Delta-ML · zh794390558 · Nov 10, 2020 · Aug 27, 2020 · Aug 27, 2020 · Aug 28, 2020
diff --git a/.travis.yml b/.travis.yml
@@ -6,7 +6,7 @@ services:
 before_install:
   - export DELTA_PATH=`pwd`; echo $DELTA_PATH
   - export DOCKER_DELTA="/home/gitlab-runner/delta"; echo $DOCKER_DELTA
-  - export CI_IMAGE=zh794390558/delta:ci-cpu-py3
+  - export CI_IMAGE=zh794390558/delta:2.3.0-ci-cpu-py3
   - docker pull ${CI_IMAGE}
   - docker run -it -d --name travis_con --user root -v ${DELTA_PATH}:${DOCKER_DELTA} ${CI_IMAGE} bash
   - docker exec travis_con bash -c "gcc -v && g++ -v"

diff --git a/README.md b/README.md
@@ -118,9 +118,9 @@ To verify the installation, run:
 
 ```shell
 # Activate conda environment
-conda activate delta-py3.6-tf2.0.0
+conda activate delta-py3.6-tf2.3.0
 # Or use the following command if your conda version is < 4.6
-# source activate delta-py3.6-tf2.0.0
+# source activate delta-py3.6-tf2.3.0
 
 # Add DELTA environment
 source env.sh

diff --git a/core/ops/Makefile b/core/ops/Makefile
@@ -8,7 +8,7 @@ endif
 #$(info $(MAKEFILE_DIR))
 #$(info $(MAIN_ROOT))
 
-CXX := g++ 
+CXX := g++-7
 NVCC := nvcc
 PYTHON_BIN_PATH= python3
 CC :=

diff --git a/delta/data/feat/speech_ops_test.py b/delta/data/feat/speech_ops_test.py
@@ -209,13 +209,13 @@ def test_splice(self):
           self.assertTupleEqual(out.eval().shape, (1, 3, 2 * ctx))
           self.assertAllEqual(out, tf.ones([1, 3, 2 * ctx]))
 
-      with self.assertRaises(ValueError):
+      with self.assertRaises(tf.errors.InvalidArgumentError):
         out = tffeat.splice(feat, left_context=-2, right_context=-2).eval()
 
-      with self.assertRaises(ValueError):
+      with self.assertRaises(tf.errors.InvalidArgumentError):
         out = tffeat.splice(feat, left_context=2, right_context=-2).eval()
 
-      with self.assertRaises(ValueError):
+      with self.assertRaises(tf.errors.InvalidArgumentError):
         out = tffeat.splice(feat, left_context=-2, right_context=2).eval()
 
 

diff --git a/delta/data/preprocess/base_preparer.py b/delta/data/preprocess/base_preparer.py
@@ -144,7 +144,7 @@ def prepare_one_raw_data(self, one_path, one_path_after, mode,
       if self.multi_output:
         for i in range(self.output_num):
           label_ds = label[i].batch(self.batch_size)
-          label_iterator = label_ds.make_initializable_iterator()
+          label_iterator = tf.data.make_initializable_iterator(label_ds)
           label_after_arr = self.run_dataset(label_iterator, batch_num)
           label_after_one = [
               one_line.decode("utf-8") for one_line in label_after_arr
@@ -154,7 +154,7 @@ def prepare_one_raw_data(self, one_path, one_path_after, mode,
       else:
         label = label[0]
         label_ds = label.batch(self.batch_size)
-        label_iterator = label_ds.make_initializable_iterator()
+        label_iterator = tf.data.make_initializable_iterator(label_ds)
         label_after_arr = self.run_dataset(label_iterator, batch_num)
         one_label_after = [
             one_line.decode("utf-8") for one_line in label_after_arr

diff --git a/delta/data/preprocess/text_ops.py b/delta/data/preprocess/text_ops.py
@@ -101,7 +101,7 @@ def load_textline_dataset(paths, column_num):
   """Load raw data for text task."""
