Closed zrl4836 closed 5 years ago
dbolya
commented
5 years ago This is a pretty weird error so I'll need some more information:
zrl4836
commented
5 years ago 1,The NVIDIA Tesla K80 GPU 2, pytorch 1.0.1,CUDA 8.0 3, coco 4, CUDA_VISIBLE_DEVICES='2,4,6' python train.py --config=yolact_base_config --batch_size=5
When I training again,the bug appears
[ 0] 7080 || B: 6.776 | C: 8.805 | M: 8.457 | S: 2.717 | T: 26.756 || ETA: 7 days, 19:58:55 || timer: 0.834
[ 0] 7090 || B: 15.026 | C: 17.954 | M: 8.922 | S: 3.083 | T: 44.985 || ETA: 7 days, 19:56:06 || timer: 0.836
THCudaCheck FAIL file=/opt/conda/conda-bld/pytorch_1550813258230/work/aten/src/THC/generated/../THCReduceAll.cuh line=317 error=59 : device-side assert triggered
/opt/conda/conda-bld/pytorch_1550813258230/work/aten/src/THCUNN/BCECriterion.cu:57: void bce_updateOutput_no_reduce_functor<Dtype, Acctype>::operator()(const Dtype , const Dt
ype , Dtype ) [with Dtype = float, Acctype = float]/opt/conda/conda-bld/pytorch_1550813258230/work/aten/src/THCUNN/BCECriterion.cu:57: block: [38,0,0: void bceupdateOutput
no_reduce_functor<Dtype, Acctype>::operator()(const Dtype , const Dtype , Dtype ) [with Dtype = float, Acctype = float]: block: [45,0,0], thread: [384,0,0] Assertion *inpu t >= 0. && *input <= 1. failed.
...............................
More info :
Traceback (most recent call last):
File "train.py", line 343, in
And thanks your great job.
dbolya
commented
5 years ago Did you solve the issue or was closing the issue by accident?
If you haven't solved it, could this be because you're using a batch size of 5 with 3 GPUs? It's always best to have the batch size divisible by the number of GPUs. Does it run out of memory with a batch size of 6 or what?
zrl4836
commented
5 years ago I closed it accidentally. I'll test it with batchsize=8 on two GPU right away, and i found a similar mistake about loss in https://github.com/pytorch/pytorch/issues/2209.
dbolya
commented
5 years ago Oh you know what, I know what the problem is. With a batch size of 5 across 3 GPUs, there's too little information to compute batch norm statistics (and thus will output NaN). I wonder if you froze batch norm if it'd work. To do that, add
'freeze_bn': True,to yolact_base_config in data/config.py.
zrl4836
commented
5 years ago Maybe you are right. when I training with batchsize=8 on two GPU, this information is displayed. Warning: Moving average ignored a value of nan Warning: Moving average ignored a value of nan [ 0] 500 || B: 30644214.152 | C: 637367212153369816721387360425607168.000 | M: 59.406 | S: 41098620849.102 | T: 637367212153369816721387360425607168.000 || ETA: 13 days, 12:48:24 || timer: 1.409 Warning: Moving average ignored a value of nan Warning: Moving average ignored a value of nan Warning: Moving average ignored a value of nan Warning: Moving average ignored a value of nan Warning: Moving average ignored a value of nan Warning: Moving average ignored a value of nan Warning: Moving average ignored a value of nan Warning: Moving average ignored a value of nan Warning: Moving average ignored a value of nan Warning: Moving average ignored a value of nan [ 0] 510 || B: 30644214.152 | C: 762162918339897924231348422549962752.000 | M: 69.052 | S: 49108673907.415 | T: 762162918339897924231348422549962752.000 || ETA: 13 days, 12:33:12 || timer: 1.422 Warning: Moving average ignored a value of nan /opt/conda/conda-bld/pytorch_1550813258230/work/aten/src/THCUNN/BCECriterion.cu:57: void bce_updateOutput_no_reduce_functor<Dtype, Acctype>::operator()(const Dtype , const Dt ype , Dtype ) [with Dtype = float, Acctype = float]: block: [33,0,0], thread: [256,0,0] Assertion `input >= 0. && *input <= 1.` failed. OK. I will train again about 'freeze_bn': True,
ridasalam
commented
5 years ago Hi, i had the same problem even after i tried 'freeze_bn': True.
dbolya
commented
5 years ago The loss exploding happens for some initial conditions. It usually only happens within the first epoch, though so after that it's good. Does this happen consistently for you?
ridasalam
commented
5 years ago It actually happened after a few epochs, but not always.
