PyTorch

Debug CUDA error for PyTorch

After I changed my dataset for my code, the training failed:

/tmp/pip-req-build-_tx3iysr/aten/src/ATen/native/cuda/ScatterGatherKernel.cu:310: operator(): block: [0,0,0], thread: [59,0,0] Assertion `idx_dim >= 0 && idx_dim < index_size && "index out of bounds"` failed.
/tmp/pip-req-build-_tx3iysr/aten/src/ATen/native/cuda/ScatterGatherKernel.cu:310: operator(): block: [0,0,0], thread: [60,0,0] Assertion `idx_dim >= 0 && idx_dim < index_size && "index out of bounds"` failed.
/tmp/pip-req-build-_tx3iysr/aten/src/ATen/native/cuda/ScatterGatherKernel.cu:310: operator(): block: [0,0,0], thread: [61,0,0] Assertion `idx_dim >= 0 && idx_dim < index_size && "index out of bounds"` failed.
/tmp/pip-req-build-_tx3iysr/aten/src/ATen/native/cuda/ScatterGatherKernel.cu:310: operator(): block: [0,0,0], thread: [62,0,0] Assertion `idx_dim >= 0 && idx_dim < index_size && "index out of bounds"` failed.
/tmp/pip-req-build-_tx3iysr/aten/src/ATen/native/cuda/ScatterGatherKernel.cu:310: operator(): block: [0,0,0], thread: [63,0,0] Assertion `idx_dim >= 0 && idx_dim < index_size && "index out of bounds"` failed.
Traceback (most recent call last):
  File "train.py", line 337, in <module>
    train(args, train_loader, eval_loader)
  File "train.py", line 189, in train
    sounds = aug(sounds)
  File "/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 881, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/sanbai/birds_sound_classification/utils/augment.py", line 13, in forward
    image = (image - image.mean()) / image.std()
RuntimeError: CUDA error: device-side assert triggered

It’s terribly hard to find out the reason for this common error “RuntimeError: CUDA error: device-side assert triggered”. But someone on Github recommends a method: adding CUDA_LAUNCH_BLOCKING=1 before the program.

Now the real error behind RuntimeError shows up: it’s the wrong number of categories I set to the model.

A struggle to keep the accuracy

In this August, we have got 0.83 evaluation accuracy for DIB-10K dataset. But since last month, we have updated the dataset and the accuracy could only get to 0.82.

The first doubtful point is the Weight Standardization method we used for micro-batch (since the model is too big). So I turned to try gradient-accumulation and use this snippet as an example because it won’t need me to change my code heavily:

model.zero_grad()                                   # Reset gradients tensors
for i, (inputs, labels) in enumerate(training_set):
    predictions = model(inputs)                     # Forward pass
    loss = loss_function(predictions, labels)
    loss = loss / accumulation_steps
    loss.backward()                                 # Backward pass
    if (i+1) % accumulation_steps == 0:
        optimizer.step()
        model.zero_grad()
        if (i+1) % evaluation_steps == 0:
            evaluate_model()

But after changing my code and retrain the model, the accuracy still keep around 0.82:

Epoch     4: reducing learning rate of group 0 to 1.0000e-01.
[2020-12-16 05:53:29] Eval accuracy: 0.8283 | Train accuracy: 0.8187
[2020-12-16 10:01:40] Eval accuracy: 0.8284 | Train accuracy: 0.8938
[2020-12-16 14:11:35] Eval accuracy: 0.8284 | Train accuracy: 0.8313
Epoch     7: reducing learning rate of group 0 to 5.0000e-02.
[2020-12-16 18:21:47] Eval accuracy: 0.8285 | Train accuracy: 0.8750
[2020-12-16 22:31:19] Eval accuracy: 0.8285 | Train accuracy: 0.8313
[2020-12-17 02:41:37] Eval accuracy: 0.8284 | Train accuracy: 0.8625
Epoch    10: reducing learning rate of group 0 to 2.5000e-02.
[2020-12-17 06:52:05] Eval accuracy: 0.8286 | Train accuracy: 0.8500
[2020-12-17 11:02:11] Eval accuracy: 0.8285 | Train accuracy: 0.8063
[2020-12-17 15:12:23] Eval accuracy: 0.8286 | Train accuracy: 0.8375
Epoch    13: reducing learning rate of group 0 to 1.2500e-02.
[2020-12-17 19:22:04] Eval accuracy: 0.8285 | Train accuracy: 0.8313

This makes me really desperate. Maybe I should temporarily put this task aside and go on other works.

