def mse_vs_epochs(coeff, key, data_loaders, data_lengths, only_test=True):
mse = []
f_list, epochs = get_fnames(coeff)
for fname, e in zip(f_list[20:], epochs[20:]):
print(fname, e)
torch.cuda.empty_cache()
model = rUNet(out_size=1)
checkpoint = torch.load(os.path.join(saved_models,
fname))['model_state_dict']
model.load_state_dict(checkpoint)
y_true, y_pred = inference_phase_rUNet(model,
data_loaders,
data_lengths,
batch_size=16,
notebook=False,
test=only_test)
np.savez_compressed(os.path.join(
saved_models, '_'.join([
'predicted', key,
fname.split('_')[3],
fname.split('_')[5],
fname.split('_')[6]
]) + '.npz'),
true=y_true,
pred=y_pred)
mse.append(mean_squared_error(y_true, y_pred))
#plt.plot(epochs, mse)
return mse, epochs
def predict_dist(coeff,
epoch,
data_loaders,
data_lengths,
key,
only_test=True):
filelist, epochs = get_fnames(coeff)
try:
sel_epoch = np.argwhere(epochs == epoch)[0]
model_fname = filelist[int(sel_epoch)]
except IndexError as ie:
print(str(ie), coeff, epoch)
return
print(model_fname)
torch.cuda.empty_cache()
model = rUNet(out_size=1)
try:
checkpoint_file = torch.load(root_path(model_fname))
checkpoint = checkpoint_file['model_state_dict']
except RuntimeError as rte:
print(str(rte), root_path(model_fname))
except FileNotFoundError as fnf:
print(str(fnf), root_path(model_fname))
except KeyError as ke:
print(str(ke), root_path(model_fname))
else:
out_filename = '_'.join([
'predicted', key,
model_fname.split('_')[3],
model_fname.split('_')[5],
model_fname.split('_')[6]
]) + '.npz'
out_filepath = os.path.join(saved_models, out_filename)
if os.path.exists(out_filepath):
print('out file {} exists!'.format(out_filepath))
return
model.load_state_dict(checkpoint)
y_true, y_pred = inference_phase_rUNet(model,
data_loaders,
data_lengths,
batch_size=8,
notebook=False,
test=only_test)
print('saving to {}'.format(out_filename))
np.savez_compressed(out_filepath, true=y_true, pred=y_pred)
mse = mean_squared_error(y_true, y_pred)
print(mse)
ToTensor()
]
print("Load dataset")
data_loaders, data_length = define_dataset(
root_folder=ROOT_DIR,
base_transformers=base_transformers,
train_transformers=train_transformers,
batch_size=16,
excluded_list=EXCLUDED,
alldata=False,
multi_processing=4)
print("combined loss: {}*dice_loss + {} mse".format(coef, 1.0 - coef))
torch.cuda.empty_cache()
print("Train model")
model = rUNet(out_size=1)
optimizer = optim.Adam(model.parameters(), lr=1e-4)
checkpoint_file = os.path.join(
SRC_DIR, 'saved_models', 'trained_6positions_multi_loss',
'Trained_rUNet_pytorch_6positions_dataset_100epochs_{}coeff_mask.pkl'
.format(coef))
print(torch.load(checkpoint_file).keys())
history = retrain_rUNet_multi_loss(
model=model,
optimizer=optimizer,
criterion_dist=nn.MSELoss(),
criterion_mask=dice_loss,
loss_coeff=coef,
data_loaders=data_loaders,
data_lengths=data_length,
checkpoint_file=checkpoint_file,
