File size: 12,116 Bytes
4d4dd90
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
# Copyright (C) 2022-present Naver Corporation. All rights reserved.
# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).

# --------------------------------------------------------
# Main function for training one epoch or testing
# --------------------------------------------------------

import math
import sys
from typing import Iterable
import numpy as np
import torch
import torchvision

from utils import misc as misc


def split_prediction_conf(predictions, with_conf=False):
    if not with_conf:
        return predictions, None
    conf = predictions[:,-1:,:,:]
    predictions = predictions[:,:-1,:,:]
    return predictions, conf

def train_one_epoch(model: torch.nn.Module, criterion: torch.nn.Module, metrics: torch.nn.Module,
                    data_loader: Iterable, optimizer: torch.optim.Optimizer,
                    device: torch.device, epoch: int, loss_scaler,
                    log_writer=None, print_freq = 20,
                    args=None):
    model.train(True)
    metric_logger = misc.MetricLogger(delimiter="  ")
    metric_logger.add_meter('lr', misc.SmoothedValue(window_size=1, fmt='{value:.6f}'))
    header = 'Epoch: [{}]'.format(epoch)

    accum_iter = args.accum_iter

    optimizer.zero_grad()

    details = {}

    if log_writer is not None:
        print('log_dir: {}'.format(log_writer.log_dir))

    if args.img_per_epoch:
        iter_per_epoch = args.img_per_epoch // args.batch_size + int(args.img_per_epoch % args.batch_size > 0)
        assert len(data_loader) >= iter_per_epoch, 'Dataset is too small for so many iterations'
        len_data_loader = iter_per_epoch
    else:
        len_data_loader, iter_per_epoch = len(data_loader), None

    for data_iter_step, (image1, image2, gt, pairname) in enumerate(metric_logger.log_every(data_loader, print_freq, header, max_iter=iter_per_epoch)):
        
        image1 = image1.to(device, non_blocking=True)
        image2 = image2.to(device, non_blocking=True)
        gt = gt.to(device, non_blocking=True)
        
        # we use a per iteration (instead of per epoch) lr scheduler
        if data_iter_step % accum_iter == 0:
            misc.adjust_learning_rate(optimizer, data_iter_step / len_data_loader + epoch, args)

        with torch.cuda.amp.autocast(enabled=bool(args.amp)):
            prediction = model(image1, image2)
            prediction, conf = split_prediction_conf(prediction, criterion.with_conf)
            batch_metrics = metrics(prediction.detach(), gt)
            loss = criterion(prediction, gt) if conf is None else criterion(prediction, gt, conf)
            
        loss_value = loss.item()
        if not math.isfinite(loss_value):
            print("Loss is {}, stopping training".format(loss_value))
            sys.exit(1)

        loss /= accum_iter
        loss_scaler(loss, optimizer, parameters=model.parameters(),
                    update_grad=(data_iter_step + 1) % accum_iter == 0)
        if (data_iter_step + 1) % accum_iter == 0:
            optimizer.zero_grad()

        torch.cuda.synchronize()
        
        metric_logger.update(loss=loss_value)
        for k,v in batch_metrics.items():
            metric_logger.update(**{k: v.item()})
        lr = optimizer.param_groups[0]["lr"]
        metric_logger.update(lr=lr)

        #if args.dsitributed: loss_value_reduce = misc.all_reduce_mean(loss_value)
        time_to_log = ((data_iter_step + 1) % (args.tboard_log_step * accum_iter) == 0 or data_iter_step == len_data_loader-1)
        loss_value_reduce = misc.all_reduce_mean(loss_value)
        if log_writer is not None and time_to_log:
            epoch_1000x = int((data_iter_step / len_data_loader + epoch) * 1000)
            # We use epoch_1000x as the x-axis in tensorboard. This calibrates different curves when batch size changes.
            log_writer.add_scalar('train/loss', loss_value_reduce, epoch_1000x)
            log_writer.add_scalar('lr', lr, epoch_1000x)
            for k,v in batch_metrics.items():
                log_writer.add_scalar('train/'+k, v.item(), epoch_1000x)

    # gather the stats from all processes
    #if args.distributed: metric_logger.synchronize_between_processes()
    print("Averaged stats:", metric_logger)
    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}