   ds = tf.data.TextLineDataset(paths)
   ds = ds.map(
-      lambda x: tf.strings.split(x, sep="\t", result_type="RaggedTensor"))
+    lambda x: tf.squeeze(tf.strings.split(x, sep="\t", result_type="RaggedTensor"), axis = 0))
   ds = ds.filter(lambda line: tf.equal(tf.size(line), column_num))
   ds_list = []
   for i in range(column_num):

diff --git a/delta/data/preprocess/text_ops_test.py b/delta/data/preprocess/text_ops_test.py
@@ -121,7 +121,7 @@ def test_process_one_label_dataset(self):
     true_res = [0, 0, 0, 8]
     label_ds = process_one_label_dataset(label_ds, self.config)
 
-    iterator = label_ds.make_initializable_iterator()
+    iterator = tf.data.make_initializable_iterator(label_ds)
     label_res = iterator.get_next()
 
     with tf.Session() as sess:
@@ -139,7 +139,7 @@ def test_process_multi_label_dataset(self):
     label_ds = tf.data.TextLineDataset(label_filepath)
     true_res = [[0, 8, 8], [0, 7, 8]]
     label_ds = process_multi_label_dataset(label_ds, self.config)
-    iterator = label_ds.make_initializable_iterator()
+    iterator = tf.data.make_initializable_iterator(label_ds)
     label_res = iterator.get_next()
 
     with tf.Session() as sess:

diff --git a/delta/data/preprocess/utils.py b/delta/data/preprocess/utils.py
@@ -37,7 +37,7 @@ def get_pre_process_text_ds_iter(
 
   text_ds = text_ds.batch(batch_size)
 
-  iterator = text_ds.make_initializable_iterator()
+  iterator = tf.data.make_initializable_iterator(text_ds)
 
   return iterator
 

diff --git a/delta/data/task/text_cls_task.py b/delta/data/task/text_cls_task.py
@@ -184,7 +184,7 @@ def dataset(self):
 
     data_set = data_set.prefetch(self.num_prefetch_batch)
 
-    iterator = data_set.make_initializable_iterator()
+    iterator = tf.data.make_initializable_iterator(data_set)
 
     # pylint: disable=unused-variable
     if self.infer_without_label:

diff --git a/delta/data/task/text_match_task.py b/delta/data/task/text_match_task.py
@@ -170,7 +170,7 @@ def dataset(self):
 
     text_ds_left_right = text_ds_left_right.prefetch(self.num_prefetch_batch)
 
-    iterator = text_ds_left_right.make_initializable_iterator()
+    iterator = tf.data.make_initializable_iterator(text_ds_left_right)
     # pylint: disable=unused-variable
     if self.infer_without_label:
       (input_x_left, input_x_right), (input_x_left_len,

diff --git a/delta/data/task/text_nlu_joint_task.py b/delta/data/task/text_nlu_joint_task.py
@@ -160,7 +160,7 @@ def dataset(self):
 
     data_set = data_set.prefetch(self.num_prefetch_batch)
 
-    iterator = data_set.make_initializable_iterator()
+    iterator = tf.data.make_initializable_iterator(data_set)
 
     if self.infer_without_label:
       input_x, input_x_len = iterator.get_next()

diff --git a/delta/data/task/text_seq2seq_task.py b/delta/data/task/text_seq2seq_task.py
@@ -224,7 +224,7 @@ def dataset(self):
 
     data_set = data_set.prefetch(self.num_prefetch_batch)
 
-    iterator = data_set.make_initializable_iterator()
+    iterator = tf.data.make_initializable_iterator(data_set)
 
     # pylint: disable=unused-variable
     if self.infer_without_label:

diff --git a/delta/data/task/text_seq_label_task.py b/delta/data/task/text_seq_label_task.py
@@ -139,7 +139,7 @@ def dataset(self):
 
     data_set = data_set.prefetch(self.num_prefetch_batch)
 
-    iterator = data_set.make_initializable_iterator()
