From: https://www.kaggle.com/mathormad/resnet50-v2-keras-focal-loss-mix-up
Author: KeepLearning
Score: 0.56
# This Python 3 environment comes with many helpful analytics libraries installed
# It is defined by the kaggle/python docker image: https://github.com/kaggle/docker-python
# For example, here's several helpful packages to load in
import numpy as np # linear algebra
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)
# Input data files are available in the "../input/" directory.
# For example, running this (by clicking run or pressing Shift+Enter) will list the files in the input directory
import os
print(os.listdir("../input"))
# Any results you write to the current directory are saved as output.
import os, sys
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import skimage.io
from skimage.transform import resize
from imgaug import augmenters as iaa
from tqdm import tqdm
import PIL
from PIL import Image, ImageOps
import cv2
from sklearn.utils import class_weight, shuffle
from keras.losses import binary_crossentropy
from keras.applications.resnet50 import preprocess_input
import keras.backend as K
import tensorflow as tf
from sklearn.metrics import f1_score, fbeta_score
from keras.utils import Sequence
WORKERS = 2
CHANNEL = 3
import warnings
warnings.filterwarnings("ignore")
SIZE = 300
NUM_CLASSES = 1103
beta_f2=2
from keras.legacy import interfaces
from keras.optimizers import Optimizer
from keras import backend as K
class AdamAccumulate_v1(Optimizer):
def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999,
epsilon=None, decay=0., amsgrad=False, accum_iters=20, **kwargs):
super(AdamAccumulate, self).__init__(**kwargs)
with K.name_scope(self.__class__.__name__):
self.iterations = K.variable(0, dtype='int64', name='iterations')
self.effective_iterations = K.variable(0, dtype='int64', name='effective_iterations')
self.lr = K.variable(lr, name='lr')
self.beta_1 = K.variable(beta_1, name='beta_1')
self.beta_2 = K.variable(beta_2, name='beta_2')
self.decay = K.variable(decay, name='decay')
if epsilon is None:
epsilon = K.epsilon()
self.epsilon = epsilon
self.initial_decay = decay
self.amsgrad = amsgrad
self.accum_iters = K.variable(accum_iters, dtype='int64')
@interfaces.legacy_get_updates_support
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [K.update(self.iterations, self.iterations + 1)]
flag = K.equal(self.iterations % self.accum_iters, self.accum_iters - 1)
flag = K.cast(flag, K.floatx())
self.updates.append(K.update(self.effective_iterations,
self.effective_iterations + K.cast(flag, 'int64')))
lr = self.lr
if self.initial_decay > 0:
lr = lr * (1. / (1. + self.decay * K.cast(self.effective_iterations,
K.dtype(self.decay))))
t = K.cast(self.effective_iterations, K.floatx()) + 1
lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
(1. - K.pow(self.beta_1, t)))
ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
gs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
if self.amsgrad:
vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
else:
vhats = [K.zeros(1) for _ in params]
self.weights = [self.iterations] + ms + vs + vhats
for p, g, m, v, vhat, gg in zip(params, grads, ms, vs, vhats, gs):
gg_t = (1 - flag) * (gg + g)
m_t = (self.beta_1 * m) + (1. - self.beta_1) * (gg + flag * g) / K.cast(self.accum_iters, K.floatx())
v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(
(gg + flag * g) / K.cast(self.accum_iters, K.floatx()))
if self.amsgrad:
vhat_t = K.maximum(vhat, v_t)
p_t = p - flag * lr_t * m_t / (K.sqrt(vhat_t) + self.epsilon)
self.updates.append(K.update(vhat, vhat_t))
else:
p_t = p - flag * lr_t * m_t / (K.sqrt(v_t) + self.epsilon)
self.updates.append((m, flag * m_t + (1 - flag) * m))
self.updates.append((v, flag * v_t + (1 - flag) * v))
self.updates.append((gg, gg_t))
new_p = p_t
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
return self.updates
def get_config(self):
config = {'lr': float(K.get_value(self.lr)),
'beta_1': float(K.get_value(self.beta_1)),
'beta_2': float(K.get_value(self.beta_2)),
'decay': float(K.get_value(self.decay)),
'epsilon': self.epsilon,
'amsgrad': self.amsgrad}
base_config = super(AdamAccumulate, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
class AdamAccumulate(Optimizer):
