Ejemplo n.º 1
0
 def test_pull_box_sparse_op(self):
     paddle.enable_static()
     program = fluid.Program()
     with fluid.program_guard(program):
         x = fluid.layers.data(name='x',
                               shape=[1],
                               dtype='int64',
                               lod_level=0)
         y = fluid.layers.data(name='y',
                               shape=[1],
                               dtype='int64',
                               lod_level=0)
         emb_x, emb_y = _pull_box_sparse([x, y], size=1)
Ejemplo n.º 2
0
    def run_boxps_preload(self, is_cpu=True):
        x = fluid.layers.data(name='x', shape=[1], dtype='int64', lod_level=0)
        y = fluid.layers.data(name='y', shape=[1], dtype='int64', lod_level=0)
        emb_x, emb_y = _pull_box_sparse([x, y], size=2)
        emb_xp = _pull_box_sparse(x, size=2)
        layers.Print(emb_xp)
        concat = layers.concat([emb_x, emb_y], axis=1)
        fc = layers.fc(input=concat,
                       name="fc",
                       size=1,
                       num_flatten_dims=1,
                       bias_attr=False)
        loss = layers.reduce_mean(fc)
        layers.Print(loss)
        place = fluid.CPUPlace(
        ) if is_cpu or not core.is_compiled_with_cuda() else fluid.CUDAPlace(0)
        exe = fluid.Executor(place)
        optimizer = fluid.optimizer.SGD(learning_rate=0.5)
        batch_size = 2

        def binary_print(slot, fout):
            fout.write(str(len(slot)) + " ")
            for e in slot:
                fout.write(str(e) + " ")

        batch1 = np.ones(
            (batch_size, 2, 1)).astype("int64").reshape(batch_size, 2, 1)
        filelist = []
        place_str = "cpu" if is_cpu else "gpu"
        for i in range(2):
            filelist.append("test_hdfs_" + place_str + "_" + str(i))
        for f in filelist:
            with open(f, "w") as fout:
                for ins in batch1:
                    for slot in ins:
                        binary_print(slot, fout)
                fout.write("\n")

        def create_dataset():
            dataset = fluid.DatasetFactory().create_dataset("BoxPSDataset")
            dataset.set_use_var([x, y])
            dataset.set_batch_size(2)
            dataset.set_thread(1)
            dataset.set_filelist(filelist)
            return dataset

        datasets = []
        datasets.append(create_dataset())
        datasets.append(create_dataset())
        optimizer.minimize(loss)
        exe.run(fluid.default_startup_program())
        datasets[0].load_into_memory()
        datasets[0].begin_pass()
        datasets[1].preload_into_memory()
        exe.train_from_dataset(program=fluid.default_main_program(),
                               dataset=datasets[0],
                               print_period=1)
        datasets[0].end_pass()
        datasets[1].wait_preload_done()
        datasets[1].begin_pass()
        exe.train_from_dataset(program=fluid.default_main_program(),
                               dataset=datasets[1],
                               print_period=1)
        datasets[1].end_pass()
        for f in filelist:
            os.remove(f)
Ejemplo n.º 3
0
    def run_dataset(self, is_cpu):
        x = fluid.layers.data(name='x', shape=[1], dtype='int64', lod_level=0)
        y = fluid.layers.data(name='y', shape=[1], dtype='int64', lod_level=0)
        rank_offset = fluid.layers.data(name="rank_offset",
                                        shape=[-1, 7],
                                        dtype="int32",
                                        lod_level=0,
                                        append_batch_size=False)

        emb_x, emb_y = _pull_box_sparse([x, y], size=2)
        emb_xp = _pull_box_sparse(x, size=2)
        concat = layers.concat([emb_x, emb_y], axis=1)
        fc = layers.fc(input=concat,
                       name="fc",
                       size=1,
                       num_flatten_dims=1,
                       bias_attr=False)
        loss = layers.reduce_mean(fc)
        place = fluid.CPUPlace(
        ) if is_cpu or not core.is_compiled_with_cuda() else fluid.CUDAPlace(0)
        exe = fluid.Executor(place)

        with open("test_run_with_dump_a.txt", "w") as f:
            data = "1 1702f830eee19501ad7429505f714c1d 1 1 1 9\n"
            data += "1 1702f830eee19502ad7429505f714c1d 1 2 1 8\n"
            data += "1 1702f830eee19503ad7429505f714c1d 1 3 1 7\n"
            data += "1 1702f830eee0de01ad7429505f714c2d 1 4 1 6\n"
            data += "1 1702f830eee0df01ad7429505f714c3d 1 5 1 5\n"
            data += "1 1702f830eee0df02ad7429505f714c3d 1 6 1 4\n"
            f.write(data)
        with open("test_run_with_dump_b.txt", "w") as f:
            data = "1 1702f830fff22201ad7429505f715c1d 1 1 1 1\n"
            data += "1 1702f830fff22202ad7429505f715c1d 1 2 1 2\n"
            data += "1 1702f830fff22203ad7429505f715c1d 1 3 1 3\n"
            data += "1 1702f830fff22101ad7429505f714ccd 1 4 1 4\n"
            data += "1 1702f830fff22102ad7429505f714ccd 1 5 1 5\n"
            data += "1 1702f830fff22103ad7429505f714ccd 1 6 1 6\n"
            data += "1 1702f830fff22104ad7429505f714ccd 1 6 1 7\n"
            f.write(data)

