Optimize ram usage

This commit is contained in:
2026-07-06 09:23:36 +02:00
parent b728668440
commit 0c771f89c3
2 changed files with 118 additions and 58 deletions
+27 -23
View File
@@ -81,32 +81,36 @@ def run_inference(model, data_loader, conf):
def accumulate_hits(predictions: np.ndarray, reference_points: np.ndarray,
binning_factor: int, number_of_subframes: int = 20):
ml_super_frames = np.zeros((number_of_subframes, NY * binning_factor, NX * binning_factor))
count_frame = np.zeros((NY, NX))
subpixel_dist = np.zeros((binning_factor, binning_factor))
ml_super_frames = np.zeros((number_of_subframes, NY * binning_factor, NX * binning_factor), dtype=np.uint32)
count_frame = np.zeros((NY, NX), dtype=np.uint32)
subpixel_dist = np.zeros((binning_factor, binning_factor), dtype=np.uint32)
# absolute coordinate = predicted subpixel + reference point
absolute_positions = predictions + reference_points[:, :2]
# super resolution frames (binning)
hit_x = np.floor(absolute_positions[:, 0] * binning_factor).astype(int)
hit_y = np.floor(absolute_positions[:, 1] * binning_factor).astype(int)
chunk_size = len(predictions) // number_of_subframes
for i in range(number_of_subframes):
start_idx = i * len(predictions) // number_of_subframes
end_idx = min((i + 1) * len(predictions) // number_of_subframes, len(predictions))
np.add.at(ml_super_frames[i], (hit_y[start_idx:end_idx], hit_x[start_idx:end_idx]), 1)
# count frame (by reference point pixel index)
ref_x = (reference_points[:, 0] + 1).astype(int) # reference point is lower-left corner, +1 to get pixel index
ref_y = (reference_points[:, 1] + 1).astype(int)
np.add.at(count_frame, (ref_y, ref_x), 1)
# subpixel distribution
sub_x = np.floor((absolute_positions[:, 0] % 1) * binning_factor).astype(int)
sub_y = np.floor((absolute_positions[:, 1] % 1) * binning_factor).astype(int)
np.add.at(subpixel_dist, (sub_y, sub_x), 1)
start_idx = i * chunk_size
end_idx = (i + 1) * chunk_size if i < number_of_subframes - 1 else len(predictions)
pred_chunk = predictions[start_idx:end_idx]
ref_chunk = reference_points[start_idx:end_idx]
abs_pos_chunk = pred_chunk + ref_chunk[:, :2]
# --- Super resolution frames ---
hit_x = np.floor(abs_pos_chunk[:, 0] * binning_factor).astype(np.int32)
hit_y = np.floor(abs_pos_chunk[:, 1] * binning_factor).astype(np.int32)
np.add.at(ml_super_frames[i], (hit_y, hit_x), 1)
# --- Count frame ---
ref_x = (ref_chunk[:, 0] + 1).astype(np.int32)
ref_y = (ref_chunk[:, 1] + 1).astype(np.int32)
np.add.at(count_frame, (ref_y, ref_x), 1)
# --- Subpixel distribution ---
sub_x = np.floor((abs_pos_chunk[:, 0] % 1) * binning_factor).astype(np.int32)
sub_y = np.floor((abs_pos_chunk[:, 1] % 1) * binning_factor).astype(np.int32)
np.add.at(subpixel_dist, (sub_y, sub_x), 1)
return ml_super_frames, count_frame, subpixel_dist
def save_results(ml_super_frames, ml_super_frames_eta,
+91 -35
View File
@@ -3,6 +3,8 @@ sys.path.append('./src')
import matplotlib
matplotlib.use('Agg')
from etaInterpolationFromPoints import interpolate_eta_from_points
from pathlib import Path
from omegaconf import OmegaConf
import torch
@@ -47,11 +49,7 @@ def apply_inverse_transforms(predictions: torch.Tensor, numberOfAugOps: int) ->
return corrected.mean(dim=1)
def prepare_output_folder(conf):
if conf.data.normalize:
normalize_suffix = '_normalized'
else:
normalize_suffix = ''
output_base = Path(conf.experiment.output_base) / conf.experiment.name / conf.model.experiment_name / f'augX{conf.inference.num_aug_ops}{normalize_suffix}'
output_base = Path(conf.experiment.output_base) / conf.experiment.name / (f'3ph_{conf.model.experiment_name}') / (f'augX{conf.inference.num_aug_ops}')
output_base.mkdir(parents=True, exist_ok=True)
OmegaConf.save(conf, output_base / 'config.yaml')
return output_base
@@ -75,45 +73,52 @@ def run_inference(model, data_loader, conf):
all_predictions = torch.cat(all_predictions, dim=0)
# all_predictions = apply_inverse_transforms(all_predictions, conf.inference.num_aug_ops)
all_predictions += torch.tensor([NSIZE/2., NSIZE/2.]).unsqueeze(0) # adjust back to original coordinate system