def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999,
epsilon=None, decay=0., amsgrad=False, accum_iters=1, **kwargs):
if accum_iters < 1:
raise ValueError('accum_iters must be >= 1')
super(AdamAccumulate, self).__init__(**kwargs)
with K.name_scope(self.__class__.__name__):
self.iterations = K.variable(0, dtype='int64', name='iterations')
self.lr = K.variable(lr, name='lr')
self.beta_1 = K.variable(beta_1, name='beta_1')
self.beta_2 = K.variable(beta_2, name='beta_2')
self.decay = K.variable(decay, name='decay')
if epsilon is None:
epsilon = K.epsilon()
self.epsilon = epsilon
self.initial_decay = decay
self.amsgrad = amsgrad
self.accum_iters = K.variable(accum_iters, K.dtype(self.iterations))
self.accum_iters_float = K.cast(self.accum_iters, K.floatx())
@interfaces.legacy_get_updates_support
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [K.update_add(self.iterations, 1)]
lr = self.lr
completed_updates = K.cast(K.tf.floor(self.iterations / self.accum_iters), K.floatx())
if self.initial_decay > 0:
lr = lr * (1. / (1. + self.decay * completed_updates))
t = completed_updates + 1
lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) / (1. - K.pow(self.beta_1, t)))
# self.iterations incremented after processing a batch
# batch: 1 2 3 4 5 6 7 8 9
# self.iterations: 0 1 2 3 4 5 6 7 8
# update_switch = 1: x x (if accum_iters=4)
update_switch = K.equal((self.iterations + 1) % self.accum_iters, 0)
update_switch = K.cast(update_switch, K.floatx())
ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
gs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
if self.amsgrad:
vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
else:
vhats = [K.zeros(1) for _ in params]
self.weights = [self.iterations] + ms + vs + vhats
for p, g, m, v, vhat, tg in zip(params, grads, ms, vs, vhats, gs):
sum_grad = tg + g
avg_grad = sum_grad / self.accum_iters_float
m_t = (self.beta_1 * m) + (1. - self.beta_1) * avg_grad
v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(avg_grad)
if self.amsgrad:
vhat_t = K.maximum(vhat, v_t)
p_t = p - lr_t * m_t / (K.sqrt(vhat_t) + self.epsilon)
self.updates.append(K.update(vhat, (1 - update_switch) * vhat + update_switch * vhat_t))
else:
p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)
self.updates.append(K.update(m, (1 - update_switch) * m + update_switch * m_t))
self.updates.append(K.update(v, (1 - update_switch) * v + update_switch * v_t))
self.updates.append(K.update(tg, (1 - update_switch) * sum_grad))
new_p = p_t
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, (1 - update_switch) * p + update_switch * new_p))
return self.updates
def get_config(self):
config = {'lr': float(K.get_value(self.lr)),
'beta_1': float(K.get_value(self.beta_1)),
'beta_2': float(K.get_value(self.beta_2)),
'decay': float(K.get_value(self.decay)),
'epsilon': self.epsilon,
'amsgrad': self.amsgrad}
base_config = super(AdamAccumulate, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
# Load dataset info
path_to_train = '../input/imet-2019-fgvc6/train/'
data = pd.read_csv('../input/imet-2019-fgvc6/train.csv')
train_dataset_info = []
for name, labels in zip(data['id'], data['attribute_ids'].str.split(' ')):
train_dataset_info.append({
'path':os.path.join(path_to_train, name),
'labels':np.array([int(label) for label in labels])})
train_dataset_info = np.array(train_dataset_info)
gamma = 2.0
epsilon = K.epsilon()
def focal_loss(y_true, y_pred):
pt = y_pred * y_true + (1-y_pred) * (1-y_true)
pt = K.clip(pt, epsilon, 1-epsilon)
CE = -K.log(pt)
FL = K.pow(1-pt, gamma) * CE
loss = K.sum(FL, axis=1)
return loss
sometimes = lambda aug: iaa.Sometimes(0.5, aug)
class data_generator(Sequence):
def mix_up(x, y):
x = np.array(x, np.float32)
lam = np.random.beta(0.2, 0.4)
ori_index = np.arange(int(len(x)))
index_array = np.arange(int(len(x)))
np.random.shuffle(index_array)
mixed_x = lam * x[ori_index] + (1 - lam) * x[index_array]
mixed_y = lam * y[ori_index] + (1 - lam) * y[index_array]
return mixed_x, mixed_y
def create_train(dataset_info, batch_size, shape, augument=True, mix=False):
assert shape[2] == 3
while True:
dataset_info = shuffle(dataset_info)
for start in range(0, len(dataset_info), batch_size):
end = min(start + batch_size, len(dataset_info))
batch_images = []
