rsomani95

import timm
import torch
import torch.distributed as dist
import torch.nn as nn
from upyog.imports import *

from yolox.exp.yolox_base import Exp as DefaultBaseExp
from yolox.models import YOLOPAFPN, YOLOX, YOLOXHead
from yolox.utils import get_local_rank, wait_for_the_master

rsomani95

rsomani95

import fastcore.all as fastcore
import PIL

@fastcore.patch
def __or__(self: PIL.Image.Image, other: PIL.Image.Image):
    "Horizontally stack two PIL Images"
    assert isinstance(other, PIL.Image.Image)
    widths, heights = zip(*(i.size for i in [self, other]))
    
    new_img = PIL.Image.new("RGB", (sum(widths), max(heights)))

rsomani95

from PIL import Image, ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True

img_path = "image.jpg"
lut_path = "lut.cube"

img = Image.open(img_path)  # .convert("RGB")
lut = read_lut(lut_path)

# This returns a PIL Image with the LUT applied

rsomani95

import torch, torchvision
from onnxruntime import InferenceSession

model = torchvision.models.detection.maskrcnn_resnet50_fpn(pretrained=True)
x = torch.rand(1, 3, 360, 640)

torch.onnx.export(
    model,
    x,  # ONNX requires fixed input size
    "test-mask-rcnn-export.onnx",

rsomani95

# !pip install pyav
import av

PATH_VIDEO = "..."

container = av.open(PATH_VIDEO)
video_stream = container.streams.video[0]
video_decoder = container.decode(video_stream)

for frame in video_decoder:

rsomani95

from icevision.visualize.draw_data import draw_bbox
# import a bunch of other stuff here

cmap_path = "zz_color_map_coco.json"
with open(cmap_path) as f:
    COLOR_MAP_COCO = json.load(f)
    COLOR_MAP_COCO = {
        k: np.array(v).astype(np.float) for k, v in COLOR_MAP_COCO.items()
    }

rsomani95

import torch.nn as nn
import torch

class FocalLoss(nn.Module):
    # implementation adapted from https://amaarora.github.io/2020/06/29/FocalLoss.html
    # paper: https://arxiv.org/abs/1708.02002
    "Focal Loss"
    def __init__(self, alpha=.25, gamma=2, reduction='none', pos_weight=None):
        super(FocalLoss, self).__init__()
        self.alpha = torch.tensor([alpha, 1-alpha])#.cuda()

rsomani95

[['color.key.shot.na: 0.9225'],
 ['color.saturation.neutral: 0.9356'],
 ['color.theory.analagous: 0.8844'],
 ['color.tones.warm: 0.8524'],
 ['shot.angle.low: 0.8402'],
 ['shot.composition.center: 0.9619'],
 ['shot.focus.na: 0.7580'],
 ['shot.framing.wide: 0.8396'],
 ['shot.level.level: 0.9755'],
 ['shot.lighting.cast.hard: 0.8004'],

rsomani95

{'color_key': {'color.key.blue': 0.0003,
               'color.key.green': 0.0001,
               'color.key.luma': 0.0636,
               'color.key.matte': 0.0087,
               'color.key.shot.na': 0.9225,
               'color.key.texture.na': 0.0049},
 'color_saturation': {'color.saturation.desaturated': 0.0453,
                      'color.saturation.neutral': 0.9356,
                      'color.saturation.pastel': 0.001,
                      'color.saturation.saturated': 0.018},