General Information¶
This notebook demonstrates how the fastai_sparse library can be used in semantic segmentation tasks using the example of the ShapeNet Core55 3D semantic segmentation solution presented in SparseConvNet example.
Initial data is subset of ShapeNetCore containing about 17,000 models from 16 shape categories Each category is annotated with 2 to 6 parts and there are 50 different parts annotated in total. 3D shapes are represented as point clouds uniformly sampled from 3D surfaces
Evaluation metric: weighted IoU (see https://arxiv.org/pdf/1711.10275.pdf)
Firstly, it is necessary to upload and prepare the initial data. See examples/shapenet_iccv17
Imports¶
In [1]:
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
In [2]:
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import sparseconvnet as scn
import time
import os, sys
import math
import numpy as np
import pandas as pd
import datetime
import glob
from IPython.display import display, HTML, FileLink
from os.path import join, exists, basename, splitext
from pathlib import Path
from matplotlib import pyplot as plt
from matplotlib import cm
#from tensorboardX import SummaryWriter
from joblib import cpu_count
from tqdm import tqdm
#from tqdm import tqdm_notebook as tqdm
#import fastai
# autoreload python modules on the fly when its source is changed
%load_ext autoreload
%autoreload 2
In [3]:
from fastai_sparse import utils
from fastai_sparse.utils import log, log_dict, print_random_states
from fastai_sparse.datasets import find_files
from fastai_sparse.datasets import PointsDataset
#, SparseDataBunch
from fastai_sparse import visualize
from datasets import DataSourceConfig, reader_fn
import transform as T
#from data import merge_fn
In [4]:
assert torch.cuda.is_available()
Experiment environment / system metrics¶
In [5]:
experiment_name = 'unet_24_detailed'
In [6]:
utils.watermark()
virtualenv: (fastai_sparse) python: 3.6.8 nvidia driver: b'384.130' nvidia cuda: 9.0, V9.0.176 cudnn: 7.1.4 torch: 1.0.1.post2 fastai: 1.0.48 fastai_sparse: 0.0.3.dev0
Notebook options¶
In [7]:
utils.wide_notebook()
# uncomment this lines if you want switch off interactive and save visaulisation as screenshoots:
# For rendering run command in terminal: `chromium-browser --remote-debugging-port=9222`
if True:
visualize.options.interactive = False
visualize.options.save_images = True
visualize.options.verbose = True
visualize.options.filename_pattern_image = Path('images', experiment_name, 'fig_{fig_number}')
Sourse¶
Create DataFrames¶
In [8]:
SOURCE_DIR = Path('data').expanduser()
#SOURCE_DIR = Path('/home/ssd/shapenet_data').expanduser()
assert SOURCE_DIR.exists()
DIR_TRAIN_VAL = SOURCE_DIR / 'train_val'
assert DIR_TRAIN_VAL.exists(), "Hint: run `download_and_split_data.sh` then `convert_to_numpy.ipynb`"
print(SOURCE_DIR)
print(DIR_TRAIN_VAL)
data data/train_val
In [9]:
categories = [
"02691156", "02773838", "02954340", "02958343", "03001627", "03261776",
"03467517", "03624134", "03636649", "03642806", "03790512", "03797390",
"03948459", "04099429", "04225987", "04379243"
]
classes = [
'Airplane', 'Bag', 'Cap', 'Car', 'Chair', 'Earphone', 'Guitar', 'Knife',
'Lamp', 'Laptop', 'Motorbike', 'Mug', 'Pistol', 'Rocket', 'Skateboard',
'Table'
]
num_classes_by_category = [4, 2, 2, 4, 4, 3, 3, 2, 4, 2, 6, 2, 3, 3, 3, 3]
