"""External configuration .gin"""
from gin import config
import jax
config.external_configurable(jax.nn.initializers.zeros,
'jax.nn.initializers.zeros')
config.external_configurable(jax.nn.initializers.ones,
'jax.nn.initializers.ones')
config.external_configurable(jax.nn.initializers.variance_scaling,
'jax.nn.initializers.variance_scaling')
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from gin import config
from gin.tf import external_configurables # pylint: disable=unused-import
import tensorflow as tf
# Necessary for AdagradaDAOptimizer test.
config.external_configurable(tf.compat.v1.train.get_global_step)
@config.configurable
def fake_train_model(learning_rate, optimizer):
global_step = tf.compat.v1.train.get_or_create_global_step()
lr = learning_rate(global_step=global_step)
opt = optimizer(learning_rate=lr)
return lr, opt
@config.configurable
def configurable(**kwargs):
return kwargs
"""Other gin configurables
"""
import os
import gin
import modisco
import modisco.tfmodisco_workflow.workflow
from gin import config
import keras
# keras.optimizers
config.external_configurable(keras.optimizers.Adam, module='keras.optimizers')
config.external_configurable(keras.optimizers.RMSprop, module='keras.optimizers')
config.external_configurable(keras.optimizers.Adagrad, module='keras.optimizers')
config.external_configurable(keras.optimizers.Adadelta, module='keras.optimizers')
config.external_configurable(keras.optimizers.Adamax, module='keras.optimizers')
config.external_configurable(keras.optimizers.Nadam, module='keras.optimizers')
config.external_configurable(keras.optimizers.SGD, module='keras.optimizers')
# modisco
config.external_configurable(modisco.tfmodisco_workflow.workflow.TfModiscoWorkflow)
config.external_configurable(modisco.tfmodisco_workflow.seqlets_to_patterns.TfModiscoSeqletsToPatternsFactory)
@gin.configurable
def report_template(name, raise_error=True):
"""Evaluation report template found in ../templates/
"""
import inspect
filename = inspect.getframeinfo(inspect.currentframe()).filename
this_dir = os.path.dirname(os.path.abspath(filename))
from gin import config import tensorflow as tf config.external_configurable(tf.keras.layers.Input, name='Input') config.external_configurable(tf.keras.layers.Conv2D, name='Conv2D') config.external_configurable(tf.keras.layers.Dense, name='Dense') config.external_configurable(tf.keras.layers.Flatten, name='Flatten') config.external_configurable(tf.keras.layers.Dropout, name='Dropout')
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Supplies a default set of configurables from core TensorFlow."""
from gin import config
import torch
# Losses.
config.external_configurable(torch.nn.modules.loss.BCELoss,
module='torch.nn.modules.loss')
config.external_configurable(torch.nn.modules.loss.BCEWithLogitsLoss,
module='torch.nn.modules.loss')
config.external_configurable(torch.nn.modules.loss.CosineEmbeddingLoss,
module='torch.nn.modules.loss')
config.external_configurable(torch.nn.modules.loss.CrossEntropyLoss,
module='torch.nn.modules.loss')
config.external_configurable(torch.nn.modules.loss.CTCLoss,
module='torch.nn.modules.loss')
config.external_configurable(torch.nn.modules.loss.HingeEmbeddingLoss,
module='torch.nn.modules.loss')
config.external_configurable(torch.nn.modules.loss.KLDivLoss,
module='torch.nn.modules.loss')
config.external_configurable(torch.nn.modules.loss.L1Loss,
module='torch.nn.modules.loss')
config.external_configurable(torch.nn.modules.loss.MarginRankingLoss,
"""Supplies a default set of configurables from core TensorFlow."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import functools
from gin import config
import numpy as np
import tensorflow as tf
# Learning rate decays.
config.external_configurable(tf.compat.v1.train.exponential_decay,
module='tf.train')
config.external_configurable(tf.compat.v1.train.inverse_time_decay,
module='tf.train')
config.external_configurable(tf.compat.v1.train.natural_exp_decay,
module='tf.train')
config.external_configurable(tf.compat.v1.train.polynomial_decay,
module='tf.train')
@config.configurable(module='tf.train')
@functools.wraps(tf.compat.v1.train.piecewise_constant)
def piecewise_constant(global_step, *args, **kwargs):
if 'boundaries' in kwargs:
kwargs['boundaries'] = list(np.int64(kwargs['boundaries']))
return tf.compat.v1.train.piecewise_constant(global_step, *args, **kwargs)
"""Supplies a default set of configurables from core TensorFlow."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import functools
