def cli_preproc(command, raw_args):
"""Preprocess:
- Run the dataloader and store the results to a (hdf5) file
"""
assert command == "preproc"
parser = argparse.ArgumentParser(
'kipoi {}'.format(command),
description='Run the dataloader and save the output to an hdf5 file.')
add_dataloader_main(parser, with_args=True)
parser.add_argument('--batch_size',
type=int,
default=32,
help='Batch size to use in data loading')
parser.add_argument("-i",
"--install_req",
action='store_true',
help="Install required packages from requirements.txt")
parser.add_argument(
"-n",
"--num_workers",
type=int,
default=0,
help="Number of parallel workers for loading the dataset")
parser.add_argument("-o",
"--output",
required=True,
help="Output hdf5 file")
args = parser.parse_args(raw_args)
dataloader_kwargs = parse_json_file_str(args.dataloader_args)
dir_exists(os.path.dirname(args.output), logger)
# --------------------------------------------
# install args
if args.install_req:
kipoi.pipeline.install_dataloader_requirements(args.dataloader,
args.source)
Dataloader = kipoi.get_dataloader_factory(args.dataloader, args.source)
dataloader_kwargs = kipoi.pipeline.validate_kwargs(Dataloader,
dataloader_kwargs)
dataloader = Dataloader(**dataloader_kwargs)
it = dataloader.batch_iter(batch_size=args.batch_size,
num_workers=args.num_workers)
logger.info("Writing to the hdf5 file: {0}".format(args.output))
writer = writers.HDF5BatchWriter(file_path=args.output)
for i, batch in enumerate(tqdm(it)):
# check that the first batch was indeed correct
if i == 0 and not Dataloader.output_schema.compatible_with_batch(
batch):
logger.warn(
"First batch of data is not compatible with the dataloader schema."
)
writer.batch_write(batch)
writer.close()
logger.info("Done!")
def test_gradplotter():
from kipoi.postprocessing.gradient_vis.vis import GradPlotter, get_selector
example = "rbp"
if example in {"rbp", "non_bedinput_model", "iris_model_template"} and sys.version_info[0] == 2:
pytest.skip("rbp example not supported on python 2 ")
example_dir = "tests/models/{0}".format(example)
output = os.path.realpath(example_dir + "/grad_outputs.hdf5")
try:
os.unlink(output)
except:
pass
writer = writers.HDF5BatchWriter(file_path=output)
get_example_data(example, predict_activation_layers[example], writer=writer)
gp = GradPlotter.from_hdf5(output, example_dir, source="dir")
# test get_num_samples
assert gp.data['inputs']['seq'].shape[0] == gp.get_num_samples("seq")
# once we have a gp instance:
exp_ret_fns = [gp._select_ds_dict, gp._select_ds_list, gp._select_ds_ndarray]
for model_schema_yaml, exp_ret_fn in zip(MODEL_SCHEMA_EXAMPLES, exp_ret_fns):
schema = ModelSchema.from_config(from_yaml(model_schema_yaml))
ret_fn, ret_lables = gp._get_ds_extractor(schema.inputs)
assert ret_fn == exp_ret_fn
assert ret_lables == ['seq']
try:
os.unlink(output)
except:
pass
def test_deeplift():
# return True
example = "tal1_model"
layer = predict_activation_layers[example]
example_dir = "tests/models/{0}".format(example)
if INSTALL_REQ:
install_model_requirements(example_dir, "dir", and_dataloaders=True)
model = kipoi.get_model(example_dir, source="dir")
# The preprocessor
Dataloader = kipoi.get_dataloader_factory(example_dir, source="dir")
#
with open(example_dir + "/example_files/test.json", "r") as ifh:
dataloader_arguments = json.load(ifh)
for k in dataloader_arguments:
dataloader_arguments[k] = "example_files/" + dataloader_arguments[k]
d = DeepLift(model,
output_layer=-2,
task_idx=0,
preact=None,
mxts_mode='grad_times_inp')
new_ofname = model.source_dir + "/example_files/deeplift_grads_pred.hdf5"
if os.path.exists(new_ofname):
os.unlink(new_ofname)
writer = writers.HDF5BatchWriter(file_path=new_ofname)
with kipoi.utils.cd(model.source_dir):
dl = Dataloader(**dataloader_arguments)
it = dl.batch_iter(batch_size=32, num_workers=0)
# Loop through the data, make predictions, save the output
for i, batch in enumerate(tqdm(it)):
# make the prediction
pred_batch = d.score(batch['inputs'], None)