assert len(categories) == len(classes)
print("Number of categories:", len(categories))
Number of categories: 16
In [10]:
df_train = find_files(path=SOURCE_DIR / 'npy' / 'train', ext='.points.npy', ext_labels='.labels.npy', categories=categories)
df_valid = find_files(path=SOURCE_DIR / 'npy' / 'valid', ext='.points.npy', ext_labels='.labels.npy', categories=categories)
In [11]:
print(len(df_train))
df_train.head()
6955
Out[11]:
| example_id | subdir | categ_idx | ext | ext_labels | |
|---|---|---|---|---|---|
| 0 | 000908 | 02691156 | 0 | .points.npy | .labels.npy |
| 1 | 010886 | 02691156 | 0 | .points.npy | .labels.npy |
| 2 | 013973 | 02691156 | 0 | .points.npy | .labels.npy |
| 3 | 007190 | 02691156 | 0 | .points.npy | .labels.npy |
| 4 | 010360 | 02691156 | 0 | .points.npy | .labels.npy |
In [12]:
print(len(df_valid))
df_valid.head()
7052
Out[12]:
| example_id | subdir | categ_idx | ext | ext_labels | |
|---|---|---|---|---|---|
| 0 | 005663 | 02691156 | 0 | .points.npy | .labels.npy |
| 1 | 011957 | 02691156 | 0 | .points.npy | .labels.npy |
| 2 | 009038 | 02691156 | 0 | .points.npy | .labels.npy |
| 3 | 009906 | 02691156 | 0 | .points.npy | .labels.npy |
| 4 | 004778 | 02691156 | 0 | .points.npy | .labels.npy |
DataSets config¶
You can create PointsDataset using the configuration.
In [13]:
train_source_config = DataSourceConfig(
root_dir=SOURCE_DIR / 'npy' / 'train',
df=df_train,
batch_size=16,
num_workers=12,
num_classes=50,
num_classes_by_category=num_classes_by_category,
)
train_source_config
Out[13]:
DataSourceConfig; root_dir: data/npy/train batch_size: 16 num_workers: 12 init_numpy_random_seed: True num_classes: 50 num_classes_by_category: [4, 2, 2, 4, 4, 3, 3, 2, 4, 2, 6, 2, 3, 3, 3, 3] class_offsets: [ 0 4 6 8 12 16 19 22 24 28 30 36 38 41 44 47 50] Items count: 6955
In [14]:
valid_source_config = DataSourceConfig(
root_dir=SOURCE_DIR / 'npy' / 'valid',
df=df_valid,
batch_size=16,
num_workers=12,
num_classes=50,
num_classes_by_category=num_classes_by_category,
init_numpy_random_seed=False,
)
valid_source_config
Out[14]:
DataSourceConfig; root_dir: data/npy/valid batch_size: 16 num_workers: 12 init_numpy_random_seed: False num_classes: 50 num_classes_by_category: [4, 2, 2, 4, 4, 3, 3, 2, 4, 2, 6, 2, 3, 3, 3, 3] class_offsets: [ 0 4 6 8 12 16 19 22 24 28 30 36 38 41 44 47 50] Items count: 7052
In [15]:
train_source_config.check_accordance(valid_source_config)
Datasets¶
In [16]:
train_items = PointsDataset.from_source_config(train_source_config, reader_fn=reader_fn)
valid_items = PointsDataset.from_source_config(valid_source_config, reader_fn=reader_fn)
Load file names: 100%|██████████| 6955/6955 [00:01<00:00, 6525.49it/s] Load file names: 100%|██████████| 7052/7052 [00:01<00:00, 6516.53it/s]
In [17]:
train_items.check()
valid_items.check()
#train_items.check_num_classes(max_num_examples=100)
Check files exist: 100%|██████████| 6955/6955 [00:00<00:00, 188045.98it/s] Check files exist: 100%|██████████| 7052/7052 [00:00<00:00, 184664.28it/s]
In [18]:
train_items
Out[18]:
PointsDataset (6955 items)
('000908', n: 2460),('010886', n: 2407),('013973', n: 2463),('007190', n: 2690),('010360', n: 2438)
Path: data/npy/train
In [19]:
o = train_items.get(5)
o
Out[19]:
PointsItem ('002337', n: 2712)
Let's see what we've done with one example.