+    iterator = tf.data.make_initializable_iterator(data_set)
 
     if self.infer_without_label:
       input_x, input_x_len = iterator.get_next()

diff --git a/delta/utils/metrics/py_metrics.py b/delta/utils/metrics/py_metrics.py
@@ -220,8 +220,8 @@ def call(self, y_true=None, y_pred=None, arguments=None):
         in_f.write(ref_sent)
         in_f.write("\n")
 
-    files_rouge = FilesRouge(self.hyp_path, self.ref_path)
-    scores = files_rouge.get_scores(avg=True)
+    files_rouge = FilesRouge()
+    scores = files_rouge.get_scores(self.hyp_path, self.ref_path, avg=True)
     return self.get_scores_output(scores)
 
   @staticmethod

diff --git a/delta/utils/solver/asr_solver.py b/delta/utils/solver/asr_solver.py
@@ -25,7 +25,7 @@
 from tensorflow.keras import backend as K
 from tensorflow.keras.models import Model
 from tensorflow.keras.layers import Lambda
-from tensorflow.keras.experimental import export_saved_model
+#from tensorflow.keras.experimental import export_saved_model
 
 from delta import utils
 from delta.utils.decode import py_ctc
@@ -265,10 +265,11 @@ def ctc_greedy_decode_lambda_func(args):
         inputs=[input_feat, input_length], outputs=greedy_decode)
 
     model_export_path = Path(self._model_path).joinpath("export")
-    export_saved_model(
-        model=model_to_export,
-        saved_model_path=str(model_export_path),
-        custom_objects=None,
-        as_text=False,
-        input_signature=None,
-        serving_only=False)
+    model_to_export.save(
+        filepath=str(model_export_path),
+        overwrite=True,
+        include_optimizer=True,
+        save_format='tf',
+        signatures=None,
+        options=None,
+    )
diff --git a/docker/build.sh b/docker/build.sh
@@ -34,7 +34,7 @@ fi
 
 TAG=${TF_VER}-${TARGET}-${DEVICE}-py3
 DOCKER='sudo docker'
-PIP_INSTALL="pip --no-cache-dir install -i https://pypi.tuna.tsinghua.edu.cn/simple/"
+PIP_INSTALL="pip --no-cache-dir install -i https://mirrors.aliyun.com/pypi/simple"
 
 set -e
 set -u
@@ -48,14 +48,19 @@ on_exit() {
 trap on_exit HUP INT PIPE QUIT TERM EXIT
 
 
+# https://hub.docker.com/r/tensorflow/tensorflow
+# Versioned images <= 1.15.0 (1.x) and <= 2.1.0 (2.x) have Python 3 
+# (3.5 for Ubuntu 16-based images; 3.6 for Ubuntu 18-based images) in images tagged "-py3"
+# and Python 2.7 in images without "py" in the tag. 
+# All newer images are Python 3 only. Tags containing -py3 are deprecated.
 if [ ${DEVICE} == 'cpu' ] && [ ${TARGET} == 'deltann' ];then
-  IMAGE=tensorflow/tensorflow:devel-py3
+  IMAGE=tensorflow/tensorflow:devel
 elif [ ${DEVICE} == 'gpu' ] && [ ${TARGET} == 'deltann' ];then
-  IMAGE=tensorflow/tensorflow:devel-gpu-py3
+  IMAGE=tensorflow/tensorflow:devel-gpu
 elif [ ${DEVICE} == 'cpu' ] && [ ${TARGET} == 'delta' ] || [ ${TARGET} == 'ci' ];then
-  IMAGE=tensorflow/tensorflow:${TF_VER}-py3
-elif [ ${DEVICE} == 'gpu' ] && [ ${TARGET} == 'delta' ] || [ ${TARGET} == 'ci' ];then
-  IMAGE=tensorflow/tensorflow:${TF_VER}-gpu-py3
+  IMAGE=tensorflow/tensorflow:${TF_VER}
+elif [ ${DEVICE} == 'gpu' ] && [ ${TARGET} == 'delta' ];then
+  IMAGE=tensorflow/tensorflow:${TF_VER}-gpu
 else
   echo "no support target or device"
   exit -1
@@ -134,11 +139,11 @@ if [ $MODE == 'push' ] || [ $MODE == 'build' ];then
   $DOCKER pull $IMAGE
 
   # build image
-  $DOCKER build --no-cache=false -t delta:$TAG -f $DOCKERFILE . || { echo "build ${TARGET} ${DEVICE} error"; exit 1; }
