cyberzhg / keras-transformer-xl

Transformer-XL with checkpoint loader

Keras Transformer-XL

[中文|English]

Unofficial implementation of Transformer-XL.

Install

pip install keras-transformer-xl

Usage

Load Pretrained Weights

Several configuration files can be found at the info directory.

import os
from keras_transformer_xl import load_trained_model_from_checkpoint

checkpoint_path = 'foo/bar/sota/enwiki8'
model = load_trained_model_from_checkpoint(
    config_path=os.path.join(checkpoint_path, 'config.json'),
    checkpoint_path=os.path.join(checkpoint_path, 'model.ckpt')
)
model.summary()

About IO

The generated model has two inputs, and the second input is the lengths of memories.

You can use MemorySequence wrapper for training and prediction:

from tensorflow import keras
import numpy as np
from keras_transformer_xl import MemorySequence, build_transformer_xl


class DummySequence(keras.utils.Sequence):

    def __init__(self):
        pass

    def __len__(self):
        return 10

    def __getitem__(self, index):
        return np.ones((3, 5 * (index + 1))), np.ones((3, 5 * (index + 1), 3))


model = build_transformer_xl(
    units=4,
    embed_dim=4,
    hidden_dim=4,
    num_token=3,
    num_block=3,
    num_head=2,
    batch_size=3,
    memory_len=20,
    target_len=10,
)
seq = MemorySequence(
    model=model,
    sequence=DummySequence(),
    target_len=10,
)

model.predict(model, seq, verbose=True)

keras-transformer-xl's Issues

Transformer-Xl for a Classification Task

Is your feature request related to a problem? Please describe.
A clear and concise description of what the problem is. Ex. I'm always frustrated when [...]

Describe the solution you'd like
A clear and concise description of what you want to happen.

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.

Memory-Test does not evaluate a single model

Describe the Bug

I believe that test is intended to evaluate the Memory function for 10 batches of sentences (or the same sentence for 10 epochs).

The current implementation initializes the Model 10x instead of initializing one model and evaluating 10 batches.

Version Info

• I'm using the latest version

Minimal Codes To Reproduce
NA

I think the recommended change is to shift the model instantiation outside of the for loop.

        for _ in range(10):
            input_data = keras.layers.Input(shape=(3, 3))
            input_length = keras.layers.Input(shape=(1,))
            output = Memory(3, 5, 3, 3)([input_data, input_length])
            model = keras.models.Model(
                inputs=[input_data, input_length],
                outputs=output,
            )
            ....

        input_data = keras.layers.Input(shape=(3, 3))
        input_length = keras.layers.Input(shape=(1,))
        output = Memory(3, 5, 3, 3)([input_data, input_length])
        model = keras.models.Model(
            inputs=[input_data, input_length],
            outputs=output,
        )
        for _ in range(10):
            ...

Transformer-Xl for a Classification Task

I would like to use the model for a text classification task on sequences of length less than 60 tokens.
Also I have trouble understanding the architecture of Transformer XL and relate it with Encoder Decoder structure of Transformers.

Visualization

Make documentation clearer about the MemorySequence class and the expected inputs and outputs passed.

Is your feature request related to a problem? Please describe.
It is not clear from the documentation how to pass data for a real world model language modeling example.

Describe the solution you'd like
I am trying make sense of the DummySequence class provided in the example but I cannot figure out how to pass data for a language modeling task. I see that the DummySequence class returns samples of the following form.

# input for batch 0
array([[1., 1., 1., 1., 1.],
       [1., 1., 1., 1., 1.],
       [1., 1., 1., 1., 1.]])

# output for batch 0

array([[[1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.]],

       [[1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.]],

       [[1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.]]])

# input for batch 9
array([[1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
        1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
        1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
        1., 1.],
       [1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
        1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
        1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
        1., 1.],
       [1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
        1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
        1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
        1., 1.]])

# output for batch 9

array([[[1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.]],

       [[1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.]],

       [[1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.]]])

What does the input and output represent and why do they keep growing by 5 elements each time ? Can you provide an example on what would a Sequence class be returning in order to implement a real world language model ? Also what is the connection to the parameters of the MemorySequence like target_len, memory_len and units to the parameters of the transformer model like units, num_block e.t.c.

Lets say for example that I have the following corpus of 10_000 tokens and want to train a language model on it. Token_1, Token_2 ... Token_10_000. What would I be passing as input for each batch.

Describe alternatives you've considered
None

Additional context
None

By the way thanks for this great library and all the other fantastic libraries you have open sourced. Everything is very modular and very well designed. Truly a step above of any other ml library of model implementation on github.

can you reproduce the sota transformer-xl results on wikitext-103

Describe the Bug

A clear and concise description of what the bug is.

Version Info

• I'm using the latest version

Minimal Codes To Reproduce

import keras_transformer_xl

pass

can you reproduce the sota transformer-xl results on wt103?

with the sota settings, I use the pretrained_xl weights to build my graph, but my test ppl is around 35, and when set the tgt_len to 384 and mem_len to 384 to eval, the test ppl is 24.2, when tgt_len and mem_len all set to 2048, test ppl reduce to 20.34, why the sota setting can not reproduce the 18.3 test ppl?

I find the structure has a little difference to official model.

I use the latest code to build a transformer-xl model and load the checkpoints trained from official code. It predicts right, but gets a much bigger loss when continue training.

Thus, I make a compare with official code. I find that there are two difference in structure:

1. There is a wrong order between dropout layer and add layer. dropout is first in official code both in attention layer and feed forward layer.
2. There is a dropout layer missing before softmax

Then I try modify the code and it seams run correctly.

Could you please make a double check?