add all files

README.md

@@ -1,6 +1,6 @@
 ---
 title: Visdif
-emoji: 💩
+emoji: 🚀
 colorFrom: green
 colorTo: yellow
 sdk: gradio

app.py

@@ -3,23 +3,69 @@ import torch
 import requests
 from torchvision import transforms
 
-model = torch.hub.load('pytorch/vision:v0.6.0', 'resnet18', pretrained=True).eval()
-response = requests.get("https://git.io/JJkYN")
-labels = response.text.split("\n")
+from sampling_util import furthest_neighbours
+from video_reader import video_reader
 
-torch.hub.download_url_to_file("https://github.com/pytorch/hub/raw/master/images/dog.jpg", "dog.jpg")
+model = torch.load("model").eval()
+avg_pool = nn.AdaptiveAvgPool2d((1, 1))
 
-def predict(inp):
-  inp = transforms.ToTensor()(inp).unsqueeze(0)
-  with torch.no_grad():
-    prediction = torch.nn.functional.softmax(model(inp)[0], dim=0)
-    confidences = {labels[i]: float(prediction[i]) for i in range(1000)}    
-  return confidences
 
-demo = gr.Interface(fn=predict, 
-             inputs=gr.inputs.Image(type="pil"),
-             outputs=gr.outputs.Label(num_top_classes=3),
-             examples=[["dog.jpg"]],
-             )
+def predict(input_file):
+	base_directory = os.getcwd()
+    selected_directory = os.path.join(base_directory, "selected_images")
+    if os.path.isdir(selected_directory):
+        shutil.rmtree(selected_directory)
+    os.mkdir(selected_directory)
+    
+    zip_path = os.path.join(input_file.split('/')[-1][:-4] + ".zip")
+    
+    mean = [0.3156024, 0.33569682, 0.34337464]
+    std = [0.16568947, 0.17827448, 0.18925823]
+    
+    img_vecs = []
+    with torch.no_grad():
+		for fp_i, file_path in enumerate([input_file]):
+		    for i, in_img in enumerate(video_reader(file_path,
+		                                         targetFPS=9,
+		                                         targetWidth=100,
+		                                         to_rgb=True)):
+		        in_img = (in_img.astype(np.float32) / 255.)                                 
+		        in_img = (in_img - mean) / std
+		        in_img = np.transpose(in_img, (0, 3, 1, 2))
+		        in_img = torch.from_numpy(in_img)
+		        encoded = avg_pool(model(in_img))[0, :, 0, 0].cpu().numpy()
+		        img_vecs += [encoded]
+
+	img_vecs = np.asarray(img_vecs)
+	rv_indices, _ = furthest_neighbours(
+        img_vecs,
+        downsample_size,
+        seed=0)
+    indices = np.zeros((img_vecs.shape[0],))
+    indices[np.asarray(rv_indices)] = 1
+    
+    global_ctr = 0
+    for fp_i, file_path in enumerate([input_file]):
+        for i, img in enumerate(video_reader(file_path,
+                                             targetFPS=9,
+                                             targetWidth=None,
+                                             to_rgb=False)):
+            if indices[global_ctr] == 1:
+                cv2.imwrite(join(selected_directory, str(global_ctr) + ".jpg"), img)
+            global_ctr += 1
+            
+    all_selected_imgs_path = [join(selected_directory, f) for f in listdir(selected_directory) if isfile(join(selected_directory, f))]
+    if 0 < len(all_file_paths):
+    zipf = zipfile.ZipFile(zip_path, 'w', zipfile.ZIP_DEFLATED)
+    for i, f in enumerate(all_selected_imgs_path):
+        zipf.write(f, basename(f))
+    zipf.close()
+	
+	return zip_path
+
+demo = gr.Interface(
+	fn=predict, 
+	inputs=gr.inputs.Video(label="Upload Video File"),
+	outputs=gr.outputs.File(label="Zip"))
              
 demo.launch()

