Paano i-compress ang mga imahe gamit ang SVD at TensorFlow Core APIs

Paglalarawan ng Content Introduction ang ang setup Ang mga pangunahing Low rank approximation sa pamamagitan ng SVD Pag-load ng Imahe Mga algorithm ng kompresyon ng imahe Paglalarawan ng Approximations Mga pahinang tumuturo sa Rank Kumulatifong halaga ng mga singular na halaga Ang mga error at singular na mga halaga Konklusyon Introduction ang Ang tutorial na ito ay nagpapakita ng teknolohiya ng (SVD) at ang kanyang mga application para sa mga problema ng low-rank approximation. Ang SVD ay ginagamit upang factorize real o complex matris at may isang iba't ibang mga kaso ng paggamit sa data science, tulad ng pag-compression ng imahe. Ang mga imahe para sa tutorial na ito ay dumating mula sa Google Brain's ang project. Ang mga sumusunod na halaga ang imahe ang setup import matplotlib from matplotlib.image import imread from matplotlib import pyplot as plt import requests # Preset Matplotlib figure sizes. matplotlib.rcParams['figure.figsize'] = [16, 9] import tensorflow as tf print(tf.__version__) 2024-08-15 02:40:31.007111: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:485] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered 2024-08-15 02:40:31.028202: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:8454] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered 2024-08-15 02:40:31.034794: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1452] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered 2.17.0 Ang mga pangunahing Ang singular value decomposition ng isang matrix, A, ay itinatag sa pamamagitan ng mga sumusunod na factorization: A = sa loob Nasaan ang Am×n: input matrix kung saan m≥n Um×n: orthogonal matrix, UTU=I, na may bawat column, ui, na nagpapakita ng isang solong singular na vector ng A Σn×n: diagonal na matrix na may bawat diagonal na input, σi, na nagpapakita ng isang singular na halaga ng A VTn×n: orthogonal matrix, VTV=I, na may bawat linya, vi, na nagpapakita ng isang kanan na singular na vector ng A Kapag m<n, U at Σ ay may dimension (m×m), at ang VT ay may dimension (m×n). Ang linear algebra package ng TensorFlow ay may isang function, , na maaaring gamitin upang i-calculate ang singular value decomposition ng isang o higit pa ng mga matrices. Magsisimula sa pamamagitan ng defining isang simpleng matrice at i-calculate ang kanyang SVD factorization. tf.linalg.svd A = tf.random.uniform(shape=[40,30]) # Compute the SVD factorization s, U, V = tf.linalg.svd(A) # Define Sigma and V Transpose S = tf.linalg.diag(s) V_T = tf.transpose(V) # Reconstruct the original matrix A_svd = U@S@V_T # Visualize plt.bar(range(len(s)), s); plt.xlabel("Singular value rank") plt.ylabel("Singular value") plt.title("Bar graph of singular values"); WARNING: All log messages before absl::InitializeLog() is called are written to STDERR I0000 00:00:1723689633.424919 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689633.428762 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689633.432420 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689633.437707 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689633.449819 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689633.453529 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689633.457079 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689633.460600 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689633.464019 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689633.467451 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689633.470857 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689633.474376 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.696864 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.698995 