        self.set_data_config()
        self.dataset.set_use_var([x, y])
        self.dataset.set_filelist(
            ["test_run_with_dump_a.txt", "test_run_with_dump_b.txt"])

        optimizer = fluid.optimizer.SGD(learning_rate=0.5)
        optimizer = fluid.optimizer.PipelineOptimizer(optimizer,
                                                      cut_list=[],
                                                      place_list=[place],
                                                      concurrency_list=[1],
                                                      queue_size=1,
                                                      sync_steps=-1)
        optimizer.minimize(loss)
        exe.run(fluid.default_startup_program())
        self.dataset.set_current_phase(1)
        self.dataset.load_into_memory()
        self.dataset.preprocess_instance()
        self.dataset.begin_pass()
        pv_num = self.dataset.get_pv_data_size()

        exe.train_from_dataset(program=fluid.default_main_program(),
                               dataset=self.dataset,
                               print_period=1)
        self.dataset.set_current_phase(0)
        self.dataset.postprocess_instance()
        exe.train_from_dataset(program=fluid.default_main_program(),
                               dataset=self.dataset,
                               print_period=1)
        self.dataset.end_pass(True)
        os.remove("test_run_with_dump_a.txt")
        os.remove("test_run_with_dump_b.txt")
Ejemplo n.º 4
0
    def run_boxps_preload(self, is_cpu=True, random_with_lineid=False):
        program = fluid.Program()
        with fluid.program_guard(program):
            x = fluid.layers.data(name='x',
                                  shape=[1],
                                  dtype='int64',
                                  lod_level=0)
            y = fluid.layers.data(name='y',
                                  shape=[1],
                                  dtype='int64',
                                  lod_level=0)
            emb_x, emb_y = _pull_box_sparse([x, y], size=2)
            emb_xp = _pull_box_sparse(x, size=2)
            concat = layers.concat([emb_x, emb_y], axis=1)
            fc = layers.fc(input=concat,
                           name="fc",
                           size=1,
                           num_flatten_dims=1,
                           bias_attr=False)
            loss = layers.reduce_mean(fc)
            layers.Print(loss)
            place = fluid.CPUPlace(
            ) if is_cpu or not core.is_compiled_with_cuda(
            ) else fluid.CUDAPlace(0)
            exe = fluid.Executor(place)
            batch_size = 100

            def binary_print(slot, fout):
                fout.write(str(len(slot)) + " ")
                for e in slot:
                    fout.write(str(e) + " ")

            batch1 = np.ones(
                (batch_size, 2, 1)).astype("int64").reshape(batch_size, 2, 1)
            filelist = []
            place_str = "cpu" if is_cpu else "gpu"
            for i in range(2):
                filelist.append("test_hdfs_" + place_str + "_" + str(i))
            for f in filelist:
                with open(f, "w") as fout:
                    for ins in batch1:
                        for slot in ins:
                            binary_print(slot, fout)
                    fout.write("\n")

            def create_dataset():
                dataset = fluid.DatasetFactory().create_dataset("BoxPSDataset")
                dataset.set_date("20190930")
                dataset.set_use_var([x, y])
                dataset.set_batch_size(2)
                dataset.set_thread(1)
                dataset.set_filelist(filelist)
                return dataset

            datasets = []
            datasets.append(create_dataset())
            datasets.append(create_dataset())
            optimizer = fluid.optimizer.SGD(learning_rate=0.5)
            optimizer = fluid.optimizer.PipelineOptimizer(optimizer,
                                                          cut_list=[],
                                                          place_list=[place],
                                                          concurrency_list=[1],
                                                          queue_size=1,
                                                          sync_steps=-1)
            optimizer.minimize(loss)

            program._pipeline_opt["dump_fields"] = [
                "fc.tmp_0", "fc.tmp_0@GRAD", "hehe"
            ]
            program._pipeline_opt["dump_fields_path"] = "./dump_log/"
            program._pipeline_opt["dump_param"] = ["fc.w_0"]
            program._pipeline_opt["enable_random_dump"] = True
            program._pipeline_opt["dump_interval"] = 10
            program._pipeline_opt["random_with_lineid"] = random_with_lineid

            exe.run(fluid.default_startup_program())
            datasets[0].load_into_memory()
            datasets[0].begin_pass()
            datasets[1].preload_into_memory()
            exe.train_from_dataset(program=fluid.default_main_program(),
                                   dataset=datasets[0],
                                   print_period=1)
            datasets[0].end_pass(True)
            datasets[1].wait_preload_done()
            datasets[1].begin_pass()
            exe.train_from_dataset(program=fluid.default_main_program(),
                                   dataset=datasets[1],
                                   print_period=1,
                                   debug=True)
            datasets[1].end_pass(False)
            for f in filelist:
                os.remove(f)
            if os.path.isdir("dump_log"):
                shutil.rmtree("dump_log")