def test_split_train_test():
data_path = pkg_resources.resource_filename('janggu', 'resources/')
bed_file = os.path.join(data_path, 'sample.bed')
refgenome = os.path.join(data_path, 'sample_genome.fa')
dna = Bioseq.create_from_refgenome('dna',
refgenome=refgenome,
storage='ndarray',
roi=bed_file,
binsize=200,
stepsize=200,
order=1,
store_whole_genome=True)
traindna, testdna = split_train_test(dna, holdout_chroms='chr2')
assert len(traindna) == 50
assert len(testdna) == 50
assert len(dna) == len(traindna) + len(testdna)
traindna, testdna = split_train_test([dna, dna], holdout_chroms='chr2')
assert len(traindna[0]) == 50
assert len(testdna[0]) == 50
assert len(dna) == len(traindna[0]) + len(testdna[0])
def fit(self, # pylint: disable=too-many-locals
inputs=None,
outputs=None,
batch_size=None,
epochs=1,
verbose=1,
callbacks=None,
validation_data=None,
shuffle=True,
class_weight=None,
sample_weight=None,
initial_epoch=0,
steps_per_epoch=None,
use_multiprocessing=False,
workers=1):
"""Model fitting.
This method is used to fit a given model.
Most of parameters are directly delegated the
fit_generator of the keras model.
Parameters
----------
inputs : :code:`Dataset`, list(Dataset) or Sequence (keras.utils.Sequence)
Input Dataset or Sequence to use for fitting the model.
outputs : :code:`Dataset`, list(Dataset) or None
Output Dataset containing the training targets. If a Sequence
is used for inputs, outputs will have no effect.
batch_size : int or None
Batch size. If set to None a batch size of 32 is used.
epochs : int
Number of epochs. Default: 1.
verbose : int
Verbosity level. See https://keras.io.
callbacks : List(keras.callbacks.Callback)
Callbacks to be applied during training. See https://keras.io/callbacks
validation_data : tuple, Sequence or None
Validation data can be a tuple (input_dataset, output_dataset),
or (input_dataset, output_dataset, sample_weights) or
a keras.utils.Sequence instance or a list of validation chromsomoes.
The latter choice only works with when using Cover and Bioseq dataset.
This allows you to train on a dedicated set of chromosomes
and to validate the performance on respective heldout chromosomes.
If None, validation is not applied.
shuffle : boolean
shuffle batches. Default: True.
class_weight : dict
Class weights. See https://keras.io.
sample_weight : np.array or None
Sample weights. See https://keras.io.
initial_epoch : int
Initial epoch at which to start training.
steps_per_epoch : int, None.
Number of steps per epoch. If None, this value is determined from
the dataset size and the batch_size.
use_multiprocessing : boolean
Whether to use multiprocessing. See https://keras.io. Default: False.
workers : int
Number of workers to use in multiprocessing mode. Default: 1.
Examples
--------
.. code-block:: python
model.fit(DATA, LABELS)
"""
if not isinstance(inputs, Sequence):
inputs = _convert_data(self.kerasmodel, inputs, 'input_layers')
outputs = _convert_data(self.kerasmodel, outputs, 'output_layers')
hyper_params = {
'epochs': epochs,
'batch_size': batch_size,
'shuffle': shuffle,
'class_weight': class_weight,
'initial_epoch': initial_epoch,
'steps_per_epoch': steps_per_epoch,
'use_multiprocessing': use_multiprocessing,
'workers': workers
}
self.logger.info('Fit: %s', self.name)
if isinstance(inputs, Sequence):
self.logger.info('using custom Sequence')
else:
self.logger.info("Input:")
self.__dim_logging(inputs)
self.logger.info("Output:")
self.__dim_logging(outputs)
self.timer = time.time()
history = None
self.logger.info("Hyper-parameters:")
for par_ in hyper_params:
self.logger.info('%s: %s', par_, str(hyper_params[par_]))
callbacks = [] if callbacks is None else callbacks
callbacks.append(LambdaCallback(on_epoch_end=lambda epoch, logs: self.logger.info(
"epoch %s: %s",
epoch + 1,
' '.join(["{}=".format(k) +
('{:.4f}' if
abs(logs[k]) > 1e-3
else '{:.4e}').format(logs[k]) for k in logs]))))
if not os.path.exists(os.path.join(self.outputdir, 'evaluation')):
os.mkdir(os.path.join(self.outputdir, 'evaluation'))
if not os.path.exists(os.path.join(self.outputdir, 'evaluation', self.name)):
os.mkdir(os.path.join(self.outputdir, 'evaluation', self.name))
callbacks.append(CSVLogger(os.path.join(self.outputdir,
'evaluation',
self.name,
'training.log')))
if not batch_size:
batch_size = 32
if isinstance(inputs, Sequence):
# input could be a sequence
jseq = inputs
else:
jseq = JangguSequence(batch_size, inputs, outputs, sample_weight, shuffle=shuffle)
if isinstance(validation_data, tuple):
valinputs = _convert_data(self.kerasmodel, validation_data[0],
'input_layers')
valoutputs = _convert_data(self.kerasmodel, validation_data[1],
'output_layers')
sweights = validation_data[2] if len(validation_data) == 3 else None
valjseq = JangguSequence(batch_size, valinputs, valoutputs, sweights, shuffle=False)
elif isinstance(validation_data, Sequence):
valjseq = validation_data
elif isinstance(validation_data, list) and isinstance(validation_data[0], str):