dbolya
commented
5 years ago Hmm I've tried lr warmup before and that seemed to alleviate the early loss explosion issue. I originally didn't consider putting lr warmup in v1.0 because I didn't see it as a big deal, but I'll put it in for v1.1. Hopefully that makes the early stages of training smoother.
ridasalam
commented
5 years ago My batch size is 4, and 1 GPU. I would like to modify the learning rate, decay schedule, and number of iterations but I don't know which parameters to use. I want to fine tune on my custom dataset. Could those also require different warmup?
dbolya
commented
5 years ago Definitely freeze batch norm for a batch size of 4. Then for fine tuning, it depends on the size of your dataset. If it's relatively small (like <10k images), change / add these settings and it should probably be fine (note: I'm pulling these numbers out of my ass so they might take some tuning):
'lr': 5e-4
'lr_steps': (100000, 150000, 175000),
'max_iter': 200000,If you want smoother training to start, enable learning rate warmup with this:
'lr_warmup_init': 1e-6,
'lr_warmup_until': 1000,and you can change this parameters as well (this will start at 1e-6 then warm up to 5e-4 after 1000 iteratons).
andyhahaha
commented
5 years ago You can use torch.clamp(x, 0, 1) before F.binary_cross_entropy. Reference from here https://github.com/NVIDIA/pix2pixHD/issues/9
dbolya
commented
5 years ago I think the problem is more ingrained than that, since the loss often explodes (and then devolves in to NaN) before the assert gets triggered. Nevertheless, I've added a clamp before every F.binary_cross_entropy call. Hopefully this alleviates the issue.
I'll close this issue for now. If this problem persists, feel free to reopen it.
ridasalam
commented
5 years ago Hi, I'm getting the same error again.
I have frozen batch norm and use the lr warm up.
Pytorch 1.1, batch 8, 2 GPUs
pytorch/aten/src/THCUNN/BCECriterion.cu:57: void bce_updateOutput_no_reduce_functor<Dtype, Acctype>::operator()(const Dtype , const Dtype , Dtype ) [with Dtype = float, Acctype = float]: block: [8,0,0], thread: [436,0,0] Assertion `input >= 0. && *input <= 1.` failed.
dbolya
commented
5 years ago Are you on the latest commit? I have a feeling that this is happening because the input is NaN which means there are other problems. What does the loss look like before this error, and are you using a custom dataset?
ridasalam
commented
5 years ago Yes, latest commit. However, in my lastest trial, I so far have not seen that error. Instead:
result = self.forward(*input, **kwargs) File "/hdd_mount/DOCKERSHARE/pytorch/External_Repos/yolact/layers/modules/multibox_loss.py", line 141, in forward pos_idx = pos.unsqueeze(pos.dim()).expand_as(loc_data) RuntimeError: The expanded size of the tensor (19248) must match the existing size (9624) at non-singleton dimension 1. Target sizes: [4, 19248, 4]. Tensor sizes: [4, 9624, 1]
Could it be something to do with using multiple GPUs?
dbolya
commented
5 years ago I'm not entirely sure how that could have happened. Are you using a custom config, or custom dataset? Did you change anything between these two runs? I don't think this issue in particular is caused by multiple GPUs.
ridasalam
commented
5 years ago Im using a custom dataset. I noticed that the tensor size is 1/2 the target size
dbolya
commented
5 years ago I see how this could happen if pytorch changed how dataparallel worked in 1.1, but I just tested this on 2 GPUs using pytorch 1.1 and everything works fine. Can you send me the config you're using (or have you edited the config beyond the two custom dataset lines?)
