github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-04-29 03:00:14 -04:00
Code Issues Packages Projects Releases Wiki Activity

revert helper changes

Browse Source
This commit is contained in:
Francis Lata
2024-12-23 05:16:13 +00:00
parent fb689f7097
commit 0e4fd5c573
2 changed files with 16 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
extra/models/helpers.py
View File

@@ -1,23 +0,0 @@
import numpy as np
from tinygrad import Tensor
def nms(boxes, scores, thresh=0.5):
x1, y1, x2, y2 = np.rollaxis(boxes, 1)
areas = (x2 - x1 + 1) * (y2 - y1 + 1)
to_process, keep = scores.argsort()[::-1], []
while to_process.size > 0:
cur, to_process = to_process[0], to_process[1:]
keep.append(cur)
inter_x1 = np.maximum(x1[cur], x1[to_process])
inter_y1 = np.maximum(y1[cur], y1[to_process])
inter_x2 = np.minimum(x2[cur], x2[to_process])
inter_y2 = np.minimum(y2[cur], y2[to_process])
inter_area = np.maximum(0, inter_x2 - inter_x1 + 1) * np.maximum(0, inter_y2 - inter_y1 + 1)
iou = inter_area / (areas[cur] + areas[to_process] - inter_area)
to_process = to_process[np.where(iou <= thresh)[0]]
return keep
def meshgrid(x, y):
grid_x = Tensor.cat(*[x[idx:idx+1].expand(y.shape).unsqueeze(0) for idx in range(x.shape[0])])
grid_y = Tensor.cat(*[y.unsqueeze(0)]*x.shape[0])
return grid_x.reshape(-1, 1), grid_y.reshape(-1, 1)

17
extra/models/retinanet.py
View File

@@ -7,11 +7,26 @@ import tinygrad.nn as nn
from examples.mlperf.helpers import generate_anchors
from examples.mlperf.initializers import Conv2dNormal, Conv2dKaimingUniform
from examples.mlperf.losses import sigmoid_focal_loss, l1_loss
from extra.models.helpers import nms
from extra.models.resnet import ResNet
from extra.models.mask_rcnn import BoxCoder
import numpy as np
def nms(boxes, scores, thresh=0.5):
x1, y1, x2, y2 = np.rollaxis(boxes, 1)
areas = (x2 - x1 + 1) * (y2 - y1 + 1)
to_process, keep = scores.argsort()[::-1], []
while to_process.size > 0:
cur, to_process = to_process[0], to_process[1:]
keep.append(cur)
inter_x1 = np.maximum(x1[cur], x1[to_process])
inter_y1 = np.maximum(y1[cur], y1[to_process])
inter_x2 = np.minimum(x2[cur], x2[to_process])
inter_y2 = np.minimum(y2[cur], y2[to_process])
inter_area = np.maximum(0, inter_x2 - inter_x1 + 1) * np.maximum(0, inter_y2 - inter_y1 + 1)
iou = inter_area / (areas[cur] + areas[to_process] - inter_area)
to_process = to_process[np.where(iou <= thresh)[0]]
return keep
def decode_bbox(offsets, anchors):
dx, dy, dw, dh = np.rollaxis(offsets, 1)
widths, heights = anchors[:, 2] - anchors[:, 0], anchors[:, 3] - anchors[:, 1]