Created by: phonism

text classification, using CNN.

def convolution_net(dict_dim, class_dim=2, emb_dim=128, hid_dim=128):
    # input layers
    title_data = paddle.layer.data("title", paddle.data_type.integer_value_sequence(dict_dim))
    lbl = paddle.layer.data("label", paddle.data_type.integer_value(class_dim))

    #embedding layer
    title_emb = paddle.layer.embedding(input=title_data, size=emb_dim)

    # convolution layers with max pooling
    title_conv_max_2 = paddle.networks.sequence_conv_pool(
        input=title_emb, context_len=2, hidden_size=hid_dim)#, pool_type=paddle.pooling.Max())
    title_conv_max_3 = paddle.networks.sequence_conv_pool(
        input=title_emb, context_len=3, hidden_size=hid_dim)#, pool_type=paddle.pooling.Max())
    title_conv_max_4 = paddle.networks.sequence_conv_pool(
        input=title_emb, context_len=4, hidden_size=hid_dim)#, pool_type=paddle.pooling.Max())

    concat_layer = paddle.layer.concat(
            input=[
                title_conv_max_2, title_conv_max_3, title_conv_max_4,
            ])

    dropout_layer = paddle.layer.dropout(input=concat_layer, dropout_rate=0.5)

    # fc and output layer
    output = paddle.layer.fc(input=dropout_layer, size=class_dim, act=paddle.activation.Softmax())

    cost = paddle.layer.classification_cost(input=output, label=lbl)

    return cost, output, lbl

def train_cnn_model(num_pass):
    dict_dim = 640000
    class_dim = 4225
    # define data reader
    train_reader = paddle.batch(
        paddle.reader.shuffle(
            lambda: data_reader(flag=True), buf_size=1000),
        batch_size=64)
    test_reader = paddle.batch(
        lambda: data_reader(flag=False), batch_size=64)

    # network config
    [cost, output, label] = convolution_net(dict_dim, class_dim=class_dim)
    # create parameters
    parameters = paddle.parameters.create(cost)
    # create optimizer
    adam_optimizer = paddle.optimizer.Adam(
        learning_rate=1e-3,
        regularization=paddle.optimizer.L2Regularization(rate=1e-3),
        model_average=paddle.optimizer.ModelAverage(average_window=0.5))

    # add auc evaluator
    paddle.evaluator.auc(input=output, label=label)

    # create trainer
    trainer = paddle.trainer.SGD(
        cost=cost, parameters=parameters, update_equation=adam_optimizer)

    feeding = {'title': 0, 'label': 1}
    trainer.train(
        reader=train_reader,
        event_handler=event_handler,
        feeding=feeding,
        # use_sparse_updater=1,
        num_passes=num_pass)

if __name__ == "__main__":
    paddle.init(use_gpu=True, trainer_count=1)
    # paddle.init(use_gpu=True, trainer_count=4)
    num_pass = 500
    # train_cnn_model(num_pass=num_pass)
    cnn_infer("in")

GPU info: Tesla K40m * 4

run with trainer_count=1, 100 batch cost 6s.

+------------------------------------------------------+
| NVIDIA-SMI 352.39     Driver Version: 352.39         |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Tesla K40m          On   | 0000:03:00.0     Off |                    0 |
| N/A   37C    P0    98W / 235W |   2723MiB / 11519MiB |     30%      Default |
+-------------------------------+----------------------+----------------------+
|   1  Tesla K40m          On   | 0000:04:00.0     Off |                    0 |
| N/A   32C    P0    61W / 235W |    143MiB / 11519MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
|   2  Tesla K40m          On   | 0000:83:00.0     Off |                    0 |
| N/A   32C    P0    62W / 235W |    143MiB / 11519MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
|   3  Tesla K40m          On   | 0000:84:00.0     Off |                    0 |
| N/A   32C    P0    62W / 235W |    143MiB / 11519MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID  Type  Process name                               Usage      |
|=============================================================================|
|    0     16645    C   python                                        2697MiB |
|    1     16645    C   python                                         118MiB |
|    2     16645    C   python                                         118MiB |
|    3     16645    C   python                                         118MiB |
+-----------------------------------------------------------------------------+

using trainer_count=4 100 batch cost 25s。

+------------------------------------------------------+
| NVIDIA-SMI 352.39     Driver Version: 352.39         |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Tesla K40m          On   | 0000:03:00.0     Off |                    0 |
| N/A   26C    P0    66W / 235W |   2683MiB / 11519MiB |     35%      Default |
+-------------------------------+----------------------+----------------------+
|   1  Tesla K40m          On   | 0000:04:00.0     Off |                    0 |
| N/A   25C    P0    70W / 235W |   1139MiB / 11519MiB |     28%      Default |
+-------------------------------+----------------------+----------------------+
|   2  Tesla K40m          On   | 0000:83:00.0     Off |                    0 |
| N/A   26C    P0    66W / 235W |   1117MiB / 11519MiB |     44%      Default |
+-------------------------------+----------------------+----------------------+
|   3  Tesla K40m          On   | 0000:84:00.0     Off |                    0 |
| N/A   25C    P0    63W / 235W |   1153MiB / 11519MiB |     47%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID  Type  Process name                               Usage      |
|=============================================================================|
|    0      9520    C   python                                        2658MiB |
|    1      9520    C   python                                        1113MiB |
|    2      9520    C   python                                        1091MiB |
|    3      9520    C   python                                        1127MiB |
+-----------------------------------------------------------------------------+