# Using Avanti's recommendation to check whether the model conversion has worked.
pred_batch_fwd = d.predict_on_batch(batch['inputs'])
orig_pred_batch_fwd = model.predict_on_batch(batch['inputs'])
assert np.all(pred_batch_fwd == orig_pred_batch_fwd)
output_batch = batch
output_batch["input_grad"] = pred_batch
writer.batch_write(output_batch)
writer.close()
new_res = readers.HDF5Reader.load(new_ofname)
ref_res = readers.HDF5Reader.load(model.source_dir +
"/example_files/grads.hdf5")
assert np.all(
np.isclose(new_res['input_grad'],
(ref_res['inputs'] * ref_res['grads'])))
if os.path.exists(new_ofname):
os.unlink(new_ofname)
def test_score():
example = "tal1_model"
layer = predict_activation_layers[example]
example_dir = "example/models/{0}".format(example)
if INSTALL_REQ:
install_model_requirements(example_dir, "dir", and_dataloaders=True)
model = kipoi.get_model(example_dir, source="dir")
# The preprocessor
Dataloader = kipoi.get_dataloader_factory(example_dir, source="dir")
#
with open(example_dir + "/example_files/test.json", "r") as ifh:
dataloader_arguments = json.load(ifh)
for k in dataloader_arguments:
dataloader_arguments[k] = "example_files/" + dataloader_arguments[k]
g = Gradient(model, None, layer=layer, avg_func="sum")
if os.path.exists(model.source_dir + "/example_files/grads_pred.hdf5"):
os.unlink(model.source_dir + "/example_files/grads_pred.hdf5")
writer = writers.HDF5BatchWriter(file_path=model.source_dir + "/example_files/grads_pred.hdf5")
with kipoi_utils.utils.cd(model.source_dir):
dl = Dataloader(**dataloader_arguments)
it = dl.batch_iter(batch_size=32, num_workers=0)
# Loop through the data, make predictions, save the output
for i, batch in enumerate(tqdm(it)):
# make the prediction
pred_batch = g.score(batch['inputs'])
output_batch = batch
output_batch["grads"] = pred_batch
writer.batch_write(output_batch)
writer.close()
obj1 = readers.HDF5Reader.load(model.source_dir + "/example_files/grads_pred.hdf5")
obj2 = readers.HDF5Reader.load(model.source_dir + "/example_files/grads.hdf5")
kipoi_utils.utils.compare_numpy_dict(obj1, obj2)
if os.path.exists(model.source_dir + "/example_files/grads_pred.hdf5"):
os.unlink(model.source_dir + "/example_files/grads_pred.hdf5")
def cli_predict(command, raw_args):
"""CLI interface to predict
"""
assert command == "predict"
parser = argparse.ArgumentParser('kipoi {}'.format(command),
description='Run the model prediction.')
