app.py

import gradio as gr

"""
=====================================================
Optical Flow: Predicting movement with the RAFT model
=====================================================

Optical flow is the task of predicting movement between two images, usually two
consecutive frames of a video. Optical flow models take two images as input, and
predict a flow: the flow indicates the displacement of every single pixel in the
first image, and maps it to its corresponding pixel in the second image. Flows
are (2, H, W)-dimensional tensors, where the first axis corresponds to the
predicted horizontal and vertical displacements.

The following example illustrates how torchvision can be used to predict flows
using our implementation of the RAFT model. We will also see how to convert the
predicted flows to RGB images for visualization.
"""

import cv2
import numpy as np
import os
import sys
import torch
import matplotlib.pyplot as plt
import torchvision.transforms.functional as F
from torchvision.io import read_video
from torchvision.models.optical_flow import Raft_Large_Weights
from torchvision.models.optical_flow import raft_large
from torchvision.io import write_jpeg

import tempfile
from pathlib import Path
from urllib.request import urlretrieve

def write_flo(flow, filename):
    """
    Write optical flow in Middlebury .flo format
    
    :param flow: optical flow map
    :param filename: optical flow file path to be saved
    :return: None
    
    from https://github.com/liruoteng/OpticalFlowToolkit/
    
    """
    # forcing conversion to float32 precision
    flow = flow.cpu().data.numpy()
    flow = flow.astype(np.float32)
    f = open(filename, 'wb')
    magic = np.array([202021.25], dtype=np.float32)
    (height, width) = flow.shape[0:2]
    w = np.array([width], dtype=np.int32)
    h = np.array([height], dtype=np.int32)
    magic.tofile(f)
    w.tofile(f)
    h.tofile(f)
    flow.tofile(f)
    f.close()

def get_pixel_value(img, x, y):
    """
    Utility function to get pixel value for coordinate
    vectors x and y from a  4D tensor image.
    Input
    -----
    - img: tensor of shape (B, H, W, C)
    - x: flattened tensor of shape (B*H*W, )
    - y: flattened tensor of shape (B*H*W, )
    Returns
    -------
    - output: tensor of shape (B, H, W, C)
    """
    shape = tf.shape(x)
    batch_size = shape[0]
    height = shape[1]
    width = shape[2]

    batch_idx = tf.range(0, batch_size)
    batch_idx = tf.reshape(batch_idx, (batch_size, 1, 1))
    b = tf.tile(batch_idx, (1, height, width))

    indices = tf.stack([b, y, x], 3)

    return tf.gather_nd(img, indices)

def tf_warp(img, flow, H, W):
#    H = 256
#    W = 256
    x,y = tf.meshgrid(tf.range(W), tf.range(H))
    x = tf.expand_dims(x,0)
    x = tf.expand_dims(x,0)

    y  =tf.expand_dims(y,0)
    y = tf.expand_dims(y,0)

    x = tf.cast(x, tf.float32)
    y = tf.cast(y, tf.float32)
    grid  = tf.concat([x,y],axis = 1)
#    print grid.shape
    flows = grid+flow
    print(flows.shape)
    max_y = tf.cast(H - 1, tf.int32)
    max_x = tf.cast(W - 1, tf.int32)
    zero = tf.zeros([], dtype=tf.int32)

    x = flows[:,0,:,:]
    y = flows[:,1,:,:]
    x0 = x
    y0 = y
    x0 = tf.cast(x0, tf.int32)
    x1 = x0 + 1
    y0 = tf.cast(y0,  tf.int32)
    y1 = y0 + 1

    # clip to range [0, H/W] to not violate img boundaries
    x0 = tf.clip_by_value(x0, zero, max_x)
    x1 = tf.clip_by_value(x1, zero, max_x)
    y0 = tf.clip_by_value(y0, zero, max_y)
    y1 = tf.clip_by_value(y1, zero, max_y)

    # get pixel value at corner coords
    Ia = get_pixel_value(img, x0, y0)
    Ib = get_pixel_value(img, x0, y1)
    Ic = get_pixel_value(img, x1, y0)
    Id = get_pixel_value(img, x1, y1)

    # recast as float for delta calculation
    x0 = tf.cast(x0, tf.float32)
    x1 = tf.cast(x1, tf.float32)
    y0 = tf.cast(y0, tf.float32)
    y1 = tf.cast(y1, tf.float32)


    # calculate deltas
    wa = (x1-x) * (y1-y)
    wb = (x1-x) * (y-y0)
    wc = (x-x0) * (y1-y)
    wd = (x-x0) * (y-y0)

    # add dimension for addition
    wa = tf.expand_dims(wa, axis=3)
    wb = tf.expand_dims(wb, axis=3)
    wc = tf.expand_dims(wc, axis=3)
    wd = tf.expand_dims(wd, axis=3)

    # compute output
    out = tf.add_n([wa*Ia, wb*Ib, wc*Ic, wd*Id])
    return out

def infer():
    video_url = "https://download.pytorch.org/tutorial/pexelscom_pavel_danilyuk_basketball_hd.mp4"
    video_path = Path(tempfile.mkdtemp()) / "basketball.mp4"
    _ = urlretrieve(video_url, video_path)



    
    frames, _, _ = read_video(str(video_path), output_format="TCHW")

    img1_batch = torch.stack([frames[100]])
    img2_batch = torch.stack([frames[101]])

