def check(model_name, fold, checkpoint):
model_info = MODELS[model_name]
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = torch.load(checkpoint, map_location=device)
model = model.to(device)
model.eval()
dataset_valid = NihDataset(fold=fold, img_size=model_info.img_size, is_training=False)
dataloader_valid = DataLoader(dataset_valid,
num_workers=1,
batch_size=1,
shuffle=False)
data_iter = tqdm(enumerate(dataloader_valid), total=len(dataloader_valid))
for iter_num, data in data_iter:
labels = data['categories'].cuda().float()
outputs = model(data['img'].cuda().float())
outputs = outputs.cpu().detach().numpy()
print(outputs, labels)
plt.cla()
plt.imshow(data['img'][0, 0].cpu().detach().numpy())
plt.show()
def generate_predictions(model_name, run, fold, from_epoch=0, to_epoch=100):
run_str = '' if run is None or run == '' else f'_{run}'
predictions_dir = f'../output/oof2/{model_name}{run_str}_fold_{fold}'
os.makedirs(predictions_dir, exist_ok=True)
model_info = MODELS[model_name]
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
for epoch_num in range(from_epoch, to_epoch):
prediction_fn = f'{predictions_dir}/{epoch_num:03}.pkl'
if os.path.exists(prediction_fn):
continue
print('epoch', epoch_num)
checkpoint = f'checkpoints/{model_name}{run_str}_fold_{fold}/{model_name}_{epoch_num:03}.pt'
print('load', checkpoint)
try:
model = torch.load(checkpoint, map_location=device)
except FileNotFoundError:
break
model = model.to(device)
model.eval()
dataset_valid = DetectionDataset(fold=fold,
img_size=model_info.img_size,
is_training=False,
images={})
dataloader_valid = DataLoader(
dataset_valid,
num_workers=2,
batch_size=1,
shuffle=False,
collate_fn=pytorch_retinanet.dataloader.collater2d)
oof = collections.defaultdict(list)
# for iter_num, data in tqdm(enumerate(dataloader_valid), total=len(dataloader_valid)):
for iter_num, data in tqdm(enumerate(dataset_valid),
total=len(dataloader_valid)):
data = pytorch_retinanet.dataloader.collater2d([data])
img = data['img'].to(device).float()
nms_scores, global_classification, transformed_anchors = \
model(img, return_loss=False, return_boxes=True)
nms_scores = nms_scores.cpu().detach().numpy()
global_classification = global_classification.cpu().detach().numpy(
)
transformed_anchors = transformed_anchors.cpu().detach().numpy()
oof['gt_boxes'].append(data['annot'].cpu().detach().numpy())
oof['gt_category'].append(data['category'].cpu().detach().numpy())
oof['boxes'].append(transformed_anchors)
oof['scores'].append(nms_scores)
oof['category'].append(global_classification)
pickle.dump(oof, open(prediction_fn, 'wb'))
def load_model(checkpoint: str) -> nn.Module:
"""
Helper to load model weihts
"""
print(f"Loading model from: {checkpoint}")
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# load model
model = torch.load(checkpoint)
model = model.to(device)
model.eval()
# model = torch.nn.DataParallel(model).cuda()
return model
def check(model_name, fold, checkpoint):
model_info = MODELS[model_name]
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = torch.load(checkpoint, map_location=device)
model = model.to(device)
model.eval()
dataset_valid = DetectionDataset(fold=fold,
img_size=model_info.img_size,
is_training=False,
images={})
dataloader_valid = DataLoader(
dataset_valid,
num_workers=1,
batch_size=1,
shuffle=False,
collate_fn=pytorch_retinanet.dataloader.collater2d)
data_iter = tqdm(enumerate(dataloader_valid), total=len(dataloader_valid))
for iter_num, data in data_iter:
classification_loss, regression_loss, global_classification_loss, nms_scores, nms_class, transformed_anchors = \
model([data['img'].to(device).float(), data['annot'].to(device).float(), data['category'].cuda()],
return_loss=True, return_boxes=True)
nms_scores = nms_scores.cpu().detach().numpy()
nms_class = nms_class.cpu().detach().numpy()
transformed_anchors = transformed_anchors.cpu().detach().numpy()
print(nms_scores, transformed_anchors.shape)
print('cls loss:', float(classification_loss), 'global cls loss:',
global_classification_loss, ' reg loss:', float(regression_loss))
print('cat:', data['category'].numpy()[0], np.exp(nms_class[0]),
dataset_valid.categories[data['category'][0]])
plt.cla()
plt.imshow(data['img'][0, 0].cpu().detach().numpy())
gt = data['annot'].cpu().detach().numpy()[0]
for i in range(gt.shape[0]):
if np.all(np.isfinite(gt[i])):
p0 = gt[i, 0:2]
p1 = gt[i, 2:4]
plt.gca().add_patch(
plt.Rectangle(p0,
width=(p1 - p0)[0],
height=(p1 - p0)[1],
fill=False,
edgecolor='b',
linewidth=2))
for i in range(len(nms_scores)):
nms_score = nms_scores[i]
if nms_score < 0.1:
break
# print(transformed_anchors[i, :])
p0 = transformed_anchors[i, 0:2]
p1 = transformed_anchors[i, 2:4]
color = 'g'
if nms_score < 0.4:
color = 'y'
if nms_score < 0.25:
color = 'r'
# print(p0, p1)
plt.gca().add_patch(
plt.Rectangle(p0,
width=(p1 - p0)[0],
height=(p1 - p0)[1],
fill=False,
edgecolor=color,
linewidth=2))
plt.gca().text(
p0[0], p0[1], f'{nms_score:.3f}',
color=color) # , bbox={'facecolor': color, 'alpha': 0.5})
plt.show()
print(nms_scores)
def generate_predictions(
model_name: str,
fold: int,
debug: bool,
weights_dir: str,
from_epoch: int = 0,
to_epoch: int = 10,
save_oof: bool = True,
run: str = None,
):
"""
Loads model weights the epoch checkpoints,
calculates oof predictions for and saves them to pickle
Args:
model_name : string name from the models configs listed in models.py file
fold : evaluation fold number, 0-3
debug : if True, runs debugging on few images
weights_dir: directory qith model weigths
from_epoch : the first epoch for predicitions generation
to_epoch : the last epoch for predicitions generation
save_oof : boolean flag weathe rto save precitions
run : string name to be added in the experinet name
"""
predictions_dir = f"{RESULTS_DIR}/test1/{model_name}_fold_{fold}"
os.makedirs(predictions_dir, exist_ok=True)
model_info = MODELS[model_name]
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print("device: ", device)
for epoch_num in range(from_epoch, to_epoch):
prediction_fn = f"{predictions_dir}/{epoch_num:03}.pkl"
if os.path.exists(prediction_fn):
continue
print("epoch", epoch_num)
# load model checkpoint
checkpoint = (
f"{weights_dir}/{model_name}_fold_{fold}/{model_name}_{epoch_num:03}.pt"
)
print("load", checkpoint)
try:
model = torch.load(checkpoint)
except FileNotFoundError:
break
model = model.to(device)
model.eval()
# load data
dataset_valid = DatasetValid(
is_training=False,
meta_file="stage_1_test_meta.csv",
debug=debug,
img_size=512,
)
dataloader_valid = DataLoader(
dataset_valid,
num_workers=2,
batch_size=4,
shuffle=False,
collate_fn=pytorch_retinanet.dataloader.collater2d,
)
oof = collections.defaultdict(list)
for iter_num, data in tqdm(enumerate(dataset_valid),
total=len(dataloader_valid)):
data = pytorch_retinanet.dataloader.collater2d([data])
img = data["img"].to(device).float()
nms_scores, global_classification, transformed_anchors = model(
img, return_loss=False, return_boxes=True)