fastdist2.py

@@ -0,0 +1,332 @@
+import math
+
+import numpy as np
+from numba import jit, prange, cuda, float32
+
+
+# https://github.com/talboger/fastdist
+
+@jit(nopython=True, fastmath=True)
+def cosine(u, v, w=None):
+    """
+    :purpose:
+    Computes the cosine similarity between two 1D arrays
+    Unlike scipy's cosine distance, this returns similarity, which is 1 - distance
+
+    :params:
+    u, v   : input arrays, both of shape (n,)
+    w      : weights at each index of u and v. array of shape (n,)
+             if no w is set, it is initialized as an array of ones
+             such that it will have no impact on the output
+
+    :returns:
+    cosine  : float, the cosine similarity between u and v
+
+    :example:
+    >>> from fastdist import fastdist
+    >>> import numpy as np
+    >>> u, v, w = np.random.RandomState(seed=0).rand(10000, 3).T
+    >>> fastdist.cosine(u, v, w)
+    0.7495065944399267
+    """
+    n = len(u)
+    num = 0
+    u_norm, v_norm = 0, 0
+    for i in range(n):
+        num += u[i] * v[i] * w[i]
+        u_norm += abs(u[i]) ** 2 * w[i]
+        v_norm += abs(v[i]) ** 2 * w[i]
+
+    denom = (u_norm * v_norm) ** (1 / 2)
+    return num / denom
+
+
+@jit(nopython=True, fastmath=True)
+def cosine_vector_to_matrix(u, m):
+    """
+    :purpose:
+    Computes the cosine similarity between a 1D array and rows of a matrix
+
+    :params:
+    u      : input vector of shape (n,)
+    m      : input matrix of shape (m, n)
+
+    :returns:
+    cosine vector  : np.array, of shape (m,) vector containing cosine similarity between u
+                     and the rows of m
+
+    :example:
+    >>> from fastdist import fastdist
+    >>> import numpy as np
+    >>> u = np.random.RandomState(seed=0).rand(10)
+    >>> m = np.random.RandomState(seed=0).rand(100, 10)
+    >>> fastdist.cosine_vector_to_matrix(u, m)
+    (returns an array of shape (100,))
+    """
+    norm = 0
+    for i in range(len(u)):
+        norm += abs(u[i]) ** 2
+    u = u / norm ** (1 / 2)
+    for i in range(m.shape[0]):
+        norm = 0
+        for j in range(len(m[i])):
+            norm += abs(m[i][j]) ** 2
+        m[i] = m[i] / norm ** (1 / 2)
+    return np.dot(u, m.T)
+
+
+@jit(nopython=True, fastmath=True)
+def cosine_matrix_to_matrix(a, b):
+    """
+    :purpose:
+    Computes the cosine similarity between the rows of two matrices
+
+    :params:
+    a, b   : input matrices of shape (m, n) and (k, n)
+             the matrices must share a common dimension at index 1
+
+    :returns:
+    cosine matrix  : np.array, an (m, k) array of the cosine similarity
+                     between the rows of a and b
+
+    :example:
+    >>> from fastdist import fastdist
+    >>> import numpy as np
+    >>> a = np.random.RandomState(seed=0).rand(10, 50)
+    >>> b = np.random.RandomState(seed=0).rand(100, 50)
+    >>> fastdist.cosine_matrix_to_matrix(a, b)
+    (returns an array of shape (10, 100))
+    """
+    for i in range(a.shape[0]):
+        norm = 0
+        for j in range(len(a[i])):
+            norm += abs(a[i][j]) ** 2
+        a[i] = a[i] / norm ** (1 / 2)
+    for i in range(b.shape[0]):
+        norm = 0
+        for j in range(len(b[i])):
+            norm += abs(b[i][j]) ** 2
+        b[i] = b[i] / norm ** (1 / 2)
+    return np.dot(a, b.T)
+
+
+@jit(nopython=True, fastmath=True)
+def euclidean(u, v):
+    """
+    :purpose:
+    Computes the Euclidean distance between two 1D arrays
+
+    :params:
+    u, v   : input arrays, both of shape (n,)
+    w      : weights at each index of u and v. array of shape (n,)
+             if no w is set, it is initialized as an array of ones
+             such that it will have no impact on the output
+
+    :returns:
+    euclidean : float, the Euclidean distance between u and v
+
+    :example:
+    >>> from fastdist import fastdist
+    >>> import numpy as np
+    >>> u, v, w = np.random.RandomState(seed=0).rand(10000, 3).T
+    >>> fastdist.euclidean(u, v, w)
+    28.822558591834163
+    """
+    n = len(u)
+    dist = 0
+    for i in range(n):
+        dist += abs(u[i] - v[i]) ** 2
+    return dist ** (1 / 2)
+
+
+@jit(nopython=True, fastmath=True)
+def euclidean_vector_to_matrix_distance(u, m):
+    """
+    :purpose:
+    Computes the distance between a vector and the rows of a matrix using any given metric
+
+    :params:
+    u      : input vector of shape (n,)
+    m      : input matrix of shape (m, n)
+
+    distance vector  : np.array, of shape (m,) vector containing the distance between u
+                       and the rows of m
+
+    :example:
+    >>> from fastdist import fastdist
+    >>> import numpy as np
+    >>> u = np.random.RandomState(seed=0).rand(10)
+    >>> m = np.random.RandomState(seed=0).rand(100, 10)
+    >>> fastdist.vector_to_matrix_distance(u, m)
+    (returns an array of shape (100,))
+
+    :note:
+    the cosine similarity uses its own function, cosine_vector_to_matrix.
+    this is because normalizing the rows and then taking the dot product
+    of the vector and matrix heavily optimizes the computation. the other similarity
+    metrics do not have such an optimization, so we loop through them