import numpy as np
import tensorflow as tf
from gin import config
# Learning rate decays.
config.external_configurable(tf.train.exponential_decay, module='tf.train')
config.external_configurable(tf.train.inverse_time_decay, module='tf.train')
config.external_configurable(tf.train.natural_exp_decay, module='tf.train')
config.external_configurable(tf.train.polynomial_decay, module='tf.train')
@config.configurable(module='tf.train')
@functools.wraps(tf.train.piecewise_constant)
def piecewise_constant(global_step, *args, **kwargs):
if 'boundaries' in kwargs:
kwargs['boundaries'] = list(np.int64(kwargs['boundaries']))
return tf.train.piecewise_constant(global_step, *args, **kwargs)
# Losses.
import torch as t
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from gin import config
# Prefix these with '@' in gin config files.
# optimizers
config.external_configurable(optim.SGD, 'sgd', module='t.optim')
config.external_configurable(optim.Adam, 'adam', module='t.optim')
config.external_configurable(optim.Adadelta, 'adadelta', module='t.optim')
config.external_configurable(optim.Adagrad, 'adagrad', module='t.optim')
config.external_configurable(optim.SparseAdam, 'sparse_adam', module='t.optim')
config.external_configurable(optim.Adamax, 'adamax', module='t.optim')
config.external_configurable(optim.AdamW, 'adamw', module='t.optim')
config.external_configurable(optim.ASGD, 'asgd', module='t.optim')
config.external_configurable(optim.LBFGS, 'lbfgs', module='t.optim')
config.external_configurable(optim.RMSprop, 'rmsprop', module='t.optim')
config.external_configurable(optim.Rprop, 'rprop', module='t.optim')
# lr scheduler
config.external_configurable(optim.lr_scheduler.LambdaLR,
'lambda_lr',
module='t.optim.lr_scheduler')
config.external_configurable(optim.lr_scheduler.StepLR,
'step_lr',
module='t.optim.lr_scheduler')
config.external_configurable(optim.lr_scheduler.MultiStepLR,
'multistep_lr',
from gin import config
import keras
# keras.optimizers
config.external_configurable(keras.optimizers.Adam, module='keras.optimizers')
config.external_configurable(keras.optimizers.RMSprop,
module='keras.optimizers')
config.external_configurable(keras.optimizers.Adagrad,
module='keras.optimizers')
config.external_configurable(keras.optimizers.Adadelta,
module='keras.optimizers')
config.external_configurable(keras.optimizers.Adamax,
module='keras.optimizers')
config.external_configurable(keras.optimizers.Nadam, module='keras.optimizers')
config.external_configurable(keras.optimizers.SGD, module='keras.optimizers')
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Declaration of a few external GIN configurable functions."""
from gin import config
# pylint: disable=unused-import
from gin.tf import external_configurables
# pylint: enable=unused-import
import tensorflow as tf
# Activation function.
config.external_configurable(tf.keras.activations.relu,
'tf.keras.activations.relu')
# Layers
config.external_configurable(tf.keras.layers.Add, 'tf.keras.layers.Add')
config.external_configurable(tf.keras.layers.Concatenate,
'tf.keras.layers.Concatenate')
# Loss functions
config.external_configurable(tf.keras.losses.MeanSquaredError,
'tf.keras.losses.MeanSquaredError')
config.external_configurable(tf.keras.losses.CategoricalCrossentropy,
'tf.keras.losses.CategoricalCrossentropy')
def __call__(self, true_counts, logits):
for i in range(self.n):
loss = multinomial_nll(true_counts[..., i], logits[..., i])
if i == 0:
total = loss
else:
total += loss
return total
def get_config(self):
return {"n": self.n}
mc_multinomial_nll_1 = MultichannelMultinomialNLL(1)
mc_multinomial_nll_1.__name__ = "mc_multinomial_nll_1"
config.external_configurable(mc_multinomial_nll_1)
mc_multinomial_nll_2 = MultichannelMultinomialNLL(2)
mc_multinomial_nll_2.__name__ = "mc_multinomial_nll_2"
config.external_configurable(mc_multinomial_nll_2)
twochannel_multinomial_nll = mc_multinomial_nll_2
twochannel_multinomial_nll.__name__ = "twochannel_multinomial_nll"
AVAILABLE = [
"multinomial_nll", "twochannel_multinomial_nll", "mc_multinomial_nll_1",
"mc_multinomial_nll_2", "MultichannelMultinomialNLL",
"CountsMultinomialNLL", "PoissonMultinomialNLL"
]
import torch as T
