From 4311142624c678468a925952b6e157673b0f1771 Mon Sep 17 00:00:00 2001 From: liuqi Date: Tue, 6 Mar 2018 20:50:43 +0800 Subject: [PATCH] Support multiple input or output API. --- benchmark.sh | 8 ++- env.sh | 3 +- example.yaml | 8 +-- generate_data.py | 58 ++++++++++++++++++ tuning_run.sh | 31 +++++++--- validate.py | 119 +++++++++++++++++++++++------------- validate_caffe.py | 152 ---------------------------------------------- validate_tools.sh | 44 ++++++++++---- 8 files changed, 199 insertions(+), 224 deletions(-) create mode 100644 generate_data.py delete mode 100644 validate_caffe.py diff --git a/benchmark.sh b/benchmark.sh index ca450b22..7e404642 100644 --- a/benchmark.sh +++ b/benchmark.sh @@ -58,7 +58,11 @@ else cp bazel-bin/benchmark/benchmark_model $MODEL_OUTPUT_DIR adb shell "mkdir -p ${PHONE_DATA_DIR}" || exit 1 - adb push ${MODEL_OUTPUT_DIR}/${INPUT_FILE_NAME} ${PHONE_DATA_DIR} || exit 1 + IFS=',' read -r -a INPUT_NAMES <<< "${INPUT_NODE}" + for NAME in "${INPUT_NAMES[@]}";do + FORMATTED_NAME=$(sed s/[^[:alnum:]]/_/g <<< ${NAME}) + adb push ${MODEL_OUTPUT_DIR}/${INPUT_FILE_NAME}_${FORMATTED_NAME} ${PHONE_DATA_DIR} || exit 1 + done adb push ${MODEL_OUTPUT_DIR}/benchmark_model ${PHONE_DATA_DIR} || exit 1 if [ "$EMBED_MODEL_DATA" = 0 ]; then adb push ${MODEL_OUTPUT_DIR}/${MODEL_TAG}.data ${PHONE_DATA_DIR} || exit 1 @@ -73,7 +77,9 @@ else ${PHONE_DATA_DIR}/benchmark_model \ --model_data_file=${PHONE_DATA_DIR}/${MODEL_TAG}.data \ --device=${DEVICE_TYPE} \ + --input_node="${INPUT_NODE}" \ --input_shape="${INPUT_SHAPE}"\ + --output_node="${OUTPUT_NODE}" \ --output_shape="${OUTPUT_SHAPE}"\ --input_file=${PHONE_DATA_DIR}/${INPUT_FILE_NAME} || exit 1 fi diff --git a/env.sh b/env.sh index 9e7896ce..d7e46fb9 100644 --- a/env.sh +++ b/env.sh @@ -3,8 +3,7 @@ LIBMACE_TAG=`git describe --abbrev=0 --tags` LIBMACE_SOURCE_DIR=`/bin/pwd` INPUT_FILE_NAME="model_input" -OUTPUT_FILE_NAME="model.out" -OUTPUT_LIST_FILE="model.list" +OUTPUT_FILE_NAME="model_out" PHONE_DATA_DIR="/data/local/tmp/mace_run" KERNEL_DIR="${PHONE_DATA_DIR}/cl/" CODEGEN_DIR=${LIBMACE_SOURCE_DIR}/codegen diff --git a/example.yaml b/example.yaml index f34c1298..d24870f2 100644 --- a/example.yaml +++ b/example.yaml @@ -25,10 +25,10 @@ models: weight_file_path: path/to/weight.caffemodel model_sha256_checksum: 05d92625809dc9edd6484882335c48c043397aed450a168d75eb8b538e86881a weight_sha256_checksum: 05d92625809dc9edd6484882335c48c043397aed450a168d75eb8b538e86881a - input_node: input_node - output_node: output_node - input_shape: 1,256,256,3 - output_shape: 1,256,256,2 + input_node: input_node0,input_node1 + output_node: output_node0,output_node1 + input_shape: 1,256,256,3:1,128,128,3 + output_shape: 1,256,256,2:1,1,1,2 runtime: cpu limit_opencl_kernel_time: 1 dsp_mode: 0 diff --git a/generate_data.py b/generate_data.py new file mode 100644 index 00000000..24ac6c23 --- /dev/null +++ b/generate_data.py @@ -0,0 +1,58 @@ +import argparse +import sys +import os +import os.path +import numpy as np +import re +from scipy import spatial + +# Validation Flow: +# 1. Generate input data +# python generate_data.py \ +# --input_node input_node \ +# --input_shape 1,64,64,3 \ +# --input_file input_file +# + +def generate_data(name, shape): + np.random.seed() + data = np.random.random(shape) * 2 - 1 + input_file_name = FLAGS.input_file + "_" + re.sub('[^0-9a-zA-Z]+', '_', name) + print 'Generate input file: ', input_file_name + data.astype(np.float32).tofile(input_file_name) + +def main(unused_args): + input_names = [name for name in FLAGS.input_node.split(',')] + input_shapes = [shape for shape in FLAGS.input_shape.split(':')] + assert len(input_names) == len(input_shapes) + for i in range(len(input_names)): + shape = [int(x) for x in input_shapes[i].split(',')] + generate_data(input_names[i], shape) + print "Generate input file done." + +def parse_args(): + """Parses command line arguments.""" + parser = argparse.ArgumentParser() + parser.register("type", "bool", lambda v: v.lower() == "true") + parser.add_argument( + "--input_file", + type=str, + default="", + help="input file.") + parser.add_argument( + "--input_node", + type=str, + default="input_node", + help="input node") + parser.add_argument( + "--input_shape", + type=str, + default="1,64,64,3", + help="input shape.") + + return parser.parse_known_args() + +if __name__ == '__main__': + FLAGS, unparsed = parse_args() + main(unused_args=[sys.argv[0]] + unparsed) + diff --git a/tuning_run.sh b/tuning_run.sh index 0f35dc0a..0203b5a2 100644 --- a/tuning_run.sh +++ b/tuning_run.sh @@ -20,13 +20,15 @@ PRODUCTION_MODE=$4 if [ x"$TARGET_ABI" = x"host" ]; then MACE_CPP_MIN_VLOG_LEVEL=$VLOG_LEVEL \ ${MODEL_OUTPUT_DIR}/mace_run \ - --input_shape="${INPUT_SHAPE}"\ - --output_shape="${OUTPUT_SHAPE}"\ - --input_file=${MODEL_OUTPUT_DIR}/${INPUT_FILE_NAME} \ - --output_file=${MODEL_OUTPUT_DIR}/${OUTPUT_FILE_NAME} \ - --model_data_file=${MODEL_OUTPUT_DIR}/${MODEL_TAG}.data \ - --device=${DEVICE_TYPE} \ - --round=1 || exit 1 + --input_node="${INPUT_NODE}" \ + --input_shape="${INPUT_SHAPE}"\ + --output_node="${OUTPUT_NODE}" \ + --output_shape="${OUTPUT_SHAPE}"\ + --input_file=${MODEL_OUTPUT_DIR}/${INPUT_FILE_NAME} \ + --output_file=${MODEL_OUTPUT_DIR}/${OUTPUT_FILE_NAME} \ + --model_data_file=${MODEL_OUTPUT_DIR}/${MODEL_TAG}.data \ + --device=${DEVICE_TYPE} \ + --round=1 || exit 1 else if [[ "${TUNING_OR_NOT}" != "0" && "$PRODUCTION_MODE" != 1 ]];then tuning_flag=1 @@ -38,7 +40,14 @@ else if [ "$PRODUCTION_MODE" = 0 ]; then adb shell "mkdir -p ${KERNEL_DIR}" || exit 1 fi - adb push ${MODEL_OUTPUT_DIR}/${INPUT_FILE_NAME} ${PHONE_DATA_DIR} || exit 1 + + IFS=',' read -r -a INPUT_NAMES <<< "${INPUT_NODE}" + for NAME in "${INPUT_NAMES[@]}";do + FORMATTED_NAME=$(sed s/[^[:alnum:]]/_/g <<< ${NAME}) + echo $FORMATTED_NAME + adb push ${MODEL_OUTPUT_DIR}/${INPUT_FILE_NAME}_${FORMATTED_NAME} ${PHONE_DATA_DIR} || exit 1 + done + adb push ${MODEL_OUTPUT_DIR}/mace_run ${PHONE_DATA_DIR} || exit 1 if [ "$EMBED_MODEL_DATA" = 0 ]; then adb push ${MODEL_OUTPUT_DIR}/${MODEL_TAG}.data ${PHONE_DATA_DIR} || exit 1 @@ -53,8 +62,10 @@ else MACE_KERNEL_PATH=$KERNEL_DIR \ MACE_LIMIT_OPENCL_KERNEL_TIME=${LIMIT_OPENCL_KERNEL_TIME} \ ${PHONE_DATA_DIR}/mace_run \ - --input_shape=${INPUT_SHAPE}\ - --output_shape=${OUTPUT_SHAPE}\ + --input_node="${INPUT_NODE}" \ + --input_shape="${INPUT_SHAPE}"\ + --output_node="${OUTPUT_NODE}" \ + --output_shape="${OUTPUT_SHAPE}"\ --input_file=${PHONE_DATA_DIR}/${INPUT_FILE_NAME} \ --output_file=${PHONE_DATA_DIR}/${OUTPUT_FILE_NAME} \ --model_data_file=${PHONE_DATA_DIR}/${MODEL_TAG}.data \ diff --git a/validate.py b/validate.py index 62401489..6f27383f 100644 --- a/validate.py +++ b/validate.py @@ -2,18 +2,12 @@ import argparse import sys import os import os.path -import tensorflow as tf import numpy as np +import re from scipy import spatial -from tensorflow import gfile - # Validation Flow: # 1. Generate input data -# python validate.py --generate_data true \ -# --input_file input_file -# --input_shape 1,64,64,3 -# # 2. Use mace_run to run model on phone. # 3. adb pull the result. # 4. Compare output data of mace and tf @@ -25,23 +19,18 @@ from tensorflow import gfile # --input_shape 1,64,64,3 \ # --output_shape 1,64,64,2 -def generate_data(shape): - np.random.seed() - data = np.random.random(shape) * 2 - 1 - print FLAGS.input_file - data.astype(np.float32).tofile(FLAGS.input_file) - print "Generate input file