# if the validation data is a list of chromosomes that should
# be used as validation dataset we end up here.
# This is only possible, however, if all input and output datasets
# are Cover or Bioseq dataset.
if not all(hasattr(datum, 'gindexer') \
for datum in [jseq.inputs[k] for k in jseq.inputs] +
[jseq.outputs[k] for k in jseq.outputs]):
raise ValueError("Not all dataset are Cover or Bioseq dataset"
" which is required for this options.")
# then split the original dataset into training and validation set.
train, val = split_train_test((jseq.inputs, jseq.outputs), validation_data)
traininp, trainoup = train
valinp, valoup = val
self.logger.info("Split in training and validation set:")
self.logger.info("Training-Input:")
self.__dim_logging(traininp)
self.logger.info("Training-Output:")
self.__dim_logging(trainoup)
self.logger.info("Validation-Input:")
self.__dim_logging(valinp)
self.logger.info("Validation-Output:")
self.__dim_logging(valoup)
jseq = JangguSequence(jseq.batch_size,
_convert_data(self.kerasmodel, traininp,
'input_layers'),
_convert_data(self.kerasmodel, trainoup,
'output_layers'),
sample_weights=None, shuffle=jseq.shuffle)
valjseq = JangguSequence(jseq.batch_size,
_convert_data(self.kerasmodel, valinp,
'input_layers'),
_convert_data(self.kerasmodel, valoup,
'output_layers'),
sample_weights=None, shuffle=False)
else:
valjseq = None
try:
history = self.kerasmodel.fit_generator(
jseq,
epochs=epochs,
validation_data=valjseq,
class_weight=class_weight,
initial_epoch=initial_epoch,
shuffle=shuffle,
use_multiprocessing=use_multiprocessing,
max_queue_size=50,
workers=workers,
verbose=verbose,
callbacks=callbacks)
except Exception: # pragma: no cover
# ignore the linter warning, the exception
# is reraised anyways.
self.logger.exception('fit_generator failed:')
raise
self.logger.info('#' * 40)
for k in history.history:
self.logger.info('%s: %f', k, history.history[k][-1])
self.logger.info('#' * 40)
self.save()
self._save_hyper(hyper_params)
self.logger.info("Training finished after %1.3f s",
time.time() - self.timer)
return history
def objective(params):
print(params)
try:
train_data = get_data(params)
train_data, test = split_train_test(train_data, [test_chrom])
train, val = split_train_test(train_data, [params['val_chrom']])
# define a keras model only based on DNA
K.clear_session()
if params['inputs'] == 'epi_dna':
dnam = Janggu.create_by_name('cage_promoters_dna_only')
epim = Janggu.create_by_name('cage_promoters_epi_only')
layer = Concatenate()([
dnam.kerasmodel.layers[-2].output,
epim.kerasmodel.layers[-2].output
])
layer = Dense(1, name='geneexpr')(layer)
model = Janggu([dnam.kerasmodel.input] + epim.kerasmodel.input,
layer,
name='cage_promoters_epi_dna')
if not params['pretrained']:
# This part randomly reinitializes the network
# so that we can train it from scratch
newjointmodel = model_from_json(model.kerasmodel.to_json())
newjointmodel = Janggu(
newjointmodel.inputs,
newjointmodel.outputs,
name='cage_promoters_epi_dna_randominit')
model = newjointmodel
else:
model = Janggu.create(get_model,
params,
train_data[0],
train_data[1],
name='cage_promoters_{}'.format(
params['inputs']))
except ValueError:
main_logger.exception('objective:')
return {'status': 'fail'}
model.compile(optimizer=get_opt(params['opt']),
loss='mae',
metrics=['mse'])
hist = model.fit(
train_data[0],
train_data[1],
epochs=params['epochs'],
batch_size=64,
validation_data=[params['val_chrom']],
callbacks=[EarlyStopping(patience=5, restore_best_weights=True)])
print('#' * 40)
for key in hist.history:
print('{}: {}'.format(key, hist.history[key][-1]))
print('#' * 40)
pred_train = model.predict(train[0])
pred_val = model.predict(val[0])
pred_test = model.predict(test[0])
model.evaluate(train[0],
train[1],
callbacks=['var_explained', 'mse', 'mae', 'cor'],
datatags=['train'])
mae_val = model.evaluate(val[0],
val[1],
callbacks=['var_explained', 'mse', 'mae', 'cor'],