X_train_batch = dataset_info[start:end]
batch_labels = np.zeros((len(X_train_batch), NUM_CLASSES))
for i in range(len(X_train_batch)):
image = data_generator.load_image(
X_train_batch[i]['path'], augument)
batch_images.append(preprocess_input(image))
batch_labels[i][X_train_batch[i]['labels']] = 1
if(mix):
batch_images, batch_labels = data_generator.mix_up(batch_images, batch_labels)
yield np.array(batch_images, np.float32), batch_labels
def create_valid(dataset_info, batch_size, shape, augument=False):
assert shape[2] == 3
while True:
# dataset_info = shuffle(dataset_info)
for start in range(0, len(dataset_info), batch_size):
end = min(start + batch_size, len(dataset_info))
batch_images = []
X_train_batch = dataset_info[start:end]
batch_labels = np.zeros((len(X_train_batch), NUM_CLASSES))
for i in range(len(X_train_batch)):
image = data_generator.load_image(
X_train_batch[i]['path'], augument)
batch_images.append(preprocess_input(image))
batch_labels[i][X_train_batch[i]['labels']] = 1
yield np.array(batch_images, np.float32), batch_labels
def load_image(path, augument=False):
image = cv2.imread(path+'.png')
h, w, _ = image.shape
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
if augument:
image = iaa.CropToFixedSize(height=min(SIZE*2, h),
width=min(SIZE*2, w),
position='uniform').augment_image(image)
image = data_generator.augment(image)
else:
image = iaa.CropToFixedSize(height=min(SIZE*2, h),
width=min(SIZE*2, w),
position='center').augment_image(image)
return cv2.resize(image, (SIZE, SIZE))
def augment(image):
augment_img = iaa.Sequential([
iaa.OneOf([
iaa.Fliplr(0.5),
])], random_order=True)
image_aug = augment_img.augment_image(image)
return image_aug
from keras.preprocessing.image import ImageDataGenerator
from keras.models import Sequential, load_model
from keras.layers import (Activation, Dropout, Flatten, Dense, GlobalMaxPooling2D,
BatchNormalization, Input, Conv2D, GlobalAveragePooling2D)
from keras.applications.resnet50 import ResNet50
from keras.callbacks import ModelCheckpoint
from keras import metrics
from keras.optimizers import Adam
from keras import backend as K
import keras
from keras.models import Model
from keras_applications import imagenet_utils as utils
def ResNet(stack_fn,
preact,
use_bias,
model_name='resnet',
include_top=True,
weights='imagenet',
input_tensor=None,
input_shape=None,
pooling=None,
classes=1000,
**kwargs):
"""Instantiates the ResNet, ResNetV2, and ResNeXt architecture.
Optionally loads weights pre-trained on ImageNet.
Note that the data format convention used by the model is
the one specified in your Keras config at `~/.keras/keras.json`.
# Arguments
stack_fn: a function that returns output tensor for the
stacked residual blocks.
preact: whether to use pre-activation or not
(True for ResNetV2, False for ResNet and ResNeXt).
use_bias: whether to use biases for convolutional layers or not
(True for ResNet and ResNetV2, False for ResNeXt).
model_name: string, model name.
include_top: whether to include the fully-connected
layer at the top of the network.
weights: one of `None` (random initialization),
'imagenet' (pre-training on ImageNet),
or the path to the weights file to be loaded.
input_tensor: optional Keras tensor
(i.e. output of `layers.Input()`)
to use as image input for the model.
input_shape: optional shape tuple, only to be specified
if `include_top` is False (otherwise the input shape
has to be `(224, 224, 3)` (with `channels_last` data format)
or `(3, 224, 224)` (with `channels_first` data format).
It should have exactly 3 inputs channels.
pooling: optional pooling mode for feature extraction
when `include_top` is `False`.
- `None` means that the output of the model will be
the 4D tensor output of the
last convolutional layer.
- `avg` means that global average pooling
will be applied to the output of the
last convolutional layer, and thus
the output of the model will be a 2D tensor.
- `max` means that global max pooling will
be applied.
classes: optional number of classes to classify images
into, only to be specified if `include_top` is True, and
if no `weights` argument is specified.
# Returns
A Keras model instance.
# Raises
ValueError: in case of invalid argument for `weights`,
or invalid input shape.
