def predict(model: nn.Module,
image: np.ndarray,
image_size,
normalize=A.Normalize(),
batch_size=1) -> np.ndarray:
tile_step = (image_size[0] // 2, image_size[1] // 2)
tile_slicer = ImageSlicer(image.shape, image_size, tile_step)
tile_merger = CudaTileMerger(tile_slicer.target_shape, 1,
tile_slicer.weight)
patches = tile_slicer.split(image)
transform = A.Compose([normalize, A.Lambda(image=_tensor_from_rgb_image)])
data = list({
"image": patch,
"coords": np.array(coords, dtype=np.int)
} for (patch, coords) in zip(patches, tile_slicer.crops))
for batch in DataLoader(InMemoryDataset(data, transform),
pin_memory=True,
batch_size=batch_size):
image = batch["image"].cuda(non_blocking=True)
coords = batch["coords"]
mask_batch = model(image)
tile_merger.integrate_batch(mask_batch, coords)
mask = tile_merger.merge()
mask = np.moveaxis(to_numpy(mask), 0, -1)
mask = tile_slicer.crop_to_orignal_size(mask)
return mask
def test_tiles_split_merge_cuda():
if not torch.cuda.is_available():
return
class MaxChannelIntensity(nn.Module):
def __init__(self):
super().__init__()
def forward(self, input):
max_channel, _ = torch.max(input, dim=1, keepdim=True)
return max_channel
image = np.random.random((5000, 5000, 3)).astype(np.uint8)
tiler = ImageSlicer(image.shape,
tile_size=(512, 512),
tile_step=(256, 256),
weight='pyramid')
tiles = [tensor_from_rgb_image(tile) for tile in tiler.split(image)]
model = MaxChannelIntensity().eval().cuda()
merger = CudaTileMerger(tiler.target_shape, 1, tiler.weight)
for tiles_batch, coords_batch in DataLoader(list(zip(tiles, tiler.crops)),
batch_size=8,
pin_memory=True):
tiles_batch = tiles_batch.float().cuda()
pred_batch = model(tiles_batch)
merger.integrate_batch(pred_batch, coords_batch)
merged = np.moveaxis(to_numpy(merger.merge()), 0, -1).astype(np.uint8)
merged = tiler.crop_to_orignal_size(merged)
np.testing.assert_equal(merged, image.max(axis=2, keepdims=True))
def infer_one(model,
mask,
tile_size=(512, 512),
tile_step=(256, 256),
weight='mean'):
image = mask.cpu().numpy()
image = np.moveaxis(image, 0, -1)
with torch.no_grad():
tiler = ImageSlicer((900, 900),
tile_size=tile_size,
tile_step=tile_step,
weight=weight)
tiles = [np.moveaxis(tile, -1, 0) for tile in tiler.split(image)]
merger = CudaTileMerger(tiler.target_shape, 1, tiler.weight)
for tiles_batch, coords_batch in DataLoader(list(
zip(tiles, tiler.crops)),
batch_size=10,
pin_memory=False):
tiles_batch = tiles_batch.float().cuda()
pred_batch = model(tiles_batch)
tiles_batch.cpu().detach()
merger.integrate_batch(pred_batch, coords_batch)
merged_mask = np.moveaxis(to_numpy(merger.merge()), 0, -1)
merged_mask = tiler.crop_to_orignal_size(merged_mask)
m = merged_mask[..., 0].copy()
return m
def inference_tiles(inference_model,
img_full,
device='cuda',
shape=(32, 1, 768, 448),
weight='mean',
mean=88.904434,
std=62.048634,
plot=False):
bs = shape[0]
input_x = shape[2]
input_y = shape[3]
# Cut large image into overlapping tiles
tiler = ImageSlicer(img_full.shape,
tile_size=(input_x, input_y),
tile_step=(input_x // 2, input_y // 2),
weight=weight)
# HCW -> CHW. Optionally, do normalization here
tiles = [
tensor_from_rgb_image(tile)
for tile in tiler.split(cv2.cvtColor(img_full, cv2.COLOR_GRAY2RGB))
]
# Allocate a CUDA buffer for holding entire mask
merger = CudaTileMerger(tiler.target_shape,
channels=1,
weight=tiler.weight)
# Run predictions for tiles and accumulate them
for tiles_batch, coords_batch in DataLoader(list(zip(tiles, tiler.crops)),
batch_size=bs,
pin_memory=True):
# Move tile to GPU
tiles_batch = ((tiles_batch.float() - mean) / std).to(device)
# Predict
pred_batch = inference_model(tiles_batch)
# Merge on GPU
merger.integrate_batch(pred_batch, coords_batch)
if plot:
for i in range(pred_batch.to('cpu').numpy().shape[0]):
plt.imshow(tiles_batch.to('cpu').numpy()[i, 0, :, :])