+  $DOCKER build --no-cache=false -t zh794390558/delta:$TAG -f $DOCKERFILE . || { echo "build ${TARGET} ${DEVICE} error"; exit 1; }
 
   #push image
   if [ $MODE == 'push' ];then
-    $DOCKER tag delta:${TAG} zh794390558/delta:${TAG}
+    #$DOCKER tag delta:${TAG} zh794390558/delta:${TAG}
     $DOCKER push zh794390558/delta:$TAG
 
     if [ $? == 0  ]; then

diff --git a/docker/dockerfile.ci.cpu b/docker/dockerfile.ci.cpu
@@ -1,4 +1,4 @@
-FROM tensorflow/tensorflow:2.0.0-py3
+FROM tensorflow/tensorflow:2.3.0
 COPY sources.list.ubuntu18.04 /etc/apt/sources.list
 
 # install tools 
@@ -7,6 +7,6 @@ RUN /bin/bash /install.sh
 
 COPY requirements.txt /ci/requirements.txt
 WORKDIR /ci
-RUN sudo pip --no-cache-dir install -i https://pypi.tuna.tsinghua.edu.cn/simple/ --upgrade pip && pip --no-cache-dir install -i https://pypi.tuna.tsinghua.edu.cn/simple/ --user -r requirements.txt
+RUN sudo pip --no-cache-dir install -i https://mirrors.aliyun.com/pypi/simple --upgrade pip && pip --no-cache-dir install -i https://mirrors.aliyun.com/pypi/simple --user -r requirements.txt
 
 CMD ["/bin/bash", "-c"]
diff --git a/docker/dockerfile.delta.cpu b/docker/dockerfile.delta.cpu
@@ -1,4 +1,4 @@
-FROM tensorflow/tensorflow:2.0.0-py3
+FROM tensorflow/tensorflow:2.3.0
 COPY sources.list.ubuntu18.04 /etc/apt/sources.list
 
 # install tools 

diff --git a/docker/dockerfile.delta.gpu b/docker/dockerfile.delta.gpu
@@ -1,4 +1,4 @@
-FROM tensorflow/tensorflow:2.0.0-gpu-py3
+FROM tensorflow/tensorflow:2.3.0-gpu
 COPY sources.list.ubuntu18.04 /etc/apt/sources.list
 
 # install tools 

diff --git a/docker/dockerfile.deltann.cpu b/docker/dockerfile.deltann.cpu
@@ -1,4 +1,4 @@
-FROM tensorflow/tensorflow:devel-py3
+FROM tensorflow/tensorflow:devel
 COPY sources.list.ubuntu18.04 /etc/apt/sources.list
 
 # install tools 

diff --git a/docker/dockerfile.deltann.gpu b/docker/dockerfile.deltann.gpu
@@ -1,4 +1,4 @@
-FROM tensorflow/tensorflow:devel-gpu-py3
+FROM tensorflow/tensorflow:devel-gpu
 COPY sources.list.ubuntu18.04 /etc/apt/sources.list
 
 # install tools 

diff --git a/docker/run.sh → docker/gen_dockerfile.sh b/docker/run.sh → docker/gen_dockerfile.sh
@@ -1,5 +1,7 @@
 #!/bin/bash
 
+rm dockerfile.delta*
+
 bash build.sh ci cpu  dockerfile
 bash build.sh delta cpu  dockerfile
 bash build.sh delta gpu  dockerfile

diff --git a/docker/install.sh b/docker/install.sh
@@ -12,7 +12,6 @@ apt-get update && apt-get install -y --no-install-recommends \
         vim \
         zlib1g-dev \
         wget \
-        gfortran \
         subversion \
         ca-certificates \
         unzip \

diff --git a/docker/requirements.txt b/docker/requirements.txt
@@ -5,9 +5,11 @@
 matplotlib
 sklearn
 pandas
-librosa
+librosa>=0.7.2
+numba==0.48.0  #must be this by librosa>=0.7.2
 absl-py
 jieba
+wget
 kaldiio
 soundfile
 textgrid
@@ -28,4 +30,10 @@ pyAudioAnalysis
 flake8
 sentencepiece
 deepdiff
-tensorflow_addons
+tensorflow-addons==0.11.1
+tensorflow-model-optimization
+
+
+# TODO: rm below
+# https://github.com/tensorflow/addons/issues/864
+tensorflow-cpu==2.3.0
diff --git a/docs/released_models.md b/docs/released_models.md
@@ -50,6 +50,3 @@ Recently several deep learning based approaches have been successfully used in s
 ### Textual+numeric
 
 We implement the direct concatenation data fusion in data processing stage,therefore this type of multimodal training can be directly used for existing models in DELTA.