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.701043 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.703182 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.705235 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.707223 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.709163 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.711162 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.713114 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.715090 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.717036 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.719041 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.757916 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.759983 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.761969 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.764105 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.766234 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.768229 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.770180 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.772176 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.774158 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.776614 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.779022 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689634.781451 126589 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 ang Dapat gamitin ang matrix reconstruction mula sa mga output. . tf.einsum tf.linalg.svd A_svd = tf.einsum('s,us,vs -> uv',s,U,V) print('\nReconstructed Matrix, A_svd', A_svd) Reconstructed Matrix, A_svd tf.Tensor( [[0.819255 0.95082176 0.30085772 ... 0.34306473 0.9112067 0.45277762] [0.71988434 0.20786765 0.3098724 ... 0.82478774 0.82917416 0.19994356] [0.23278469 0.17072585 0.10488278 ... 0.60508984 0.8121238 0.10594818] ... [0.10953172 0.72077066 0.30288345 ... 0.5790409 0.15389094 0.11265922] [0.44894537 0.04313891 0.6947845 ... 0.41589245 0.6668094 0.53494996] [0.79140556 0.47814542 0.24980234 ... 0.32851136 0.86007696 0.32692394]], shape=(40, 30), dtype=float32) Low rank approximation sa pamamagitan ng SVD Ang ranking ng isang matrix, A, ay itinuturing sa pamamagitan ng dimensyon ng vector space na tinatawag sa pamamagitan ng mga columns nito. Ang SVD ay maaaring gamitin upang i-approximate ang isang matrix na may isang mas mababang ranking, na kung saan sa katapusan ay hinihintay ang dimensionality ng data na kinakailangan upang i-storage ang impormasyon na ibinibigay ng matrix. Ang ranking-r approximation ng A sa ilalim ng SVD ay defined sa pamamagitan ng formula: Ar=UrΣrVrT Nasaan ang Urm×r: matrix na binubuo ng mga unang r columns ng U Σrr×r: ang diagonal na matrix na binubuo ng mga unang r singular na mga halaga sa Σ Vrr×nT: matrix na binubuo ng mga unang r row ng VT Magsisimula sa pag-script ng isang function upang i-calculate ang rank-r approximation ng isang given matrix. Ang low-range approximation procedure na ito ay ginagamit para sa pag-compression ng imahe; kaya, ito ay madaling din upang i-calculate ang physical data sizes para sa bawat approximation. Para sa simplicity, itinuturing na ang data size para sa isang rank-r approximated matrix ay katumbas sa kabuuang bilang ng mga elemento na kinakailangan upang i-calculate ang approximation. Pagkatapos, i-script ang isang function upang i-visualize ang orihinal na matrix, A ang rank-r approximation, Ar at ang error matrix, I-A-Ar. def rank_r_approx(s, U, V, r, verbose=False): # Compute the matrices necessary for a rank-r approximation s_r, U_r, V_r = s[..., :r], U[..., :, :r], V[..., :, :r] # ... implies any number of extra batch axes # Compute the low-rank approximation and its size A_r = tf.einsum('...s,...us,...vs->...uv',s_r,U_r,V_r) A_r_size = tf.size(U_r) + tf.size(s_r) + tf.size(V_r) if