I used Actnn in Deeplab v3 plus ResNet50, and the it degrade the iou by %2

[hewiew]

[hewiew]

Here is my codes (part):

import actnn from actnn import QModule, QScheme

actnn.set_optimization_level("L3")

def main():

os.environ['CUDA_VISIBLE_DEVICES'] = opts.gpu_id
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print("Device: %s" % device)

# Setup random seed
torch.manual_seed(opts.random_seed)
torch.cuda.manual_seed(opts.random_seed)
# torch.cuda.manual_seed_all(opts.random_seed)
np.random.seed(opts.random_seed)
random.seed(opts.random_seed)

# Setup dataloader
if opts.dataset=='voc' and not opts.crop_val:
    opts.val_batch_size = 1

train_dst, val_dst = get_dataset(opts)
train_loader = data.DataLoader(
    train_dst, batch_size=opts.batch_size, shuffle=True, num_workers=4)
val_loader = data.DataLoader(
    val_dst, batch_size=opts.val_batch_size, shuffle=True, num_workers=4)
print("Dataset: %s, Train set: %d, Val set: %d" %
      (opts.dataset, len(train_dst), len(val_dst)))

step_per_epoch = int(len(train_dst) / opts.batch_size) + 1
if opts.lr_policy == 'cos':
    T_max = int(opts.batch_size * opts.total_itrs / (len(train_dst) + 1e-6)) + 1
    print("lr_policy: cos, T_max: %d" % (T_max))

# Set up model
model_map = {
    'deeplabv3_resnet50': network.deeplabv3_resnet50,
    'deeplabv3plus_resnet50': network.deeplabv3plus_resnet50,
    'deeplabv3_resnet101': network.deeplabv3_resnet101,
    'deeplabv3plus_resnet101': network.deeplabv3plus_resnet101,
    'deeplabv3_mobilenet': network.deeplabv3_mobilenet,
    'deeplabv3plus_mobilenet': network.deeplabv3plus_mobilenet,
    'deeplabv3plus_fuse_mobilenet': network.deeplabv3plus_fuse_mobilenet
}
if opts.norm_layer == 'GN':
    # norm_layer = nn.GroupNorm
    norm_layer = network.GroupNorm32
elif opts.norm_layer == 'FRN':
    norm_layer = network.FilterResponseNorm2d
else:
    norm_layer = None
model = model_map[opts.model](input_channels=opts.input_channels,
                              num_classes=opts.num_classes,
                              output_stride=opts.output_stride,
                              crop_size=opts.crop_size,
                              stochastic_depth_rate=opts.stochastic_depth_rate,
                              norm_layer=norm_layer)

if opts.use_actnn:
    model = actnn.QModule(model)
    # if isinstance(model, QModule):
        # model = model.model

if opts.separable_conv and 'plus' in opts.model:
    network.convert_to_separable_conv(model.model.classifier)
utils.set_bn_momentum(model.model.backbone, momentum=0.01)

# Set up metrics
metrics = StreamSegMetrics(opts.num_classes)

# Set up optimizer
optimizer = torch.optim.SGD(params=[
    {'params': model.model.backbone.parameters(), 'lr': 0.1*opts.lr},
    {'params': model.model.classifier.parameters(), 'lr': opts.lr},
], lr=opts.lr, momentum=0.9, weight_decay=opts.weight_decay)
#optimizer = torch.optim.SGD(params=model.parameters(), lr=opts.lr, momentum=0.9, weight_decay=opts.weight_decay)
#torch.optim.lr_scheduler.StepLR(optimizer, step_size=opts.lr_decay_step, gamma=opts.lr_decay_factor)

if opts.lr_policy=='poly':
    scheduler = utils.PolyLR(optimizer, opts.total_itrs - opts.warm_up_epoch * step_per_epoch, power=0.9)
    scheduler_warmup = GradualWarmupScheduler(optimizer, multiplier=1,
                                              total_epoch=opts.warm_up_epoch * step_per_epoch + 1, after_scheduler=scheduler)
elif opts.lr_policy=='step':
    scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=opts.step_size * (
                opts.total_itrs - opts.warm_up_epoch * step_per_epoch - 1) / opts.step_size, gamma=0.1)
    scheduler_warmup = GradualWarmupScheduler(optimizer, multiplier=1, total_epoch=opts.warm_up_epoch * step_per_epoch + 1,
                                              after_scheduler=scheduler)
elif opts.lr_policy=='cos':
    scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=opts.total_itrs - opts.warm_up_epoch * step_per_epoch)
    scheduler_warmup = GradualWarmupScheduler(optimizer, multiplier=1, total_epoch=opts.warm_up_epoch * step_per_epoch + 1, after_scheduler=scheduler)

after_scheduler=scheduler)

