 jobs:
@@ -168,11 +168,15 @@ jobs:
     name: Upload dev container to DockerHub
     needs: [release-wheel, test-with-bazel]
     runs-on: ubuntu-18.04
+    strategy:
+      matrix:
+        py-version: ['3.7', '3.8', '3.9']
     if: (github.event_name == 'push' && github.ref == 'refs/heads/master')
     steps:
       - uses: actions/checkout@v2
       - run: |
           set -e -x
           echo ${{ secrets.DOCKER_PW }} | docker login --username ${{ secrets.DOCKER_USER }} --password-stdin
+          export PY_VERSION=${{ matrix.py-version }}
           bash .github/workflows/github_build_dev_container.sh
-          docker push tfra/dev_container:latest-cpu
+          docker push tfra/dev_container:latest-python${{ matrix.py-version }}

README.md

@@ -134,8 +134,15 @@ but maybe this doc can help you : [Extract headers from TensorFlow compiling dir
 At the same time, we find some OPs used by TRFA have better performance, so we highly recommend you update TensorFlow to 2.x.
 
 #### Installing from Source
+
+For all developers, we recommend you use the development docker containers which are all GPU enabled:
+```shell
+docker pull tfra/dev_container:latest-python3.8  # "3.7", "3.9" are all avaliable.
+docker run --privileged --gpus all -it --rm -v $(pwd):$(pwd) tfra/dev_container:latest-3.8
+```
+
 ##### CPU Only
-You can also install from source. This requires the [Bazel](https://bazel.build/) build system (version == 3.7.2).
+You can also install from source. This requires the [Bazel](https://bazel.build/) build system (version == 5.1.1).
 Please install a TensorFlow on your compiling machine, The compiler needs to know the version of Tensorflow and 
 its headers according to the installed TensorFlow. 
 
@@ -158,18 +165,20 @@ pip install artifacts/tensorflow_recommenders_addons-*.whl
 ##### GPU Support
 Only `TF_NEED_CUDA=1` is required and other environment variables are optional:
 ```shell
-PY_VERSION="3.7" \
-TF_NEED_CUDA=1 \
-TF_CUDA_VERSION=11.2 \
-TF_CUDNN_VERSION=8.1 \
-CUDA_TOOLKIT_PATH="/usr/local/cuda" \
-CUDNN_INSTALL_PATH="/usr/lib/x86_64-linux-gnu" \
+export TF_VERSION="2.5.1"  # "2.7.0", "2.5.1" are well tested.
+export PY_VERSION="3.8" 
+export TF_NEED_CUDA=1
+export TF_CUDA_VERSION=11.2
+export TF_CUDNN_VERSION=8.1
+export CUDA_TOOLKIT_PATH="/usr/local/cuda"
+export CUDNN_INSTALL_PATH="/usr/lib/x86_64-linux-gnu"
+
 python configure.py
 ```
 And then build the pip package and install:
 ```shell
 bazel build --enable_runfiles build_pip_pkg
-TF_NEED_CUDA=1 bazel-bin/build_pip_pkg artifacts
+bazel-bin/build_pip_pkg artifacts
 pip install artifacts/tensorflow_recommenders_addons_gpu-*.whl
 ```
 ##### Apple Silicon Support (Beta Release)

configure.py

@@ -40,10 +40,10 @@ def _VALID_BAZEL_VERSION(tf_version):
         'refering to the previous COMMIT to compile properly by themselves.')
     return target_bazel
   elif tf_version >= "2.0.0":
-    target_bazel = "3.7.2"
+    target_bazel = "5.1.1"
     logging.info(
         'To ensure code compatibility with Bazel rules_foreign_cc component, '
-        'we specify Bazel version greater than 3.7.2 '
+        'we specify Bazel version greater than 5.1.1 '
         'for Tensorflow versions greater than 2.0.0.')
     return target_bazel
   else:

tensorflow_recommenders_addons/dynamic_embedding/core/kernels/cuckoo_hashtable_op_gpu.cu.cc

@@ -115,7 +115,6 @@ class CuckooHashTableOfTensorsGpu final : public LookupInterface {
               const Tensor& default_value) override {
     size_t len = d_keys.flat<K>().size();
     bool* d_status;
-    gpu::ValueArrayBase<V>* d_default_value;
 
     auto value_flat = value->flat_inner_dims<V, 2>();
     const auto default_flat = default_value.flat<V>();
@@ -140,7 +139,6 @@ class CuckooHashTableOfTensorsGpu final : public LookupInterface {
         CUDA_CHECK(cudaStreamSynchronize(_stream));
       }
       CUDA_CHECK(cudaFree(d_status));
-      CUDA_CHECK(cudaFree(d_default_value));
       CUDA_CHECK(cudaStreamDestroy(_stream));
     }
     return Status::OK();
@@ -150,7 +148,6 @@ class CuckooHashTableOfTensorsGpu final : public LookupInterface {
                         Tensor* value, const Tensor& default_value,
                         Tensor* exists) {
     size_t len = d_keys.flat<K>().size();
-    gpu::ValueArrayBase<V>* d_default_value;
 
     auto value_flat = value->flat_inner_dims<V, 2>();
     const auto default_flat = default_value.flat<V>();
@@ -173,7 +170,6 @@ class CuckooHashTableOfTensorsGpu final : public LookupInterface {
                     _stream, is_full_default);
         CUDA_CHECK(cudaStreamSynchronize(_stream));
       }
-      CUDA_CHECK(cudaFree(d_default_value));
       CUDA_CHECK(cudaStreamDestroy(_stream));
     }
     return Status::OK();

tensorflow_recommenders_addons/dynamic_embedding/core/kernels/lookup_impl/lookup_table_op_cpu.h

