Created by: gmcather

*** Aborted at 1521028443 (unix time) try "date -d @1521028443" if you are using GNU date ***
PC: @                0x0 (unknown)
*** SIGSEGV (@0x7f0f4d031b64) received by PID 2524 (TID 0x7f16a98b0700) from PID 1292049252; stack trace: ***
    @     0x7f16a948d390 (unknown)
    @     0x7f1671535b83 mkl_blas_avx_sgemm_kernel_0_b0
Segmentation fault

from __future__ import print_function
import numpy as np
import paddle.v2 as paddle
import paddle.fluid as fluid
import paddle.fluid.layers as layers
import sys
import time
from functools import partial

batch_size = 300
d_model = 64

def position_encoding_init(n_position, d_pos_vec):
    """
    Generate the initial values for the sinusoid position encoding table.
    """
    position_enc = np.array([[
        pos / np.power(10000, 2 * (j // 2) / d_pos_vec)
        for j in range(d_pos_vec)
    ] if pos != 0 else np.zeros(d_pos_vec) for pos in range(n_position)])
    position_enc[1:, 0::2] = np.sin(position_enc[1:, 0::2])  # dim 2i
    position_enc[1:, 1::2] = np.cos(position_enc[1:, 1::2])  # dim 2i+1
    return position_enc.astype("float32")


def multi_head_attention(queries,
                         keys,
                         values,
                         attn_bias,
                         d_key,
                         d_value,
                         d_model,
                         n_head=1,
                         dropout_rate=0.):
    """
    Multi-Head Attention. Note that attn_bias is added to the logit before
    computing softmax activiation to mask certain selected positions so that
    they will not considered in attention weights.
    """
    if not (len(queries.shape) == len(keys.shape) == len(values.shape) == 3):
        raise ValueError(
            "Inputs: quries, keys and values should all be 3-D tensors.")

    def __compute_qkv(queries, keys, values, n_head, d_key, d_value):
        """
        Add linear projection to queries, keys, and values.
        """
        q = layers.fc(input=queries,
                      size=d_key * n_head,
                      param_attr=fluid.initializer.Xavier(
                          uniform=False,
                          fan_in=d_model * d_key,
                          fan_out=n_head * d_key),
                      bias_attr=False,
                      num_flatten_dims=2)
        k = layers.fc(input=keys,
                      size=d_key * n_head,
                      param_attr=fluid.initializer.Xavier(
                          uniform=False,
                          fan_in=d_model * d_key,
                          fan_out=n_head * d_key),
                      bias_attr=False,
                      num_flatten_dims=2)
        v = layers.fc(input=values,
                      size=d_value * n_head,
                      param_attr=fluid.initializer.Xavier(
                          uniform=False,
                          fan_in=d_model * d_value,
                          fan_out=n_head * d_value),
                      bias_attr=False,
                      num_flatten_dims=2)
        return q, k, v

    def __split_heads(x, n_head):
        """
        Reshape the last dimension of inpunt tensor x so that it becomes two
        dimensions and then transpose. Specifically, input a tensor with shape
        [bs, max_sequence_length, n_head * hidden_dim] then output a tensor
        with shape [bs, n_head, max_sequence_length, hidden_dim].
        """
        if n_head == 1:
            return x

        hidden_size = x.shape[-1]
        # FIXME(guosheng): Decouple the program desc with batch_size.
        reshaped = layers.reshape(
            x=x, shape=[batch_size, -1, n_head, hidden_size // n_head])

        # permuate the dimensions into:
        # [batch_size, n_head, max_sequence_len, hidden_size_per_head]
        return layers.transpose(x=reshaped, perm=[0, 2, 1, 3])

    def __combine_heads(x):
        """
        Transpose and then reshape the last two dimensions of inpunt tensor x
        so that it becomes one dimension, which is reverse to __split_heads.
        """
        if len(x.shape) == 3: return x
        if len(x.shape) != 4:
            raise ValueError("Input(x) should be a 4-D Tensor.")

        trans_x = layers.transpose(x, perm=[0, 2, 1, 3])
        # FIXME(guosheng): Decouple the program desc with batch_size.
        return layers.reshape(
            x=trans_x,
            shape=map(int,
                      [batch_size, -1, trans_x.shape[2] * trans_x.shape[3]]))

    def scaled_dot_product_attention(q, k, v, attn_bias, d_model, dropout_rate):
        """
        Scaled Dot-Product Attention
        """