plt.show()
plt.imshow(pred_batch.to('cpu').numpy()[i, 0, :, :])
plt.colorbar()
plt.show()
# Normalize accumulated mask and convert back to numpy
merged_mask = np.moveaxis(to_numpy(merger.merge()), 0,
-1).astype('float32')
merged_mask = tiler.crop_to_orignal_size(merged_mask)
torch.cuda.empty_cache()
return merged_mask.squeeze()
def predict_mask(image,
model,
dims=3,
size=394,
step=192,
batch_size=8,
plot_image=False,
dstdir=None,
img_name='image1.png'):
if image.ndim == 2:
image = np.expand_dims(image, 2)
if image.shape[-1] != dims:
if image.shape[-1] == 1:
image = np.repeat(image, 3, axis=2)
elif image.shape[-1] == 3:
image = np.expand_dims(image[:, :, 0], 2)
print(image.shape)
# Cut large image into overlapping tiles
tiler = ImageSlicer(image.shape,
tile_size=(size, size),
tile_step=(step, step),
weight='pyramid')
# HCW -> CHW. Optionally, do normalization here
tiles = [tensor_from_rgb_image(tile) for tile in tiler.split(image)]
# Allocate a CUDA buffer for holding entire mask
merger = CudaTileMerger(tiler.target_shape, 1, tiler.weight)
# Run predictions for tiles and accumulate them
with torch.no_grad():
for tiles_batch, coords_batch in DataLoader(list(
zip(tiles, tiler.crops)),
batch_size=batch_size,
pin_memory=True):
# print(tiles_batch.shape)
tiles_batch = tiles_batch.float().cuda()
pred_batch = model(tiles_batch)
pred_mask = pred_batch.max(dim=1)[1].float()
merger.integrate_batch(pred_mask, coords_batch)
# Normalize accumulated mask and convert back to numpy
merged_mask = np.moveaxis(to_numpy(merger.merge()), 0, -1).astype(np.uint8)
merged_mask = tiler.crop_to_orignal_size(merged_mask)
if plot_image:
assert dstdir is not None, 'dstdir should be passed'
fig, ax = plt.subplots(ncols=2, figsize=(20, 10))
ax[0].imshow(image[:, :, 0], cmap='gray')
ax[1].imshow(merged_mask[:, :, 0], alpha=0.3)
fig.savefig(osp.join(dstdir, img_name),
bbox_inches='tight',
pad_inches=0)
print(osp.join(dstdir, img_name))
return merged_mask
def predict(model: nn.Module,
image: np.ndarray,
image_size,
tta=None,
normalize=A.Normalize(),
batch_size=1,
activation='sigmoid') -> np.ndarray:
model.eval()
tile_step = (image_size[0] // 2, image_size[1] // 2)
tile_slicer = ImageSlicer(image.shape,
image_size,
tile_step,
weight='pyramid')
tile_merger = CudaTileMerger(tile_slicer.target_shape, 1,
tile_slicer.weight)
patches = tile_slicer.split(image)
transform = A.Compose([normalize, A.Lambda(image=_tensor_from_rgb_image)])
if tta == 'fliplr':
model = TTAWrapperFlipLR(model)
print('Using FlipLR TTA')
if tta == 'd4':
model = TTAWrapperD4(model)
print('Using D4 TTA')
with torch.no_grad():
data = list({
'image': patch,
'coords': np.array(coords, dtype=np.int)
} for (patch, coords) in zip(patches, tile_slicer.crops))
for batch in DataLoader(InMemoryDataset(data, transform),
pin_memory=True,
batch_size=batch_size):
image = batch['image'].cuda(non_blocking=True)
coords = batch['coords']
mask_batch = model(image)
tile_merger.integrate_batch(mask_batch, coords)
mask = tile_merger.merge()
if activation == 'sigmoid':
mask = mask.sigmoid()
if isinstance(activation, float):
mask = F.relu(mask_batch - activation, inplace=True)
mask = np.moveaxis(to_numpy(mask), 0, -1)
mask = tile_slicer.crop_to_orignal_size(mask)
return mask
def inference(inference_model, img_full, device='cuda'):
x, y, ch = img_full.shape
input_x = config['training']['crop_size'][0]
input_y = config['training']['crop_size'][1]
# Cut large image into overlapping tiles
tiler = ImageSlicer(img_full.shape, tile_size=(input_x, input_y),
tile_step=(input_x // 2, input_y // 2), weight=args.weight)
# HCW -> CHW. Optionally, do normalization here
tiles = [tensor_from_rgb_image(tile) for tile in tiler.split(img_full)]
# Allocate a CUDA buffer for holding entire mask
merger = CudaTileMerger(tiler.target_shape, channels=1, weight=tiler.weight)