+    """
+
+    n = m.shape[0]
+    out = np.zeros((n), dtype=np.float32)
+    for i in prange(n):
+        dist = 0
+        for l in range(len(u)):
+            dist += abs(u[l] - m[i][l]) ** 2
+        out[i] = dist ** (1 / 2)
+
+    return out
+
+
+@cuda.jit
+def gpu_kernel_euclidean_vector_to_matrix_distance(u, m, u_dim0, m_dim0, out):
+    # Thread id in a 1D block
+    tx = cuda.threadIdx.x
+    # Block id in a 1D grid
+    ty = cuda.blockIdx.x
+    # Block width, i.e. number of threads per block
+    bw = cuda.blockDim.x
+    # Compute flattened index inside the array
+    pos = tx + ty * bw
+    if pos < m_dim0:  # Check array boundaries
+        dist = 0
+        for l in range(u_dim0):
+            d = abs(u[l] - m[pos][l])
+            dist += d * d
+        out[pos] = dist ** (1 / 2)
+
+
+def euclidean_vector_to_matrix_distance_gpu(u, m):
+    m_dim0 = m.shape[0]
+    u_dim0 = u.shape[0]
+    out = np.zeros((m_dim0), dtype=np.float32)
+
+    threadsperblock = 16
+    blockspergrid = (m_dim0 + (threadsperblock - 1)) // threadsperblock
+    gpu_kernel_euclidean_vector_to_matrix_distance[blockspergrid, threadsperblock](u, m, u_dim0, m_dim0, out)
+
+    return out
+
+
+# https://numba.readthedocs.io/en/stable/cuda/examples.html
+@cuda.jit
+def gpu_kernel_euclidean_matrix_to_matrix_distance_fast(A, B, C):
+    TPB = 16
+
+    # Define an array in the shared memory
+    # The size and type of the arrays must be known at compile time
+    sA = cuda.shared.array(shape=(TPB, TPB), dtype=float32)
+
+    sB = cuda.shared.array(shape=(TPB, TPB), dtype=float32)
+
+    x, y = cuda.grid(2)
+
+    tx = cuda.threadIdx.x
+
+    ty = cuda.threadIdx.y
+
+    bpg = cuda.gridDim.x  # blocks per grid
+
+    # Each thread computes one element in the result matrix.
+
+    # The dot product is chunked into dot products of TPB-long vectors.
+
+    tmp = float32(0.)
+
+    for i in range(bpg):
+
+        # Preload data into shared memory
+
+        sA[ty, tx] = 0
+
+        sB[ty, tx] = 0
+
+        if y < A.shape[0] and (tx + i * TPB) < A.shape[1]:
+            sA[ty, tx] = A[y, tx + i * TPB]
+
+        if x < B.shape[1] and (ty + i * TPB) < B.shape[0]:
+            sB[ty, tx] = B[ty + i * TPB, x]
+
+        # Wait until all threads finish preloading
+
+        cuda.syncthreads()
+
+        # Computes partial product on the shared memory
+
+        for j in range(TPB):
+            d = abs(sA[ty, j] - sB[j, tx])
+            tmp += d * d
+        # Wait until all threads finish computing
+
+        cuda.syncthreads()
+
+    if y < C.shape[0] and x < C.shape[1]:
+        C[y, x] = tmp ** (1 / 2)
+
+
+def euclidean_matrix_to_matrix_distance_gpu_fast(u, m):
+    u_dim0 = u.shape[0]
+    m_dim1 = m.shape[1]
+
+    # vec_dim = u.shape[1]
+    # assert vec_dim == m.shape[1]
+    out = np.zeros((u_dim0, m_dim1), dtype=np.float32)
+
+    threadsperblock = (16, 16)
+    grid_y_max = max(u.shape[0], m.shape[0])
+    grid_x_max = max(u.shape[1], m.shape[1])
+    blockspergrid_x = math.ceil(grid_x_max / threadsperblock[0])
+    blockspergrid_y = math.ceil(grid_y_max / threadsperblock[1])
+
+    blockspergrid = (blockspergrid_x, blockspergrid_y)
+
+    u_d = cuda.to_device(u)
+    m_d = cuda.to_device(m)
+    out_d = cuda.to_device(out)
+
+    gpu_kernel_euclidean_matrix_to_matrix_distance_fast[blockspergrid, threadsperblock](u_d, m_d, out_d)
+    out = out_d.copy_to_host()
+    return out
+
+
+@jit(cache=True, nopython=True, parallel=True, fastmath=True, boundscheck=False, nogil=True)
+def euclidean_matrix_to_matrix_distance(a, b):
+    """
+    :purpose:
+    Computes the distance between the rows of two matrices using any given metric
+
+    :params:
+    a, b   : input matrices either of shape (m, n) and (k, n)
+             the matrices must share a common dimension at index 1
+    metric : the function used to calculate the distance
+    metric_name : str of the function name. this is only used for
+                  the if statement because cosine similarity has its
+                  own function
+
+    :returns:
+    distance matrix  : np.array, an (m, k) array of the distance
+                       between the rows of a and b
+
+    :example:
+    >>> from fastdist import fastdist
+    >>> import numpy as np
+    >>> a = np.random.RandomState(seed=0).rand(10, 50)
+    >>> b = np.random.RandomState(seed=0).rand(100, 50)
+    >>> fastdist.matrix_to_matrix_distance(a, b, fastdist.cosine, "cosine")
+    (returns an array of shape (10, 100))
+
+    :note:
+    the cosine similarity uses its own function, cosine_matrix_to_matrix.
+    this is because normalizing the rows and then taking the dot product
+    of the two matrices heavily optimizes the computation. the other similarity
+    metrics do not have such an optimization, so we loop through them
+    """
+    n, m = a.shape[0], b.shape[0]
+    out = np.zeros((n, m), dtype=np.float32)
+    for i in prange(n):
+        for j in range(m):
+            dist = 0
+            for l in range(len(a[i])):
+                dist += abs(a[i][l] - b[j][l]) ** 2
+            out[i][j] = dist ** (1 / 2)
+    return out