add_model(parser)
add_dataloader(parser, with_args=True)
parser.add_argument('--batch_size', type=int, default=32,
help='Batch size to use in prediction')
parser.add_argument("-n", "--num_workers", type=int, default=0,
help="Number of parallel workers for loading the dataset")
parser.add_argument("-i", "--install_req", action='store_true',
help="Install required packages from requirements.txt")
parser.add_argument("-k", "--keep_inputs", action='store_true',
help="Keep the inputs in the output file. ")
parser.add_argument("-l", "--layer",
help="Which output layer to use to make the predictions. If specified," +
"`model.predict_activation_on_batch` will be invoked instead of `model.predict_on_batch`")
parser.add_argument('-o', '--output', required=True, nargs="+",
help="Output files. File format is inferred from the file path ending. Available file formats are: " +
", ".join(["." + k for k in writers.FILE_SUFFIX_MAP]))
args = parser.parse_args(raw_args)
dataloader_kwargs = parse_json_file_str(args.dataloader_args)
# setup the files
if not isinstance(args.output, list):
args.output = [args.output]
for o in args.output:
ending = o.split('.')[-1]
if ending not in writers.FILE_SUFFIX_MAP:
logger.error("File ending: {0} for file {1} not from {2}".
format(ending, o, writers.FILE_SUFFIX_MAP))
sys.exit(1)
dir_exists(os.path.dirname(o), logger)
# --------------------------------------------
# install args
if args.install_req:
kipoi.pipeline.install_model_requirements(args.model,
args.source,
and_dataloaders=True)
# load model & dataloader
model = kipoi.get_model(args.model, args.source)
if args.dataloader is not None:
Dl = kipoi.get_dataloader_factory(args.dataloader, args.dataloader_source)
else:
Dl = model.default_dataloader
dataloader_kwargs = kipoi.pipeline.validate_kwargs(Dl, dataloader_kwargs)
dl = Dl(**dataloader_kwargs)
# setup batching
it = dl.batch_iter(batch_size=args.batch_size,
num_workers=args.num_workers)
# Setup the writers
use_writers = []
for output in args.output:
ending = output.split('.')[-1]
W = writers.FILE_SUFFIX_MAP[ending]
logger.info("Using {0} for file {1}".format(W.__name__, output))
if ending == "tsv":
assert W == writers.TsvBatchWriter
use_writers.append(writers.TsvBatchWriter(file_path=output, nested_sep="/"))
elif ending == "bed":
assert W == writers.BedBatchWriter
use_writers.append(writers.BedBatchWriter(file_path=output,
dataloader_schema=dl.output_schema.metadata,
header=True))
elif ending in ["hdf5", "h5"]:
assert W == writers.HDF5BatchWriter
use_writers.append(writers.HDF5BatchWriter(file_path=output))
else:
logger.error("Unknown file format: {0}".format(ending))
sys.exit(1)
# Loop through the data, make predictions, save the output
for i, batch in enumerate(tqdm(it)):
# validate the data schema in the first iteration
if i == 0 and not Dl.output_schema.compatible_with_batch(batch):
logger.warn("First batch of data is not compatible with the dataloader schema.")
# make the prediction
if args.layer is None:
pred_batch = model.predict_on_batch(batch['inputs'])
else:
pred_batch = model.predict_activation_on_batch(batch['inputs'], layer=args.layer)
# write out the predictions, metadata (, inputs, targets)
output_batch = prepare_batch(batch, pred_batch, keep_inputs=args.keep_inputs)
for writer in use_writers:
writer.batch_write(output_batch)
for writer in use_writers:
writer.close()
logger.info('Done! Predictions stored in {0}'.format(",".join(args.output)))
def cli_grad(command, raw_args):
"""CLI interface to predict
"""
from .main import prepare_batch
from kipoi.model import GradientMixin
assert command == "grad"
from tqdm import tqdm
parser = argparse.ArgumentParser(
'kipoi {}'.format(command),
description='Save gradients and inputs to a hdf5 file.')