ridasalam
commented
5 years ago from math import sqrt
import torch
# for making bounding boxes pretty
COLORS = ((244, 67, 54),
(233, 30, 99),
(156, 39, 176),
(103, 58, 183),
( 63, 81, 181),
( 33, 150, 243),
( 3, 169, 244),
( 0, 188, 212),
( 0, 150, 136),
( 76, 175, 80),
(139, 195, 74),
(205, 220, 57),
(255, 235, 59),
(255, 193, 7),
(255, 152, 0),
(255, 87, 34),
(121, 85, 72),
(158, 158, 158),
( 96, 125, 139))
# These are in BGR and are for ImageNet
MEANS = (103.94, 116.78, 123.68)
STD = (57.38, 57.12, 58.40)
COCO_CLASSES = ('person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus',
'train', 'truck', 'boat', 'traffic light', 'fire hydrant',
'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog',
'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra', 'giraffe',
'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat',
'baseball glove', 'skateboard', 'surfboard', 'tennis racket',
'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl',
'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot',
'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop',
'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven',
'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase',
'scissors', 'teddy bear', 'hair drier', 'toothbrush')
COCO_LABEL_MAP = { 1: 1, 2: 2, 3: 3, 4: 4, 5: 5, 6: 6, 7: 7, 8: 8,
9: 9, 10: 10, 11: 11, 13: 12, 14: 13, 15: 14, 16: 15, 17: 16,
18: 17, 19: 18, 20: 19, 21: 20, 22: 21, 23: 22, 24: 23, 25: 24,
27: 25, 28: 26, 31: 27, 32: 28, 33: 29, 34: 30, 35: 31, 36: 32,
37: 33, 38: 34, 39: 35, 40: 36, 41: 37, 42: 38, 43: 39, 44: 40,
46: 41, 47: 42, 48: 43, 49: 44, 50: 45, 51: 46, 52: 47, 53: 48,
54: 49, 55: 50, 56: 51, 57: 52, 58: 53, 59: 54, 60: 55, 61: 56,
62: 57, 63: 58, 64: 59, 65: 60, 67: 61, 70: 62, 72: 63, 73: 64,
74: 65, 75: 66, 76: 67, 77: 68, 78: 69, 79: 70, 80: 71, 81: 72,
82: 73, 84: 74, 85: 75, 86: 76, 87: 77, 88: 78, 89: 79, 90: 80}
# ----------------------- CONFIG CLASS ----------------------- #
class Config(object):
"""
Holds the configuration for anything you want it to.
To get the currently active config, call get_cfg().
To use, just do cfg.x instead of cfg['x'].
I made this because doing cfg['x'] all the time is dumb.
"""
def __init__(self, config_dict):
for key, val in config_dict.items():
self.__setattr__(key, val)
def copy(self, new_config_dict={}):
"""
Copies this config into a new config object, making
the changes given by new_config_dict.
"""
ret = Config(vars(self))
for key, val in new_config_dict.items():
ret.__setattr__(key, val)
return ret
def replace(self, new_config_dict):
"""
Copies new_config_dict into this config object.
Note: new_config_dict can also be a config object.
"""
if isinstance(new_config_dict, Config):
new_config_dict = vars(new_config_dict)
for key, val in new_config_dict.items():
self.__setattr__(key, val)
def print(self):
for k, v in vars(self).items():
print(k, ' = ', v)
# ----------------------- DATASETS ----------------------- #
dataset_base = Config({
'name': 'Base Dataset',
# Training images and annotations
'train_images': './data/coco/images/',
'train_info': 'path_to_annotation_file',
# Validation images and annotations.
'valid_images': './data/coco/images/',
'valid_info': 'path_to_annotation_file',
# Whether or not to load GT. If this is False, eval.py quantitative evaluation won't work.
'has_gt': True,
# A list of names for each of you classes.
'class_names': COCO_CLASSES,
# COCO class ids aren't sequential, so this is a bandage fix. If your ids aren't sequential,
# provide a map from category_id -> index in class_names + 1 (the +1 is there because it's 1-indexed).
# If not specified, this just assumes category ids start at 1 and increase sequentially.
'label_map': None
})
coco2014_dataset = dataset_base.copy({
'name': 'COCO 2014',
'train_info': './data/coco/annotations/instances_train2014.json',
'valid_info': './data/coco/annotations/instances_val2014.json',
'label_map': COCO_LABEL_MAP
})
coco2017_dataset = dataset_base.copy({
'name': 'COCO 2017',
'train_info': './data/coco/annotations/instances_train2017.json',
'valid_info': './data/coco/annotations/instances_val2017.json',
'label_map': COCO_LABEL_MAP
})
coco2017_testdev_dataset = dataset_base.copy({
'name': 'COCO 2017 Test-Dev',
'valid_info': './data/coco/annotations/image_info_test-dev2017.json',
'has_gt': False,
'label_map': COCO_LABEL_MAP
})
MY_dataset = dataset_base.copy({
'name': 'MYDataset',
'train_images': '/hdd_mount/train/',
'train_info': '/hdd_mount/annotations_train_45k.json',
'valid_images': '/hdd_mount/val/',
'valid_info': '/hdd_mount/annotations_val_1k.json',
'has_gt': True,
'class_names': ('person')
})
# ----------------------- TRANSFORMS ----------------------- #
resnet_transform = Config({
'channel_order': 'RGB',
'normalize': True,
'subtract_means': False,
'to_float': False,
})