verbose: print(f"Approximation Size: {A_r_size}") return A_r, A_r_size def viz_approx(A, A_r): # Plot A, A_r, and A - A_r vmin, vmax = 0, tf.reduce_max(A) fig, ax = plt.subplots(1,3) mats = [A, A_r, abs(A - A_r)] titles = ['Original A', 'Approximated A_r', 'Error |A - A_r|'] for i, (mat, title) in enumerate(zip(mats, titles)): ax[i].pcolormesh(mat, vmin=vmin, vmax=vmax) ax[i].set_title(title) ax[i].axis('off') print(f"Original Size of A: {tf.size(A)}") s, U, V = tf.linalg.svd(A) Original Size of A: 1200 # Rank-15 approximation A_15, A_15_size = rank_r_approx(s, U, V, 15, verbose = True) viz_approx(A, A_15) Approximation Size: 1065 # Rank-3 approximation A_3, A_3_size = rank_r_approx(s, U, V, 3, verbose = True) viz_approx(A, A_3) Approximation Size: 213 Sa katunayan, ang paggamit ng mga mas mababa na antas ay nagdudulot sa mas mababa na antas ng antas. Gayunpaman, ang kalidad ng mga low-ranking antas ay karaniwang mababago sa real world scenarios. Tandaan din na ang pangunahing target ng low-ranking antas na may SVD ay upang mabawasan ang dimensionality ng data ngunit hindi upang mabawasan ang disk space ng data mismo. Gayunpaman, dahil ang mga input matris ay mas mataas-dimensional, maraming mga low-ranking antas ay nagsisimula din na may kapaki-pakinabang sa size ng data. Ito ay ang diskarte ng reduction na dahilan kung bakit ang proseso ay magagamit para sa mga problema ng pag-compression ng imahe. Pag-load ng Imahe Ang mga sumusunod na imahe ay magagamit sa Home page. Imagen ay isang text-to-image diffusion model na binuo sa pamamagitan ng Google Research's Brain team. Ang isang AI ay binuo ng larawan na ito batay sa prompt: "A photo of a Corgi dog riding a bike in Times Square. It's wearing sunglasses and a beach hat." How cool is that! You can also change the url below to any .jpg link to load in a custom image of choice. ang imahe Pagkatapos ng pagbabasa ng isang JPEG file, Matplotlib output ang isang matrix, I, ng shape (m×n×3) na nagpapakita ng isang 2-dimensional na imahe na may 3 color channels para sa red, green at blue respectively. img_link = "https://imagen.research.google/main_gallery_images/a-photo-of-a-corgi-dog-riding-a-bike-in-times-square.jpg" img_path = requests.get(img_link, stream=True).raw I = imread(img_path, 0) print("Input Image Shape:", I.shape) Input Image Shape: (1024, 1024, 3) def show_img(I): # Display the image in matplotlib img = plt.imshow(I) plt.axis('off') return show_img(I) Mga algorithm ng kompresyon ng imahe Ngayon, gamitin ang SVD upang i-calculate low-range approximations ng sample image. Mangyaring i-recall na ang imahe ay ng shape (1024×1024×3) at na ang teorya ng SVD ay ginagamit lamang para sa 2-dimensional na matrices. Ito ay nangangahulugan na ang sample image ay dapat na batch sa 3 equal-size matrices na katumbas sa bawat ng 3 color channels. Ito ay maaaring gawin ito sa pamamagitan ng transposing ang matrix upang maging ng shape (3×1024×1024). Upang malinaw ang mga error ng approximation, i-rescale ang mga halaga ng RGB ng imahe mula sa [0,255] sa [0,1]. Mangyaring i-clip ang mga halaga ng approximation upang pumunta sa antas na ito bago Ang function ay useful para sa ito. tf.clip_by_value def compress_image(I, r, verbose=False): # Compress an image with the SVD given a rank I_size = tf.size(I) print(f"Original size of image: {I_size}") # Compute SVD of image I = tf.convert_to_tensor(I)/255 I_batched = tf.transpose(I, [2, 0, 1]) # einops.rearrange(I, 'h w c -> c h w') s, U, V = tf.linalg.svd(I_batched) # Compute low-rank approximation of image across each RGB channel I_r, I_r_size = rank_r_approx(s, U, V, r) I_r = tf.transpose(I_r, [1, 2, 0]) # einops.rearrange(I_r, 'c h w -> h w c') I_r_prop = (I_r_size / I_size) if verbose: # Display