:jC ddlmZmZddlZddlmZddlmZmZer ddlZ ddl m Z ndgZ erddl mZded efd Z dd ed ed ed eeeeefffdZy))DictUnionN)Tensor)_EINOPS_AVAILABLE_TORCH_VMAF_AVAILABLE)VMAF$video_multi_method_assessment_fusion) rearrangevideoreturnc|dddddddddf}|dddddddddf}|dddddddddf}d|zd|zzd|zzjddzS) z/Calculate the luma component of a video tensor.NrgA`"?gbX9?gv/?) unsqueeze)r rgbs w/home/mcse/projects/srt_converter/srt-converter-venv/lib/python3.12/site-packages/torchmetrics/functional/video/vmaf.pycalculate_lumarsj aAq!mA aAq!mA aAq!mA AI !EAI - 8 8 ;c AApredstargetfeaturesc Pts td|jd}|j|j}}t |}t |}t j|}|svt|D cgc].} |jt|| dt|| d0} } tj| djj|St|D cgc].} |jt|| dt|| d0} } t|D cgc]&} | | jt j"d(} } t|D cgc]K} | | j$D cic]0} | dk7s | tj&| | | j(| 2c} M}} } |dD cic]5} | tj*t|D cgc] } || |  c} 7c} } Scc} wcc} wcc} wcc} wcc} } wcc} wcc} } w) aCalculates Video Multi-Method Assessment Fusion (VMAF) metric. VMAF is a full-reference video quality assessment algorithm that combines multiple quality assessment features such as detail loss, motion, and contrast using a machine learning model to predict human perception of video quality more accurately than traditional metrics like PSNR or SSIM. The metric works by: 1. Converting input videos to luma component (grayscale) 2. Computing multiple elementary features: - Additive Detail Measure (ADM): Evaluates detail preservation at different scales - Visual Information Fidelity (VIF): Measures preservation of visual information across frequency bands - Motion: Quantifies the amount of motion in the video 3. Combining these features using a trained SVM model to predict quality .. note:: This implementation requires you to have vmaf-torch installed: https://github.com/alvitrioliks/VMAF-torch. Install either by cloning the repository and running `pip install .` or with `pip install torchmetrics[video]`. Args: preds: Video tensor of shape (batch, channels, frames, height, width). Expected to be in RGB format with values in range [0, 1]. target: Video tensor of shape (batch, channels, frames, height, width). Expected to be in RGB format with values in range [0, 1]. features: If True, all the elementary features (ADM, VIF, motion) are returned along with the VMAF score in a dictionary. This corresponds to the output you would get from the VMAF command line tool with the `--csv` option enabled. If False, only the VMAF score is returned as a tensor. Returns: - If `features` is False, returns a tensor with shape (batch, frame) of VMAF score for each frame in each video. Higher scores indicate better quality, with typical values ranging from 0 to 100. - If `features` is True, returns a dictionary where each value is a (batch, frame) tensor of the corresponding feature. The keys are: - 'integer_motion2': Integer motion feature - 'integer_motion': Integer motion feature - 'integer_adm2': Integer ADM feature - 'integer_adm_scale0': Integer ADM feature at scale 0 - 'integer_adm_scale1': Integer ADM feature at scale 1 - 'integer_adm_scale2': Integer ADM feature at scale 2 - 'integer_adm_scale3': Integer ADM feature at scale 3 - 'integer_vif_scale0': Integer VIF feature at scale 0 - 'integer_vif_scale1': Integer VIF feature at scale 1 - 'integer_vif_scale2': Integer VIF feature at scale 2 - 'integer_vif_scale3': Integer VIF feature at scale 3 - 'vmaf': VMAF score for each frame in each video Example: >>> import torch >>> from torchmetrics.functional.video import video_multi_method_assessment_fusion >>> # 2 videos, 3 channels, 10 frames, 32x32 resolution >>> preds = torch.rand(2, 3, 10, 32, 32, generator=torch.manual_seed(42)) >>> target = torch.rand(2, 3, 10, 32, 32, generator=torch.manual_seed(43)) >>> vmaf_score = video_multi_method_assessment_fusion(preds, target) >>> torch.round(vmaf_score, decimals=2) tensor([[ 9.9900, 15.9000, 14.2600, 16.6100, 15.9100, 14.3000, 13.5800, 13.4900, 15.4700, 20.2800], [ 6.2500, 11.3000, 17.3000, 11.4600, 19.0600, 14.9300, 14.0500, 14.4100, 12.4700, 14.8200]]) >>> vmaf_dict = video_multi_method_assessment_fusion(preds, target, features=True) >>> # show a couple of features, more features are available >>> vmaf_dict['vmaf'].round(decimals=2) tensor([[ 9.9900, 15.9000, 14.2600, 16.6100, 15.9100, 14.3000, 13.5800, 13.4900, 15.4700, 20.2800], [ 6.2500, 11.3000, 17.3000, 11.4600, 19.0600, 14.9300, 14.0500, 14.4100, 12.4700, 14.8200]]) >>> vmaf_dict['integer_adm2'].round(decimals=2) tensor([[0.4500, 0.4500, 0.3600, 0.4700, 0.4300, 0.3600, 0.3900, 0.4100, 0.3700, 0.4700], [0.4200, 0.3900, 0.4400, 0.3700, 0.4500, 0.3900, 0.3800, 0.4800, 0.3900, 0.3900]]) zSvmaf-torch is not installed. Please install with `pip install torchmetrics[video]`.rzc f h w -> f c h wr)dimcoerce)errorsFrame)dtype)r RuntimeErrorshaper devicerrtorangecompute_vmaf_scorer torchcatttableapplypd to_numericcolumnstensorvaluesstack)rrrr orig_dtyper# preds_luma target_lumavmafr scoresscores_and_featuresdfscolresults rr r 'sR !pqq AAellJ&J (K 699V D  q    # #+e,.BCYzZ_O`bvEw   yyQ'))+..z:: 1X   k%(*> ?:V[K\^rAs  Z__`Ya bPU u % + +BMM( + K bC b1X QTTYPZPbPbufimtftell3u:c?11D DuFU[[\T] ^SCeAhGUfUmC0GH H ^^'  cuH ^sH73H-3H/+H *H H *H 5H"H"'H6 H"HH")F)typingrrr'rtorchmetrics.utilities.importsrrpandasr, vmaf_torchr__doctest_skip__einopsr rboolstrr rrrDs S>? B&BVBg_ g_ g_g_ 64V $ $% g_r