# Run predictions for tiles and accumulate them
for tiles_batch, coords_batch in DataLoader(list(zip(tiles, tiler.crops)), batch_size=args.bs, pin_memory=True):
# Move tile to GPU
tiles_batch = (tiles_batch.float() / 255.).to(device)
# Predict and move back to CPU
pred_batch = inference_model(tiles_batch)
# Merge on GPU
merger.integrate_batch(pred_batch, coords_batch)
# Plot
if args.plot:
for i in range(args.bs):
if args.bs != 1:
plt.imshow(pred_batch.cpu().detach().numpy().astype('float32').squeeze()[i, :, :])
else:
plt.imshow(pred_batch.cpu().detach().numpy().astype('float32').squeeze())
plt.show()
# Normalize accumulated mask and convert back to numpy
merged_mask = np.moveaxis(to_numpy(merger.merge()), 0, -1).astype('float32')
merged_mask = tiler.crop_to_orignal_size(merged_mask)
# Plot
if args.plot:
for i in range(args.bs):
if args.bs != 1:
plt.imshow(merged_mask)
else:
plt.imshow(merged_mask.squeeze())
plt.show()
torch.cuda.empty_cache()
gc.collect()
return merged_mask.squeeze()
def run_validation(data_df, model, data_folder, augmentation, tiles=False):
total_dice_coeffs = []
mean_dice_per_image = []
for image_n in tqdm(range(data_df.shape[0])):
image = cv2.imread(
os.path.join(data_folder, data_df.index.values[image_n]))
augmented = augmentation(image=image)
image_processed = augmented['image']
if tiles:
tiler = ImageSlicer(image_processed.shape[:2],
tile_size=(224, 224),
tile_step=(56, 56),
weight='mean')
merger = CudaTileMerger(tiler.target_shape, 4, tiler.weight)
tiles = [
tensor_from_rgb_image(tile)
for tile in tiler.split(image_processed)
]
for tiles_batch, coords_batch in DataLoader(list(
zip(tiles, tiler.crops)),
batch_size=16,
pin_memory=True):
tiles_batch = tiles_batch.float().cuda()
pred_batch = torch.nn.Sigmoid()(model(tiles_batch))
merger.integrate_batch(pred_batch, coords_batch)
predictions = np.moveaxis(to_numpy(merger.merge()), 0, -1)
predictions = tiler.crop_to_orignal_size(predictions)
else:
image_processed = torch.from_numpy(
np.expand_dims(image_processed.transpose((2, 0, 1)),
0)).float()
predictions = torch.nn.Sigmoid()(model(
image_processed.cuda())[0]).detach().cpu().numpy()
predictions = np.moveaxis(predictions, 0, -1)
predictions_bin = (predictions > 0.5).astype(int)
fname, masks = make_mask(image_n, data_df)
dices_image = []
for defect_type in range(4):
computed_dice = dice(masks[:, :, defect_type],
predictions_bin[:, :, defect_type])
total_dice_coeffs.append(computed_dice)
dices_image.append(computed_dice)
mean_dice_per_image.append(np.mean(dices_image))
return np.mean(total_dice_coeffs), mean_dice_per_image
def predict_on_zslice_tiles(model,
zimage,
tile_size=(512, 512),
tile_step=(256, 256)):
image = zimage[0, 0, :, :]
print(f'Stack shape:{zimage.shape}')
print(f'Slice shape:{image.shape}')
# Cut large image into overlapping tiles
tiler = ImageSlicer(image.shape,
tile_size=(512, 512),
tile_step=(256, 256))
print(tiler.crops)
# HCW -> CHW. Optionally, do normalization here
tiles = [tensor_from_mask_image(tile) for tile in tiler.split(image)]
# Allocate a CUDA buffer for holding entire mask
merger = CudaTileMerger(tiler.target_shape, 1, tiler.weight)
# Run predictions for tiles and accumulate them
for tiles_batch, coords_batch in DataLoader(list(zip(tiles, tiler.crops)),
batch_size=1,
pin_memory=True):
# for x, y, tile_width, tile_height in coords_batch:
# tile = image[y : y + tile_height, x : x + tile_width].copy()
tiles_batch = tiles_batch.float().cuda()
pred_batch = model(tiles_batch)
merger.integrate_batch(pred_batch, coords_batch)
# Normalize accumulated mask and convert back to numpy
# merged_mask = np.moveaxis(to_numpy(merger.merge()), 0, -1).astype(np.uint8)
merged_mask = np.moveaxis(to_numpy(merger.merge()), 0, -1)