add_model(parser)
add_dataloader(parser, with_args=True)
parser.add_argument('--batch_size',
type=int,
default=32,
help='Batch size to use in prediction')
parser.add_argument(
"-n",
"--num_workers",
type=int,
default=0,
help="Number of parallel workers for loading the dataset")
parser.add_argument("-i",
"--install_req",
action='store_true',
help="Install required packages from requirements.txt")
parser.add_argument(
"-l",
"--layer",
default=None,
help="Which output layer to use to make the predictions. If specified,"
+
"`model.predict_activation_on_batch` will be invoked instead of `model.predict_on_batch`",
required=False)
parser.add_argument(
"--final_layer",
help=
"Alternatively to `--layer` this flag can be used to indicate that the last layer should "
"be used.",
action='store_true')
parser.add_argument(
"--pre_nonlinearity",
help=
"Flag indicating that it should checked whether the selected output is post activation "
"function. If a non-linear activation function is used attempt to use its input. This "
"feature is not available for all models.",
action='store_true')
parser.add_argument(
"-f",
"--filter_idx",
help=
"Filter index that should be inspected with gradients. If not set all filters will "
+ "be used.",
default=None)
parser.add_argument(
"-a",
"--avg_func",
help=
"Averaging function to be applied across selected filters (`--filter_idx`) in "
+ "layer `--layer`.",
choices=GradientMixin.allowed_functions,
default="sum")
parser.add_argument(
'--selected_fwd_node',
help="If the selected layer has multiple inbound connections in "
"the graph then those can be selected here with an integer "
"index. Not necessarily supported by all models.",
default=None,
type=int)
parser.add_argument(
'-o',
'--output',
required=True,
nargs="+",
help=
"Output files. File format is inferred from the file path ending. Available file formats are: "
+ ", ".join(["." + k for k in writers.FILE_SUFFIX_MAP]))
args = parser.parse_args(raw_args)
dataloader_kwargs = parse_json_file_str(args.dataloader_args)
# setup the files
if not isinstance(args.output, list):
args.output = [args.output]
for o in args.output:
ending = o.split('.')[-1]
if ending not in writers.FILE_SUFFIX_MAP:
logger.error("File ending: {0} for file {1} not from {2}".format(
ending, o, writers.FILE_SUFFIX_MAP))
sys.exit(1)
dir_exists(os.path.dirname(o), logger)
# --------------------------------------------
# install args
if args.install_req:
kipoi.pipeline.install_model_requirements(args.model,
args.source,
and_dataloaders=True)
layer = args.layer
if layer is None and not args.final_layer:
raise Exception(
"A layer has to be selected explicitely using `--layer` or implicitely by using the"
"`--final_layer` flag.")
# Not a good idea
# if layer is not None and isint(layer):
# logger.warn("Interpreting `--layer` value as integer layer index!")
# layer = int(args.layer)
# load model & dataloader
model = kipoi.get_model(args.model, args.source)
if not isinstance(model, GradientMixin):
raise Exception("Model does not support gradient calculation.")
if args.dataloader is not None:
Dl = kipoi.get_dataloader_factory(args.dataloader,
args.dataloader_source)
else:
Dl = model.default_dataloader
dataloader_kwargs = kipoi.pipeline.validate_kwargs(Dl, dataloader_kwargs)
dl = Dl(**dataloader_kwargs)
filter_idx_parsed = None
if args.filter_idx is not None:
filter_idx_parsed = parse_filter_slice(args.filter_idx)
# setup batching
it = dl.batch_iter(batch_size=args.batch_size,
num_workers=args.num_workers)
# Setup the writers
use_writers = []
for output in args.output:
ending = output.split('.')[-1]
W = writers.FILE_SUFFIX_MAP[ending]
logger.info("Using {0} for file {1}".format(W.__name__, output))
if ending == "tsv":
assert W == writers.TsvBatchWriter
use_writers.append(
writers.TsvBatchWriter(file_path=output, nested_sep="/"))
elif ending == "bed":
raise Exception("Please use tsv or hdf5 output format.")
elif ending in ["hdf5", "h5"]:
assert W == writers.HDF5BatchWriter
use_writers.append(writers.HDF5BatchWriter(file_path=output))
else:
logger.error("Unknown file format: {0}".format(ending))
sys.exit(1)
# Loop through the data, make predictions, save the output
for i, batch in enumerate(tqdm(it)):
# validate the data schema in the first iteration
if i == 0 and not Dl.output_schema.compatible_with_batch(batch):
logger.warn(
"First batch of data is not compatible with the dataloader schema."