compressed image and attributes print(f"Number of singular values used in compression: {r}") print(f"Compressed image size: {I_r_size}") print(f"Proportion of original size: {I_r_prop:.3f}") ax_1 = plt.subplot(1,2,1) show_img(tf.clip_by_value(I_r,0.,1.)) ax_1.set_title("Approximated image") ax_2 = plt.subplot(1,2,2) show_img(tf.clip_by_value(0.5+abs(I-I_r),0.,1.)) ax_2.set_title("Error") return I_r, I_r_prop Ngayon, i-calculate ang ranking-r approximations para sa mga sumusunod na mga ranking : 100, 50, 10 I_100, I_100_prop = compress_image(I, 100, verbose=True) Original size of image: 3145728 Number of singular values used in compression: 100 Compressed image size: 614700 Proportion of original size: 0.195 I_50, I_50_prop = compress_image(I, 50, verbose=True) Original size of image: 3145728 Number of singular values used in compression: 50 Compressed image size: 307350 Proportion of original size: 0.098 I_10, I_10_prop = compress_image(I, 10, verbose=True) Original size of image: 3145728 Number of singular values used in compression: 10 Compressed image size: 61470 Proportion of original size: 0.020 Paglalarawan ng Approximations Mayroong isang iba't-ibang mga paraan na interesado upang matukoy ang katotohanan at magkaroon ng mas mababang kontrol sa mga matrix approximations. Mga pahinang tumuturo sa Rank Para sa bawat ng mga halimbawa sa itaas, tingnan kung paano ang mga size ng data ay nagbabago sa ranking. plt.figure(figsize=(11,6)) plt.plot([100, 50, 10], [I_100_prop, I_50_prop, I_10_prop]) plt.xlabel("Rank") plt.ylabel("Proportion of original image size") plt.title("Compression factor vs rank"); Upang i-explore ang mga ito sa karagdagang, tinatanggap na ang size ng data ng isang pamantayan na matrix, Ar, ay defined bilang ang kabuuang bilang ng mga elemento na kinakailangan para sa kanyang kalkulasyon. Ang mga sumusunod na equations ay maaaring gamitin upang makikita ang relasyon sa pagitan ng pamantayan na factor at ranking: x=(m×r)+r+(r×n)=r×(m+n+1) c=xy=r×(m+n+1)m×n Nasaan ang x: ang size ng air Y: ang size ng A C = XY: ang kompresyon factor r: ang rank ng approximation m at n: mga dimensyon ng row at column ng A Upang makakuha ng ang ranking, r, na kailangan upang i-compress ang isang imahe sa isang kinakailangan na factor, c, ang mga sumusunod sa itaas ay maaaring rearranged upang i-solve para sa r: r=c×m×nm+n+1 Tingnan na ang formula na ito ay independiyenteng sa color channel dimension dahil ang bawat ng mga RGB approximations ay hindi nakakaapekto sa ibang tao. Ngayon, mag-sign ng isang function upang compress ang isang input na imahe ayon sa isang kinakailangan na compression factor. def compress_image_with_factor(I, compression_factor, verbose=False): # Returns a compressed image based on a desired compression factor m,n,o = I.shape r = int((compression_factor * m * n)/(m + n + 1)) I_r, I_r_prop = compress_image(I, r, verbose=verbose) return I_r I-compress ang isang imahe sa 15% ng kanyang orihinal na size. compression_factor = 0.15 I_r_img = compress_image_with_factor(I, compression_factor, verbose=True) Original size of image: 3145728 Number of singular values used in compression: 76 Compressed image size: 467172 Proportion of original size: 0.149 Kumulatifong halaga ng mga singular na halaga Ang kumulatifong halaga ng mga singular na halaga ay maaaring maging isang useful indicator para sa halaga ng enerhiya na nakuha sa pamamagitan ng isang ranking-r approximation. Visualize ang RGB-average kumulative proportion ng singular na halaga sa sample image. Ang function ay maaaring magagamit para sa ito. tf.cumsum def viz_energy(I): # Visualize the energy captured based on rank # Computing SVD I = tf.convert_to_tensor(I)/255 I_batched = tf.transpose(I, [2, 0, 1]) s, U, V = tf.linalg.svd(I_batched) # Plotting