    

    weights = Raft_Large_Weights.DEFAULT
    transforms = weights.transforms()


    def preprocess(img1_batch, img2_batch):
        img1_batch = F.resize(img1_batch, size=[520, 960])
        img2_batch = F.resize(img2_batch, size=[520, 960])
        return transforms(img1_batch, img2_batch)


    img1_batch, img2_batch = preprocess(img1_batch, img2_batch)

    print(f"shape = {img1_batch.shape}, dtype = {img1_batch.dtype}")


####################################
# Estimating Optical flow using RAFT
# ----------------------------------
# We will use our RAFT implementation from
# :func:`~torchvision.models.optical_flow.raft_large`, which follows the same
# architecture as the one described in the `original paper <https://arxiv.org/abs/2003.12039>`_.
# We also provide the :func:`~torchvision.models.optical_flow.raft_small` model
# builder, which is smaller and faster to run, sacrificing a bit of accuracy.

    

# If you can, run this example on a GPU, it will be a lot faster.
    device = "cuda" if torch.cuda.is_available() else "cpu"

    model = raft_large(weights=Raft_Large_Weights.DEFAULT, progress=False).to(device)
    model = model.eval()

    list_of_flows = model(img1_batch.to(device), img2_batch.to(device))
    print(f"type = {type(list_of_flows)}")
    print(f"length = {len(list_of_flows)} = number of iterations of the model")

####################################
# The RAFT model outputs lists of predicted flows where each entry is a
# (N, 2, H, W) batch of predicted flows that corresponds to a given "iteration"
# in the model. For more details on the iterative nature of the model, please
# refer to the `original paper <https://arxiv.org/abs/2003.12039>`_. Here, we
# are only interested in the final predicted flows (they are the most acccurate
# ones), so we will just retrieve the last item in the list.
#
# As described above, a flow is a tensor with dimensions (2, H, W) (or (N, 2, H,
# W) for batches of flows) where each entry corresponds to the horizontal and
# vertical displacement of each pixel from the first image to the second image.
# Note that the predicted flows are in "pixel" unit, they are not normalized
# w.r.t. the dimensions of the images.
    predicted_flows = list_of_flows[-1]
    print(f"dtype = {predicted_flows.dtype}")
    print(f"shape = {predicted_flows.shape} = (N, 2, H, W)")
    print(f"min = {predicted_flows.min()}, max = {predicted_flows.max()}")


####################################
# Visualizing predicted flows
# ---------------------------
# Torchvision provides the :func:`~torchvision.utils.flow_to_image` utlity to
# convert a flow into an RGB image. It also supports batches of flows.
# each "direction" in the flow will be mapped to a given RGB color. In the
# images below, pixels with similar colors are assumed by the model to be moving
# in similar directions. The model is properly able to predict the movement of
# the ball and the player. Note in particular the different predicted direction
# of the ball in the first image (going to the left) and in the second image
# (going up).

    from torchvision.utils import flow_to_image

    #flow_imgs = flow_to_image(predicted_flows)

    #print(flow_imgs)

    predicted_flow = list_of_flows[-1][0]
    flow_img = flow_to_image(predicted_flow).to("cpu")
#     output_folder = "/tmp/"  # Update this to the folder of your choice
    write_jpeg(flow_img, f"predicted_flow.jpg")
    flo_file = write_flo(predicted_flow, "flofile.flo")
    
    with tf.Session() as sess:
        a = tf.placeholder(tf.float32, shape = [None,None,None,3])
        flow_vec = tf.placeholder(tf.float32, shape = [None, 2, None, None])
        init = tf.global_variables_initializer()
        sess.run(init)

        output = tf_warp(a, predicted_flow, 520, 960)
        out = sess.run(output, feed_dict = {a:img, flow_vec:flow})
        out = np.clip(out,0,255).astype('uint8')
#        print out.shape
        im = Image.fromarray(out[0].astype('uint8'))
        im.save('output.jpg')    
    return "done", "predicted_flow.jpg", ["flofile.flo"]
####################################
# Bonus: Creating GIFs of predicted flows
# ---------------------------------------
# In the example above we have only shown the predicted flows of 2 pairs of
# frames. A fun way to apply the Optical Flow models is to run the model on an
# entire video, and create a new video from all the predicted flows. Below is a
# snippet that can get you started with this. We comment out the code, because
# this example is being rendered on a machine without a GPU, and it would take
# too long to run it.

# from torchvision.io import write_jpeg
# for i, (img1, img2) in enumerate(zip(frames, frames[1:])):
#     # Note: it would be faster to predict batches of flows instead of individual flows
#     img1, img2 = preprocess(img1, img2)

#     list_of_flows = model(img1.to(device), img2.to(device))
#     predicted_flow = list_of_flows[-1][0]
#     flow_img = flow_to_image(predicted_flow).to("cpu")
#     output_folder = "/tmp/"  # Update this to the folder of your choice
#     write_jpeg(flow_img, output_folder + f"predicted_flow_{i}.jpg")

####################################
# Once the .jpg flow images are saved, you can convert them into a video or a
# GIF using ffmpeg with e.g.:
#
# ffmpeg -f image2 -framerate 30 -i predicted_flow_%d.jpg -loop -1 flow.gif

      
gr.Interface(fn=infer, inputs=[], outputs=[gr.Textbox(), gr.Image(), gr.Files()]).launch()