)
# make the prediction
pred_batch = model.input_grad(batch['inputs'],
filter_idx=filter_idx_parsed,
avg_func=args.avg_func,
layer=layer,
final_layer=args.final_layer,
selected_fwd_node=args.selected_fwd_node,
pre_nonlinearity=args.pre_nonlinearity)
# write out the predictions, metadata (, inputs, targets)
# always keep the inputs so that input*grad can be generated!
# output_batch = prepare_batch(batch, pred_batch, keep_inputs=True)
output_batch = batch
output_batch["grads"] = pred_batch
for writer in use_writers:
writer.batch_write(output_batch)
for writer in use_writers:
writer.close()
logger.info('Done! Gradients stored in {0}'.format(",".join(args.output)))
def cli_ism(command, raw_args):
# TODO: find a way to define the model output selection
"""CLI interface to predict
"""
# from .main import prepare_batch
assert command == "ism"
from tqdm import tqdm
from .ism import Mutation
parser = argparse.ArgumentParser('kipoi interpret {}'.format(command),
description='Calculate DeepLIFT scores.')
add_model(parser)
add_dataloader(parser, with_args=True)
parser.add_argument('--batch_size',
type=int,
default=32,
help='Batch size to use in prediction')
parser.add_argument(
"-n",
"--num_workers",
type=int,
default=0,
help="Number of parallel workers for loading the dataset")
parser.add_argument("--model_input",
help="Name of the model input that should be scored.",
required=True)
parser.add_argument(
'-s',
"--scores",
default="diff",
nargs="+",
help=
"Scoring method to be used. Only scoring methods selected in the model yaml file are"
"available except for `diff` which is always available. Select scoring function by the"
"`name` tag defined in the model yaml file.")
parser.add_argument(
'-k',
"--score_kwargs",
default=None,
nargs="+",
help=
"JSON definition of the kwargs for the scoring functions selected in --scores. The "
"definiton can either be in JSON in the command line or the path of a .json file. The "
"individual JSONs are expected to be supplied in the same order as the labels defined in "
"--scores. If the defaults or no arguments should be used define '{}' for that respective "
"scoring method.")
parser.add_argument(
"-c",
"--category_axis",
help="Using the selected model input with `--model_input`: Which "
"dimension of that array contains the one-hot encoded categories?"
" e.g. for a one-hot encoded DNA-sequence"
"array with input shape (1000, 4) for a single sample, "
"`category_dim` is 1, for (4, 1000) `category_dim`"
"is 0.",
default=1,
type=int,
required=False)
parser.add_argument(
"-f",
"--output_sel_fn",
help="Define an output selection function in order to return effects"
"on the output of the function. example definitoin: "
"`--output_sel_fn my_file.py::my_sel_fn`",
default=None,
required=False)
parser.add_argument(
'-o',
'--output',
required=True,
nargs="+",
help="Output files. File format is inferred from the file path ending. "
"Available file formats are: " +
", ".join(["." + k for k in writers.FILE_SUFFIX_MAP]))
args = parser.parse_args(raw_args)
dataloader_kwargs = parse_json_file_str(args.dataloader_args)
# setup the files
if not isinstance(args.output, list):
args.output = [args.output]
for o in args.output:
ending = o.split('.')[-1]
if ending not in writers.FILE_SUFFIX_MAP:
logger.error("File ending: {0} for file {1} not from {2}".format(
ending, o, writers.FILE_SUFFIX_MAP))
sys.exit(1)
dir_exists(os.path.dirname(o), logger)
# --------------------------------------------
if not isinstance(args.scores, list):
args.scores = [args.scores]
# load model & dataloader
model = kipoi.get_model(args.model, args.source)
if args.dataloader is not None:
Dl = kipoi.get_dataloader_factory(args.dataloader,
args.dataloader_source)
else:
Dl = model.default_dataloader
dataloader_kwargs = kipoi.pipeline.validate_kwargs(Dl, dataloader_kwargs)
dl = Dl(**dataloader_kwargs)
# setup batching
it = dl.batch_iter(batch_size=args.batch_size,
num_workers=args.num_workers)
# Setup the writers
use_writers = []
for output in args.output:
ending = output.split('.')[-1]
W = writers.FILE_SUFFIX_MAP[ending]
logger.info("Using {0} for file {1}".format(W.__name__, output))
if ending == "tsv":
assert W == writers.TsvBatchWriter
use_writers.append(
writers.TsvBatchWriter(file_path=output, nested_sep="/"))
elif ending == "bed":
raise Exception("Please use tsv or hdf5 output format.")
