"tfms": tfms.__str__(),
"TRAIN": TRAIN,
"SAVE_DIR": SAVE_DIR,
"model_file_prefix": "",
"is_pretrained": is_pretrained,
"pretrained_dir": pretrained_dir
}
# Training data
sampler = FoldSampler(TRAIN,
LABELS,
mean,
std,
N,
tfms=tfms,
sz=sz,
bs=bs,
n_folds=n_folds,
model_name=model_name,
uniform_augmentations=uniform_augmentations,
shuffle_nonempty_imgs=shuffle_nonempty_imgs,
is_ordinal=is_ordinal,
num_workers=n_workers)
# Model
model_func = get_panda_model(model_name,
arch=arch,
n=model_n_out,
final_dropout=final_dropout,
num_tiles=N)
# create the model config which is saved along the models
def evaluate_model_dir(model_dir,
sampler=None,
TRAIN=None,
LABELS=None,
**kwargs):
"""
Evaluates CV models in out-of-fold fashion and saves some stats to the model dir
Provide either sampler or TRAIN and LABELS.
model_dir: directory containing models
sampler (FoldSampler): optional data sampler instance
TRAIN: optional training images folder
LABELS: optional train.csv path
"""
# load config
config = ModelConfig.fromDir(model_dir)
# load models
models = load_models_from_dir(model_dir)
model_name = config.getField('model_name')
regr = "regr" in model_name
n_folds = len(models)
sz = config.getField('sz')
mean = torch.tensor(np.array(config.getField('mean')).astype(np.float32))
std = torch.tensor(np.array(config.getField('std')).astype(np.float32))
N = config.getField('N')
is_ordinal = config.getMetaField('is_ordinal')
if sampler is None:
assert (TRAIN is not None and LABELS is not None
), "Either sampler or TRAIN + LABELS must be provided"
sampler = FoldSampler(TRAIN,
LABELS,
mean,
std,
N,
tfms=[],
sz=sz,
bs=1,
n_folds=n_folds,
is_ordinal=is_ordinal,
model_name=model_name)
# evaluate out of fold
val_qwks = []
karolinska_preds = []
karolinska_targets = []
radboud_preds = []
radboud_targets = []
all_preds = []
all_targets = []
score_dict = {}
for fold, model in zip(range(n_folds), models):
data = sampler.get_data(fold)
default_metrics, monitor_metric = get_default_metrics(
model_name, data=data, is_ordinal=is_ordinal)
learn = Learner(data,
model,
metrics=default_metrics,
opt_func=Over9000).to_fp16()
learn.create_opt(1e-3, 0.9)
# calculate data provider specific scores
preds, targets, losses = learn.get_preds(with_loss=True)
targets = targets.numpy()
if is_ordinal:
targets = ordinalRegs2cat(targets)
losses = torch.sum(losses.view(preds.shape[0], preds.shape[1]),
axis=1)
if not regr:
if is_ordinal:
preds = ordinalRegs2cat((preds > 0.5).numpy())
else:
preds = np.argmax(preds.numpy(), axis=1)
else:
# convert to categories
preds = regrPreds2cat(preds)
all_preds += list(preds)
all_targets += list(targets)
# fold qwk
val_qwk = cohen_kappa_score(preds, targets, weights="quadratic")
val_qwks.append(val_qwk)
score_dict[f'{fold}_qwk'] = str(val_qwk)
# get 'karolinska' 'radboud' labels
data_providers = [
sampler.df[sampler.df.image_id == os.path.basename(
_id)].data_provider.values[0] for _id in data.valid_ds.items
]
for pred, target, provider in zip(preds, targets, data_providers):
if provider == "karolinska":
karolinska_preds.append(pred)
karolinska_targets.append(target)
else:
radboud_preds.append(pred)
radboud_targets.append(target)
# plot top and min losses
plot_samples(data, losses, preds,
sampler.df[sampler.df.split == fold].image_id.values)
plt.savefig(os.path.join(model_dir,
"losses_fold-{0}.png".format(fold)),
transparent=False)
# confusion matrices
if not regr:
_ = plot_confusion_matrix_scipy(
preds,
targets,
normalize=False,
title='fold:{0} - qwk:{1:.3f}'.format(fold, val_qwk))
plt.savefig(os.path.join(model_dir,
"cm_fold-{0}.png".format(fold)),
transparent=False)
cm = plot_confusion_matrix_scipy(
preds,
targets,
normalize=True,
title='Norm. fold:{0} - qwk:{1:.3f}'.format(fold, val_qwk))
plt.savefig(os.path.join(model_dir,
"cm_fold-{0}-norm.png".format(fold)),
transparent=False)
else:
_ = plot_confusion_matrix_scipy(
preds,
targets,
normalize=False,
title='fold:{0} - qwk:{1:.3f}'.format(fold, val_qwk))
plt.savefig(os.path.join(model_dir,
"cm_fold-{0}.png".format(fold)),
transparent=False)
cm = plot_confusion_matrix_scipy(
preds,
targets,
normalize=True,
title='Norm. fold:{0} - qwk:{1:.3f}'.format(fold, val_qwk))
plt.savefig(os.path.join(model_dir,
"cm_fold-{0}-norm.png".format(fold)),
transparent=False)
# save confusion matrix values
np.save(os.path.join(model_dir, "cm_fold-{0}.npy".format(fold)), cm)
cv_qwk = cohen_kappa_score(np.array(all_preds),
np.array(all_targets),
weights="quadratic")
score_dict['cv_qwk'] = str(cv_qwk)
score_dict['karolinska_qwk'] = str(
cohen_kappa_score(karolinska_preds,
karolinska_targets,
weights="quadratic"))
score_dict['radboud_qwk'] = str(
cohen_kappa_score(radboud_preds, radboud_targets, weights="quadratic"))
# save out-of-fold predictions
np.save(os.path.join(model_dir, 'oof_preds.npy'), np.array(all_preds))
np.save(os.path.join(model_dir, 'oof_trues.npy'), np.array(all_targets))
with open(os.path.join(model_dir, 'eval.json'), 'w') as outfile:
json.dump(score_dict, outfile, indent=4)
# record for the notebook
print(score_dict)
plt.close('all')