TabNet: a new neural-network architecture for tabular data

The neural network seems mostly to be used on Computer Vision and Natural Language Processing scenarios, while tree-models like GBDT are mainly used for tabular data.

But why?

Although this article tries to give an explanation of this, it hasn’t been so promising to me. In my humble opinion, the neural network could finally surpass, or at least be competitive, to the GBDT model.

For example, the paper <TabNet: Attentive Interpretable Tabular Learning> describe a Transformer-like model to simulate the tree-model. The PyTorch implementation is here. I have used it on our own data and it finally reached 90% accuracy ( the accuracy of LightGBM is 93%). In spite of the lower accuracy, this is the first neural model reached 90% accuracy when used on our private data. The author has already done a great job.

The nn.Sigmoid() of PyTorch on Android device

I have trained an EfficientNet model to classify more than ten thousand different categories of birds, by using PyTorch. To run this model on the mobile device, I built a program by learning the PyTorch iOS demo at first, make sure it runs well, and then try to build another Android program by learning the PyTorch Android demo. But after replacing the demo model with my model, the Android program prints out the result as all ‘NaN’.
To figure out the reason step by step. I tested a few widely used deep learning models:

Model Name	Result on Android demo
EfficientNet	Nan
ResNet-50	Normal
ResNet-101	Normal
RegNet+SE	NaN
RegNet	Normal

Seems that the only difference between ‘Nan’ and ‘Normal’ is the Squeeze-Excitation module. But Squeeze-Excitation module is quite simple. It’s just:

        self.f_ex = nn.Sequential(
            nn.Conv2d(w_in, w_se, 1, bias=True),
            nn.ReLU(inplace=cfg.MEM.RELU_INPLACE),
            nn.Conv2d(w_se, w_in, 1, bias=True),
            nn.Sigmoid(),
        )

Aha, I saw the nn.Sigmoid. That’s the only layer which haven’t been used in ResNet.
Let’s dive into Sigmoid function

Looks the $latex e^{-x}$ is the problem: it will became underflow if the X is too big!
Until now, I still have another question: why the ‘NaN’ appears in the Android program but not the iOS program. Perhaps it’s about JVM or the emulator that Android Studio used.

Grab a hands-on realtime-object-detection tool

Try to get a fast (what I mean is detecting in lesss than 1 second on mainstream CPU) object-detection tool from Github, I experiment with some repositories written by PyTorch (because I am familiar with it). Below are some conclusions:
1. detectron2
This the official tool from Facebook Corporation. I download and installed it successfully. The test python code is:

import detectron2
from detectron2.utils.logger import setup_logger
from detectron2.config import get_cfg
from detectron2.engine import DefaultPredictor
from detectron2 import model_zoo
setup_logger()
# import some common libraries
import numpy as np
import cv2
import sys
import time
cfg = get_cfg()
# add project-specific config (e.g., TensorMask) here if you're not running a model in detectron2's core library
cfg.merge_from_file(model_zoo.get_config_file("COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml"))
cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5  # set threshold for this model
# Find a model from detectron2's model zoo. You can use the https://dl.fbaipublicfiles... url as well
cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url("COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml")
cfg.MODEL.DEVICE = "cpu"
predictor = DefaultPredictor(cfg)
img = cv2.imread(sys.argv[1])
begin = time.time()
outputs = predictor(img)
print("time:", time.time() - begin)
print(outputs)

Although can’t recognize all birds in below image, it will cost more than 5 seconds on CPU (my MackbookPro). Performance is not as good as my expectation.