-
-
-
diff --git a/docs/version.md b/docs/version.md
@@ -20,5 +20,3 @@ Make sure all PRs under milestone `v0.3.2` are closed, then close the milestone.
 Using below command to generate relase note.
 
 `python tools/release_notes.py -c didi delta v0.3.2`
-
-
diff --git a/egs/mini_an4/asr/v1/conf/asr-ctc.yml b/egs/mini_an4/asr/v1/conf/asr-ctc.yml
@@ -88,7 +88,7 @@ solver:
   metrics:
     pos_label: 1 # int, same to sklearn
     metrics_used : null
-    monitor_used : val_token_err
+    monitor_used : val_loss
     cals:
     - name: AccuracyCal
       arguments: null

diff --git a/env.sh b/env.sh
@@ -14,7 +14,7 @@ else
 fi
 
 # need version
-export TF_VER=2.0.0
+export TF_VER=2.3.0
 export PY_VER=3.6
 
 # root dir
@@ -51,6 +51,7 @@ if [ -e $MAIN_ROOT/tools/go.env ];then
   source $MAIN_ROOT/tools/go.env
 fi
 
+# maybe used by deltann
 # tensorflow lib path
-TF_LIB_PATH=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_lib())')
-export LD_LIBRARY_PATH=${TF_LIB_PATH}:${LD_LIBRARY_PATH}
+#TF_LIB_PATH=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_lib())')
+#export LD_LIBRARY_PATH=${TF_LIB_PATH}:${LD_LIBRARY_PATH}
diff --git a/tools/install/prepare_kaldi.sh b/tools/install/prepare_kaldi.sh
@@ -40,8 +40,9 @@ pushd ${MAIN_ROOT}/tools/kaldi/tools
 extras/check_dependencies.sh || ${SUDO} apt-get install -y zlib1g-dev wget gfortran subversion
 
 SPH2PIPE_VERSION=v2.5
+test -e sph2pipe_${SPH2PIPE_VERSION}.tar.gz && rm sph2pipe_${SPH2PIPE_VERSION}.tar.gz
 wget -T 10 -t 3 https://www.openslr.org/resources/3/sph2pipe_${SPH2PIPE_VERSION}.tar.gz || wget -T 10 https://sourceforge.net/projects/kaldi/files/sph2pipe_${SPH2PIPE_VERSION}.tar.gz || exit 1
 tar --no-same-owner -xzf sph2pipe_v2.5.tar.gz
-cd sph2pipe_v2.5/
-gcc -o sph2pipe  *.c -lm
+cd sph2pipe_${SPH2PIPE_VERSION}; patch -p1 < ${MAIN_ROOT}/tools/install/sph2pipe.patch; gcc -o sph2pipe  *.c -lm; cd -
+touch  sph2pipe_${SPH2PIPE_VERSION}/.patched
 popd
diff --git a/tools/install/sph2pipe.patch b/tools/install/sph2pipe.patch
@@ -0,0 +1,23 @@
+diff --git a/sph2pipe.c b/sph2pipe.c
+index d0d502b..928739c 100644
+--- a/sph2pipe.c
++++ b/sph2pipe.c
+@@ -103,6 +103,7 @@
+
+ #define _SPH_CONVERT_MAIN_
+
++#include <unistd.h>
+ #include "sph_convert.h"
+ #include "ulaw.h"
+
+diff --git a/ulaw.h b/ulaw.h
+index 990c833..62ff96e 100644
+--- a/ulaw.h
++++ b/ulaw.h
+@@ -82,3 +82,6 @@ GLOBAL short int alaw2pcm[256]
+ }
+ #endif
+ ;
++
++uchar pcm2ulaw( short int sample );
++uchar pcm2alaw( short int pcmval );
Original file line number	Diff line number	Diff line change
Expand Up		@@ -50,6 +50,3 @@ Recently several deep learning based approaches have been successfully used in s
		### Textual+numeric

		We implement the direct concatenation data fusion in data processing stage,therefore this type of multimodal training can be directly used for existing models in DELTA.
Original file line number	Diff line number	Diff line change
Expand Up		@@ -20,5 +20,3 @@ Make sure all PRs under milestone `v0.3.2` are closed, then close the milestone.
		Using below command to generate relase note.

		`python tools/release_notes.py -c didi delta v0.3.2`