I have problems for 4channel dataset to run mask rcnn about medicaldetectiontoolkit

[pfjaeger]

hi. What is the shape of the data array, which your 'generate_train_batch' method returns?
it should be (batch, channel, x, y, (z)). If you are using the lidc data loader as a starting point, note, that it has a dummy dimension for channels already (np.newaxis in line 232).

[cammu00]

Hi, Thank you for your response.
The shape of my data is (4, 240,240,155)
I modified the lidc data loader like following this:
#data = np.transpose(np.load(patient['data'], mmap_mode='r'), axes=(1, 2, 0))[np.newaxis]
data = np.transpose(np.load(patient['data'], mmap_mode='r'), axes=(0, 1, 2, 3))[np.newaxis]
and generate_train_batch returns (4, 1, 4, 156, 96, 240).

Best,
David.

[cammu00]

Sorry, just in case.
If I run n_channel dataset, do I have to match data shape with seg_data(roi mask)?

best,
David.

[pfjaeger]

you need to get rid of the [np.newaxis]. This introduces a dummy channel dimension for lidc data, which you dont need. As I said, check that generate_train_batch returns (batch, channel, x, y, (z))

[pfjaeger]

the seg should have shape (batch, 1, x, y, (z)) and contain class labels as integer values.

[cammu00]

Thank you. training code now is work!
But I met an error after first epoch following like this:
Traceback (most recent call last):
File "exec.py", line 172, in
train(logger)
File "exec.py", line 85, in train
, monitor_metrics['train'] = train_evaluator.evaluate_predictions(train_results_list, monitor_metrics['train'])
File "/data2/project/keras/medicaldetectiontoolkit/evaluator.py", line 191, in evaluate_predictions
return self.return_metrics(monitor_metrics)
File "/data2/project/keras/medicaldetectiontoolkit/evaluator.py", line 298, in return_metrics
plotting.plot_stat_curves(all_stats, out_filename)
File "/data2/project/keras/medicaldetectiontoolkit/plotting.py", line 263, in plot_stat_curves
plt.plot(s[c][0], s[c][1], label=s['name'] + '' + c)
TypeError: 'float' object is not subscriptable

Do you comment for this how to resolve it?
Many thanks,

Best,
David.

[pfjaeger]

hi by default the plot_stat_curves are switched off in configs, because in most scenarios they are less informative than the prediction histograms. Did you intentionally switch on this flag in the configs? As for your error I was not able to reproduce this. For Debugging I recommend to print our your all_stats list and check if it has the expected structure.

[cammu00]

Hi, Thank you. I didn't realize that I did switch on it in configs. I did back to off and now it's working well.
Thank you for your help :).

Best,
David.

[pfjaeger]

glad I could help. Feel free to join our new slack channel for further discussion: https://join.slack.com/t/mdtoolkit/shared_invite/enQtNTQ3MjY2MzE0MDg2LWNjY2I2Njc5MTY0NmM0ZWIxNmQwZDRhYzk2MDdhM2QxYjliYTcwYzhkNTAxYmRkMDA0MjcyNDMyYjllNTZhY2M