In [20]:
o.describe()
o.show()
PointsItem (002337) points shape: (2712, 3) dtype: float32 min: -0.35228, max: 0.35657, mean: -0.01345 labels shape: (2712,) dtype: int64 min: 1, max: 4, mean: 1.74410
Saved to file: images/unet_24_detailed/fig_1.png
Out[20]:
Transforms¶
Define transforms¶
In order to reproduce the example of SparseConvNet, the same transformations have been redone, but in the manner of fast.ai transformations.
The following cells define the transformations: preprocessing (PRE_TFMS); augmentation (AUGS_); and transformation to convert the points cloud to a sparse representation (SPARSE_TFMS). Sparse representation is the input format for the SparseConvNet model and contains a list of voxels and their features
In [21]:
resolution = 24
PRE_TFMS = [
T.fit_to_sphere(center=False),
T.shift_labels(offset=-1)
]
AUGS_TRAIN = [
T.rotate(),
T.flip_x(p=0.5),
]
AUGS_VALID = [
T.rotate(),
T.flip_x(p=0.5),
]
SPARSE_TFMS = [
T.translate(offset=2), # segment [-1, 1] ---> segment [1, 3]
T.scale(scale=resolution),
T.merge_features(ones=True),
T.to_sparse_voxels(),
]
Let's see what we got in results of train and valid tranformations for the first example:
initial representation:
In [22]:
o = train_items[-1]
o.describe()
o.show()
PointsItem (012843) points shape: (2640, 3) dtype: float32 min: -0.45134, max: 0.44987, mean: 0.01532 labels shape: (2640,) dtype: int64 min: 48, max: 49, mean: 48.21932
Saved to file: images/unet_24_detailed/fig_2.png
Out[22]:
transformed:
In [23]:
np.random.seed(42)
b = o.copy().apply_tfms(PRE_TFMS + AUGS_TRAIN + SPARSE_TFMS)
b.describe()
b.show()
id: 012843 coords shape: (2640, 3) dtype: int64 min: 27, max: 70, mean: 47.64508 features shape: (2640, 1) dtype: float32 min: 1.00000, max: 1.00000, mean: 1.00000 x shape: (2640,) dtype: int64 min: 27, max: 70, mean: 47.19735 y shape: (2640,) dtype: int64 min: 34, max: 59, mean: 44.41023 z shape: (2640,) dtype: int64 min: 40, max: 59, mean: 51.32765 labels shape: (2640,) dtype: int64 min: 47, max: 48, mean: 47.21932 voxels: 1011 points / voxels: 2.6112759643916914
Saved to file: images/unet_24_detailed/fig_3.png
Out[23]:
Apply transforms to datasets¶
In [24]:
tfms = (
PRE_TFMS + AUGS_TRAIN + SPARSE_TFMS,
PRE_TFMS + AUGS_VALID + SPARSE_TFMS,
)
train_items.transform(tfms[0])
pass
valid_items.transform(tfms[1])
pass
DataBunch¶
In fast.ai the data is represented DataBunch which contains train, valid and optionally test data loaders.