import torch.optim as optim
import torch.nn as nn
from gin import config
from .. import optim as nnt_optim
from .. import zoo
# optimizers
config.external_configurable(optim.SGD, 'sgd', module='T.optim')
config.external_configurable(optim.Adam, 'adam', module='T.optim')
config.external_configurable(optim.Adadelta, 'adadelta', module='T.optim')
config.external_configurable(optim.Adagrad, 'adagrad', module='T.optim')
config.external_configurable(optim.SparseAdam, 'sparse_adam', module='T.optim')
config.external_configurable(optim.Adamax, 'adamax', module='T.optim')
config.external_configurable(optim.AdamW, 'adamw', module='T.optim')
config.external_configurable(optim.ASGD, 'asgd', module='T.optim')
config.external_configurable(optim.LBFGS, 'lbfgs', module='T.optim')
config.external_configurable(optim.RMSprop, 'rmsprop', module='T.optim')
config.external_configurable(optim.Rprop, 'rprop', module='T.optim')
config.external_configurable(nnt_optim.AdaBound, 'adabound', module='nnt')
config.external_configurable(nnt_optim.Lookahead, 'lookahead', module='nnt')
config.external_configurable(nnt_optim.NAdam, 'nadam', module='nnt')
try:
import apex
config.external_configurable(apex.optimizers.FusedAdam, 'fusedadam', module='apex.optimizers')
config.external_configurable(apex.optimizers.FusedSGD, 'fusedsgd', module='apex.optimizers')
config.external_configurable(apex.optimizers.FusedNovoGrad, 'fusednovograd', module='apex.optimizers')
config.external_configurable(apex.optimizers.FusedLAMB, 'fusedlamb', module='apex.optimizers')
print('Apex Fused Optimizers is availble for GIN')
"""External configuration .gin"""
from gin import config
from flax import linen as nn
config.external_configurable(nn.initializers.zeros, 'nn.initializers.zeros')
config.external_configurable(nn.initializers.ones, 'nn.initializers.ones')
config.external_configurable(nn.initializers.orthogonal,
'nn.initializers.orthogonal')
config.external_configurable(nn.initializers.variance_scaling,
'nn.initializers.variance_scaling')
out[k] = mean(v)
# flatten everything
out = flatten(out, separator='/')
return out
def save(self, file_path):
"""Save model to a file
"""
from bpnet.utils import write_pkl, SerializableLock
# fix the serialization of _OPERATIVE_CONFIG_LOCK
gin.config._OPERATIVE_CONFIG_LOCK = SerializableLock()
write_pkl(self, file_path)
@classmethod
def load(cls, file_path):
"""Load model from a file
"""
from bpnet.utils import read_pkl
return read_pkl(file_path)
@classmethod
def from_mdir(cls, model_dir):
"""Load the model from pkl
"""
return cls.load(os.path.join(model_dir, 'seq_model.pkl'))
# avoid the decorator so that we can pickle it
config.external_configurable(SeqModel)
def _register_schedule(module): config.external_configurable(module, module='tf.keras.optimizers.schedules')
def _register_callables(package, module, blacklist):
for k in dir(package):
if k not in blacklist:
v = getattr(package, k)
if callable(v):
config.external_configurable(v, name=k, module=module)
from torch import optim
from torch.optim import lr_scheduler
from torch import nn
from gin import config
import gin
# Optimizer
config.external_configurable(optim.Adam, "optim.Adam")
config.external_configurable(optim.Adadelta, "optim.Adadelta")
config.external_configurable(optim.Adagrad, "optim.Adagrad")
config.external_configurable(optim.Adamax, "optim.Adamax")
config.external_configurable(optim.RMSprop, "optim.RMSprop")
config.external_configurable(optim.SGD, "optim.SGD")
# Learning rate decay
config.external_configurable(lr_scheduler.StepLR, "lr_scheduler.StepLR")
config.external_configurable(lr_scheduler.MultiStepLR,
"lr_scheduler.MultiStepLR")
config.external_configurable(lr_scheduler.ExponentialLR,
"lr_scheduler.ExponentialLR")
config.external_configurable(lr_scheduler.CosineAnnealingLR,
"lr_scheduler.CosineAnnealingLR")
config.external_configurable(lr_scheduler.ReduceLROnPlateau,
"lr_scheduler.ReduceLROnPlateau")
@gin.configurable("lr_scheduler.LambdaLR", blacklist=["optimizer"])
def create_LambdaLR(optimizer, epoch_total, epoch_decay):
def lambda_rule(epoch):
epoch_stable = epoch_total - epoch_decay
lr_l = 1.0 - max(0, epoch - epoch_stable) / float(epoch_decay + 1)