elif ending in ["hdf5", "h5"]:
assert W == writers.HDF5BatchWriter
use_writers.append(writers.HDF5BatchWriter(file_path=output))
else:
logger.error("Unknown file format: {0}".format(ending))
sys.exit(1)
output_sel_fn = None
if args.output_sel_fn is not None:
file_path, obj_name = tuple(args.output_sel_fn.split("::"))
output_sel_fn = getattr(load_module(file_path), obj_name)
m = Mutation(model,
args.model_input,
scores=args.scores,
score_kwargs=args.score_kwargs,
batch_size=args.batch_size,
output_sel_fn=output_sel_fn,
category_axis=args.category_axis,
test_ref_ref=True)
out_batches = {}
# Loop through the data, make predictions, save the output..
# TODO: batch writer fails because it tries to concatenate on highest dimension rather than the lowest!
for i, batch in enumerate(tqdm(it)):
# validate the data schema in the first iteration
if i == 0 and not Dl.output_schema.compatible_with_batch(batch):
logger.warn(
"First batch of data is not compatible with the dataloader schema."
)
# calculate scores without reference for the moment.
pred_batch = m.score(batch['inputs'])
# with the current writers it's not possible to store the scores and the model inputs in the same file
output_batch = {}
output_batch["scores"] = pred_batch
for k in output_batch:
if k not in out_batches:
out_batches[k] = []
out_batches[k].append(output_batch[k])
# concatenate batches:
full_output = {
k: np.concatenate([np.array(el) for el in v])
for k, v in out_batches.items()
}
logger.info('Full output shape: {0}'.format(
str(full_output["scores"].shape)))
for writer in use_writers:
writer.batch_write(full_output)
for writer in use_writers:
writer.close()
logger.info('Done! ISM stored in {0}'.format(",".join(args.output)))
def cli_feature_importance(command, raw_args):
"""CLI interface to predict
"""
# from .main import prepare_batch
assert command == "feature_importance"
parser = argparse.ArgumentParser(
'kipoi {}'.format(command),
description='Save gradients and inputs to a hdf5 file.')
add_model(parser)
add_dataloader(parser, with_args=True)
parser.add_argument("--imp_score",
help="Importance score name",
choices=available_importance_scores())
parser.add_argument("--imp_score_kwargs", help="Importance score kwargs")
parser.add_argument('--batch_size',
type=int,
default=32,
help='Batch size to use in prediction')
parser.add_argument(
"-n",
"--num_workers",
type=int,
default=0,
help="Number of parallel workers for loading the dataset")