In [25]:
from fastai_sparse.data import SparseDataBunch
from data import merge_fn
In [26]:
data = SparseDataBunch.create(train_ds=train_items,
valid_ds=valid_items,
collate_fn=merge_fn,)
data.describe()
Train: 6955, shuffle: True, batch_size: 16, num_workers: 12, num_batches: 434, drop_last: True Valid: 7052, shuffle: False, batch_size: 16, num_workers: 12, num_batches: 441, drop_last: False
Dataloader idle run speed measurement¶
In [27]:
from fastai_sparse.core import num_cpus
print("num_cpus:", num_cpus())
!lscpu | grep "Model"
print()
#data.describe()
num_cpus: 16 Model name: AMD Ryzen 7 1700 Eight-Core Processor
In [28]:
# train
t = tqdm(enumerate(data.train_dl), total=len(data.train_dl))
for i, batch in t:
pass
100%|██████████| 434/434 [00:04<00:00, 106.94it/s]
In [29]:
# valid
t = tqdm(enumerate(data.valid_dl), total=len(data.valid_dl))
for i, batch in t:
pass
100%|██████████| 441/441 [00:02<00:00, 184.60it/s]
Model¶
U-Net SparseConvNet implemenation (link):
In [30]:
from fastai_sparse.learner import SparseModelConfig
model_config = SparseModelConfig(spatial_size=24 * 8, num_input_features=1)
model_config.check_accordance(data.train_ds.source_config, sparse_item=data.train_ds[0])
model_config
Out[30]:
SparseModelConfig; spatial_size: 192 dimension: 3 block_reps: 1 m: 32 num_planes: [32, 64, 96, 128, 160] residual_blocks: False num_classes: 50 num_input_features: 1 mode: 3 downsample: [2, 2] bias: False
In [31]:
class Model(nn.Module):
def __init__(self, cfg):
nn.Module.__init__(self)
spatial_size = torch.LongTensor([cfg.spatial_size]*3)
self.sparseModel = scn.Sequential(
scn.InputLayer(cfg.dimension, spatial_size, mode=cfg.mode),
scn.SubmanifoldConvolution(cfg.dimension, nIn=cfg.num_input_features, nOut=cfg.m, filter_size=3, bias=cfg.bias),
scn.UNet(cfg.dimension, cfg.block_reps, cfg.num_planes, residual_blocks=cfg.residual_blocks, downsample=cfg.downsample),
scn.BatchNormReLU(cfg.m),
scn.OutputLayer(cfg.dimension),
)
self.linear = nn.Linear(cfg.m, cfg.num_classes)
def forward(self, xb):
coords = xb['coords']
features = xb['features']
x = [coords, features]
x = self.sparseModel(x)
x = self.linear(x)
return x
model = Model(model_config)
In [32]:
model
Out[32]:
Model(
(sparseModel): Sequential(
(0): InputLayer()
(1): SubmanifoldConvolution 1->32 C3
(2): Sequential(
(0): Sequential(
(0): BatchNormLeakyReLU(32,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(1): SubmanifoldConvolution 32->32 C3
)
(1): ConcatTable(
(0): Identity()
(1): Sequential(
(0): BatchNormLeakyReLU(32,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(1): Convolution 32->64 C2/2
(2): Sequential(
(0): Sequential(
(0): BatchNormLeakyReLU(64,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(1): SubmanifoldConvolution 64->64 C3
)
(1): ConcatTable(
(0): Identity()
(1): Sequential(
(0): BatchNormLeakyReLU(64,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(1): Convolution 64->96 C2/2
(2): Sequential(
(0): Sequential(
(0): BatchNormLeakyReLU(96,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(1): SubmanifoldConvolution 96->96 C3
)
(1): ConcatTable(
(0): Identity()
