connecting_the_dots/model/exp_synph_real.py

import torch
import numpy as np
import time
from pathlib import Path
import logging
import sys
import itertools
import json
import matplotlib.pyplot as plt
import cv2
import torchvision.transforms as transforms


import co
import torchext
from model import networks
from data import dataset


class Worker(torchext.Worker):
    def __init__(self, args, num_workers=18, train_batch_size=4, test_batch_size=4, save_frequency=1, **kwargs):
        super().__init__(args.output_dir, args.exp_name, epochs=args.epochs, num_workers=num_workers,
                         train_batch_size=train_batch_size, test_batch_size=test_batch_size,
                         save_frequency=save_frequency, **kwargs)

        self.ms = args.ms
        self.pattern_path = args.pattern_path
        self.lcn_radius = args.lcn_radius
        self.dp_weight = args.dp_weight
        self.data_type = args.data_type

        self.imsizes = [(488, 648)]
        for iter in range(3):
            self.imsizes.append((int(self.imsizes[-1][0] / 2), int(self.imsizes[-1][1] / 2)))

        with open('config.json') as fp:
            config = json.load(fp)
            data_root = Path(config['DATA_ROOT'])
        self.settings_path = data_root / self.data_type / 'settings.pkl'
        sample_paths = sorted((data_root / self.data_type).glob('0*/'))

        self.train_paths = sample_paths[2 ** 10:]
        self.test_paths = sample_paths[:2 ** 8]

        # supervise the edge encoder with only 2**8 samples
        self.train_edge = len(self.train_paths) - 2 ** 8

        self.lcn_in = networks.LCN(self.lcn_radius, 0.05)
        self.disparity_loss = networks.DisparityLoss()
        # self.sup_disp_loss = torch.nn.CrossEntropyLoss()
        self.sup_disp_loss = torch.nn.MSELoss()
        self.edge_loss = torch.nn.BCEWithLogitsLoss(pos_weight=torch.Tensor([0.1]).to(self.train_device))

        # evaluate in the region where opencv Block Matching has valid values
        self.eval_mask = np.zeros(self.imsizes[0])
        self.eval_mask[13:self.imsizes[0][0] - 13, 140:self.imsizes[0][1] - 13] = 1
        self.eval_mask = self.eval_mask.astype(np.bool)
        self.eval_h = self.imsizes[0][0] - 2 * 13
        self.eval_w = self.imsizes[0][1] - 13 - 140

    def get_train_set(self):
        train_set = dataset.TrackSynDataset(self.settings_path, self.train_paths, train=True, data_aug=True,
                                            track_length=1)

        return train_set

    def get_test_sets(self):
        test_sets = torchext.TestSets()
        test_set = dataset.TrackSynDataset(self.settings_path, self.test_paths, train=False, data_aug=True,
                                           track_length=1)
        test_sets.append('simple', test_set, test_frequency=1)

        # initialize photometric loss modules according to image sizes
        self.losses = []
        for imsize, pat in zip(test_set.imsizes, test_set.patterns):
            pat = pat.mean(axis=2)
            pat = torch.from_numpy(pat[None][None].astype(np.float32))
            pat = pat.to(self.train_device)
            self.lcn_in = self.lcn_in.to(self.train_device)
            pat, _ = self.lcn_in(pat)
            pat = torch.cat([pat for idx in range(3)], dim=1)
            self.losses.append(networks.RectifiedPatternSimilarityLoss(imsize[0], imsize[1], pattern=pat))

        return test_sets

    def copy_data(self, data, device, requires_grad, train):
        self.lcn_in = self.lcn_in.to(device)

        self.data = {}
        for key, val in data.items():
            grad = 'im' in key and requires_grad
            self.data[key] = val.to(device).requires_grad_(requires_grad=grad)

            # apply lcn to IR input
            # concatenate the normalized IR input and the original IR image
            if 'im' in key and 'blend' not in key:
                im = self.data[key]
                im_lcn, im_std = self.lcn_in(im)
                im_cat = torch.cat((im_lcn, im), dim=1)
                key_std = key.replace('im', 'std')
                self.data[key] = im_cat
                self.data[key_std] = im_std.to(device).detach()

    def net_forward(self, net, train):
        # FIXME hier schnibbeln?
        out = net(self.data['im0'])
        return out