# TODO - handle the reference-based importance scores...
# io
parser.add_argument(
'-o',
'--output',
required=True,
nargs="+",
help=
"Output files. File format is inferred from the file path ending. Available file formats are: "
+ ", ".join(["." + k for k in writers.FILE_SUFFIX_MAP]))
args = parser.parse_args(raw_args)
dataloader_kwargs = parse_json_file_str(args.dataloader_args)
imp_score_kwargs = parse_json_file_str(args.imp_score_kwargs)
# setup the files
if not isinstance(args.output, list):
args.output = [args.output]
for o in args.output:
ending = o.split('.')[-1]
if ending not in writers.FILE_SUFFIX_MAP:
logger.error("File ending: {0} for file {1} not from {2}".format(
ending, o, writers.FILE_SUFFIX_MAP))
sys.exit(1)
dir_exists(os.path.dirname(o), logger)
# --------------------------------------------
# install args
if args.install_req:
kipoi.pipeline.install_model_requirements(args.model,
args.source,
and_dataloaders=True)
# load model & dataloader
model = kipoi.get_model(args.model,
args.source,
with_dataloader=args.dataloader is None)
if args.dataloader is not None:
Dl = kipoi.get_dataloader_factory(args.dataloader,
args.dataloader_source)
else:
Dl = model.default_dataloader
dataloader_kwargs = kipoi.pipeline.validate_kwargs(Dl, dataloader_kwargs)
dl = Dl(**dataloader_kwargs)
# get_importance_score
ImpScore = get_importance_score(args.imp_score)
if not ImpScore.is_compatible(model):
raise ValueError("model not compatible with score: {0}".format(
args.imp_score))
impscore = ImpScore(model, **imp_score_kwargs)
# setup batching
it = dl.batch_iter(batch_size=args.batch_size,
num_workers=args.num_workers)
# Setup the writers
use_writers = []
for output in args.output:
ending = output.split('.')[-1]
W = writers.FILE_SUFFIX_MAP[ending]
logger.info("Using {0} for file {1}".format(W.__name__, output))
if ending == "tsv":
assert W == writers.TsvBatchWriter
use_writers.append(
writers.TsvBatchWriter(file_path=output, nested_sep="/"))
elif ending == "bed":
raise Exception("Please use tsv or hdf5 output format.")
elif ending in ["hdf5", "h5"]:
assert W == writers.HDF5BatchWriter
use_writers.append(writers.HDF5BatchWriter(file_path=output))
else:
logger.error("Unknown file format: {0}".format(ending))
sys.exit(1)
# Loop through the data, make predictions, save the output
for i, batch in enumerate(tqdm(it)):
# validate the data schema in the first iteration
if i == 0 and not Dl.output_schema.compatible_with_batch(batch):
logger.warn(
"First batch of data is not compatible with the dataloader schema."
)
# make the prediction
# TODO - handle the reference-based importance scores...
importance_scores = impscore.score(batch['inputs'])
# write out the predictions, metadata (, inputs, targets)
# always keep the inputs so that input*grad can be generated!
# output_batch = prepare_batch(batch, pred_batch, keep_inputs=True)
output_batch = batch
output_batch["importance_scores"] = importance_scores
for writer in use_writers:
writer.batch_write(output_batch)
for writer in use_writers:
writer.close()
logger.info('Done! Importance scores stored in {0}'.format(",".join(
args.output)))
def cli_deeplift(command, raw_args):
"""CLI interface to predict
"""
# TODO: find a way to define the "reference" for a scored sequence.
# from .main import prepare_batch
assert command == "deeplift"
from tqdm import tqdm
from .referencebased import DeepLift
from .referencebased import get_mxts_modes
parser = argparse.ArgumentParser('kipoi interpret {}'.format(command),
description='Calculate DeepLIFT scores.')