        # FIXME(guosheng): Optimize the shape in reshape_op or softmax_op.

        # The current implementation of softmax_op only supports 2D tensor,
        # consequently it cannot be directly used here.
        # If to use the reshape_op, Besides, the shape of product inferred in
        # compile-time is not the actual shape in run-time. It cann't be used
        # to set the attribute of reshape_op.
        # So, here define the softmax for temporary solution.

        def __softmax(x, eps=1e-9):
            exp_out = layers.exp(x=x)
            sum_out = layers.reduce_sum(exp_out, dim=-1, keep_dim=False)
            return layers.elementwise_div(x=exp_out, y=sum_out, axis=0)

        scaled_q = layers.scale(x=q, scale=d_model**-0.5)
        product = layers.matmul(x=scaled_q, y=k, transpose_y=True)
        weights = __softmax(layers.elementwise_add(x=product, y=attn_bias))
        if dropout_rate:
            weights = layers.dropout(
                weights, dropout_prob=dropout_rate, is_test=False)
        out = layers.matmul(weights, v)
        return out

    q, k, v = __compute_qkv(queries, keys, values, n_head, d_key, d_value)

    q = __split_heads(q, n_head)
    k = __split_heads(k, n_head)
    v = __split_heads(v, n_head)

    ctx_multiheads = scaled_dot_product_attention(q, k, v, attn_bias, d_model,
                                                  dropout_rate)

    out = __combine_heads(ctx_multiheads)

    # Project back to the model size.
    proj_out = layers.fc(input=out,
                         size=d_model,
                         param_attr=fluid.initializer.Xavier(uniform=False),
                         bias_attr=False,
                         num_flatten_dims=2)
    return proj_out


def positionwise_feed_forward(x, d_inner_hid, d_hid):
    """
    Position-wise Feed-Forward Networks.
    This module consists of two linear transformations with a ReLU activation
    in between, which is applied to each position separately and identically.
    """
    hidden = layers.fc(input=x,
                       size=d_inner_hid,
                       num_flatten_dims=2,
                       param_attr=fluid.initializer.Uniform(
                           low=-(d_hid**-0.5), high=(d_hid**-0.5)),
                       act="relu")
    out = layers.fc(input=hidden,
                    size=d_hid,
                    num_flatten_dims=2,
                    param_attr=fluid.initializer.Uniform(
                        low=-(d_inner_hid**-0.5), high=(d_inner_hid**-0.5)))
    return out


def pre_post_process_layer(prev_out, out, process_cmd, dropout=0.):
    """
    Add residual connection, layer normalization and droput to the out tensor
    optionally according to the value of process_cmd.

    This will be used before or after multi-head attention and position-wise
    feed-forward networks.
    """
    for cmd in process_cmd:
        if cmd == "a":  # add residual connection
            out = out + prev_out if prev_out else out
        elif cmd == "n":  # add layer normalization
            out = layers.layer_norm(
                out,
                begin_norm_axis=len(out.shape) - 1,
                param_attr=fluid.initializer.Constant(1.),
                bias_attr=fluid.initializer.Constant(0.))
        elif cmd == "d":  # add dropout
            if dropout:
                out = layers.dropout(out, dropout_prob=dropout, is_test=False)
    return out


pre_process_layer = partial(pre_post_process_layer, None)
post_process_layer = pre_post_process_layer


def prepare_encoder(src_word,
                    src_pos,
                    src_vocab_size,
                    src_emb_dim,
                    src_pad_idx,
                    src_max_len,
                    dropout=0.,
                    pos_pad_idx=0,
                    pos_enc_param_name=None):
    """Add word embeddings and position encodings.
    The output tensor has a shape of:
    [batch_size, max_src_length_in_batch, d_model].

    This module is used at the bottom of the encoder stacks.
    """
    src_word_emb = layers.embedding(
        src_word,
        size=[src_vocab_size, src_emb_dim],
        padding_idx=src_pad_idx,
        param_attr=fluid.initializer.Normal(0., 1.))
    src_pos_enc = layers.embedding(
        src_pos,
        size=[src_max_len, src_emb_dim],
        padding_idx=pos_pad_idx,
        param_attr=fluid.ParamAttr(
            name=pos_enc_param_name, trainable=False))
    enc_input = src_word_emb + src_pos_enc

    # FIXME(guosheng): Decouple the program desc with batch_size.
    enc_input = layers.reshape(x=enc_input, shape=[batch_size, -1, src_emb_dim])
    return layers.dropout(
        enc_input, dropout_prob=dropout,
        is_test=False) if dropout else enc_input


#prepare_encoder = partial(
#    prepare_encoder, pos_enc_param_name=pos_enc_param_names[0])


def encoder_layer(enc_input,
                  attn_bias,
                  n_head,
                  d_key,
                  d_value,
                  d_model,
                  d_inner_hid,
                  dropout_rate=0.):
    """The encoder layers that can be stacked to form a deep encoder.