datatags=['val'])
mae_val = mae_val[0]
model.evaluate(test[0],
test[1],
callbacks=['var_explained', 'mse', 'mae', 'cor'],
datatags=['test'])
cor_train = np.corrcoef(train[1][:][:, 0], pred_train[:, 0])[0, 1]
cor_val = np.corrcoef(val[1][:][:, 0], pred_val[:, 0])[0, 1]
cor_test = np.corrcoef(test[1][:][:, 0], pred_test[:, 0])[0, 1]
model.summary()
main_logger.info('cor [train/val/test]: {:.2f}/{:.2f}/{:.2f}'.format(
cor_train, cor_val, cor_test))
return {
'loss': mae_val,
'status': 'ok',
'all_losses': hist.history,
'cor_train': cor_train,
'cor_val': cor_val,
'cor_test': cor_test,
'model_config': model.kerasmodel.to_json(),
'model_weights': model.kerasmodel.get_weights(),
'concrete_params': params
}
shared_space['inputs'] = 'epi_dna'
shared_space['pretrained'] = False
res = objective(shared_space)
write_results(shared_space, res)
else:
print('no training')
shared_space['val_chrom'] = "chr22"
shared_space['order'] = dnaorder
shared_space['pretrained'] = False
shared_space['seq_dropout'] = 0.2
shared_space['inputs'] = 'epi_dna'
params = shared_space
train_data = get_data(params)
train, test = split_train_test(train_data, [test_chrom])
model = Janggu.create_by_name('cage_promoters_epi_dna')
testpred = model.predict(test[0])
fig, ax = plt.subplots()
ax.scatter(test[1][:], testpred)
ax.set_xlabel('Observed normalized CAGE signal')
ax.set_ylabel('Predicted normalized CAGE signal')
fig.savefig(
os.path.join(os.environ['JANGGU_OUTPUT'],
'cage_promoter_testchrom_agreement.png'))
fig, ax = plt.subplots()
ax.scatter(test[1][:], testpred)
def get_data(params):
binsize = params['binsize']
# PEAKS
LABELS = ReduceDim(Cover.create_from_bed('peaks',
bedfiles=PEAKS,
roi=ROI,
binsize=binsize,
conditions=['JunD'],
resolution=binsize,
store_whole_genome=True,
storage='sparse',
cache=True),
aggregator='max')
# training on chr1, validation on chr2, test on chr3 with swapped Dnase samples
LABELS, LABELS_TEST = split_train_test(LABELS, 'chr3')
LABELS_TRAIN, LABELS_VAL = split_train_test(LABELS, 'chr2')
if params['type'] in ['dna_only', 'dnase_dna']:
dnaflank = params['dnaflank']
order = params['order']
# DNA
DNA = Bioseq.create_from_refgenome('dna',
refgenome=REFGENOME,
roi=ROI,
binsize=binsize,
flank=dnaflank,
order=order,
cache=True,
store_whole_genome=True)
DNA, DNA_TEST = split_train_test(DNA, 'chr3')
DNA_TRAIN, DNA_VAL = split_train_test(DNA, 'chr2')
if params['type'] in ['dnase_bam_only', 'dnase_dna']:
dnaseflank = params['dnaseflank']
# ACCESSIBILITY
ACCESS_TEST = Cover.create_from_bam(
'dnase',
bamfiles=[DNASE_STAM_ENCODE, DNASE_STAM_ROADMAP],
roi=ROI,
binsize=binsize,
conditions=['Encode', 'Roadmap'],
flank=dnaseflank,
resolution=50,
normalizer=params['normalize'],
store_whole_genome=True,
cache=True)
ACCESS = Cover.create_from_bam(
'dnase',
roi=ROI,
bamfiles=[DNASE_STAM_ROADMAP, DNASE_STAM_ENCODE],
binsize=binsize,
conditions=['Roadmap', 'Encode'],
resolution=50,
flank=dnaseflank,
normalizer=params['normalize'],
store_whole_genome=True,
cache=True)
_, ACCESS_TEST = split_train_test(ACCESS_TEST, 'chr3')
ACCESS, _ = split_train_test(ACCESS, 'chr3')
ACCESS_TRAIN, ACCESS_VAL = split_train_test(ACCESS, 'chr2')
if params['type'] in ['dna_dnase', 'dnase_bam_only']:
if params['augment'] == 'orient':
ACCESS_TRAIN = RandomOrientation(ACCESS_TRAIN)
if params['augment'] == 'scale':
ACCESS_TRAIN = RandomSignalScale(ACCESS_TRAIN, 0.1)
if params['augment'] == 'both':
ACCESS_TRAIN = RandomSignalScale(RandomOrientation(ACCESS_TRAIN),
0.1)
if params['type'] == 'dna_only':
return (DNA_TRAIN, LABELS_TRAIN), (DNA_VAL, LABELS_VAL), \
(DNA_TEST, LABELS_TEST)
elif params['type'] == 'dnase_dna':
return ([DNA_TRAIN, ACCESS_TRAIN], LABELS_TRAIN), \
([DNA_VAL, ACCESS_VAL], LABELS_VAL),\
([DNA_TEST, ACCESS_TEST], LABELS_TEST)
elif params['type'] in ['dnase_bam_only']:
return ([ACCESS_TRAIN], LABELS_TRAIN), \
([ACCESS_VAL], LABELS_VAL), \
([ACCESS_TEST], LABELS_TEST)