add_model(parser)
add_dataloader(parser, with_args=True)
parser.add_argument('--batch_size',
type=int,
default=32,
help='Batch size to use in prediction')
parser.add_argument(
"-n",
"--num_workers",
type=int,
default=0,
help="Number of parallel workers for loading the dataset")
parser.add_argument(
"-l",
"--layer",
type=int,
default=None,
help="With respect to which layer the scores should be calculated.",
required=True)
parser.add_argument(
"--pre_nonlinearity",
help=
"Flag indicating that it should checked whether the selected output is post activation "
"function. If a non-linear activation function is used attempt to use its input. This "
"feature is not available for all models.",
action='store_true')
parser.add_argument(
"-f",
"--filter_idx",
help="Filter index that should be inspected with gradients",
default=None,
required=True,
type=int)
parser.add_argument("-m",
"--mxts_mode",
help="Deeplift score, allowed values are: %s" %
str(list(get_mxts_modes().keys())),
default='rescale_conv_revealcancel_fc')
parser.add_argument(
'-o',
'--output',
required=True,
nargs="+",
help=
"Output files. File format is inferred from the file path ending. Available file formats are: "
+ ", ".join(["." + k for k in writers.FILE_SUFFIX_MAP]))
args = parser.parse_args(raw_args)
dataloader_kwargs = parse_json_file_str(args.dataloader_args)
# setup the files
if not isinstance(args.output, list):
args.output = [args.output]
for o in args.output:
ending = o.split('.')[-1]
if ending not in writers.FILE_SUFFIX_MAP:
logger.error("File ending: {0} for file {1} not from {2}".format(
ending, o, writers.FILE_SUFFIX_MAP))
sys.exit(1)
dir_exists(os.path.dirname(o), logger)
# --------------------------------------------
layer = args.layer
if layer is None and not args.final_layer:
raise Exception(
"A layer has to be selected explicitely using `--layer` or implicitely by using the"
"`--final_layer` flag.")
# Not a good idea
# if layer is not None and isint(layer):
# logger.warn("Interpreting `--layer` value as integer layer index!")
# layer = int(args.layer)
# load model & dataloader
model = kipoi.get_model(args.model, args.source)
if args.dataloader is not None:
Dl = kipoi.get_dataloader_factory(args.dataloader,
args.dataloader_source)
else:
Dl = model.default_dataloader
dataloader_kwargs = kipoi.pipeline.validate_kwargs(Dl, dataloader_kwargs)
dl = Dl(**dataloader_kwargs)
# setup batching
it = dl.batch_iter(batch_size=args.batch_size,
num_workers=args.num_workers)
# Setup the writers
use_writers = []
for output in args.output:
ending = output.split('.')[-1]
W = writers.FILE_SUFFIX_MAP[ending]
logger.info("Using {0} for file {1}".format(W.__name__, output))
if ending == "tsv":
assert W == writers.TsvBatchWriter
use_writers.append(
writers.TsvBatchWriter(file_path=output, nested_sep="/"))
elif ending == "bed":
raise Exception("Please use tsv or hdf5 output format.")
elif ending in ["hdf5", "h5"]:
assert W == writers.HDF5BatchWriter
use_writers.append(writers.HDF5BatchWriter(file_path=output))
else:
logger.error("Unknown file format: {0}".format(ending))
sys.exit(1)
d = DeepLift(model,
output_layer=args.layer,
task_idx=args.filter_idx,
preact=args.pre_nonlinearity,
mxts_mode=args.mxts_mode,
batch_size=args.batch_size)
# Loop through the data, make predictions, save the output
for i, batch in enumerate(tqdm(it)):
# validate the data schema in the first iteration
if i == 0 and not Dl.output_schema.compatible_with_batch(batch):
logger.warn(
"First batch of data is not compatible with the dataloader schema."
)
# calculate scores without reference for the moment.
pred_batch = d.score(batch['inputs'], None)
# write out the predictions, metadata (, inputs, targets)
# always keep the inputs so that input*grad can be generated!
# output_batch = prepare_batch(batch, pred_batch, keep_inputs=True)
output_batch = batch
output_batch["scores"] = pred_batch
for writer in use_writers:
writer.batch_write(output_batch)
for writer in use_writers:
writer.close()
logger.info('Done! Gradients stored in {0}'.format(",".join(args.output)))