    This module consits of a multi-head (self) attention followed by
    position-wise feed-forward networks and both the two components companied
    with the post_process_layer to add residual connection, layer normalization
    and droput.
    """
    attn_output = multi_head_attention(enc_input, enc_input, enc_input,
                                       attn_bias, d_key, d_value, d_model,
                                       n_head, dropout_rate)
    attn_output = post_process_layer(enc_input, attn_output, "dan",
                                     dropout_rate)
    ffd_output = positionwise_feed_forward(attn_output, d_inner_hid, d_model)
    return post_process_layer(attn_output, ffd_output, "dan", dropout_rate)


def encoder(enc_input,
            attn_bias,
            n_layer,
            n_head,
            d_key,
            d_value,
            d_model,
            d_inner_hid,
            dropout_rate=0.):
    """
    The encoder is composed of a stack of identical layers returned by calling
    encoder_layer.
    """
    for i in range(n_layer):
        enc_output = encoder_layer(enc_input, attn_bias, n_head, d_key, d_value,
                                   d_model, d_inner_hid, dropout_rate)
        enc_input = enc_output
    return enc_output

def to_lodtensor(data, place):
    """
    load LODtensor
    """
    seq_lens = [len(seq) for seq in data]
    cur_len = 0
    lod = [cur_len]
    for l in seq_lens:
        cur_len += l
        lod.append(cur_len)
    flattened_data = np.concatenate(data, axis=0).astype("int64")
    flattened_data = flattened_data.reshape([len(flattened_data), 1])
    res = fluid.LoDTensor()
    res.set(flattened_data, place)
    res.set_lod([lod])
    return res

def load_vocab(filename):
    """
    load vocabulary
    """
    vocab = {}
    with open(filename) as f:
        wid = 0
        for line in f:
            vocab[line.strip()] = wid
            wid += 1
    vocab["<unk>"] = len(vocab)
    return vocab

def prepare_batch_input(data, pad_idx, n_head, place):
    """
    Pad the instances to the max sequence length in batch, and generate the
    corresponding position data and attention bias. Then, convert the numpy
    data to tensors and return a dict mapping names to tensors.
    """
    input_dict = {}

    def __pad_batch_data(data,
                         pad_idx,
                        n_head):
        """
        Pad the instances to the max sequence length in batch, and generate the
        corresponding position data and attention bias.
        """
        insts = [x[0] for x in data]

        return_list = []
        max_len = max(len(inst) for inst in insts)

        inst_data = np.array(
            [inst + [pad_idx] * (max_len - len(inst)) for inst in insts])
        return_list += [inst_data.astype("int64").reshape([-1, 1])]

        src_mask = np.array(
            [ [ [1] * d_model ] * len(inst) + [ [0] * d_model ] * (max_len - len(inst)) for inst in insts])
        return_list += [src_mask.astype("float32").reshape([batch_size, -1, d_model])]

        inst_pos = np.array([[
                        pos_i + 1 if w_i != pad_idx else 0
                        for pos_i, w_i in enumerate(inst)
                        ] for inst in inst_data])

        return_list += [inst_pos.astype("int64").reshape([-1, 1])]

        label_data = np.array(map(lambda x: x[1], data)).astype("int64")
        label_data = label_data.reshape([-1, 1])
        return_list += [label_data]
        # This is used to avoid attention on paddings.
        slf_attn_bias_data = np.array([[0] * len(inst) + [-1e9] *
                                               (max_len - len(inst))
                                               for inst in insts])
        slf_attn_bias_data = np.tile(
                            slf_attn_bias_data.reshape([-1, 1, 1, max_len]),
                            [1, n_head, max_len, 1])
        return_list += [slf_attn_bias_data.astype("float32")]

        return return_list if len(return_list) > 1 else return_list[0]

    def data_to_tensor(data_list, name_list, input_dict, place):
        assert len(data_list) == len(name_list)
        for i in range(len(name_list)):
            tensor = fluid.LoDTensor()
            tensor.set(data_list[i], place)
            input_dict[name_list[i]] = tensor

    src_word, src_mask, src_pos, src_label, src_slf_attn_bias = __pad_batch_data(data, pad_idx, n_head)