# Set up criterion
#criterion = utils.get_loss(opts.loss_type)
if opts.loss_type == 'focal_loss':
    criterion = utils.FocalLoss(ignore_index=255, size_average=True)
elif opts.loss_type == 'cross_entropy':
    criterion = nn.CrossEntropyLoss(ignore_index=255, reduction='mean')
elif opts.loss_type == 'gce_loss':
    criterion = utils.TruncatedLoss(trainset_size=len(train_dst))
elif opts.loss_type == 'sce_loss':
    criterion = utils.SCELoss()
elif opts.loss_type == 'bitl_loss':
    criterion = utils.BiTemperedLogisticLoss(t1=1.0, t2=1.0, label_smoothing=0.0)
elif opts.loss_type == 'dmi_loss':
    criterion = utils.DMI_loss()

def save_ckpt(path):
    """ save current model
    """
    torch.save({
        "cur_itrs": cur_itrs,
        "model_state": model.state_dict(),
        "optimizer_state": optimizer.state_dict(),
        "scheduler_state": scheduler_warmup.state_dict(),
        "best_score": best_score,
    }, path)
    print("Model saved as %s" % path)

utils.mkdir('checkpoints')
# Restore
best_score = 0.0
best_class_score = {'0': 0, '1': 0}
best_class_recall = {'0': 0, '1': 0}
best_class_precision = {'0': 0, '1': 0}
best_class_f1 = {'0': 0, '1': 0}
cur_itrs = 0
cur_epochs = 0

if opts.ckpt is not None and os.path.isfile(opts.ckpt):
    # https://github.com/VainF/DeepLabV3Plus-Pytorch/issues/8#issuecomment-605601402, @PytaichukBohdan
    checkpoint = torch.load(opts.ckpt, map_location=torch.device('cpu'))
    model.load_state_dict(checkpoint["model_state"])
    model = nn.DataParallel(model)
    model.to(device)
    if opts.continue_training:
        optimizer.load_state_dict(checkpoint["optimizer_state"])
        scheduler_warmup.load_state_dict(checkpoint["scheduler_state"])
        cur_itrs = checkpoint["cur_itrs"]
        best_score = checkpoint['best_score']
        print("Training state restored from %s" % opts.ckpt)
    print("Model restored from %s" % opts.ckpt)
    del checkpoint  # free memory
else:
    print("[!] Retrain")
    model = nn.DataParallel(model)
    model.to(device)

#==========   Train Loop   ==========#
vis_sample_id = np.random.randint(0, len(val_loader), opts.vis_num_samples,
                                  np.int32) if opts.enable_vis else None  # sample idxs for visualization

denorm = utils.Denormalize(mean=opts.mean, std=opts.std)

# denorm = utils.Denormalize(mean = [0.283, 0.284, 0.254], std = [0.181, 0.160, 0.144])  # denormalization for ori images

if opts.test_only:
    model.eval()
    val_score, ret_samples = validate(
        opts=opts, model=model, loader=val_loader, device=device, metrics=metrics, ret_samples_ids=vis_sample_id)
    print(metrics.to_str(val_score))
    return

interval_loss = 0
# train_start_time = time.time()
while True: #cur_itrs < opts.total_itrs:
    # =====  Train  =====
    model.train()
    cur_epochs += 1
    for (images, labels) in train_loader:
        # time1 = time.time()
        if images.shape[0] != opts.batch_size:
            continue
        start_time = time.time()
        cur_itrs += 1

        images = images.to(device, dtype=torch.float32)
        labels = labels.to(device, dtype=torch.long)

        optimizer.zero_grad()
        outputs = model(images)
        loss = criterion(outputs, labels)

        loss.backward()
        optimizer.step()
        scheduler_warmup.step()

        np_loss = loss.detach().cpu().numpy()
        interval_loss += np_loss


        if (cur_itrs) % opts.print_interval == 0:
            if vis is not None:
                vis.vis_scalar('Loss', cur_itrs, np_loss)
                vis.vis_scalar('Learning Rate', cur_itrs, optimizer.param_groups[1]['lr'])
            interval_loss = interval_loss/opts.print_interval
            print("Epoch %d, Itrs %d/%d, Loss=%f, Lr=%.6f [%.4f s]" %
                  (cur_epochs, cur_itrs, opts.total_itrs, interval_loss, optimizer.param_groups[1]['lr'], time.time() - start_time))