    @staticmethod
    def find_corr_points_and_F(left, right):
        sift = cv2.SIFT_create()
        # find the keypoints and descriptors with SIFT
        kp1, des1 = sift.detectAndCompute(cv2.normalize(left, None, 0, 255, cv2.NORM_MINMAX).astype('uint8'), None)
        kp2, des2 = sift.detectAndCompute(cv2.normalize(right, None, 0, 255, cv2.NORM_MINMAX).astype('uint8'), None)
        # FLANN parameters
        FLANN_INDEX_KDTREE = 1
        index_params = dict(algorithm=FLANN_INDEX_KDTREE, trees=5)
        search_params = dict(checks=50)
        flann = cv2.FlannBasedMatcher(index_params, search_params)
        matches = flann.knnMatch(des1, des2, k=2)
        pts1 = []
        pts2 = []
        # ratio test as per Lowe's paper
        for i, (m, n) in enumerate(matches):
            if m.distance < 0.8 * n.distance:
                pts2.append(kp2[m.trainIdx].pt)
                pts1.append(kp1[m.queryIdx].pt)

        pts1 = np.int32(pts1)
        pts2 = np.int32(pts2)
        F, mask = cv2.findFundamentalMat(pts1, pts2, cv2.FM_LMEDS)
        # We select only inlier points
        pts1 = pts1[mask.ravel() == 1]
        pts2 = pts2[mask.ravel() == 1]
        return pts1, pts2, F

    def calc_sgbm_gt(self):
        sgbm_matcher = cv2.StereoSGBM_create()
        disp_gt = []
        # cam_view = np.array(np.array_split(self.data['im0'].detach().to('cpu').numpy(), 4)[2:])
        # for i in range(self.data['im0'].shape[0]):
        for i in range(1):
            cam_view = self.data['im0'].detach().to('cpu').numpy()[i, 0]
            pattern = self.pattern_proj.to('cpu').numpy()[i, 0]
            pts_l, pts_r, F = self.find_corr_points_and_F(cam_view, pattern)
            H_l, _ = cv2.findHomography(pts_l, pts_r)
            H_r, _ = cv2.findHomography(pts_r, pts_l)

            left_rect = cv2.warpPerspective(cam_view, H_l, cam_view.shape)
            right_rect = cv2.warpPerspective(pattern, H_r, pattern.shape)

            transform = transforms.ToTensor()
            disparity_gt = transform(cv2.normalize(
                sgbm_matcher.compute(cv2.normalize(left_rect, None, 0, 255, cv2.NORM_MINMAX).astype('uint8'),
                                     cv2.normalize(right_rect, None, 0, 255, cv2.NORM_MINMAX).astype('uint8')), None,
                alpha=0, beta=1, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_32F).T)
            disp_gt.append(disparity_gt)
        return disp_gt

    def loss_forward(self, out, train):
        out, edge = out
        if not (isinstance(out, tuple) or isinstance(out, list)):
            out = [out]
        if not (isinstance(edge, tuple) or isinstance(edge, list)):
            edge = [edge]

        vals = []

        # apply photometric loss
        for s, l, o in zip(itertools.count(), self.losses, out):
            val, pattern_proj = l(o[0], self.data[f'im{s}'][:, 0:1, ...], self.data[f'std{s}'])
            if s == 0:
                self.pattern_proj = pattern_proj.detach()
            vals.append(val)

        # apply disparity loss
        # 1-edge as ground truth edge if inversed
        if isinstance(edge, tuple):
            edge0 = 1 - torch.sigmoid(edge[0][0])
        else:
            edge0 = 1 - torch.sigmoid(edge[0])
        val = 0
        if isinstance(out[0], tuple):
            # val = self.disparity_loss(out[0][1], edge0)
            # FIXME disparity_loss ist unsupervised, wir wollen supervised(?)
            #   warum nicht einfach so die GT die wir eh schon haben?
            # gt = self.data[f'disp0'].type('torch.LongTensor')

