val_noise_power_range=(5, 5),
inpainting_train_range=(70, 70),
inpainting_val_range=(70, 70),
)
]
train_eval_grid(
job_name,
train_dncnn_denoising,
eval_dncnn_denoising,
parameters,
to_grid=False,
timeout_train=20,
n_gpus_train=1,
timeout_eval=2,
n_gpus_eval=1,
# timeout=10,
# n_gpus=1,
project='mdeq',
params_to_ignore=[
'grad_clipping',
'n_val',
'train_noise_power_range',
'val_noise_power_range',
'inpainting_train_range',
'inpainting_val_range',
],
val_noise_powers=[5],
inpainting_range=(70, 70),
no_force_32=True,
)
restart_from=48,
)
parameters = []
for i_run in range(n_runs):
parameters += [
dict(seed=i_run, shine=True, **base_params),
dict(seed=i_run, fpn=True, **base_params),
]
res_all = train_eval_grid(
job_name,
train_classifier,
evaluate_classifier,
parameters,
to_grid=False,
timeout_train=100,
n_gpus_train=n_gpus,
timeout_eval=2,
n_gpus_eval=n_gpus,
project='shine',
params_to_ignore=['n_epochs', 'save_at', 'restart_from'],
torch=True,
)
perf_orig = [
res for (res, params) in zip(res_all, parameters)
if not params.get('shine', False) and not params.get('fpn', False)
]
perf_shine = [
res for (res, params) in zip(res_all, parameters)
if params.get('shine', False)
]
) for model_name, model_config in models.items()
]
eval_results, run_ids = train_eval_grid(
# run_ids = ['dncnn-relu_1607341394', 'dcnnn-relu-bn_1607346862']
# eval_results = eval_grid(
job_name,
train,
evaluate,
parameter_grid,
# run_ids=run_ids,
n_samples_eval=n_samples_eval,
timeout_train=20,
n_gpus_train=4,
timeout_eval=4,
n_gpus_eval=1,
# n_samples=n_samples_eval,
# timeout=10,
# n_gpus=1,
to_grid=False,
patch_size=patch_size_eval, # just for eval
batch_size=batch_size_eval, # just for eval
noise_config=noise_config_eval, # just for eval
mode='bsd68', # just for eval
project='soft_thresholding',
return_run_ids=True,
no_force_32=True,
)
print(eval_results)
run_ids = [
'xpdnet_sense__af8_i24_compound_mssim_rf_smb_MWCNNbig_1612431469',
'xpdnet_sense__af4_i24_compound_mssim_rf_smb_MWCNNbig_1612431769',
]
eval_results, run_ids = train_eval_grid(
# eval_results = eval_grid(
job_name,
train_xpdnet,
evaluate_xpdnet,
parameter_grid,
# run_ids=run_ids,
n_samples_eval=50,
timeout_train=20,
n_gpus_train=batch_size if batch_size else 1,
timeout_eval=20,
n_gpus_eval=1,
# n_samples=50,
# timeout=10,
# n_gpus=1,
to_grid=False,
return_run_ids=True,
checkpoints_train=19, # one checkpoint every 5 epochs
resume_checkpoint=5,
resume_run_run_ids=run_ids,
params_to_ignore=['batch_size', 'use_mixed_precision', 'model_size'],
project='fastmri4',
)
print(eval_results)
infer_grid(
refine_big=refine_big,
af=4,
loss=loss,
mask_type='equidistant',
) for _, model_size, model_fun, kwargs, _, n_scales, res in model_specs
]
eval_results = train_eval_grid(
# run_ids = [
# 'xpdnet_sense__af8_CORPD_FBK_compound_mssim_rf_smb_MWCNNmedium_1604901311',
# 'xpdnet_sense__af4_CORPD_FBK_compound_mssim_rf_smb_MWCNNmedium_1604901311',
# ]
# eval_results = eval_grid(
job_name,
train_xpdnet,
evaluate_xpdnet,
parameter_grid,
# run_ids=run_ids,
n_samples_eval=100,
timeout_train=100,
n_gpus_train=1,
timeout_eval=10,
n_gpus_eval=1,
# n_samples=100,
# timeout=10,
# n_gpus=1,
to_grid=False,
params_to_ignore=['model_size'])
print(eval_results)
from jean_zay.submitit.general_submissions import train_eval_grid, eval_grid
job_name = 'dncnn_denoise'
parameters = dict(
n_val=[20],
grad_clipping=[100],
fixed_lr=[True],
loss=['mae'],
n_epochs=[5000],
noise_conditionning=[True, False],
noise_residual_weighting=[True, False],
)
train_eval_grid(
job_name,
train_dncnn_denoising,
eval_dncnn_denoising,
parameters,
# run_ids=[None],
to_grid=True,
timeout_train=60,
n_gpus_train=1,
timeout_eval=2,
n_gpus_eval=1,
# timeout=10,
# n_gpus=1,
project='mdeq',
params_to_ignore=['grad_clipping', 'n_val'],
val_noise_powers=[15, 25, 50],
no_force_32=True,
)
lr=1e-4,
loss='mae',
) for model_name, model_config in models.items()
]
eval_results_50, run_ids = train_eval_grid(
# eval_results = eval_grid(
job_name,
train,
evaluate,
parameter_grid,
n_samples_eval=n_samples_eval,
timeout_train=30,
n_gpus_train=4,
timeout_eval=4,
n_gpus_eval=1,
# n_samples=200,
# timeout=10,
# n_gpus=1,
to_grid=False,
patch_size=patch_size_eval, # just for eval
batch_size=batch_size_eval, # just for eval
noise_config=noise_config_eval, # just for eval
mode='bsd68', # just for eval
project='soft_thresholding',
return_run_ids=True,
)
eval_res = []
for noise_level in noise_levels[:-1]:
noise_config_eval = dict(
noise_input=True,
params = [
dict(use_mixed_precision=[True], model=['pdnet'], **base_params),
]
run_ids = [
'ncpdnet_3d___radial_stacks_mse_dcomp_1614707627',
'ncpdnet_3d___spiral_stacks_mse_dcomp_1614707626',
]
eval_results = train_eval_grid(
# eval_results = eval_grid(
'3d_nc',
train_ncnet_multinet,
evaluate_nc_multinet,
params,
# run_ids=run_ids,
n_gpus_train=1,
timeout_train=60,
n_gpus_eval=1,
n_samples_eval=100,
# timeout=20,
# n_gpus=1,
params_to_ignore=['use_mixed_precision', 'scale_factor'],
checkpoints_train=7,
resume_checkpoint=6,
resume_run_run_ids=run_ids,
project='fastmri4',
)
print(eval_results)
n_primal=n_primal,
contrast=contrast,
n_epochs=n_epochs,
n_samples=n_samples,
af=4,
loss=loss,
) for _, model_size, model_fun, kwargs, _, n_scales, res in model_specs
]
eval_results = train_eval_grid(
job_name,
train_xpdnet,
evaluate_xpdnet,
parameter_grid,
n_samples_eval=100,
timeout_train=60,
n_gpus_train=1,
timeout_eval=10,
n_gpus_eval=1,
to_grid=False,
params_to_ignore=['model_size'],
)
df_results = pd.DataFrame(columns='model_name model_size psnr ssim'.split())
for (name, model_size, _, _, _, _, _), eval_res in zip(model_specs,
eval_results):
df_results = df_results.append(dict(
model_name=name,
model_size=model_size,
psnr=eval_res[0][0],