vgg_transform = Config({
# Note that though vgg is traditionally BGR,
# the channel order of vgg_reducedfc.pth is RGB.
'channel_order': 'RGB',
'normalize': False,
'subtract_means': True,
'to_float': False,
})
darknet_transform = Config({
'channel_order': 'RGB',
'normalize': False,
'subtract_means': False,
'to_float': True,
})
# ----------------------- BACKBONES ----------------------- #
backbone_base = Config({
'name': 'Base Backbone',
'path': 'path/to/pretrained/weights',
'type': object,
'args': tuple(),
'transform': resnet_transform,
'selected_layers': list(),
'pred_scales': list(),
'pred_aspect_ratios': list(),
'use_pixel_scales': False,
'preapply_sqrt': True,
'use_square_anchors': False,
})
resnet101_backbone = backbone_base.copy({
'name': 'ResNet101',
'path': 'resnet101_reducedfc.pth',
'type': ResNetBackbone,
'args': ([3, 4, 23, 3],),
'transform': resnet_transform,
'selected_layers': list(range(2, 8)),
'pred_scales': [[1]]*6,
'pred_aspect_ratios': [ [[0.66685089, 1.7073535, 0.87508774, 1.16524493, 0.49059086]] ] * 6,
})
resnet101_gn_backbone = backbone_base.copy({
'name': 'ResNet101_GN',
'path': 'R-101-GN.pkl',
'type': ResNetBackboneGN,
'args': ([3, 4, 23, 3],),
'transform': resnet_transform,
'selected_layers': list(range(2, 8)),
'pred_scales': [[1]]*6,
'pred_aspect_ratios': [ [[0.66685089, 1.7073535, 0.87508774, 1.16524493, 0.49059086]] ] * 6,
})
resnet50_backbone = resnet101_backbone.copy({
'name': 'ResNet50',
'path': 'resnet50-19c8e357.pth',
'type': ResNetBackbone,
'args': ([3, 4, 6, 3],),
'transform': resnet_transform,
})
darknet53_backbone = backbone_base.copy({
'name': 'DarkNet53',
'path': 'darknet53.pth',
'type': DarkNetBackbone,
'args': ([1, 2, 8, 8, 4],),
'transform': darknet_transform,
'selected_layers': list(range(3, 9)),
'pred_scales': [[3.5, 4.95], [3.6, 4.90], [3.3, 4.02], [2.7, 3.10], [2.1, 2.37], [1.8, 1.92]],
'pred_aspect_ratios': [ [[1, sqrt(2), 1/sqrt(2), sqrt(3), 1/sqrt(3)][:n], [1]] for n in [3, 5, 5, 5, 3, 3] ],
})
vgg16_arch = [[64, 64],
[ 'M', 128, 128],
[ 'M', 256, 256, 256],
[('M', {'kernel_size': 2, 'stride': 2, 'ceil_mode': True}), 512, 512, 512],
[ 'M', 512, 512, 512],
[('M', {'kernel_size': 3, 'stride': 1, 'padding': 1}),
(1024, {'kernel_size': 3, 'padding': 6, 'dilation': 6}),
(1024, {'kernel_size': 1})]]
vgg16_backbone = backbone_base.copy({
'name': 'VGG16',
'path': 'vgg16_reducedfc.pth',
'type': VGGBackbone,
'args': (vgg16_arch, [(256, 2), (128, 2), (128, 1), (128, 1)], [3]),
'transform': vgg_transform,
'selected_layers': [3] + list(range(5, 10)),
'pred_scales': [[5, 4]]*6,
'pred_aspect_ratios': [ [[1], [1, sqrt(2), 1/sqrt(2), sqrt(3), 1/sqrt(3)][:n]] for n in [3, 5, 5, 5, 3, 3] ],
})