@torch.no_grad()
def validate_one_epoch(model: torch.nn.Module,
                   criterion: torch.nn.Module,
                   metrics: torch.nn.Module,
                   data_loaders: list[Iterable],
                   device: torch.device,
                   epoch: int,
                   log_writer=None,
                   args=None):

    model.eval()
    metric_loggers = []
    header = 'Epoch: [{}]'.format(epoch)
    print_freq = 20

    conf_mode = args.tile_conf_mode
    crop = args.crop
    
    if log_writer is not None:
        print('log_dir: {}'.format(log_writer.log_dir))

    results = {}
    dnames = []
    image1, image2, gt, prediction = None, None, None, None
    for didx, data_loader in enumerate(data_loaders):
        dname = str(data_loader.dataset)
        dnames.append(dname)
        metric_loggers.append(misc.MetricLogger(delimiter="  "))
        for data_iter_step, (image1, image2, gt, pairname) in enumerate(metric_loggers[didx].log_every(data_loader, print_freq, header)):
            image1 = image1.to(device, non_blocking=True)
            image2 = image2.to(device, non_blocking=True)
            gt = gt.to(device, non_blocking=True)
            if dname.startswith('Spring'):
                assert gt.size(2)==image1.size(2)*2 and gt.size(3)==image1.size(3)*2
                gt = (gt[:,:,0::2,0::2] + gt[:,:,0::2,1::2] + gt[:,:,1::2,0::2] + gt[:,:,1::2,1::2] ) / 4.0 # we approximate the gt based on the 2x upsampled ones

            with torch.inference_mode():
                prediction, tiled_loss, c = tiled_pred(model, criterion, image1, image2, gt, conf_mode=conf_mode, overlap=args.val_overlap, crop=crop, with_conf=criterion.with_conf)
                batch_metrics = metrics(prediction.detach(), gt)
                loss = criterion(prediction.detach(), gt) if not criterion.with_conf else criterion(prediction.detach(), gt, c)
                loss_value = loss.item()
                metric_loggers[didx].update(loss_tiled=tiled_loss.item())
                metric_loggers[didx].update(**{f'loss': loss_value})
                for k,v in batch_metrics.items():
                    metric_loggers[didx].update(**{dname+'_' + k: v.item()})
        
    results = {k: meter.global_avg for ml in metric_loggers for k, meter in ml.meters.items()}
    if len(dnames)>1:
        for k in batch_metrics.keys():
            results['AVG_'+k] = sum(results[dname+'_'+k] for dname in dnames) / len(dnames)
            
    if log_writer is not None :
        epoch_1000x = int((1 + epoch) * 1000)
        for k,v in results.items():
            log_writer.add_scalar('val/'+k, v, epoch_1000x)

    print("Averaged stats:", results)
    return results

import torch.nn.functional as F
def _resize_img(img, new_size):
    return F.interpolate(img, size=new_size, mode='bicubic', align_corners=False)
def _resize_stereo_or_flow(data, new_size):
    assert data.ndim==4
    assert data.size(1) in [1,2]
    scale_x = new_size[1]/float(data.size(3))
    out = F.interpolate(data, size=new_size, mode='bicubic', align_corners=False)
    out[:,0,:,:] *= scale_x
    if out.size(1)==2:
        scale_y = new_size[0]/float(data.size(2))        
        out[:,1,:,:] *= scale_y
        print(scale_x, new_size, data.shape)
    return out
    

@torch.no_grad()
def tiled_pred(model, criterion, img1, img2, gt,
               overlap=0.5, bad_crop_thr=0.05,
               downscale=False, crop=512, ret='loss',
               conf_mode='conf_expsigmoid_10_5', with_conf=False, 
               return_time=False):
                     
    # for each image, we are going to run inference on many overlapping patches
    # then, all predictions will be weighted-averaged
    if gt is not None:
        B, C, H, W = gt.shape
    else:
        B, _, H, W = img1.shape
        C = model.head.num_channels-int(with_conf)
    win_height, win_width = crop[0], crop[1]
    
