def __prepare_tensor_dataset(self):
tensor_dataset_path = os.path.join(
self.metaconf["ws_path"], "tensor_datasets",
self.dataset_params.tensor_dataset_name)
# compare configs, if not same, refresh dataset
current_config_snapshot_exists = H.config_snapshot(
"dataset_params",
self.dataset_params.params,
"src/data/aux/.dataset_config_snapshot.json",
)
if not current_config_snapshot_exists:
H.makedirs(tensor_dataset_path)
_tqdm_kwargs = {
"desc": "Preparing TensorDataset",
"total": len(self.generic_dataset)
}
for i, sample in tqdm(enumerate(self.generic_dataset),
**_tqdm_kwargs):
f_folder_path = os.path.join(tensor_dataset_path,
f"{sample['texture']}")
H.makedirs(f_folder_path)
f_path = os.path.join(f_folder_path, f"nodule_{i}.pt")
save_nodules = {
"nodule": sample["nodule"],
"texture": sample["texture"]
}
torch.save(save_nodules, f_path)
return tensor_dataset_path
def patient_workflow(config, patient_id, patient_input, output_file):
workflow = pypeliner.workflow.Workflow()
patient_bam_dir = config["bam_directory"] + patient_id
patient_result_dir = config["results_dir"] + patient_id
helpers.makedirs(patient_bam_dir)
helpers.makedirs(patient_result_dir)
input_args = helpers.create_input_args(patient_input, patient_bam_dir)
workflow.setobj(obj=mgd.OutputChunks('sample_id', ),
value=input_args['all_samples'])
workflow.subworkflow(name='align_samples',
func=alignment.align_sample,
axes=('sample_id', ),
args=(
config,
mgd.InputFile('fastq_1',
'sample_id',
fnames=input_args['fastqs_r1']),
mgd.InputFile('fastq_2',
'sample_id',
fnames=input_args['fastqs_r2']),
mgd.InputInstance('sample_id'),
mgd.OutputFile('sample.bam',
'sample_id',
fnames=input_args['all_bams']),
mgd.OutputFile('sample.bam.bai',
'sample_id',
fnames=input_args['all_bais']),
))
workflow.subworkflow(name='run_analyses',
func=analysis.partition_tumour,
args=(
config,
input_args,
patient_id,
patient_result_dir,
mgd.InputFile('sample.bam',
'sample_id',
fnames=input_args['all_bams'],
axes_origin=[]),
mgd.InputFile('sample.bam.bai',
'sample_id',
fnames=input_args['all_bais'],
axes_origin=[]),
mgd.OutputFile(output_file),
))
return workflow
def ctDNA_workflow(args):
pyp = pypeliner.app.Pypeline(config=args)
workflow = pypeliner.workflow.Workflow()
config = helpers.load_yaml(args['config'])
for arg, value in args.iteritems():
config[arg] = value
helpers.makedirs(config["bam_directory"])
helpers.makedirs(config["results_dir"])
inputs = helpers.load_yaml(args['input_yaml'])
patients = inputs.keys()
workflow.setobj(obj=mgd.OutputChunks('patient_id', ), value=patients)
workflow.transform(name='get_input_by_patient',
func=helpers.get_input_by_patient,
ret=mgd.TempOutputObj('patient_input', 'patient_id'),
axes=('patient_id', ),
args=(
inputs,
mgd.InputInstance('patient_id'),
))
workflow.subworkflow(name='patient_workflow',
func=patient_workflow,
axes=('patient_id', ),
args=(
config,
mgd.InputInstance('patient_id'),
mgd.TempInputObj('patient_input', 'patient_id'),
mgd.OutputFile(
os.path.join(config['results_dir'],
'{patient_id}.log'),
'patient_id'),
))
pyp.run(workflow)
def optimizer_step(self,
current_epoch,
batch_idx,
optimizer,
optimizer_idx,
second_order_closure=None):
gradplot_savepath = H.makedirs(
os.path.join(self.metaconf["ws_path"], "artifacts",
"gradflow_plots"))
fig = H.plot_grad_flow(self.named_parameters(), "VAE",
self.global_step, gradplot_savepath)
self.logger.experiment.add_figure("gradflow_plots", fig,
self.global_step)
optimizer.step()
optimizer.zero_grad()
def optimizer_step(
self, current_epoch, batch_idx, optimizer, optimizer_idx, second_order_closure=None
):
g_norm = H.get_gradient_norm(self.parameters())
gradplot_savepath = H.makedirs(
os.path.join(self.metaconf["ws_path"], "artifacts", "gradflow_plots")
)
fig = H.plot_grad_flow(self.named_parameters(), "RLS", 0, gradplot_savepath)
self.logger.experiment.add_figure("gradflow_plots", fig, self.global_step)
if np.isnan(g_norm):
log.warning(" gradient norm is NaN -> skip")
optimizer.zero_grad()
return
elif g_norm > self.hparams.optimizer_max_grad_norm:
log.warning(f" gradient norm is too high: {g_norm:.5f} -> clip to OPTIMIZER_MAX_GRAD_NORM")
torch.nn.utils.clip_grad_norm_(self.parameters(), self.hparams.optimizer_max_grad_norm)
else:
log.info(f" gradient norm: {g_norm:.5f}")
optimizer.step()
optimizer.zero_grad()