(1): Sequential(
(0): BatchNormLeakyReLU(96,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(1): Convolution 96->128 C2/2
(2): Sequential(
(0): Sequential(
(0): BatchNormLeakyReLU(128,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(1): SubmanifoldConvolution 128->128 C3
)
(1): ConcatTable(
(0): Identity()
(1): Sequential(
(0): BatchNormLeakyReLU(128,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(1): Convolution 128->160 C2/2
(2): Sequential(
(0): Sequential(
(0): BatchNormLeakyReLU(160,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(1): SubmanifoldConvolution 160->160 C3
)
)
(3): BatchNormLeakyReLU(160,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(4): Deconvolution 160->128 C2/2
)
)
(2): JoinTable()
(3): Sequential(
(0): BatchNormLeakyReLU(256,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(1): SubmanifoldConvolution 256->128 C3
)
)
(3): BatchNormLeakyReLU(128,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(4): Deconvolution 128->96 C2/2
)
)
(2): JoinTable()
(3): Sequential(
(0): BatchNormLeakyReLU(192,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(1): SubmanifoldConvolution 192->96 C3
)
)
(3): BatchNormLeakyReLU(96,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(4): Deconvolution 96->64 C2/2
)
)
(2): JoinTable()
(3): Sequential(
(0): BatchNormLeakyReLU(128,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(1): SubmanifoldConvolution 128->64 C3
)
)
(3): BatchNormLeakyReLU(64,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(4): Deconvolution 64->32 C2/2
)
)
(2): JoinTable()
(3): Sequential(
(0): BatchNormLeakyReLU(64,eps=0.0001,momentum=0.9,affine=True,leakiness=0)
(1): SubmanifoldConvolution 64->32 C3
)
)
(3): BatchNormReLU(32,eps=0.0001,momentum=0.9,affine=True)
(4): OutputLayer()
)
(linear): Linear(in_features=32, out_features=50, bias=True)
)
In [33]:
utils.print_trainable_parameters(model)
Total: 3,841,234
| name | number | shape | |
|---|---|---|---|
| 0 | sparseModel.1.weight | 864 | 27 x 1 x 32 |
| 1 | sparseModel.2.0.0.weight | 32 | 32 |
| 2 | sparseModel.2.0.0.bias | 32 | 32 |
| 3 | sparseModel.2.0.1.weight | 27648 | 27 x 32 x 32 |
| 4 | sparseModel.2.1.1.0.weight | 32 | 32 |
| 5 | sparseModel.2.1.1.0.bias | 32 | 32 |
| 6 | sparseModel.2.1.1.1.weight | 16384 | 8 x 32 x 64 |
| 7 | sparseModel.2.1.1.2.0.0.weight | 64 | 64 |
| 8 | sparseModel.2.1.1.2.0.0.bias | 64 | 64 |
| 9 | sparseModel.2.1.1.2.0.1.weight | 110592 | 27 x 64 x 64 |
| 10 | sparseModel.2.1.1.2.1.1.0.weight | 64 | 64 |
| 11 | sparseModel.2.1.1.2.1.1.0.bias | 64 | 64 |
| 12 | sparseModel.2.1.1.2.1.1.1.weight | 49152 | 8 x 64 x 96 |
| 13 | sparseModel.2.1.1.2.1.1.2.0.0.weight | 96 | 96 |
| 14 | sparseModel.2.1.1.2.1.1.2.0.0.bias | 96 | 96 |
| 15 | sparseModel.2.1.1.2.1.1.2.0.1.weight | 248832 | 27 x 96 x 96 |
| 16 | sparseModel.2.1.1.2.1.1.2.1.1.0.weight | 96 | 96 |
| 17 | sparseModel.2.1.1.2.1.1.2.1.1.0.bias | 96 | 96 |
| 18 | sparseModel.2.1.1.2.1.1.2.1.1.1.weight | 98304 | 8 x 96 x 128 |
| 19 | sparseModel.2.1.1.2.1.1.2.1.1.2.0.0.weight | 128 | 128 |
| 20 | sparseModel.2.1.1.2.1.1.2.1.1.2.0.0.bias | 128 | 128 |