# ----------------------- MASK BRANCH TYPES ----------------------- #
mask_type = Config({
# Direct produces masks directly as the output of each pred module.
# This is denoted as fc-mask in the paper.
# Parameters: mask_size, use_gt_bboxes
'direct': 0,
# Lincomb produces coefficients as the output of each pred module then uses those coefficients
# to linearly combine features from a prototype network to create image-sized masks.
# Parameters:
# - masks_to_train (int): Since we're producing (near) full image masks, it'd take too much
# vram to backprop on every single mask. Thus we select only a subset.
# - mask_proto_src (int): The input layer to the mask prototype generation network. This is an
# index in backbone.layers. Use to use the image itself instead.
# - mask_proto_net (list<tuple>): A list of layers in the mask proto network with the last one
# being where the masks are taken from. Each conv layer is in
# the form (num_features, kernel_size, **kwdargs). An empty
# list means to use the source for prototype masks. If the
# kernel_size is negative, this creates a deconv layer instead.
# If the kernel_size is negative and the num_features is None,
# this creates a simple bilinear interpolation layer instead.
# - mask_proto_bias (bool): Whether to include an extra coefficient that corresponds to a proto
# mask of all ones.
# - mask_proto_prototype_activation (func): The activation to apply to each prototype mask.
# - mask_proto_mask_activation (func): After summing the prototype masks with the predicted
# coeffs, what activation to apply to the final mask.
# - mask_proto_coeff_activation (func): The activation to apply to the mask coefficients.
# - mask_proto_crop (bool): If True, crop the mask with the predicted bbox during training.
# - mask_proto_crop_expand (float): If cropping, the percent to expand the cropping bbox by
# in each direction. This is to make the model less reliant
# on perfect bbox predictions.
# - mask_proto_loss (str [l1|disj]): If not None, apply an l1 or disjunctive regularization
# loss directly to the prototype masks.
# - mask_proto_binarize_downsampled_gt (bool): Binarize GT after dowsnampling during training?
# - mask_proto_normalize_mask_loss_by_sqrt_area (bool): Whether to normalize mask loss by sqrt(sum(gt))
# - mask_proto_reweight_mask_loss (bool): Reweight mask loss such that background is divided by
# #background and foreground is divided by #foreground.
# - mask_proto_grid_file (str): The path to the grid file to use with the next option.
# This should be a numpy.dump file with shape [numgrids, h, w]
# where h and w are w.r.t. the mask_proto_src convout.
# - mask_proto_use_grid (bool): Whether to add extra grid features to the proto_net input.
# - mask_proto_coeff_gate (bool): Add an extra set of sigmoided coefficients that is multiplied
# into the predicted coefficients in order to "gate" them.
# - mask_proto_prototypes_as_features (bool): For each prediction module, downsample the prototypes
# to the convout size of that module and supply the prototypes as input
# in addition to the already supplied backbone features.
# - mask_proto_prototypes_as_features_no_grad (bool): If the above is set, don't backprop gradients to
# to the prototypes from the network head.
# - mask_proto_remove_empty_masks (bool): Remove masks that are downsampled to 0 during loss calculations.
# - mask_proto_reweight_coeff (float): The coefficient to multiple the forground pixels with if reweighting.
# - mask_proto_coeff_diversity_loss (bool): Apply coefficient diversity loss on the coefficients so that the same
# instance has similar coefficients.
# - mask_proto_coeff_diversity_alpha (float): The weight to use for the coefficient diversity loss.
# - mask_proto_normalize_emulate_roi_pooling (bool): Normalize the mask loss to emulate roi pooling's affect on loss.
# - mask_proto_double_loss (bool): Whether to use the old loss in addition to any special new losses.
# - mask_proto_double_loss_alpha (float): The alpha to weight the above loss.
'lincomb': 1,
})
# ----------------------- ACTIVATION FUNCTIONS ----------------------- #
activation_func = Config({
'tanh': torch.tanh,
'sigmoid': torch.sigmoid,
'softmax': lambda x: torch.nn.functional.softmax(x, dim=-1),
'relu': lambda x: torch.nn.functional.relu(x, inplace=True),
'none': lambda x: x,
})
# ----------------------- FPN DEFAULTS ----------------------- #
fpn_base = Config({
# The number of features to have in each FPN layer
'num_features': 256,
# The upsampling mode used
'interpolation_mode': 'bilinear',
# The number of extra layers to be produced by downsampling starting at P5
'num_downsample': 1,
# Whether to down sample with a 3x3 stride 2 conv layer instead of just a stride 2 selection
'use_conv_downsample': False,
# Whether to pad the pred layers with 1 on each side (I forgot to add this at the start)
# This is just here for backwards compatibility
'pad': True,
})
# ----------------------- CONFIG DEFAULTS ----------------------- #
coco_base_config = Config({
'dataset': coco2014_dataset,
'num_classes': 81, # This should include the background class
'max_iter': 400000,
# The maximum number of detections for evaluation
'max_num_detections': 100,
# dw' = momentum * dw - lr * (grad + decay * w)
'lr': 1e-3,
'momentum': 0.9,
'decay': 5e-4,
# For each lr step, what to multiply the lr with
'gamma': 0.1,
'lr_steps': (280000, 360000, 400000),
# Initial learning rate to linearly warmup from (if until > 0)
'lr_warmup_init': 1e-4,
# If > 0 then increase the lr linearly from warmup_init to lr each iter for until iters
'lr_warmup_until': 1000,
# The terms to scale the respective loss by
'conf_alpha': 1,
'bbox_alpha': 1.5,
'mask_alpha': 0.4 / 256 * 140 * 140, # Some funky equation. Don't worry about it.
