Spaces:
Running on Zero
Running on Zero
File size: 3,185 Bytes
5f5f544 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import math
from pathlib import Path
import cv2
import numpy as np
import torch
from sapiens.registry import VISUALIZERS
from torch import nn
@VISUALIZERS.register_module()
class BaseVisualizer(nn.Module):
def __init__(
self,
output_dir: str,
vis_interval: int = 100,
vis_max_samples: int = 16,
vis_downsample: int = 2,
):
super().__init__()
self.output_dir = Path(output_dir)
self.output_dir.mkdir(parents=True, exist_ok=True)
self.vis_max_samples = vis_max_samples
self.vis_interval = vis_interval
self.vis_downsample = vis_downsample
def add_batch(self, data_batch: dict, logs: dict, step: int):
images = data_batch["data_samples"]["image"].detach().cpu()
outputs = logs["outputs"].detach().cpu() # (B, C, H, W)
if outputs.dtype == torch.bfloat16:
images = images.float()
outputs = outputs.float()
batch_size = min(len(images), self.vis_max_samples)
save_images = []
for i in range(batch_size):
gt_image = images[i].permute(1, 2, 0).cpu().numpy() * 255
pred_image = outputs[i].permute(1, 2, 0).cpu().numpy() * 255
gt_image = np.clip(gt_image, 0, 255).astype(np.uint8)
pred_image = np.clip(pred_image, 0, 255).astype(np.uint8)
image_height, image_width = gt_image.shape[:2]
if self.vis_downsample > 1:
image_height = int(image_height / self.vis_downsample)
image_width = int(image_width / self.vis_downsample)
gt_image = cv2.resize(
gt_image,
(image_width, image_height),
interpolation=cv2.INTER_AREA,
)
pred_image = cv2.resize(
pred_image,
(image_width, image_height),
interpolation=cv2.INTER_AREA,
)
save_image = np.concatenate([gt_image, pred_image], axis=1)
save_images.append(save_image)
out_file = self.output_dir / f"{step:06d}.jpg"
image_height, image_width = save_images[0].shape[:2]
cols = int(math.ceil(math.sqrt(batch_size)))
rows = int(math.ceil(batch_size / cols))
canvas_height = rows * image_height
canvas_width = cols * image_width
canvas = np.zeros((canvas_height, canvas_width, 3), dtype=np.uint8)
for idx, image in enumerate(save_images):
row = idx // cols
col = idx % cols
canvas[
row * image_height : (row + 1) * image_height,
col * image_width : (col + 1) * image_width,
] = image
## downsample canvas by 2x
canvas = cv2.resize(
canvas,
(canvas_width, canvas_height),
interpolation=cv2.INTER_AREA,
)
cv2.imwrite(out_file, canvas)
return
|