# model outputs to numpy
nms_scores = nms_scores.cpu().detach().numpy()
global_classification = global_classification.cpu().detach().numpy(
)
transformed_anchors = transformed_anchors.cpu().detach().numpy()
# out-of-fold predictions
oof["gt_boxes"].append(data["annot"].cpu().detach().numpy())
oof["gt_category"].append(data["category"].cpu().detach().numpy())
oof["boxes"].append(transformed_anchors)
oof["scores"].append(nms_scores)
oof["category"].append(global_classification)
# save epoch predictions
if save_oof:
pickle.dump(oof, open(f"{predictions_dir}/{epoch_num:03}.pkl",
"wb"))
def test_model(model_name: str, fold: int, debug: bool, checkpoint: str,
pics_dir: str):
"""
Loads model weights from the checkpoint, plots ground truth and predictions
Args:
model_name : string name from the models configs listed in models.py file
fold : evaluation fold number, 0-3
debug : if True, runs debugging on few images
checkpoint : directory with weights (if avaialable)
pics_dir : directory for saving prediction images
"""
model_info = MODELS[model_name]
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# load model
model = torch.load(checkpoint)
model = model.to(device)
model.eval()
# load data
dataset_valid = DetectionDataset(fold=fold,
img_size=model_info.img_size,
is_training=False,
debug=debug)
dataloader_valid = DataLoader(
dataset_valid,
num_workers=1,
batch_size=1,
shuffle=False,
collate_fn=pytorch_retinanet.dataloader.collater2d,
)
data_iter = tqdm(enumerate(dataloader_valid), total=len(dataloader_valid))
for iter_num, data in data_iter:
(
classification_loss,
regression_loss,
global_classification_loss,
nms_scores,
nms_class,
transformed_anchors,
) = model(
[
data["img"].to(device).float(),
data["annot"].to(device).float(),
data["category"].cuda(),
],
return_loss=True,
return_boxes=True,
)
nms_scores = nms_scores.cpu().detach().numpy()
nms_class = nms_class.cpu().detach().numpy()
transformed_anchors = transformed_anchors.cpu().detach().numpy()
print("nms_scores {}, transformed_anchors.shape {}".format(
nms_scores, transformed_anchors.shape))
print(
"cls loss:",
float(classification_loss),
"global cls loss:",
global_classification_loss,
" reg loss:",
float(regression_loss),
)
print(
"category:",
data["category"].numpy()[0],
np.exp(nms_class[0]),
dataset_valid.categories[data["category"][0]],
)
# plot data and ground truth
plt.figure(iter_num, figsize=(6, 6))
plt.cla()
plt.imshow(data["img"][0, 0].cpu().detach().numpy(),
cmap=plt.cm.gist_gray)
plt.axis("off")
gt = data["annot"].cpu().detach().numpy()[0]
for i in range(gt.shape[0]):
if np.all(np.isfinite(gt[i])):
p0 = gt[i, 0:2]
p1 = gt[i, 2:4]
plt.gca().add_patch(
plt.Rectangle(
p0,
width=(p1 - p0)[0],
height=(p1 - p0)[1],
fill=False,
edgecolor="b",
linewidth=2,
))
# add predicted boxes to the plot
for i in range(len(nms_scores)):
nms_score = nms_scores[i]
if nms_score < 0.1:
break
p0 = transformed_anchors[i, 0:2]
p1 = transformed_anchors[i, 2:4]
color = "r"
if nms_score < 0.3:
color = "y"
if nms_score < 0.25:
color = "g"
plt.gca().add_patch(
plt.Rectangle(
p0,
width=(p1 - p0)[0],
height=(p1 - p0)[1],
fill=False,
edgecolor=color,
linewidth=2,
))
plt.gca().text(p0[0], p0[1], f"{nms_score:.3f}", color=color)
plt.show()
os.makedirs(pics_dir, exist_ok=True)
plt.savefig(
f"{pics_dir}/predict_{iter_num}.eps",
dpi=300,
bbox_inches="tight",
pad_inches=0,
)
plt.savefig(
f"{pics_dir}/predict_{iter_num}.png",
dpi=300,
bbox_inches="tight",
pad_inches=0,
)
plt.close()
print(nms_scores)