print(f'mean x = {torch.mean(all_predictions[:, 0])}, std x = {torch.std(all_predictions[:, 0])}')
print(f'mean y = {torch.mean(all_predictions[:, 1])}, std y = {torch.std(all_predictions[:, 1])}')
print(f'[Inference]: mean x = {torch.mean(all_predictions[:, 0])}, std x = {torch.std(all_predictions[:, 0])}')
print(f'[Inference]: mean y = {torch.mean(all_predictions[:, 1])}, std y = {torch.std(all_predictions[:, 1])}')
referencePoints = data_loader.dataset.referencePoint ### the lower-left corner of the cluster in absolute coordinate
referencePoints = np.repeat(referencePoints, 3, axis=0) ### duplicate reference points for 3-photon clusters
return all_predictions.numpy(), referencePoints
def accumulate_hits(predictions: np.ndarray, reference_points: np.ndarray,
binning_factor: int):
### ret
ml_super_frame = np.zeros((NY*binning_factor, NX*binning_factor), dtype=np.int32)
binning_factor: int, number_of_subframes: int = 20):
ml_super_frames = np.zeros((number_of_subframes, NY * binning_factor, NX * binning_factor))
count_frame = np.zeros((NY, NX), dtype=np.int32)
subpixel_dist = np.zeros((binning_factor, binning_factor), dtype=np.int32)
### absolute coordinate = predicted subpixel + reference point
absolute_positions = predictions + reference_points
hit_x_superpixel_idx = np.floor(absolute_positions[:, 0] * binning_factor).astype(int)
hit_x_superpixel_idx = np.clip(hit_x_superpixel_idx, 0, NX*binning_factor-1)
hit_y_superpixel_idx = np.floor(absolute_positions[:, 1] * binning_factor).astype(int)
hit_y_superpixel_idx = np.clip(hit_y_superpixel_idx, 0, NY*binning_factor-1)
np.add.at(ml_super_frame, (hit_y_superpixel_idx, hit_x_superpixel_idx), 1)
hit_x_pixel_idx = np.floor(absolute_positions[:, 0]).astype(int)
hit_x_pixel_idx = np.clip(hit_x_pixel_idx, 0, NX-1)
hit_y_pixel_idx = np.floor(absolute_positions[:, 1]).astype(int)
hit_y_pixel_idx = np.clip(hit_y_pixel_idx, 0, NY-1)
np.add.at(count_frame, (hit_y_pixel_idx, hit_x_pixel_idx), 1)
# super resolution frames (binning)
hit_x = np.floor(absolute_positions[:, 0] * binning_factor).astype(int)
hit_y = np.floor(absolute_positions[:, 1] * binning_factor).astype(int)
for i in range(number_of_subframes):
start_idx = i * len(predictions) // number_of_subframes
end_idx = min((i + 1) * len(predictions) // number_of_subframes, len(predictions))
np.add.at(ml_super_frames[i], (hit_y[start_idx:end_idx], hit_x[start_idx:end_idx]), 1)
# count frame (by reference point pixel index)
ref_x = (reference_points[:, 0] + 1).astype(int) # reference point is lower-left corner, +1 to get pixel index
ref_y = (reference_points[:, 1] + 1).astype(int)
np.add.at(count_frame, (ref_y, ref_x), 1)
# subpixel distribution
sub_x = np.floor((absolute_positions[:, 0] % 1) * binning_factor).astype(int)
sub_y = np.floor((absolute_positions[:, 1] % 1) * binning_factor).astype(int)
np.add.at(subpixel_dist, (sub_y, sub_x), 1)
return ml_super_frames, count_frame, subpixel_dist
subpixel_x_idx = np.floor((absolute_positions[:, 0] % 1) * binning_factor).astype(int)
subpixel_y_idx = np.floor((absolute_positions[:, 1] % 1) * binning_factor).astype(int)
np.add.at(subpixel_dist, (subpixel_y_idx, subpixel_x_idx), 1)
return ml_super_frame, count_frame, subpixel_dist
def save_results(ml_super_frame, count_frame, subpixel_dist,
def save_results(ml_super_frames, ml_super_frames_eta,
count_frame, subpixel_dist, subpixel_dist_eta,
roi: list, binning_factor: int, output_dir: Path):
x_st, x_ed, y_st, y_ed = roi
# 1. super-resolution frame
np.save(output_dir / '3Photon_ML_superFrames.npy', ml_super_frames)
ml_super_frame = np.sum(ml_super_frames, axis=0)
plt.figure(figsize=(8, 8))
plt.imshow(ml_super_frame[y_st*binning_factor:y_ed*binning_factor, x_st*binning_factor:x_ed*binning_factor], origin='lower', extent=[x_st, x_ed, y_st, y_ed])
plt.colorbar(label='Counts')
@@ -124,7 +129,21 @@ def save_results(ml_super_frame, count_frame, subpixel_dist,
plt.clf()
np.save(output_dir / '3Photon_ML_superFrame.npy', ml_super_frame)