# Eval.py sets this if you just want to run YOLACT as a detector
'eval_mask_branch': True,
# See mask_type for details.
'mask_type': mask_type.direct,
'mask_size': 16,
'masks_to_train': 100,
'mask_proto_src': None,
'mask_proto_net': [(256, 3, {}), (256, 3, {})],
'mask_proto_bias': False,
'mask_proto_prototype_activation': activation_func.relu,
'mask_proto_mask_activation': activation_func.sigmoid,
'mask_proto_coeff_activation': activation_func.tanh,
'mask_proto_crop': True,
'mask_proto_crop_expand': 0,
'mask_proto_loss': None,
'mask_proto_binarize_downsampled_gt': True,
'mask_proto_normalize_mask_loss_by_sqrt_area': False,
'mask_proto_reweight_mask_loss': False,
'mask_proto_grid_file': 'data/grid.npy',
'mask_proto_use_grid': False,
'mask_proto_coeff_gate': False,
'mask_proto_prototypes_as_features': False,
'mask_proto_prototypes_as_features_no_grad': False,
'mask_proto_remove_empty_masks': False,
'mask_proto_reweight_coeff': 1,
'mask_proto_coeff_diversity_loss': False,
'mask_proto_coeff_diversity_alpha': 1,
'mask_proto_normalize_emulate_roi_pooling': False,
'mask_proto_double_loss': False,
'mask_proto_double_loss_alpha': 1,
# SSD data augmentation parameters
# Randomize hue, vibrance, etc.
'augment_photometric_distort': True,
# Have a chance to scale down the image and pad (to emulate smaller detections)
'augment_expand': True,
# Potentialy sample a random crop from the image and put it in a random place
'augment_random_sample_crop': True,
# Mirror the image with a probability of 1/2
'augment_random_mirror': True,
# If using batchnorm anywhere in the backbone, freeze the batchnorm layer during training.
# Note: any additional batch norm layers after the backbone will not be frozen.
'freeze_bn': True,
# Set this to a config object if you want an FPN (inherit from fpn_base). See fpn_base for details.
'fpn': None,
# Use the same weights for each network head
'share_prediction_module': False,
# For hard negative mining, instead of using the negatives that are leastl confidently background,
# use negatives that are most confidently not background.
'ohem_use_most_confident': False,
# Use focal loss as described in https://arxiv.org/pdf/1708.02002.pdf instead of OHEM
'use_focal_loss': False,
'focal_loss_alpha': 0.25,
'focal_loss_gamma': 2,
# The initial bias toward forground objects, as specified in the focal loss paper
'focal_loss_init_pi': 0.01,
# Whether to use sigmoid focal loss instead of softmax, all else being the same.
'use_sigmoid_focal_loss': False,
# Use class[0] to be the objectness score and class[1:] to be the softmax predicted class.
# Note: at the moment this is only implemented if use_focal_loss is on.
'use_objectness_score': False,
# Adds a global pool + fc layer to the smallest selected layer that predicts the existence of each of the 80 classes.
# This branch is only evaluated during training time and is just there for multitask learning.
'use_class_existence_loss': False,
'class_existence_alpha': 1,
# Adds a 1x1 convolution directly to the biggest selected layer that predicts a semantic segmentations for each of the 80 classes.