    data_to_tensor([src_word, src_mask, src_pos, src_label, src_slf_attn_bias],
                    ["source_text", "source_mask", "source_pos", "label", "source_slf_attn_bias"], input_dict, place)

    return input_dict

def main(max_length = 3000,
        n_head = 2,
        n_layer = 1,
        d_key = 64,
        d_value = 64,
        d_inner_hid = 64,
        d_model = 64,
        lr = 0.001):

    start_time = time.time()

    word_dict = load_vocab("imdb.vocab")
    dict_dim = len(word_dict)
    pad_idx = dict_dim - 1

    src_word = layers.data(
        name="source_text",
        shape=[batch_size * max_length, 1],
        dtype="int64",
        append_batch_size=False)

    src_mask = layers.data(
        name="source_mask",
        shape=[batch_size, max_length, d_model],
        dtype="float32",
        append_batch_size=False)

    src_pos = layers.data(
        name="source_pos",
        shape=[batch_size * max_length, 1],
        dtype="int64",
        append_batch_size=False)

    label = layers.data(
        name="label",
        shape=[batch_size, 1],
        dtype="int64",
        append_batch_size=False)

    # The actual shape of src_slf_attn_bias in runtime is:
    # [batch_size, n_head, max_src_length_in_a_batch, max_src_length_in_a_batch].
    # This input is used to remove attention weights on paddings.
    src_slf_attn_bias = layers.data(
        name="source_slf_attn_bias",
        shape=[batch_size, n_head, max_length, max_length],
        dtype="float32",
        append_batch_size=False)

    dropout_rate = 0.0
    enc_input = prepare_encoder(
        src_word,
        src_pos,
        dict_dim,
        d_model,
        pad_idx,
        max_length,
        dropout_rate)

    enc_output = encoder(
        enc_input,
        src_slf_attn_bias,
        n_layer,
        n_head,
        d_key,
        d_value,
        d_model,
        d_inner_hid,
        dropout_rate)

    mask_data = layers.elementwise_mul(enc_output, src_mask)
    print(mask_data)
    bow_out = layers.reduce_sum(mask_data, dim = 1)
    print(bow_out)
    # TODO(guosheng): Share the weight matrix between the embedding layers and
    # the pre-softmax linear transformation.
    class_dim = 2
    prediction = layers.fc(input=[bow_out],
                    size=class_dim,
                    act="softmax",
                    num_flatten_dims=1)
    #print(prediction)
    #print(label)
    cost = layers.cross_entropy(input=prediction, label=label)
    avg_cost = fluid.layers.mean(x=cost)
    sgd_optimizer = fluid.optimizer.SGD(learning_rate=lr)
    sgd_optimizer.minimize(avg_cost)

    # accuracy metric
    accuracy = fluid.evaluator.Accuracy(input=prediction, label=label)
    #print(accuracy)
    inference_program = fluid.default_main_program().clone()
    with fluid.program_guard(inference_program):
        test_target = accuracy.metrics + accuracy.states
        inference_program = fluid.io.get_inference_program(test_target)

    # train data set
    train_reader = paddle.batch(
        paddle.reader.shuffle(
            paddle.dataset.imdb.train(word_dict), buf_size=50000),
        batch_size=batch_size)
    test_reader = paddle.batch(
        paddle.reader.shuffle(
            paddle.dataset.imdb.test(word_dict), buf_size=50000),
        batch_size=batch_size)

    # train in cpu
    place = fluid.CPUPlace()

    def test(exe):
        accuracy.reset(exe)
        for data in test_reader():
            data = prepare_batch_input(data, pad_idx, n_head, place)
            acc = exe.run(inference_program,
                      feed=data)
        return accuracy.eval(exe)

    # just like session in tensorflow
    exe = fluid.Executor(place)

    # not sure what's going on here
    exe.run(fluid.default_startup_program())

    # loop for 30 epochs, print acc ever 1000 batch
    PASS_NUM = 30

    for pass_id in xrange(PASS_NUM):
        accuracy.reset(exe)
        for data in train_reader():
            print(len(data))
            data = prepare_batch_input(data, pad_idx, n_head, place)
            cost_val, acc_val = exe.run(fluid.default_main_program(),
                                    feed=data,
                                    fetch_list=[avg_cost, accuracy.metrics[0]])
            pass_acc = accuracy.eval(exe)

        pass_test_acc = test(exe)
        print("test_acc: %f" % pass_test_acc)

    end_time = time.time()
    print(end_time - start_time)

if __name__ == "__main__":
    main()