keras-bed-sequence 1.0.11

Lazily one-hot encoding bed sequences using Keras Sequence.

Lazily one-hot encoding bed sequences using Keras Sequence.

How do I install this package?

As usual, just download it using pip:

pip install keras_bed_sequence

Tests Coverage

Since some software handling coverages sometimes get slightly different results, here’s three of them:

Usage examples

The following examples are tested within the package test suite.

Classification task example

Let’s start by building an extremely simple classification task model:

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Flatten
from keras_mixed_sequence import MixedSequence

model = Sequential([
    Flatten(),
    Dense(1)
])
model.compile(
    optimizer="nadam",
    loss="MSE"
)

We then proceed to load the training data into Keras Sequences, using, in particular, a MixedSequence object:

import numpy as np
from keras_mixed_sequence import MixedSequence
from keras_bed_sequence import BedSequence

batch_size = 32
bed_sequence = BedSequence(
    "hg19",
    "path/to/bed/files.bed",
    batch_size
)
y = the_output_values
mixed_sequence = MixedSequence(
    x=bed_sequence,
    y=y,
    batch_size=batch_size
)

Finally, we can proceed to use the obtained MixedSequence to train our model:

model.fit_generator(
    mixed_sequence,
    steps_per_epoch=mixed_sequence.steps_per_epoch,
    epochs=2,
    verbose=0,
    shuffle=True
)

Auto-encoding task example

Let’s start by building an extremely simple auto-encoding task model:

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, Reshape, Conv2DTranspose

model = Sequential([
    Reshape((200, 4, 1)),
    Conv2D(16, kernel_size=3, activation="relu"),
    Conv2DTranspose(1, kernel_size=3, activation="relu"),
    Reshape((-1, 200, 4))
])
model.compile(
    optimizer="nadam",
    loss="MSE"
)

We then proceed to load the training data into Keras Sequences, using, in particular, a MixedSequence object:

import numpy as np
from keras_mixed_sequence import MixedSequence
from keras_bed_sequence import BedSequence

batch_size = 32
bed_sequence = BedSequence(
    "hg19",
    "path/to/bed/files.bed",
    batch_size
)
mixed_sequence = MixedSequence(
    x=bed_sequence,
    y=bed_sequence,
    batch_size=batch_size
)

Finally, we can proceed to use the obtained MixedSequence to train our model:

model.fit_generator(
    mixed_sequence,
    steps_per_epoch=mixed_sequence.steps_per_epoch,
    epochs=2,
    verbose=0,
    shuffle=True
)

Multi-task example (classification + auto-encoding)

Let’s start by building an extremely simple multi-tasks model:

from tensorflow.keras.models import Model
from tensorflow.keras.layers import Dense, Conv2D, Reshape, Flatten, Conv2DTranspose, Input

inputs = Input(shape=(200, 4))

flattened = Flatten()(inputs)

output1 = Dense(
    units=1,
    activation="relu",
    name="output1"
)(flattened)

hidden = Reshape((200, 4, 1))(inputs)
hidden = Conv2D(16, kernel_size=3, activation="relu")(hidden)
hidden = Conv2DTranspose(1, kernel_size=3, activation="relu")(hidden)
output2 = Reshape((200, 4), name="output2")(hidden)

model = Model(
    inputs=inputs,
    outputs=[output1, output2],
    name="my_model"
)

model.compile(
    optimizer="nadam",
    loss="MSE"
)

We then proceed to load the training data into Keras Sequences, using, in particular, a MixedSequence object:

import numpy as np
from keras_mixed_sequence import MixedSequence
from keras_bed_sequence import BedSequence

batch_size = 32
bed_sequence = BedSequence(
    "hg19",
    "{cwd}/test.bed".format(
        cwd=os.path.dirname(os.path.abspath(__file__))
    ),
    batch_size
)
y = np.random.randint(
    2,
    size=(bed_sequence.samples_number, 1)
)
mixed_sequence = MixedSequence(
    bed_sequence,
    {
        "output1": y,
        "output2": bed_sequence
    },
    batch_size
)

Finally, we can proceed to use the obtained MixedSequence to train our model:

model.fit_generator(
    mixed_sequence,
    steps_per_epoch=mixed_sequence.steps_per_epoch,
    epochs=2,
    verbose=0,
    shuffle=True
)