# This branch is only evaluated during training time and is just there for multitask learning.
'use_semantic_segmentation_loss': False,
'semantic_segmentation_alpha': 1,
# Match gt boxes using the Box2Pix change metric instead of the standard IoU metric.
# Note that the threshold you set for iou_threshold should be negative with this setting on.
'use_change_matching': False,
# Uses the same network format as mask_proto_net, except this time it's for adding extra head layers before the final
# prediction in prediction modules. If this is none, no extra layers will be added.
'extra_head_net': None,
# What params should the final head layers have (the ones that predict box, confidence, and mask coeffs)
'head_layer_params': {'kernel_size': 3, 'padding': 1},
# Add extra layers between the backbone and the network heads
# The order is (bbox, conf, mask)
'extra_layers': (0, 0, 0),
# During training, to match detections with gt, first compute the maximum gt IoU for each prior.
# Then, any of those priors whose maximum overlap is over the positive threshold, mark as positive.
# For any priors whose maximum is less than the negative iou threshold, mark them as negative.
# The rest are neutral and not used in calculating the loss.
'positive_iou_threshold': 0.5,
'negative_iou_threshold': 0.5,
# If less than 1, anchors treated as a negative that have a crowd iou over this threshold with
# the crowd boxes will be treated as a neutral.
'crowd_iou_threshold': 1,
# This is filled in at runtime by Yolact's __init__, so don't touch it
'mask_dim': None,
# Input image size. If preserve_aspect_ratio is False, min_size is ignored.
'min_size': 200,
'max_size': 300,
# Whether or not to do post processing on the cpu at test time
'force_cpu_nms': True,
# Whether to use mask coefficient cosine similarity nms instead of bbox iou nms
'use_coeff_nms': False,
# Whether or not to have a separate branch whose sole purpose is to act as the coefficients for coeff_diversity_loss
# Remember to turn on coeff_diversity_loss, or these extra coefficients won't do anything!
# To see their effect, also remember to turn on use_coeff_nms.
'use_instance_coeff': False,
'num_instance_coeffs': 64,
# Whether or not to tie the mask loss / box loss to 0
'train_masks': True,
'train_boxes': True,
# If enabled, the gt masks will be cropped using the gt bboxes instead of the predicted ones.
# This speeds up training time considerably but results in much worse mAP at test time.
'use_gt_bboxes': False,
# Whether or not to preserve aspect ratio when resizing the image.
# If True, uses the faster r-cnn resizing scheme.
# If False, all images are resized to max_size x max_size
'preserve_aspect_ratio': False,
# Whether or not to use the prediction module (c) from DSSD
'use_prediction_module': False,
# Whether or not to use the predicted coordinate scheme from Yolo v2
'use_yolo_regressors': False,
# For training, bboxes are considered "positive" if their anchors have a 0.5 IoU overlap
# or greater with a ground truth box. If this is true, instead of using the anchor boxes
# for this IoU computation, the matching function will use the predicted bbox coordinates.
# Don't turn this on if you're not using yolo regressors!
'use_prediction_matching': False,
# A list of settings to apply after the specified iteration. Each element of the list should look like
# (iteration, config_dict) where config_dict is a dictionary you'd pass into a config object's init.
'delayed_settings': [],
# Use command-line arguments to set this.
'no_jit': False,
'backbone': None,
'name': 'base_config',
})
# ----------------------- YOLACT v1.0 CONFIGS ----------------------- #
yolact_base_config = coco_base_config.copy({
'name': 'yolact_base',
# Dataset stuff
'dataset': MY_dataset,
'num_classes': 1 + 1,
# Image Size
'max_size': 550,
# Training params
'lr_steps': (170000, 220000, 440000, 660000),
'max_iter': 800000,
'lr_warmup_init': 1e-6,
'lr_warmup_until': 5000,
# Backbone Settings
'backbone': resnet101_backbone.copy({
'selected_layers': list(range(1, 4)),
'use_pixel_scales': True,
'preapply_sqrt': False,
'use_square_anchors': True, # This is for backward compatability with a bug
'pred_aspect_ratios': [ [[1, 1/2, 2]] ]*5,
'pred_scales': [[24], [48], [96], [192], [384]],
}),
# FPN Settings
'fpn': fpn_base.copy({
'use_conv_downsample': True,
'num_downsample': 2,
}),
# Mask Settings
'mask_type': mask_type.lincomb,
'mask_alpha': 6.125,
'mask_proto_src': 0,
'mask_proto_net': [(256, 3, {'padding': 1})] * 3 + [(None, -2, {}), (256, 3, {'padding': 1})] + [(32, 1, {})],
'mask_proto_normalize_emulate_roi_pooling': True,
# Other stuff
'share_prediction_module': True,
'extra_head_net': [(256, 3, {'padding': 1})],
'positive_iou_threshold': 0.5,