# 2. count frame
# 2. super-resolution frame with eta interpolation
np.save(output_dir / '3Photon_ML_superFrames_etaInterpolated.npy', ml_super_frames_eta)
ml_super_frame_eta = np.sum(ml_super_frames_eta, axis=0)
plt.figure(figsize=(8, 8))
plt.imshow(ml_super_frame_eta[y_st*binning_factor:y_ed*binning_factor, x_st*binning_factor:x_ed*binning_factor], origin='lower', extent=[x_st, x_ed, y_st, y_ed])
plt.colorbar(label='Counts')
plt.title('ML Super-Resolution Frame with Eta Interpolation')
plt.xlabel('X (pixel)')
plt.ylabel('Y (pixel)')
plt.savefig(output_dir / '3Photon_ML_superFrame_etaInterpolated.png', dpi=300, bbox_inches='tight')
plt.clf()
np.save(output_dir / '3Photon_ML_superFrame_etaInterpolated.npy', ml_super_frame_eta)
# 3. count frame
plt.imshow(count_frame[y_st:y_ed, x_st:x_ed], origin='lower', extent=[x_st, x_ed, y_st, y_ed])
plt.colorbar(label='Counts')
plt.title('Photon Count Frame')
@@ -134,7 +153,7 @@ def save_results(ml_super_frame, count_frame, subpixel_dist,
plt.clf()
np.save(output_dir / '3Photon_count_Frame.npy', count_frame)
# 3. subpixel distribution
# 4. subpixel distribution
plt.imshow(subpixel_dist, origin='lower', extent=[0, 1, 0, 1])
plt.colorbar(label='Counts')
plt.title('Subpixel Distribution')
@@ -146,7 +165,19 @@ def save_results(ml_super_frame, count_frame, subpixel_dist,
std, mean = np.std(subpixel_dist), np.mean(subpixel_dist)
print(f"[Plotting]: Sub-pixel distribution: RMS/Mean: {std/mean:.4f}, expected value = {1/np.sqrt(mean):.4f} for uniform distribution")
print(f"Results saved to: {output_dir}")
# 5. subpixel distribution with eta interpolation
plt.imshow(subpixel_dist_eta, origin='lower', extent=[0, 1, 0, 1])
plt.colorbar(label='Counts')
plt.title('Subpixel Distribution with Eta Interpolation')
plt.xlabel('Subpixel X')
plt.ylabel('Subpixel Y')
plt.savefig(output_dir / '3Photon_subpixel_Distribution_etaInterpolated.png', dpi=300, bbox_inches='tight')
plt.close()
np.save(output_dir / '3Photon_subpixel_Distribution_etaInterpolated.npy', subpixel_dist_eta)
std_eta, mean_eta = np.std(subpixel_dist_eta), np.mean(subpixel_dist_eta)
print(f"[Plotting]: Sub-pixel distribution with eta interpolation: RMS/Mean: {std_eta/mean_eta:.4f}, expected value = {1/np.sqrt(mean_eta):.4f} for uniform distribution")
print(f"[Plotting]: Results saved to: {output_dir}")
if __name__ == "__main__":
### output folder preparation
@@ -166,6 +197,9 @@ if __name__ == "__main__":
flag_normalize = conf.data.normalize
nChunks = np.ceil(len(files_list) / conf.inference.chunk_size).astype(int)
list_of_predictions = []
list_of_reference_points = []
ml_super_frame = np.zeros((NY*BinningFactor, NX*BinningFactor), dtype=np.int32)
count_frame = np.zeros((NY, NX), dtype=np.int32)
subpixel_dist = np.zeros((BinningFactor, BinningFactor), dtype=np.int32)
@@ -191,8 +225,30 @@ if __name__ == "__main__":
)
predictions, reference_points = run_inference(model, dataLoader, conf)
ml_super_frame_chunk, count_frame_chunk, subpixel_dist_chunk = accumulate_hits(predictions, reference_points, BinningFactor)
ml_super_frame += ml_super_frame_chunk
count_frame += count_frame_chunk
subpixel_dist += subpixel_dist_chunk
save_results(ml_super_frame, count_frame, subpixel_dist, roi, BinningFactor, output_dir)
list_of_predictions.append(predictions)
list_of_reference_points.append(reference_points)
del dataset, dataLoader
torch.cuda.empty_cache()
predictions = np.concatenate(list_of_predictions, axis=0)
ref_points = np.concatenate(list_of_reference_points, axis=0)
del list_of_predictions, list_of_reference_points
ml_super_frames, count_frame, subpixel_dist = accumulate_hits(
predictions, ref_points, binning_factor=BinningFactor
)
print('[Main]: Applying eta interpolation to predictions...')
predictions_eta_interpolated = interpolate_eta_from_points(predictions)
ml_super_frames_eta, count_frame_eta, subpixel_dist_eta = accumulate_hits(
predictions_eta_interpolated, ref_points, binning_factor=BinningFactor
)
save_results(ml_super_frames, ml_super_frames_eta,
count_frame, subpixel_dist, subpixel_dist_eta,
roi=roi, binning_factor=BinningFactor,
output_dir=output_dir
)