| 21 | sparseModel.2.1.1.2.1.1.2.1.1.2.0.1.weight | 442368 | 27 x 128 x 128 |
| 22 | sparseModel.2.1.1.2.1.1.2.1.1.2.1.1.0.weight | 128 | 128 |
| 23 | sparseModel.2.1.1.2.1.1.2.1.1.2.1.1.0.bias | 128 | 128 |
| 24 | sparseModel.2.1.1.2.1.1.2.1.1.2.1.1.1.weight | 163840 | 8 x 128 x 160 |
| 25 | sparseModel.2.1.1.2.1.1.2.1.1.2.1.1.2.0.0.weight | 160 | 160 |
| 26 | sparseModel.2.1.1.2.1.1.2.1.1.2.1.1.2.0.0.bias | 160 | 160 |
| 27 | sparseModel.2.1.1.2.1.1.2.1.1.2.1.1.2.0.1.weight | 691200 | 27 x 160 x 160 |
| 28 | sparseModel.2.1.1.2.1.1.2.1.1.2.1.1.3.weight | 160 | 160 |
| 29 | sparseModel.2.1.1.2.1.1.2.1.1.2.1.1.3.bias | 160 | 160 |
| 30 | sparseModel.2.1.1.2.1.1.2.1.1.2.1.1.4.weight | 163840 | 8 x 160 x 128 |
| 31 | sparseModel.2.1.1.2.1.1.2.1.1.2.3.0.weight | 256 | 256 |
| 32 | sparseModel.2.1.1.2.1.1.2.1.1.2.3.0.bias | 256 | 256 |
| 33 | sparseModel.2.1.1.2.1.1.2.1.1.2.3.1.weight | 884736 | 27 x 256 x 128 |
| 34 | sparseModel.2.1.1.2.1.1.2.1.1.3.weight | 128 | 128 |
| 35 | sparseModel.2.1.1.2.1.1.2.1.1.3.bias | 128 | 128 |
| 36 | sparseModel.2.1.1.2.1.1.2.1.1.4.weight | 98304 | 8 x 128 x 96 |
| 37 | sparseModel.2.1.1.2.1.1.2.3.0.weight | 192 | 192 |
| 38 | sparseModel.2.1.1.2.1.1.2.3.0.bias | 192 | 192 |
| 39 | sparseModel.2.1.1.2.1.1.2.3.1.weight | 497664 | 27 x 192 x 96 |
| 40 | sparseModel.2.1.1.2.1.1.3.weight | 96 | 96 |
| 41 | sparseModel.2.1.1.2.1.1.3.bias | 96 | 96 |
| 42 | sparseModel.2.1.1.2.1.1.4.weight | 49152 | 8 x 96 x 64 |
| 43 | sparseModel.2.1.1.2.3.0.weight | 128 | 128 |
| 44 | sparseModel.2.1.1.2.3.0.bias | 128 | 128 |
| 45 | sparseModel.2.1.1.2.3.1.weight | 221184 | 27 x 128 x 64 |
| 46 | sparseModel.2.1.1.3.weight | 64 | 64 |
| 47 | sparseModel.2.1.1.3.bias | 64 | 64 |
| 48 | sparseModel.2.1.1.4.weight | 16384 | 8 x 64 x 32 |
| 49 | sparseModel.2.3.0.weight | 64 | 64 |
| 50 | sparseModel.2.3.0.bias | 64 | 64 |
| 51 | sparseModel.2.3.1.weight | 55296 | 27 x 64 x 32 |
| 52 | sparseModel.3.weight | 32 | 32 |
| 53 | sparseModel.3.bias | 32 | 32 |
| 54 | linear.weight | 1600 | 50 x 32 |
| 55 | linear.bias | 50 | 50 |
Learner creation¶
Learner is core fast.ai class which contains model architecture, databunch and optimizer options and implement train loop and prediction
In [34]:
from torch import optim
from functools import partial
from fastai.callbacks.general_sched import TrainingPhase, GeneralScheduler
#from fastai.callbacks.csv_logger import CSVLogger
#from fastai.callbacks.tracker import SaveModelCallback
from fastai.callback import annealing_exp
from fastai_sparse.learner import Learner
from fastai_sparse.callbacks import TimeLogger, SaveModelCallback, CSVLogger, CSVLoggerIouByCategory
from metrics import IouByCategories
In [35]:
learn = Learner(data, model,
opt_func=partial(optim.SGD, momentum=0.9),
wd=1e-4,
true_wd=False,
path=str(Path('results', experiment_name)))
Learning Rate finder¶
We use Learning Rate Finder provided by fast.ai library to find the optimal learning rate
In [36]:
%%time
learn.lr_find(start_lr=1e-5, end_lr=100)
learn.recorder.plot()
LR Finder is complete, type {learner_name}.recorder.plot() to see the graph.