            val += self.sup_disp_loss(out[0][1], self.data['disp0'])
            # disp_gt = self.calc_sgbm_gt()
            # if len(disp_gt) > 1:
            #     disparity_gt = torch.stack(disp_gt).to('cuda')
            #     # val += self.sup_disp_loss(out[0][1], disparity_gt)
            # else:
            #     disparity_gt = disp_gt[0].to('cuda')
            #     val += self.sup_disp_loss(out[0][1][0], disparity_gt)
            # print(disparity_gt)
            # print(disparity_gt.shape)
            # print(out[0][1])
            # print(out[0][1].shape)
        if isinstance(out[0], tuple):
            val += self.disparity_loss(out[0][0], edge0)
        else:
            val += self.disparity_loss(out[0], edge0)
        if self.dp_weight > 0:
            vals.append(val * self.dp_weight)

        # apply edge loss on a subset of training samples
        for s, e in zip(itertools.count(), edge):
            # inversed ground truth edge where 0 means edge
            grad = self.data[f'grad{s}'] < 0.2
            grad = grad.to(torch.float32)
            ids = self.data['id']
            mask = ids > self.train_edge
            if mask.sum() > 0:
                if isinstance(e, tuple):
                    val = self.edge_loss(e[0][mask], grad[mask])
                else:
                    val = self.edge_loss(e[mask], grad[mask])
            else:
                val = torch.zeros_like(vals[0])
            if s == 0:
                if isinstance(e, tuple):
                    self.edge = e[0].detach()
                else:
                    self.edge = e.detach()
                self.edge = torch.sigmoid(self.edge)
                self.edge_gt = grad.detach()
            vals.append(val)

        return vals

    def numpy_in_out(self, output):
        output, edge = output
        if not (isinstance(output, tuple) or isinstance(output, list)):
            output = [output]
        es = output[0][0].detach().to('cpu').numpy()
        gt = self.data['disp0'].to('cpu').numpy().astype(np.float32)
        im = self.data['im0'][:, 0:1, ...].detach().to('cpu').numpy()

        ma = gt > 0
        return es, gt, im, ma

    def write_img(self, out_path, es, gt, im, ma):
        logging.info(f'write img {out_path}')
        u_pos, _ = np.meshgrid(range(es.shape[1]), range(es.shape[0]))

        diff = np.abs(es - gt)

        vmin, vmax = np.nanmin(gt), np.nanmax(gt)
        vmin = vmin - 0.2 * (vmax - vmin)
        vmax = vmax + 0.2 * (vmax - vmin)

        pattern_proj = self.pattern_proj.to('cpu').numpy()[0, 0]
        im_orig = self.data['im0'].detach().to('cpu').numpy()[0, 0]
        pattern_diff = np.abs(im_orig - pattern_proj)

        fig = plt.figure(figsize=(16, 16))
        es_ = co.cmap.color_depth_map(es, scale=vmax)
        gt_ = co.cmap.color_depth_map(gt, scale=vmax)
        diff_ = co.cmap.color_error_image(diff, BGR=True)

        # plot disparities, ground truth disparity is shown only for reference
        ax = plt.subplot(3, 3, 1)
        plt.imshow(es_[..., [2, 1, 0]])
        plt.xticks([])
        plt.yticks([])
        ax.set_title(f'Disparity Est. {es.min():.4f}/{es.max():.4f}')
        ax = plt.subplot(3, 3, 2)
        plt.imshow(gt_[..., [2, 1, 0]])
        plt.xticks([])
        plt.yticks([])
        ax.set_title(f'Disparity GT {np.nanmin(gt):.4f}/{np.nanmax(gt):.4f}')
        ax = plt.subplot(3, 3, 3)
        plt.imshow(diff_[..., [2, 1, 0]])
        plt.xticks([])
        plt.yticks([])
        ax.set_title(f'Disparity Err. {diff.mean():.5f}')