2. efficientdet
From the paper, the EfficientDet should be fast and accurate. But after I wrote a test program, it totally couldn’t recognize the object at all. Then I gave up this solution.
3. EfficientDet.Pytorch
Couldn’t download models from it’s model_zoo.
4. ssd.pytorch
Finally, I came to my sweet ssd(Single Shot Detection). Since have studied it for more than half a year, I wrote below snippet quickly:

def base_transform(image, size, mean):
    x = cv2.resize(image, (size, size)).astype(np.float32)
    x -= mean
    x = x.astype(np.float32)
    return x
class BaseTransform:
    def __init__(self, size, mean):
        self.size = size
        self.mean = np.array(mean, dtype=np.float32)
    def __call__(self, image, boxes=None, labels=None):
        return base_transform(image, self.size, self.mean), boxes, labels
def detect(img, net, transform):
    FONT = cv2.FONT_HERSHEY_SIMPLEX
    COLORS = [(255, 0, 0), (0, 255, 0), (0, 0, 255)]
    height, width = img.shape[:2]
    x = torch.from_numpy(transform(img)[0]).permute(2, 0, 1)
    x = Variable(x.unsqueeze(0))
    y = net(x)  # forward pass
    detections = y.data[0]
    # scale each detection back up to the image
    scale = torch.Tensor([width, height, width, height])
    for index, loc in enumerate(detections[3]):
        score = loc.numpy()[0]
        if score >= 0.5:
            loc = loc[1:]
            pt = loc * scale
            print(score, pt)
            cv2.rectangle(
                img,
                (int(pt[0]), int(pt[1])),
                (int(pt[2]), int(pt[3])),
                COLORS[index % 3],
                2,
            )
            cv2.putText(
                img,
                str(score),
                (int(pt[0]), int(pt[1])),
                FONT,
                1,
                (255, 255, 255),
                1,
                cv2.LINE_AA,
            )
    return img
img = cv2.imread("bird_matrix.jpg")
net = build_ssd("test", 300, 21)  # initialize SSD
net.load_state_dict(torch.load("ssd300_mAP_77.43_v2.pth", map_location="cpu"))
transform = BaseTransform(net.size, (104 / 256.0, 117 / 256.0, 123 / 256.0))
img = detect(img, net, transform)
cv2.imwrite("result.jpg", img)

The result is not perfect but good enough for my current situation.

How to ignore illegal sample of dataset in PyTorch?

I have implemented a dataset class for my image samples. But it can’t handle the situation that a corrupted image has been read:

import torch.utils.data as data
class MyDataset(data.Dataset):
  ...
  def __getitem__(self, index):
    image = cv2.imread(image_list[index])
    if image is None:
      # What should we do?
...

The correct solution is in Pytorch Forum. Therefore I changed my code:

class MyDataset(data.Dataset):
  ...
  def __getitem__(self, index):
    image = cv2.imread(image_list[index])
    if image is None:
      return None
    # Other preprocessing
    ...
def my_collate(batch):
    batch = filter(lambda img: img is not None, batch)
    return data.dataloader.default_collate(list(batch))
dataset = MyDataset()
loader = data.DataLoader(dataset, collate_fn=my_collate)

But it reports:

Loading data exception: Traceback (most recent call last):
  File "/opt/conda/lib/python3.6/site-packages/torch/utils/data/_utils/worker.py", line 108, in _worker_loop
    samples = collate_fn([dataset[i] for i in batch_indices])
  File "train.py", line 197, in my_collate
    return data.dataloader.default_collate(batch)
  File "/opt/conda/lib/python3.6/site-packages/torch/utils/data/_utils/collate.py", line 34, in default_collate
    elem_type = type(batch[0])
TypeError: 'filter' object is not subscriptable

Seems default_collate() couldn’t recognize the ‘filter’ object. Don’t worry. We can just add a small function: list()

def my_collate(batch):
  ...
  return data.dataloader.default_collate(list(batch))

Books I read in year 2019

At the beginning of 2019, I finished the book “The Great Siege: Malta 1565”. The story about a few loyal knights protecting Europe from the Ottoman Empire is so extraordinary that it encouraged me to go on my learning and working about information technology.
To find a new job about Data Engineer or Data Scientist, I almost remembered the whole book of “Hundreds of interviews about machine learning” (Title translated from Chinese). Although I haven’t found a job about machine learning (actually, it’s a job about just damned PHP and Javascript), this book gave me confidence and direction before looking for a new job.
I bought the book “Rats of NIMH” at the end of 2016, and finished reading it after more than two years. In the period, life changed tremendously for me, though I hope the end of it would be as good as the Frisby family.
The most exciting new thing I learned is about NLP in deep learning. After reading the papers about Word2Vec, Transformer, Elmo, BERT, etc. I became very familiar and interesting about NLP.
After started my new job in June 2019, I read the book “Statistical Machine Learning” (Title translated from Chinese) on the commute bus. The bus was very vibrant so I have to read the book for a while and take some rest for my eyes and repeat them. Life is not easy, so I should insist further.