            interval_loss = 0.0
        if (cur_itrs) % opts.val_interval == 0:
            save_ckpt('checkpoints/latest_%s_%s_os%d.pth' %
                      (opts.model, opts.dataset, opts.output_stride))
            print("validation...")
            val_start_time = time.time()
            model.eval()
            val_score, ret_samples = validate(
                opts=opts, model=model, loader=val_loader, device=device, metrics=metrics, ret_samples_ids=vis_sample_id)
            print(metrics.to_str(val_score))
            print("Validating[%.4f s]" % (time.time() - val_start_time))

            if val_score['Class IoU'][1] > best_score:  # save best model
                best_score = val_score['Class IoU'][1]
                best_class_score = val_score['Class IoU']
                best_class_recall = val_score['Class Recall']
                best_class_precision = val_score['Class Precision']
                best_class_f1 = val_score['Class F1-score']
                save_ckpt('checkpoints/best_%s_%s_os%d.pth' %
                          (opts.model, opts.dataset,opts.output_stride))

            model.train()

        if cur_itrs >=  opts.total_itrs:
            return

    if vis.env != 'main':
        vis.vis.save(envs=[vis.env])
    

if name == 'main': main()

[cjf00000]

ActNN is a lossy algorithm, so it is possible that it does not work with 2 bits for some models.

Please try using more warmup iterations, and

actnn.set_optimization_level("L2")

[hewiew]

ActNN is a lossy algorithm, so it is possible that it does not work with 2 bits for some models.

Please try using more warmup iterations, and

actnn.set_optimization_level("L2")

I've tried set_optimization_level "L0", "L1", "L2", "L3", but the performance seems no much difference, so I'm confused

as for warmup iterations, i used 4 epoch warmup, is that enough for training?

[cjf00000]

ActNN L0 does exactly the same thing with full precision training. Is the 2% accuracy loss within random error?

[cjf00000]

Could you print(model) before the training loop, and check if the model is correctly converted?

ActNN converts nn.Modules with its own modules, and I noticed there are additional model converters after actnn.QModules. If these converters are looking for the original nn.Modules (e.g., nn.BatchNorm), they may not found the corresponding module. You can try moving actnn.QModule after these converters.

if opts.separable_conv and 'plus' in opts.model: network.convert_to_separable_conv(model.model.classifier) utils.set_bn_momentum(model.model.backbone, momentum=0.01)

[hewiew]

ActNN L0 does exactly the same thing with full precision training. Is the 2% accuracy loss within random error?

I'm afraid not, when I don't use the actnn.module wrapper, I can always get a model with 2% of iou higher than that use it. Incidentally, my model is a 2-class segmentation model, Does this have an impact?

[hewiew]

Could you print(model) before the training loop, and check if the model is correctly converted?

ActNN converts nn.Modules with its own modules, and I noticed there are additional model converters after actnn.QModules. If these converters are looking for the original nn.Modules (e.g., nn.BatchNorm), they may not found the corresponding module. You can try moving actnn.QModule after these converters.

if opts.separable_conv and 'plus' in opts.model: network.convert_to_separable_conv(model.model.classifier) utils.set_bn_momentum(model.model.backbone, momentum=0.01)

the codes you list are always skiped during my training, and I print my model, seem all layers are converted correctly:

DataParallel( (module): QModule( (model): DeepLabV3( (backbone): IntermediateLayerGetter( (conv1): QConv2d(4, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False) (bn1): QBatchNorm2d(64, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() (maxpool): QMaxPool2d(kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), dilation=(1, 1), ceil_mode=False) (layer1): Sequential( (0): Bottleneck( (conv1): QConv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(64, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): QBatchNorm2d(64, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() (downsample): Sequential( (0): QConv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) (1): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) ) ) (1): Bottleneck( (conv1): QConv2d(256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(64, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): QBatchNorm2d(64, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() ) (2): Bottleneck( (conv1): QConv2d(256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(64, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): QBatchNorm2d(64, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() ) ) (layer2): Sequential( (0): Bottleneck( (conv1): QConv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(128, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(128, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False) (bn2): QBatchNorm2d(128, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(512, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() (downsample): Sequential( (0): QConv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False) (1): QBatchNorm2d(512, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) ) ) (1): Bottleneck( (conv1): QConv2d(512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(128, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): QBatchNorm2d(128, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(512, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() ) (2): Bottleneck( (conv1): QConv2d(512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(128, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): QBatchNorm2d(128, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(512, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() ) (3): Bottleneck( (conv1): QConv2d(512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(128, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): QBatchNorm2d(128, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(512, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() ) ) (layer3): Sequential( (0): Bottleneck( (conv1): QConv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False) (bn2): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(1024, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() (downsample): Sequential( (0): QConv2d(512, 1024, kernel_size=(1, 1), stride=(2, 2), bias=False) (1): QBatchNorm2d(1024, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) ) ) (1): Bottleneck( (conv1): QConv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(1024, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() ) (2): Bottleneck( (conv1): QConv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(1024, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() ) (3): Bottleneck( (conv1): QConv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(1024, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() ) (4): Bottleneck( (conv1): QConv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(1024, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() ) (5): Bottleneck( (conv1): QConv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): QBatchNorm2d(256, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(1024, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() ) ) (layer4): Sequential( (0): Bottleneck( (conv1): QConv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(512, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): QBatchNorm2d(512, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(2048, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() (downsample): Sequential( (0): QConv2d(1024, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False) (1): QBatchNorm2d(2048, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) ) ) (1): Bottleneck( (conv1): QConv2d(2048, 512, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(512, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(2, 2), dilation=(2, 2), bias=False) (bn2): QBatchNorm2d(512, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(2048, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() ) (2): Bottleneck( (conv1): QConv2d(2048, 512, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn1): QBatchNorm2d(512, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv2): QConv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(2, 2), dilation=(2, 2), bias=False) (bn2): QBatchNorm2d(512, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (conv3): QConv2d(512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False) (bn3): QBatchNorm2d(2048, eps=1e-05, momentum=0.01, affine=True, track_running_stats=True) (relu): QReLU() ) ) ) (classifier): DeepLabHeadV3Plus( (project): Sequential( (0): QConv2d(256, 48, kernel_size=(1, 1), stride=(1, 1), bias=False) (1): QBatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): QReLU() ) (aspp): ASPP( (convs): ModuleList( (0): Sequential( (0): QConv2d(2048, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) (1): QBatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): QReLU() ) (1): ASPPConv( (0): QConv2d(2048, 256, kernel_size=(3, 3), stride=(1, 1), padding=(2, 2), dilation=(2, 2), bias=False) (1): QBatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): QReLU() ) (2): ASPPConv( (0): QConv2d(2048, 256, kernel_size=(3, 3), stride=(1, 1), padding=(5, 5), dilation=(5, 5), bias=False) (1): QBatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): QReLU() ) (3): ASPPConv( (0): QConv2d(2048, 256, kernel_size=(3, 3), stride=(1, 1), padding=(8, 8), dilation=(8, 8), bias=False) (1): QBatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): QReLU() ) (4): ASPPPooling( (0): AdaptiveAvgPool2d(output_size=1) (1): QConv2d(2048, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) (2): QBatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): QReLU() ) ) (project): Sequential( (0): QConv2d(1280, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) (1): QBatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): QReLU() (3): QDropout(p=0.1, inplace=False) ) ) (classifier): Sequential( (0): QConv2d(304, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (1): QBatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): QReLU() (3): QConv2d(256, 2, kernel_size=(1, 1), stride=(1, 1)) ) ) ) ) )

[cjf00000]

That's strange. ActNN L0 and full precision training should have identical behavior.

Could you try to debug by the following:

1. prepare a model checkpoint, and a batch of (data, label)
2. compute the gradient with full-precision training
3. compute the gradient with actnn
4. check if the gradients are identical, if not, replace actnn layers with nn.module layers one by one, and observe which layer is causing the gradient discrepancy.

If you spot a bug in our implementation, please create a PR for us.

[hewiew]

That's strange. ActNN L0 and full precision training should have identical behavior.

Could you try to debug by the following:

1. prepare a model checkpoint, and a batch of (data, label)

2. compute the gradient with full-precision training

3. compute the gradient with actnn

4. check if the gradients are identical, if not, replace actnn layers with nn.module layers one by one, and observe which layer is causing the gradient discrepancy.

If you spot a bug in our implementation, please create a PR for us.

Thanks for your advise last week. I've followed the steps you list but it seems we are getting into a trouble:

I restored my model (only a nn.Conv2D layer) from a fixed checkpoint, and I found that the gradient with actnn(set level: "L0") and full-precision training are totally different.

I've sent an email to your Gmail mailbox with my experimental code and printed results with details. You may check that during your free time.