done." - def load_data(file): if os.path.isfile(file): return np.fromfile(file=file, dtype=np.float32) else: return np.empty([0]) -def valid_output(out_shape, mace_out_file, tf_out_value): - mace_out_value = load_data(mace_out_file) +def format_output_name(name): + return re.sub('[^0-9a-zA-Z]+', '_', name) + +def compare_output(mace_out_value, out_value): if mace_out_value.size != 0: - similarity = (1 - spatial.distance.cosine(tf_out_value.flat, mace_out_value)) + similarity = (1 - spatial.distance.cosine(out_value.flat, mace_out_value.flat)) print 'MACE VS TF similarity: ', similarity if (FLAGS.mace_runtime == "cpu" and similarity > 0.999) or \ (FLAGS.mace_runtime == "gpu" and similarity > 0.995) or \ @@ -53,13 +42,14 @@ def valid_output(out_shape, mace_out_file, tf_out_value): print '=======================Skip empty node===================' -def run_model(input_shape): - if not gfile.Exists(FLAGS.model_file): +def validate_tf_model(input_names, input_shapes, output_names): + import tensorflow as tf + if not os.path.isfile(FLAGS.model_file): print("Input graph file '" + FLAGS.model_file + "' does not exist!") sys.exit(-1) input_graph_def = tf.GraphDef() - with gfile.Open(FLAGS.model_file, "rb") as f: + with open(FLAGS.model_file, "rb") as f: data = f.read() input_graph_def.ParseFromString(data) tf.import_graph_def(input_graph_def, name="") @@ -67,35 +57,85 @@ def run_model(input_shape): with tf.Session() as session: with session.graph.as_default() as graph: tf.import_graph_def(input_graph_def, name="") - input_node = graph.get_tensor_by_name(FLAGS.input_node + ':0') - output_node = graph.get_tensor_by_name(FLAGS.output_node + ':0') + input_dict = {} + for i in range(len(input_names)): + input_value = load_data(FLAGS.input_file + "_" + input_names[i]) + input_value = input_value.reshape(input_shapes[i]) + input_node = graph.get_tensor_by_name(input_names[i] + ':0') + input_dict[input_node] = input_value + + output_nodes = [] + for name in output_names: + output_nodes.extend([graph.get_tensor_by_name(name + ':0')]) + output_values = session.run(output_nodes, feed_dict=input_dict) + for i in range(len(output_names)): + output_file_name = FLAGS.mace_out_file + "_" + format_output_name(output_names[i]) + mace_out_value = load_data(output_file_name) + compare_output(mace_out_value, output_values[i]) + +def validate_caffe_model(input_names, input_shapes, output_names, output_shapes): + os.environ['GLOG_minloglevel'] = '1' # suprress Caffe verbose prints + import caffe + if not os.path.isfile(FLAGS.model_file): + print("Input graph file '" + FLAGS.model_file + "' does not exist!") + sys.exit(-1) + if not os.path.isfile(FLAGS.weight_file): + print("Input weight file '" + FLAGS.weight_file + "' does not exist!") + sys.exit(-1) - input_value = load_data(FLAGS.input_file) - input_value = input_value.reshape(input_shape) - - output_value = session.run(output_node, feed_dict={input_node: input_value}) - output_value.astype(np.float32).tofile( os.path.dirname(FLAGS.input_file) + '/tf_out') - return output_value + caffe.set_mode_cpu() -def main(unused_args): - input_shape = [int(x) for x in FLAGS.input_shape.split(',')] - output_shape = [int(x) for x in FLAGS.output_shape.split(',')] - if FLAGS.generate_data: - generate_data(input_shape) - else: - output_value = run_model(input_shape) - valid_output(output_shape, FLAGS.mace_out_file, output_value) + net = caffe.Net(FLAGS.model_file, caffe.TEST, weights=FLAGS.weight_file) + + for i in range(len(input_names)): + input_value = load_data(FLAGS.input_file + "_" + input_names[i]) + input_value = input_value.reshape(input_shapes[i]).transpose((0, 3, 1, 2)) + net.blobs[input_names[i]].data[0] = input_value + + net.forward() + for