        # plot edges
        edge = self.edge.to('cpu').numpy()[0, 0]
        edge_gt = self.edge_gt.to('cpu').numpy()[0, 0]
        edge_err = np.abs(edge - edge_gt)
        ax = plt.subplot(3, 3, 4);
        plt.imshow(edge, cmap='gray');
        plt.xticks([]);
        plt.yticks([]);
        ax.set_title(f'Edge Est. {edge.min():.5f}/{edge.max():.5f}')
        ax = plt.subplot(3, 3, 5);
        plt.imshow(edge_gt, cmap='gray');
        plt.xticks([]);
        plt.yticks([]);
        ax.set_title(f'Edge GT {edge_gt.min():.5f}/{edge_gt.max():.5f}')
        ax = plt.subplot(3, 3, 6);
        plt.imshow(edge_err, cmap='gray');
        plt.xticks([]);
        plt.yticks([]);
        ax.set_title(f'Edge Err. {edge_err.mean():.5f}')

        # plot normalized IR input and warped pattern
        ax = plt.subplot(3, 3, 7);
        plt.imshow(im, vmin=im.min(), vmax=im.max(), cmap='gray');
        plt.xticks([]);
        plt.yticks([]);
        ax.set_title(f'IR input {im.mean():.5f}/{im.std():.5f}')
        ax = plt.subplot(3, 3, 8);
        plt.imshow(pattern_proj, vmin=im.min(), vmax=im.max(), cmap='gray');
        plt.xticks([]);
        plt.yticks([]);
        ax.set_title(f'Warped Pattern {pattern_proj.mean():.5f}/{pattern_proj.std():.5f}')
        im_std = self.data['std0'].to('cpu').numpy()[0, 0]
        ax = plt.subplot(3, 3, 9);
        plt.imshow(im_std, cmap='gray');
        plt.xticks([]);
        plt.yticks([]);
        ax.set_title(f'IR std {im_std.min():.5f}/{im_std.max():.5f}')

        plt.tight_layout()
        plt.savefig(str(out_path))
        plt.close(fig)

    def callback_train_post_backward(self, net, errs, output, epoch, batch_idx, masks=[]):
        if batch_idx % 512 == 0:
            out_path = self.exp_out_root / f'train_{epoch:03d}_{batch_idx:04d}.png'
            es, gt, im, ma = self.numpy_in_out(output)
            self.write_img(out_path, es[0, 0], gt[0, 0], im[0, 0], ma[0, 0])

    def callback_test_start(self, epoch, set_idx):
        self.metric = co.metric.MultipleMetric(
            co.metric.DistanceMetric(vec_length=1),
            co.metric.OutlierFractionMetric(vec_length=1, thresholds=[0.1, 0.5, 1, 2, 5])
        )

    def callback_test_add(self, epoch, set_idx, batch_idx, n_batches, output, masks=[]):
        es, gt, im, ma = self.numpy_in_out(output)

        if batch_idx % 8 == 0:
            out_path = self.exp_out_root / f'test_{epoch:03d}_{batch_idx:04d}.png'
            self.write_img(out_path, es[0, 0], gt[0, 0], im[0, 0], ma[0, 0])

        es, gt, im, ma = self.crop_output(es, gt, im, ma)

        es = es.reshape(-1, 1)
        gt = gt.reshape(-1, 1)
        ma = ma.ravel()
        self.metric.add(es, gt, ma)

    def callback_test_stop(self, epoch, set_idx, loss):
        logging.info(f'{self.metric}')
        for k, v in self.metric.items():
            self.metric_add_test(epoch, set_idx, k, v)

    def crop_output(self, es, gt, im, ma):
        bs = es.shape[0]
        es = np.reshape(es[:, :, self.eval_mask], [bs, 1, self.eval_h, self.eval_w])
        gt = np.reshape(gt[:, :, self.eval_mask], [bs, 1, self.eval_h, self.eval_w])
        im = np.reshape(im[:, :, self.eval_mask], [bs, 1, self.eval_h, self.eval_w])
        ma = np.reshape(ma[:, :, self.eval_mask], [bs, 1, self.eval_h, self.eval_w])
        return es, gt, im, ma


if __name__ == '__main__':
    # FIXME Nicolas fixe idee
    #   SGBM nutzen, um GT zu finden
    #   bei dispnet (oder w/e) letzte paar layer 'dublizieren' (zweiten head bauen) und so mehrere Loss funktionen gleichzeitig trainieren
    #   L1 + L2 und dann im selben Backwardspass optimieren
    #   für das ganze forward pass anpassen
    pass