merged_mask = tiler.crop_to_orignal_size(merged_mask)
return merged_mask
def test_tiles_split_merge_non_dividable_cuda():
image = np.random.random((5632, 5120, 3)).astype(np.uint8)
tiler = ImageSlicer(image.shape,
tile_size=(1280, 1280),
tile_step=(1280, 1280),
weight='mean')
tiles = tiler.split(image)
merger = CudaTileMerger(tiler.target_shape,
channels=image.shape[2],
weight=tiler.weight)
for tile, coordinates in zip(tiles, tiler.crops):
# Integrate as batch of size 1
merger.integrate_batch(
tensor_from_rgb_image(tile).unsqueeze(0).float().cuda(),
[coordinates])
merged = merger.merge()
merged = rgb_image_from_tensor(merged, mean=0, std=1, max_pixel_value=1)
merged = tiler.crop_to_orignal_size(merged)
np.testing.assert_equal(merged, image)
def predict_gradcam_mask(image,
model,
dims=3,
size=(150, 300),
step=(150, 300),
batch_size=8,
grad_thr=0.6,
weight_type='mean',
plot_image=False,
dstdir=None,
img_name='image1.png'):
image = scale_img(image).astype(np.float32)
if image.ndim == 2:
image = np.expand_dims(image, 2)
if image.shape[-1] != dims:
if image.shape[-1] == 1:
image = np.repeat(image, 3, axis=2)
elif image.shape[-1] == 3:
image = np.expand_dims(image[:, :, 0], 2)
image = (image - np.min(image)) / (0.5 * np.ptp(image)) - 1
# Cut large image into overlapping tiles
tiler = ImageSlicer(image.shape,
tile_size=size,
tile_step=step,
weight=weight_type)
# HCW -> CHW. Optionally, do normalization here
tiles = [tensor_from_rgb_image(tile) for tile in tiler.split(image)]
# Allocate a CUDA buffer for holding entire mask
merger = CudaTileMerger(tiler.target_shape, 1, tiler.weight)
# Run predictions for tiles and accumulate them
for tiles_batch, coords_batch in DataLoader(list(zip(tiles, tiler.crops)),
batch_size=batch_size,
pin_memory=True):
tiles_batch = tiles_batch.float().cuda()
with torch.no_grad():
pred_batch = torch.max(F.softmax(model(tiles_batch), dim=1),
dim=1)[1].detach().cpu().numpy()
image_needed_classes = pred_batch == 1
masks = []
for tile_idx, has_target in enumerate(image_needed_classes):
if has_target:
tile = tiles_batch[tile_idx].unsqueeze(0)
heatmap, mask = show_gradcam(tile, model)
mask = mask[:, :, 0]
mask[mask < grad_thr] = 0
masks.append(torch.Tensor(mask).unsqueeze(0).unsqueeze(0))
else:
masks.append(
torch.zeros_like(tiles_batch[tile_idx,
0]).unsqueeze(0).unsqueeze(0))
masks = torch.cat([mask.cuda() for mask in masks], dim=0) * 1000
merger.integrate_batch(masks, coords_batch)
# Normalize accumulated mask and convert back to numpy
merged_mask = np.moveaxis(to_numpy(merger.merge()), 0, -1).astype(np.uint8)
merged_mask = tiler.crop_to_orignal_size(merged_mask) / 1000
if plot_image:
assert dstdir is not None, 'dstdir should be passed'
fig, ax = plt.subplots(ncols=2, figsize=(20, 10))
ax[0].imshow(image[:, :, 0], cmap='gray')
ax[1].imshow(merged_mask[:, :, 0], alpha=0.3)
fig.savefig(osp.join(dstdir, img_name),
bbox_inches='tight',
pad_inches=0)
print(osp.join(dstdir, img_name))
return merged_mask
x, y, ch = img_full.shape
mask_full = np.zeros((x, y))
# Cut large image into overlapping tiles
tiler = ImageSlicer(img_full.shape,
tile_size=(input_x, input_y),
tile_step=(input_x // 2, input_y // 2),
weight=args.weight)
# HCW -> CHW. Optionally, do normalization here
tiles = [tensor_from_rgb_image(tile) for tile in tiler.split(img_full)]
# Allocate a CUDA buffer for holding entire mask
merger = CudaTileMerger(tiler.target_shape,
channels=1,
weight=tiler.weight)
# Loop evaluating inference on every fold
masks = []
for fold in range(len(models)):
# Run predictions for tiles and accumulate them
for tiles_batch, coords_batch in DataLoader(list(
zip(tiles, tiler.crops)),
batch_size=args.bs,
pin_memory=True):
# Move tile to GPU
tiles_batch = (tiles_batch.float() / 255.).to(device)
# Predict and move back to CPU
pred_batch = torch.sigmoid(