i in range(len(output_names)): + value = net.blobs[output_names[i]].data[0] + out_shape = output_shapes[i] + out_shape[1], out_shape[2], out_shape[3] = out_shape[3], out_shape[1], out_shape[2] + value = value.reshape(out_shape).transpose((0, 2, 3, 1)) + output_file_name = FLAGS.mace_out_file + "_" + format_output_name(output_names[i]) + mace_out_value = load_data(output_file_name) + compare_output(mace_out_value, value) + +def main(unused_args): + input_names = [name for name in FLAGS.input_node.split(',')] + input_shape_strs = [shape for shape in FLAGS.input_shape.split(':')] + input_shapes = [[int(x) for x in shape.split(',')] for shape in input_shape_strs] + output_names = [name for name in FLAGS.output_node.split(',')] + assert len(input_names) == len(input_shapes) + + if FLAGS.platform == 'tensorflow': + validate_tf_model(input_names, input_shapes, output_names) + elif FLAGS.platform == 'caffe': + output_shape_strs = [shape for shape in FLAGS.output_shape.split(':')] + output_shapes = [[int(x) for x in shape.split(',')] for shape in output_shape_strs] + validate_caffe_model(input_names, input_shapes, output_names, output_shapes) def parse_args(): """Parses command line arguments.""" parser = argparse.ArgumentParser() parser.register("type", "bool", lambda v: v.lower() == "true") + parser.add_argument( + "--platform", + type=str, + default="", + help="Tensorflow or Caffe.") parser.add_argument( "--model_file", type=str, default="", - help="TensorFlow \'GraphDef\' file to load.") + help="TensorFlow or Caffe \'GraphDef\' file to load.") + parser.add_argument( + "--weight_file", + type=str, + default="", + help="caffe model file to load.") parser.add_argument( "--input_file", type=str, @@ -131,11 +171,6 @@ def parse_args(): type=str, default="output_node", help="output node") - parser.add_argument( - "--generate_data", - type='bool', - default="false", - help="Generate data or not.") return parser.parse_known_args() diff --git a/validate_caffe.py b/validate_caffe.py deleted file mode 100644 index cc50242f..00000000 --- a/validate_caffe.py +++ /dev/null @@ -1,152 +0,0 @@ -import argparse -import sys -import os -import os.path -import numpy as np -from scipy import spatial - -os.environ['GLOG_minloglevel'] = '1' # suprress Caffe verbose prints -import caffe - -# Validation Flow: -# 1. Generate input data -# python validate.py --generate_data true \ -# --input_file input_file -# --input_shape 1,64,64,3 -# -# 2. Use mace_run to run model on phone. -# 3. adb pull the result. -# 4. Compare output data of mace and tf -# python validate.py --model_file tf_model_opt.pb \ -# --input_file input_file \ -# --mace_out_file output_file \ -# --input_node input_node \ -# --output_node output_node \ -# --input_shape 1,64,64,3 \ -# --output_shape 1,64,64,2 - -def generate_data(shape): - np.random.seed() - data = np.random.random(shape) * 2 - 1 - print FLAGS.input_file - data.astype(np.float32).tofile(FLAGS.input_file) - print "Generate input file done." - -def load_data(file): - if os.path.isfile(file): - return np.fromfile(file=file, dtype=np.float32) - else: - return np.empty([0]) - -def valid_output(out_shape, mace_out_file, out_value): - mace_out_value = load_data(mace_out_file) - if mace_out_value.size != 0: - mace_out_value = mace_out_value.reshape(out_shape) - out_shape[1], out_shape[2], out_shape[3] = out_shape[3], out_shape[1], out_shape[2] - out_value = out_value.reshape(out_shape).transpose((0, 2, 3, 1)) - similarity = (1 - spatial.distance.cosine(out_value.flat, mace_out_value.flat)) - print 'MACE VS Caffe similarity: ', similarity - if (FLAGS.mace_runtime == "cpu" and similarity > 0.999) or \ - (FLAGS.mace_runtime == "gpu" and similarity > 0.995) or \ - (FLAGS.mace_runtime == "dsp" and similarity > 0.930): - print '=======================Similarity Test Passed======================' - else: - print '=======================Similarity Test Failed======================' - else: - print '=======================Skip empty node===================' - - -def run_model(input_shape): - if not os.path.isfile(FLAGS.model_file): - print("Input graph file '" + FLAGS.model_file + "' does not exist!") - sys.exit(-1) - if not os.path.isfile(FLAGS.weight_file): - print("Input weight file '" + FLAGS.weight_file + "' does not exist!") - sys.exit(-1) - - caffe.set_mode_cpu() - - net = caffe.Net(FLAGS.model_file, caffe.TEST, weights=FLAGS.weight_file) - - input_value = load_data(FLAGS.input_file) - input_value = input_value.reshape(input_shape).transpose((0, 3, 1, 2)) - net.blobs[FLAGS.input_node].data[0] = input_value - net.forward(start=FLAGS.input_node, end=FLAGS.output_node) - - result = net.blobs[FLAGS.output_node].data[0] - - return result - - -def main(unused_args): - input_shape = [int(x) for x in FLAGS.input_shape.split(',')] - output_shape = [int(x) for x in FLAGS.output_shape.split(',')] - if FLAGS.generate_data: - generate_data(input_shape) - else: - output_value = run_model(input_shape) - valid_output(output_shape, FLAGS.mace_out_file, output_value) - - -def parse_args(): - """Parses command line arguments.""" - parser = argparse.ArgumentParser() - parser.register("type", "bool", lambda v: v.lower() == "true") - parser.add_argument( - "--model_file", - type=str, - default="", - help="caffe prototxt file to load.") - parser.add_argument( - "--weight_file", - type=str, - default="", - help="caffe model file to load.") - parser.add_argument( - "--input_file", - type=str, - default="", - help="input file.") - parser.add_argument( - "--mace_out_file", - type=str, - default="", - help="mace output file to load.") - parser.add_argument( - "--mace_runtime", - type=str, - default="gpu", - help="mace runtime device.") - parser.add_argument( - "--input_shape", - type=str, - default="1,64,64,3", - help="input shape.") - parser.add_argument( - "--output_shape", - type=str, - default="1,64,64,2", - help="output shape.") - parser.add_argument( - "--input_node", - type=str, - default="input_node", - help="input node") - parser.add_argument( - "--output_node", - type=str, - default="output_node", - help="output node") - parser.add_argument( - "--generate_data", - type='bool', - default="false", - help="Generate data or not.") - - return parser.parse_known_args() - - -if __name__ == '__main__': - FLAGS, unparsed = parse_args() - main(unused_args=[sys.argv[0]] + unparsed) - diff --git a/validate_tools.sh b/validate_tools.sh index dbbb719e..fefbce48 100644 --- a/validate_tools.sh +++ b/validate_tools.sh @@ -15,21 +15,31 @@ source ${CURRENT_DIR}/env.sh MODEL_OUTPUT_DIR=$1 GENERATE_DATA_OR_NOT=$2 +IFS=',' read -r -a INPUT_NAMES <<< "${INPUT_NODE}" +IFS=',' read -r -a OUTPUT_NAMES <<< "${OUTPUT_NODE}" + echo $MODEL_OUTPUT_DIR if [ "$GENERATE_DATA_OR_NOT" = 1 ]; then - rm -rf ${MODEL_OUTPUT_DIR}/${INPUT_FILE_NAME} - python tools/validate.py --generate_data true \ - --input_file=${MODEL_OUTPUT_DIR}/${INPUT_FILE_NAME} \ - --input_shape="${INPUT_SHAPE}" || exit 1 + for NAME in "${INPUT_NAMES[@]}";do + FORMATTED_NAME=$(sed s/[^[:alnum:]]/_/g <<< ${NAME}) + rm -rf ${MODEL_OUTPUT_DIR}/${INPUT_FILE_NAME}_${FORMATTED_NAME} + done + python tools/generate_data.py --input_node=${INPUT_NODE} \ + --input_file=${MODEL_OUTPUT_DIR}/${INPUT_FILE_NAME} \ + --input_shape="${INPUT_SHAPE}" || exit 1 exit 0 fi if [ "$PLATFORM" == "tensorflow" ];then if [[ x"$TARGET_ABI" -ne x"host" ]]; then - rm -rf ${MODEL_OUTPUT_DIR}/${OUTPUT_FILE_NAME} - adb