average proportion across RGB channels props_rgb = tf.map_fn(lambda x: tf.cumsum(x)/tf.reduce_sum(x), s) props_rgb_mean = tf.reduce_mean(props_rgb, axis=0) plt.figure(figsize=(11,6)) plt.plot(range(len(I)), props_rgb_mean, color='k') plt.xlabel("Rank / singular value number") plt.ylabel("Cumulative proportion of singular values") plt.title("RGB-averaged proportion of energy captured by the first 'r' singular values") viz_energy(I) Nakikita na ang higit sa 90% ng enerhiya sa larawan na ito ay na-capture sa loob ng mga unang 100 singular na mga halaga. Ngayon, mag-sign ng isang function upang i-compress ang isang input na larawan na may isang kinakailangan na enerhiya retention factor. def compress_image_with_energy(I, energy_factor, verbose=False): # Returns a compressed image based on a desired energy factor # Computing SVD I_rescaled = tf.convert_to_tensor(I)/255 I_batched = tf.transpose(I_rescaled, [2, 0, 1]) s, U, V = tf.linalg.svd(I_batched) # Extracting singular values props_rgb = tf.map_fn(lambda x: tf.cumsum(x)/tf.reduce_sum(x), s) props_rgb_mean = tf.reduce_mean(props_rgb, axis=0) # Find closest r that corresponds to the energy factor r = tf.argmin(tf.abs(props_rgb_mean - energy_factor)) + 1 actual_ef = props_rgb_mean[r] I_r, I_r_prop = compress_image(I, r, verbose=verbose) print(f"Proportion of energy captured by the first {r} singular values: {actual_ef:.3f}") return I_r I-compress ang isang imahe upang i-return ang 75% ng enerhiya nito. energy_factor = 0.75 I_r_img = compress_image_with_energy(I, energy_factor, verbose=True) Original size of image: 3145728 Number of singular values used in compression: 35 Compressed image size: 215145 Proportion of original size: 0.068 Proportion of energy captured by the first 35 singular values: 0.753 Ang mga error at singular na mga halaga Mayroong isang interesado na relasyon sa pagitan ng error ng pag-approximation at ang mga singular na halaga. Ito ay tumutulong na ang square Frobenius norm ng pag-approximation ay katumbas ng halaga ng mga squares ng mga singular na halaga na nangangahulugan: Sino ang nagtatagumpay? I-test ang relasyon na ito na may isang ranking-10 pag-approximation ng sample matrix sa simula ng tutorial na ito. s, U, V = tf.linalg.svd(A) A_10, A_10_size = rank_r_approx(s, U, V, 10) squared_norm = tf.norm(A - A_10)**2 s_squared_sum = tf.reduce_sum(s[10:]**2) print(f"Squared Frobenius norm: {squared_norm:.3f}") print(f"Sum of squared singular values left out: {s_squared_sum:.3f}") Squared Frobenius norm: 33.021 Sum of squared singular values left out: 33.021 Konklusyon Ang notebook na ito ay inilathala ng proseso ng pag-implementasyon ng single value decomposition sa TensorFlow at pag-apply ito sa pag-script ng isang algorithm ng pag-compression ng imahe. Narito ang ilang higit pang mga tip na maaaring makatulong: Ang TensorFlow Core APIs ay maaaring gamitin para sa isang iba't ibang uri ng high-performance scientific computing mga kaso ng paggamit. Upang malaman ang higit pa tungkol sa linear algebra functionalities ng TensorFlow, bisitahin ang mga docs para sa linalg module. Ang SVD ay maaaring gamitin din upang bumuo ng mga sistema ng rekomendasyon. Para sa higit pang mga halimbawa ng paggamit ng TensorFlow Core APIs, tingnan ang Kung nais mong malaman ang higit pa tungkol sa pag-load at pag-prepare ng data, tingnan ang mga tutorial sa o sa . ang guide I-load ang data Pag-load ng data ng CSV Nai-publish sa website ng TensorFlow, ang artikulong ito ay ipinapakita dito sa ilalim ng isang bagong headline at ay may lisensya sa ilalim ng CC BY 4.0. Nai-publish sa website ng TensorFlow, ang artikulong ito ay ipinapakita dito sa ilalim ng isang bagong headline at ay may lisensya sa ilalim ng CC BY 4.0. Mga TensorFlow Mga TensorFlow