tensorflow serving 编译运行 nccl错误

梦里梦外; 2021-05-19 15:21 526阅读 0赞

由于自己编译，版本很重要。package下载出错时，多运行几遍。

**bazel安装：**  
[bazel 0.16.1][]

bash bazel-0.16.1-installer-linux-x86_64.sh
    export PATH="$PATH:$HOME/bin"

**tf serving安装：**  
[tf serving 1.12.0.zip][]

在WORKSPACE中  http_archive(  行前添加：
    load("@bazel_tools//tools/build_defs/repo:http.bzl", "http_archive")

#编译cpu版本：
    bazel build tensorflow_serving/model_servers:tensorflow_model_server

#编译gpu版本：
    export TF_NEED_CUDA=1
    export TF_CUDA_VERSION=10.0
    export TF_CUDNN_VERSION=7
    
    bazel build --config=cuda --copt="-fPIC" tensorflow_serving/model_servers:tensorflow_model_server
    
    如果nccl报错：
    export NCCL_HDR_PATH=/usr/local/nccl-2.4/include
    export NCCL_INSTALL_PATH=/usr/local/nccl-2.4/lib
    export TF_NCCL_VERSION=2.4.2
    export LD_LIBRARY_PATH=/usr/local/nccl-2.4/lib:$LD_LIBRARY_PATH

**运行服务：**

bazel-bin/tensorflow_serving/model_servers/tensorflow_model_server --port=9000 --rest_api_port=9001 --model_name=test --model_base_path=/mnt/ad_relevance/reimgdata/

**测试：**

curl http://localhost:9001/v1/models/test  # test为模型名称

import sys
    import cv2
    import time
    import numpy as np
    import tensorflow as tf
    from grpc.beta import implementations
    
    from tensorflow_serving.apis import predict_pb2
    from tensorflow_serving.apis import prediction_service_pb2
    
    num_per_request = 300
    size = 299
    central_fraction = 0.875
    HOST = "10.22.151.155"
    PORT = 9000
    channel = implementations.insecure_channel(HOST, PORT)
    stub = prediction_service_pb2.beta_create_PredictionService_stub(channel)
    
    def preprocess(img):
        print(img.shape)
        # img = img.astype(np.float32)
        h, w, c = img.shape
        # img = img * 1. / 255.
    
        bbox_start_h = int( (h - h * central_fraction) / 2 )
        bbox_start_w = int( (w - w * central_fraction) / 2 )
        bbox_end_h = bbox_start_h + h - bbox_start_h * 2
        bbox_end_w = bbox_start_w + w - bbox_start_w * 2
        img = img[bbox_start_h:bbox_end_h, bbox_start_w:bbox_end_w, :]
        print(img.shape)
        img = cv2.resize(img, (size, size), interpolation=cv2.INTER_LINEAR)
        img = img.astype(np.uint8)
        # img = (img - 0.5) * 2
        print(img.shape)
        # if use_fp:
        #     img = np.array(img).astype(np.float16)
        return img[:,:,::-1]
    
    def get(name, input_ids=0, input_mask=0, segment_ids=0, pos_input_ids=0, pos_input_mask=0,
            pos_segment_ids=0, pos_image=0.0 ):
        request = predict_pb2.PredictRequest()
        request.model_spec.name = 'test'
        request.model_spec.signature_name = name
        # request.inputs['pos_input_ids'].CopyFrom(tf.contrib.util.make_tensor_proto(pos_input_ids))
        # request.inputs['pos_input_mask'].CopyFrom(tf.contrib.util.make_tensor_proto(pos_input_mask))
        # request.inputs['pos_segment_ids'].CopyFrom(tf.contrib.util.make_tensor_proto(pos_segment_ids))
        if name=="serving_default":
            request.inputs['input_ids'].CopyFrom(tf.contrib.util.make_tensor_proto(input_ids))
            request.inputs['input_mask'].CopyFrom(tf.contrib.util.make_tensor_proto(input_mask))
            request.inputs['segment_ids'].CopyFrom(tf.contrib.util.make_tensor_proto(segment_ids))
        else:
            request.inputs['image_list'].CopyFrom(tf.contrib.util.make_tensor_proto(pos_image))
        result = stub.Predict(request, 1200)
        #print(result.outputs["image_emb"].float_val)
        return result
    
    def loop_request(func, **args):
        global total
        start = time.time()
        error = 0
        for i in range(total):
            print(i)
            try:
                func(**args)
            except Exception as e:
                print(e)
                error += 1
                continue
        end = time.time()
        print("time_rate: ", (end-start)/total, "error_rate: ", error/total)
    
    def test_cnn():
        #pos_image = np.full((1,299,299,3), 128).astype(np.float32)
        pos_image = cv2.imread("/data00/home/huangqingkang/imgs/1aacb000b0d93c755f995")
        pos_image = preprocess(pos_image)
        pos_image = np.expand_dims(pos_image, 0)
        pos_image = np.tile(pos_image, [num_per_request, 1, 1, 1])
                
        loop_request(get, name="serving_cnn", pos_image=pos_image)
    
    def test_bert():
        # pos_input_ids = np.full((num_per_request,32), 0).astype(np.int32)
        # pos_input_mask = np.full((num_per_request,32), 0).astype(np.int32)
        # pos_segment_ids = np.full((num_per_request,32), 0).astype(np.int32)
        input_ids = np.full((num_per_request+1,32), 0).astype(np.int32)
        input_mask = np.full((num_per_request+1,32), 0).astype(np.int32)
        segment_ids = np.full((num_per_request+1,32), 0).astype(np.int32)
        pos_image = np.array([0], dtype=np.uint8)
        # if use_fp:
        #     pos_image = np.array(pos_image).astype(np.float16)
    
        loop_request(get, name="serving_default", input_ids=input_ids, input_mask=input_mask, segment_ids=segment_ids, 
                pos_image = pos_image)
        # pos_input_ids = pos_input_ids, pos_input_mask = pos_input_mask, pos_segment_ids = pos_segment_ids,
    
    def get_cnn_emb(pos_image):
        result = get(name="serving_cnn", pos_image=pos_image)
        return result.output["image_emb"].float_val
    
    # def get_bert_emb(**args):
    
    
    
    if __name__ == "__main__":
        total = int(sys.argv[2])
        if int(sys.argv[3]):
            num_per_request = int(sys.argv[3])
        use_fp = int(sys.argv[4])
        if sys.argv[1] == "cnn":
            test_cnn()
        elif sys.argv[1] == "bert":
            test_bert()

[bazel 0.16.1]: https://github.com/bazelbuild/bazel/releases?after=0.19.2
[tf serving 1.12.0.zip]: https://github.com/tensorflow/serving/releases