'negative_iou_threshold': 0.4,
'crowd_iou_threshold': 0.7,
'use_semantic_segmentation_loss': True,
})
yolact_im400_config = yolact_base_config.copy({
'name': 'yolact_im400',
'max_size': 400,
'backbone': yolact_base_config.backbone.copy({
'pred_scales': [[int(x[0] / yolact_base_config.max_size * 400)] for x in yolact_base_config.backbone.pred_scales],
}),
})
yolact_im700_config = yolact_base_config.copy({
'name': 'yolact_im700',
'masks_to_train': 300,
'max_size': 700,
'backbone': yolact_base_config.backbone.copy({
'pred_scales': [[int(x[0] / yolact_base_config.max_size * 700)] for x in yolact_base_config.backbone.pred_scales],
}),
})
yolact_darknet53_config = yolact_base_config.copy({
'name': 'yolact_darknet53',
'backbone': darknet53_backbone.copy({
'selected_layers': list(range(2, 5)),
'pred_scales': yolact_base_config.backbone.pred_scales,
'pred_aspect_ratios': yolact_base_config.backbone.pred_aspect_ratios,
'use_pixel_scales': True,
'preapply_sqrt': False,
'use_square_anchors': True, # This is for backward compatability with a bug
}),
})
yolact_resnet50_config = yolact_base_config.copy({
'name': 'yolact_resnet50',
'backbone': resnet50_backbone.copy({
'selected_layers': list(range(1, 4)),
'pred_scales': yolact_base_config.backbone.pred_scales,
'pred_aspect_ratios': yolact_base_config.backbone.pred_aspect_ratios,
'use_pixel_scales': True,
'preapply_sqrt': False,
'use_square_anchors': True, # This is for backward compatability with a bug
}),
})
# Default config
cfg = yolact_base_config.copy()
def set_cfg(config_name:str):
""" Sets the active config. Works even if cfg is already imported! """
print('-----setting updated config------')
global cfg
# Note this is not just an eval because I'm lazy, but also because it can
# be used like ssd300_config.copy({'max_size': 400}) for extreme fine-tuning
cfg.replace(eval(config_name))
def set_dataset(dataset_name:str):
""" Sets the dataset of the current config. """
cfg.dataset = eval(dataset_name)
Btw, it worked fine with compute validation map, its the loss function where the problem was triggered
dbolya
commented
5 years ago Not that it would affect this, but your class_names in your dataset is incorrect (to make a tuple of one element, you need to go ('person',) If it's without the comma, it'll be parsed as a parenthesis operator and thus a string so your class list would be ['p', 'e', 'r', ...]).
Actually, can you try running it on only 1 GPU to see if it's a multi GPU problem?
ridasalam
commented
5 years ago Hi, just tested on the single GPU and there's no issue computing validation loss.
dbolya
commented
5 years ago Oh, this is about computing validation loss. Can you give me the code snippet you use do call the validation loss function? You have to make sure the net you're passing in is a dataparallel.
ridasalam
commented
5 years ago Its the exact place where you compute validation mAP
dbolya
commented
5 years ago You mean you call it like:
compute_validation_loss(yolact_net, val_dataloader, criterion)?
I'm not quite sure I fully understand you. Are you saying that you get this error when you compute validation loss, or is does this happen when computing the regular loss? I don't compute validation loss myself, so if you do compute validation loss, you'd have had to add the code to call that yourself, and if you did it in the above way then you need to change yolact_net to net.
[ 0] 3180 || B: 3.273 | C: 6.118 | M: 5.300 | S: 1.431 | T: 16.121 || ETA: 8 days, 0:27:19 || timer: 0.833 [ 0] 3190 || B: 3.251 | C: 6.134 | M: 5.046 | S: 1.343 | T: 15.774 || ETA: 8 days, 0:20:56 || timer: 0.924 [ 0] 3200 || B: 3.220 | C: 6.074 | M: 5.023 | S: 1.346 | T: 15.663 || ETA: 8 days, 0:14:25 || timer: 0.922 [ 0] 3210 || B: 3.249 | C: 6.012 | M: 4.997 | S: 1.397 | T: 15.655 || ETA: 8 days, 0:03:42 || timer: 0.824 [ 0] 3220 || B: 3.167 | C: 5.980 | M: 4.841 | S: 1.368 | T: 15.355 || ETA: 7 days, 23:56:10 || timer: 0.831 /opt/conda/conda-bld/pytorch_1550813258230/work/aten/src/THCUNN/BCECriterion.cu:57: void bce_updateOutput_no_reduce_functor<Dtype, Acctype>::operator()(const Dtype , const Dt ype , Dtype ) [with Dtype = float, Acctype = float]: block: [33,0,0], thread: [192,0,0] Assertion `input >= 0. && input <= 1.` failed. /opt/conda/conda-bld/pytorch_1550813258230/work/aten/src/THCUNN/BCECriterion.cu:57: void bce_updateOutput_no_reduce_functor<Dtype, Acctype>::operator()(const Dtype , const Dt ype , Dtype ) [with Dtype = float, Acctype = float]: block: [33,0,0], thread: [193,0,0] Assertion
*input >= 0. && *input <= 1.THCudaCheck FAIL file=/opt/conda/conda-bld/pyt orch_1550813258230/work/aten/src/THC/generated/../THCReduceAll.cuh line=317 error=59 : device-side assert triggeredfailed.can you help me solve it? Thanks