CPU times: user 27.9 s, sys: 6.98 s, total: 34.9 s
Wall time: 28.3 s
Train¶
To visualize the learning process, we specify some additional callbacks
In [37]:
learn.callbacks = []
cb_iou = IouByCategories(learn, len(categories))
learn.callbacks.append(cb_iou)
learn.callbacks.append(TimeLogger(learn))
learn.callbacks.append(CSVLogger(learn))
learn.callbacks.append(CSVLoggerIouByCategory(learn, cb_iou, categories_names=classes))
learn.callbacks.append(SaveModelCallback(learn, every='epoch', name='weights', overwrite=True))
In [38]:
learn.fit_annealing_exp(3, lr=0.1, lr_decay=4e-2, momentum=0.9)
Total time: 09:16
| epoch | train_loss | valid_loss | train_time | valid_time | train_waiou | valid_waiou | time |
|---|---|---|---|---|---|---|---|
| 0 | 2.066369 | 1.948893 | 121.718252 | 63.441886 | 0.451465 | 0.518886 | 03:05 |
| 1 | 1.672141 | 1.583488 | 121.557564 | 63.407442 | 0.548420 | 0.600601 | 03:05 |
| 2 | 1.318663 | 1.290712 | 122.697799 | 63.085566 | 0.616172 | 0.590881 | 03:05 |
In [40]:
# learn.fit_annealing_exp(100, lr=0.1, lr_decay=4e-2, momentum=0.9)
In [41]:
learn.find_callback('CSVLoggerIouByCategory').read_logged_file().tail()
Out[41]:
| epoch | datatype | average | Airplane | Bag | Cap | Car | Chair | Earphone | Guitar | Knife | Lamp | Laptop | Motorbike | Mug | Pistol | Rocket | Skateboard | Table | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | 0 | valid | 0.518886 | 0.379600 | 0.492501 | 0.414970 | 0.227367 | 0.509858 | 0.531944 | 0.484199 | 0.445833 | 0.539656 | 0.344094 | 0.231674 | 0.473616 | 0.243760 | 0.283587 | 0.391957 | 0.704618 |
| 2 | 1 | train | 0.548420 | 0.416585 | 0.472786 | 0.366650 | 0.240662 | 0.555655 | 0.513805 | 0.504931 | 0.574740 | 0.525350 | 0.424994 | 0.261024 | 0.484206 | 0.273017 | 0.287875 | 0.414793 | 0.719739 |
| 3 | 1 | valid | 0.600601 | 0.516312 | 0.492734 | 0.398441 | 0.268659 | 0.637381 | 0.516604 | 0.608734 | 0.617883 | 0.553338 | 0.431588 | 0.263243 | 0.545524 | 0.256332 | 0.275717 | 0.462712 | 0.747984 |
| 4 | 2 | train | 0.616172 | 0.543162 | 0.476210 | 0.397841 | 0.248176 | 0.669318 | 0.524780 | 0.638997 | 0.636003 | 0.541066 | 0.411982 | 0.266059 | 0.573489 | 0.281751 | 0.291917 | 0.482149 | 0.755185 |
| 5 | 2 | valid | 0.590881 | 0.523460 | 0.523732 | 0.392183 | 0.228041 | 0.613275 | 0.552095 | 0.596156 | 0.624740 | 0.534418 | 0.383885 | 0.270513 | 0.589842 | 0.294795 | 0.262375 | 0.492768 | 0.742103 |
In [42]:
learn.find_callback('CSVLogger').read_logged_file().tail()
Out[42]:
| epoch | train_loss | valid_loss | train_time | valid_time | train_waiou | valid_waiou | time | |
|---|---|---|---|---|---|---|---|---|
| 0 | 0 | 2.066369 | 1.948893 | 121.718252 | 63.441886 | 0.451465 | 0.518886 | 03:05 |
| 1 | 1 | 1.672141 | 1.583488 | 121.557564 | 63.407442 | 0.548420 | 0.600601 | 03:05 |
| 2 | 2 | 1.318663 | 1.290712 | 122.697799 | 63.085566 | 0.616172 | 0.590881 | 03:05 |
In [43]:
learn.recorder.plot_losses()
In [44]:
learn.recorder.plot_lr()
In [45]:
learn.recorder.plot_metrics()
In [46]:
learn.recorder.plot()