@@ -116,17 +116,21 @@ class TableWrapperBase {
  public:
   virtual ~TableWrapperBase() {}
   virtual bool insert_or_assign(K key, ConstTensor2D<V>& value_flat,
-                                int64 value_dim, int64 index) {}
+                                int64 value_dim, int64 index) {
+    return false;
+  }
   virtual bool insert_or_accum(K key, ConstTensor2D<V>& value_or_delta_flat,
-                               bool exist, int64 value_dim, int64 index) {}
+                               bool exist, int64 value_dim, int64 index) {
+    return false;
+  }
   virtual void find(const K& key, Tensor2D<V>& value_flat,
                     ConstTensor2D<V>& default_flat, int64 value_dim,
                     bool is_full_size_default, int64 index) const {}
   virtual void find(const K& key, Tensor2D<V>& value_flat,
                     ConstTensor2D<V>& default_flat, bool& exist,
                     int64 value_dim, bool is_full_size_default,
                     int64 index) const {}
-  virtual size_t size() const {}
+  virtual size_t size() const { return 0; }
   virtual void clear() {}
   virtual bool erase(const K& key) {}
   virtual Status export_values(OpKernelContext* ctx, int64 value_dim) {}

tensorflow_recommenders_addons/dynamic_embedding/core/kernels/lookup_impl/lookup_table_op_gpu.h

@@ -71,8 +71,8 @@ class TableWrapperBase {
   virtual void get(const K* d_keys, ValueType<V>* d_vals, bool* d_status,
                    size_t len, ValueType<V>* d_def_val, cudaStream_t stream,
                    bool is_full_size_default) const {}
-  virtual size_t get_size(cudaStream_t stream) const {}
-  virtual size_t get_capacity() const {}
+  virtual size_t get_size(cudaStream_t stream) const { return 0; }
+  virtual size_t get_capacity() const { return 0; }
   virtual void remove(const K* d_keys, size_t len, cudaStream_t stream) {}
   virtual void clear(cudaStream_t stream) {}
 };

tensorflow_recommenders_addons/dynamic_embedding/python/keras/layers/embedding.py

@@ -311,17 +311,23 @@ class FieldWiseEmbedding(BasicEmbedding):
   ```python
   nslots = 3
   @tf.function
-  def map_slot_fn(feature_id):
+  def feature_to_slot(feature_id):
     field_id = tf.math.mod(feature_id, nslots)
     return field_id
 
   ids = tf.constant([[23, 12, 0], [9, 13, 10]], dtype=tf.int64)
-  embedding = de.layers.FieldWiseEmbedding(1, nslots, map_slot_fn)
+  embedding = de.layers.FieldWiseEmbedding(2,
+                                           nslots,
+                                           slot_map_fn=feature_to_slot,
+                                           initializer=tf.keras.initializer.Zeros())
+
+  out = embedding(ids)
+  # [[[0., 0.], [0., 0.], [0., 1.]]
+  #  [[0., 0.], [0., 0.], [0., 1.]]]
 
   prepared_keys = tf.range(0, 100, dtype=tf.int64)
   prepared_values = tf.ones((100, 2), dtype=tf.float32)
   embedding.params.upsert(prepared_keys, prepared_values)
-
   out = embedding(ids)
   # [[2., 2.], [0., 0.], [1., 1.]]
   # [[1., 1.], [2., 2.], [0., 0.]]

tools/build_dev_container.sh

@@ -1,10 +1,11 @@
 #!/usr/bin/env bash
 
 set -x -e
-
-DOCKER_BUILDKIT=1 docker build \
+docker build \
     -f tools/docker/dev_container.Dockerfile \
     --build-arg TF_VERSION=2.5.1 \
-    --build-arg TF_PACKAGE=tensorflow-cpu \
-    --target dev_container_cpu \
-    -t tfaddons/dev_container:latest-cpu ./
+    --build-arg TF_PACKAGE=tensorflow-gpu \
+    --build-arg PY_VERSION=$PY_VERSION \
+    --no-cache \
+    --target dev_container \
+    -t tfra/dev_container:latest-python$PY_VERSION ./

tools/docker/cpu_tests.Dockerfile

@@ -2,7 +2,7 @@
 FROM python:3.7 as build_wheel
 
 ARG TF_VERSION=2.5.1
-ARG USE_BAZEL_VERSION=3.7.2
+ARG USE_BAZEL_VERSION=5.1.1
 ARG MPI_VERSION="4.1.1"
 ARG HOROVOD_VERSION="0.23.0"