    # upscale to be larger than the crop
    do_change_scale =  H<win_height or W<win_width
    if do_change_scale: 
        upscale_factor = max(win_width/W, win_height/W)
        original_size = (H,W)
        new_size = (round(H*upscale_factor),round(W*upscale_factor))
        img1 = _resize_img(img1, new_size)
        img2 = _resize_img(img2, new_size)
        # resize gt just for the computation of tiled losses
        if gt is not None: gt = _resize_stereo_or_flow(gt, new_size)
        H,W = img1.shape[2:4]
        
    if conf_mode.startswith('conf_expsigmoid_'): # conf_expsigmoid_30_10
        beta, betasigmoid = map(float, conf_mode[len('conf_expsigmoid_'):].split('_'))
    elif conf_mode.startswith('conf_expbeta'): # conf_expbeta3
        beta = float(conf_mode[len('conf_expbeta'):])
    else:
        raise NotImplementedError(f"conf_mode {conf_mode} is not implemented")

    def crop_generator():
        for sy in _overlapping(H, win_height, overlap):
          for sx in _overlapping(W, win_width, overlap):
            yield sy, sx, sy, sx, True

    # keep track of weighted sum of prediction*weights and weights
    accu_pred = img1.new_zeros((B, C, H, W)) # accumulate the weighted sum of predictions 
    accu_conf = img1.new_zeros((B, H, W)) + 1e-16 # accumulate the weights 
    accu_c = img1.new_zeros((B, H, W)) # accumulate the weighted sum of confidences ; not so useful except for computing some losses

    tiled_losses = []
    
    if return_time:
        start = torch.cuda.Event(enable_timing=True)
        end = torch.cuda.Event(enable_timing=True)
        start.record()

    for sy1, sx1, sy2, sx2, aligned in crop_generator():
        # compute optical flow there
        pred =  model(_crop(img1,sy1,sx1), _crop(img2,sy2,sx2))
        pred, predconf = split_prediction_conf(pred, with_conf=with_conf)
        
        if gt is not None: gtcrop = _crop(gt,sy1,sx1)
        if criterion is not None and gt is not None: 
            tiled_losses.append( criterion(pred, gtcrop).item() if predconf is None else criterion(pred, gtcrop, predconf).item() )
        
        if conf_mode.startswith('conf_expsigmoid_'):
            conf = torch.exp(- beta * 2 * (torch.sigmoid(predconf / betasigmoid) - 0.5)).view(B,win_height,win_width)
        elif conf_mode.startswith('conf_expbeta'):
            conf = torch.exp(- beta * predconf).view(B,win_height,win_width)
        else:
            raise NotImplementedError
                        
        accu_pred[...,sy1,sx1] += pred * conf[:,None,:,:]
        accu_conf[...,sy1,sx1] += conf
        accu_c[...,sy1,sx1] += predconf.view(B,win_height,win_width) * conf 
        
    pred = accu_pred / accu_conf[:, None,:,:]
    c = accu_c / accu_conf
    assert not torch.any(torch.isnan(pred))

    if return_time:
        end.record()
        torch.cuda.synchronize()
        time = start.elapsed_time(end)/1000.0 # this was in milliseconds

    if do_change_scale:
        pred = _resize_stereo_or_flow(pred, original_size)
    
    if return_time:
        return pred, torch.mean(torch.tensor(tiled_losses)), c, time
    return pred, torch.mean(torch.tensor(tiled_losses)), c


def _overlapping(total, window, overlap=0.5):
    assert total >= window and 0 <= overlap < 1, (total, window, overlap)
    num_windows = 1 + int(np.ceil( (total - window) / ((1-overlap) * window) ))
    offsets = np.linspace(0, total-window, num_windows).round().astype(int)
    yield from (slice(x, x+window) for x in offsets)

def _crop(img, sy, sx):
    B, THREE, H, W = img.shape
    if 0 <= sy.start and sy.stop <= H and 0 <= sx.start and sx.stop <= W:
        return img[:,:,sy,sx]
    l, r = max(0,-sx.start), max(0,sx.stop-W)
    t, b = max(0,-sy.start), max(0,sy.stop-H)
    img = torch.nn.functional.pad(img, (l,r,t,b), mode='constant')
    return img[:, :, slice(sy.start+t,sy.stop+t), slice(sx.start+l,sx.stop+l)]