requirements.txt

@@ -1,3 +1,7 @@
 torch 
 Pillow
 torchvision
+numpy
+opencv-python
+umap-learn
+numba

sampling_util.py

@@ -0,0 +1,30 @@
+import numpy as np
+
+from fastdist2 import euclidean_vector_to_matrix_distance
+
+
+def furthest_neighbours(x, downsampled_size, seed):
+    x = x.astype(np.float32)
+    np.random.seed(seed)
+    length = x.shape[0]
+    img_vecs_dims = x.shape[-1]
+
+    rv_indices = [np.random.randint(low=0, high=downsampled_size - 1, size=1)[0]]
+    selected_points = np.zeros((downsampled_size, img_vecs_dims), np.float32)
+    selected_points[0, :] = x[rv_indices[0], :]
+
+    distance_for_selected_min = np.ones((length,)) * 1e15
+
+    inactive_points = np.zeros(length, dtype=bool)
+    inactive_points[rv_indices[0]] = True
+
+    for i in (range(downsampled_size - 1)):
+        distance_for_selected = euclidean_vector_to_matrix_distance(selected_points[i, :], x)
+        distance_for_selected_min = np.minimum(distance_for_selected_min, distance_for_selected)
+        furthest_point_idx = np.argmax(np.ma.array(distance_for_selected_min, mask=inactive_points))
+
+        rv_indices.append(furthest_point_idx)
+        selected_points[i + 1, :] = x[furthest_point_idx, :]
+        inactive_points[furthest_point_idx] = True
+
+    return rv_indices, selected_points

video_reader.py

@@ -0,0 +1,35 @@
+import cv2
+
+
+def video_reader(file_path, targetFPS=9, targetWidth=None, to_rgb=True, prompt=None):
+    cap = cv2.VideoCapture(file_path)
+    sourceFPS = int(cap.get(cv2.CAP_PROP_FPS))
+
+    if sourceFPS < targetFPS:
+        raise ValueError("sourceFPS < targetFPS: {} < {}".format(sourceFPS, targetFPS))
+
+    fpsDiv = 3  # sourceFPS // targetFPS
+    print("sourceFPS: {}, targetFPS: {}, fpsDiv: {}".format(sourceFPS, targetFPS, fpsDiv))
+
+    frameCount = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    frameWidth = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    frameHeight = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    print("frameCount: {}, frameWidth: {}, fpsDiv: {}".format(frameCount, frameWidth, frameHeight))
+
+    if targetWidth:
+        targetHeight = int(targetWidth * frameHeight / frameWidth)
+
+    fc = 0
+    ret = True
+
+    while fc < frameCount and ret:
+        ret, img = cap.read()
+        if fc % fpsDiv == 0:
+            if targetWidth:
+                img = cv2.resize(img, (targetWidth, targetHeight), interpolation=cv2.INTER_AREA)
+            if to_rgb:
+                img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+            yield img
+        fc += 1
+
+    cap.release()