Using Single Shot Detection to detect birds (Episode four)

In the previous article, I reached mAP 0.770 for VOC2007 test.
Four months has past. After trying a lot of interesting ideas from different papers, such as FPN, celu, RFBNet, I finally realised that the data is more important than network structures. Then I use COCO2017+VOC instead of only VOC to train my model. The mAP for VOC2007 test eventually reached 0.797.
But another strange thing happens: there will be a strange big bounding box around the whole image for the 16-birds-image. After using dropout and changing augmentation policies, the strange big box still existed.
I doubt that COCO2017 dataset for birds is not general enough. Therefore I decided to use a more abundant dataset — Open Images Dataset V5. After retrieving all bird images from Open Images Dataset V5, I get 18525 images with corresponding annotations. By using them for training, I finally got a more promising bird detection result for that 16-birds-image (by using threshold 0.65):

Seems these bird images in Open Images Dataset V5 are more general than COCO2017. But the mAP of COCO evaluation is smaller for the model trained by Open Images than a model trained by COCO2017. So it looks like I need a more comprehensive evaluation metrics now.

Accelerate the speed of data loading in PyTorch

I got a desktop computer to train deep learning model last week. The GPU is GTX1050TI with 4GB memory which is enough for basic training on object detection. But the CPU is too old. Therefore when I run the training process, the idle of CPU is 0%. I need to reduce the burden of it.
I tried DALI from Nvidia. Although confessed that it is powerful, I also noticed that DALI is too specific to be used for customer dataset. For example, if I want to use complicate label structures more than just ‘bounding box’ coordinates, I can’t find any code example to in DALI to meet this requirement. By the way, the GPU memory in my computer is not big enough, so if moved computation burden from CPU to GPU, I would have to reduce batch size for training. That’s not a good option too.
Yesterday, from this post, I saw this suggestion:

You can use jpeg4py, library dedicated to encode big jpeg files much faster than PIL. Just read image using this library, then transform it to PIL.

After changed my code from using ‘cv.imread()’ to jpeg4py.JPEG().decode()’, the average training time for 1000 batches in my mode boosted improved 700 seconds to 670 seconds. This is just the simple and useful solution I need.

The uneasy way to implement SSDLite by myself

SSDLite is a variant of Single Shot Multi-box Detection. It uses MobileNetV2 instead of VGG as backbone. Thus it can make detection extremely fast. I was trying to implement SSDLite from the code base of ssd.pytorch. Although it’s not a easy work, I finally learn a lot from the entire process.
First, I just replace VGG with MobileNetV2 in the code. However the loss will stop to reduce after a while of training. Without knowing the reason, I have to compare my code with another open source project ssds.pytorch, to try to find the cause.
Very soon, I noticed that unlike VGG backbone, which built detection framework from 38×38 feature map, the MobileNetV2 use 19×19 feature map as its first detection layer.

“For MobileNetV1, we follow the setup in [33]. For MobileNetV2, the first layer of SSDLite is attached to the expansion of layer 15 (with output stride of 16).”
From: MobileNetV2: Inverted Residuals and Linear Bottlenecks

After changed my code as the description of this paper, the loss still couldn’t reduce in training.
In the next three weeks, I tried a lot of methods: change the aspect ratios, use SGDR to replace SGD, change the number of default boxes, even modifying the structure of neural network to be identical to to ssds.pytorch. But none of them solves the problem. There is another weird phenomenon: when I run prediction on my model, it usually gives random output for detection.
Just until last week, I noticed that my model size is about 10MB but the ssds.pytorch’s is 18MB. Why do they have different model size if their models is exactly the same? Through this clue, I eventually get the cause: a large part of my model hasn’t been back-propagated at all!
My old code only implements the forward() of MobileNetV2 which is not enough for the whole model. Therefore I add nn.ModuleList() to build model from a list of layers, as this patch:

self.backbone = nn.ModuleList(mobilenet.MobileNetV2(num_classes=num_classes, width_mult=1.0).features)

Only the nn.ModuleList() will take all layers into back-propagation process and keep them as model weights. Otherwise, the weights will be randomly init and just use for forwarding — that’s why I get random output before.
I think I should be more carefully when adding FPN into my model in the future.