"""
global backend, layers, models, keras_utils
# backend, layers, models, keras_utils = get_submodules_from_kwargs(kwargs)
backend, layers, models, keras_utils = keras.backend, keras.layers, keras.models, keras.utils
if not (weights in {'imagenet', None} or os.path.exists(weights)):
raise ValueError('The `weights` argument should be either '
'`None` (random initialization), `imagenet` '
'(pre-training on ImageNet), '
'or the path to the weights file to be loaded.')
if weights == 'imagenet' and include_top and classes != 1000:
raise ValueError('If using `weights` as `"imagenet"` with `include_top`'
' as true, `classes` should be 1000')
# Determine proper input shape
input_shape = utils._obtain_input_shape(input_shape,
default_size=224,
min_size=32,
data_format=backend.image_data_format(),
require_flatten=include_top,
weights=weights)
if input_tensor is None:
img_input = layers.Input(shape=input_shape)
else:
if not backend.is_keras_tensor(input_tensor):
img_input = layers.Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1
x = layers.ZeroPadding2D(padding=((3, 3), (3, 3)), name='conv1_pad')(img_input)
x = layers.Conv2D(64, 7, strides=2, use_bias=use_bias, name='conv1_conv')(x)
if preact is False:
x = layers.BatchNormalization(axis=bn_axis, epsilon=1.001e-5,
name='conv1_bn')(x)
x = layers.Activation('relu', name='conv1_relu')(x)
x = layers.ZeroPadding2D(padding=((1, 1), (1, 1)), name='pool1_pad')(x)
x = layers.MaxPooling2D(3, strides=2, name='pool1_pool')(x)
x = stack_fn(x)
if preact is True:
x = layers.BatchNormalization(axis=bn_axis, epsilon=1.001e-5,
name='post_bn')(x)
x = layers.Activation('relu', name='post_relu')(x)
if include_top:
x = layers.GlobalAveragePooling2D(name='avg_pool')(x)
x = layers.Dense(classes, activation='softmax', name='probs')(x)
else:
if pooling == 'avg':
x = layers.GlobalAveragePooling2D(name='avg_pool')(x)
elif pooling == 'max':
x = layers.GlobalMaxPooling2D(name='max_pool')(x)
# Ensure that the model takes into account
# any potential predecessors of `input_tensor`.
if input_tensor is not None:
inputs = keras_utils.get_source_inputs(input_tensor)
else:
inputs = img_input
# Create model.
model = models.Model(inputs, x, name=model_name)
# Load weights.
if (weights == 'imagenet') and (model_name in WEIGHTS_HASHES):
if include_top:
file_name = model_name + '_weights_tf_dim_ordering_tf_kernels.h5'
file_hash = WEIGHTS_HASHES[model_name][0]
else:
file_name = model_name + '_weights_tf_dim_ordering_tf_kernels_notop.h5'
file_hash = WEIGHTS_HASHES[model_name][1]
weights_path = keras_utils.get_file(file_name,
BASE_WEIGHTS_PATH + file_name,
cache_subdir='models',
file_hash=file_hash)
model.load_weights(weights_path)
elif weights is not None:
model.load_weights(weights)
return model
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
def block2(x, filters, kernel_size=3, stride=1,
conv_shortcut=False, name=None):
"""A residual block.
# Arguments
x: input tensor.
filters: integer, filters of the bottleneck layer.
kernel_size: default 3, kernel size of the bottleneck layer.
stride: default 1, stride of the first layer.
conv_shortcut: default False, use convolution shortcut if True,
otherwise identity shortcut.
name: string, block label.
# Returns
Output tensor for the residual block.
"""
bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1
preact = layers.BatchNormalization(axis=bn_axis, epsilon=1.001e-5,
name=name + '_preact_bn')(x)
preact = layers.Activation('relu', name=name + '_preact_relu')(preact)
if conv_shortcut is True:
shortcut = layers.Conv2D(4 * filters, 1, strides=stride,
name=name + '_0_conv')(preact)
else:
shortcut = layers.MaxPooling2D(1, strides=stride)(x) if stride > 1 else x
x = layers.Conv2D(filters, 1, strides=1, use_bias=False,
name=name + '_1_conv')(preact)
x = layers.BatchNormalization(axis=bn_axis, epsilon=1.001e-5,
name=name + '_1_bn')(x)
x = layers.Activation('relu', name=name + '_1_relu')(x)
x = layers.ZeroPadding2D(padding=((1, 1), (1, 1)), name=name + '_2_pad')(x)
x = layers.Conv2D(filters, kernel_size, strides=stride,
use_bias=False, name=name + '_2_conv')(x)
x = layers.BatchNormalization(axis=bn_axis, epsilon=1.001e-5,
name=name + '_2_bn')(x)
x = layers.Activation('relu', name=name + '_2_relu')(x)
x = layers.Conv2D(4 * filters, 1, name=name + '_3_conv')(x)
x = layers.Add(name=name + '_out')([shortcut, x])
return x
def stack2(x, filters, blocks, stride1=2, name=None):
"""A set of stacked residual blocks.
# Arguments
x: input tensor.
filters: integer, filters of the bottleneck layer in a block.
blocks: integer, blocks in the stacked blocks.
stride1: default 2, stride of the first layer in the first block.
name: string, stack label.
# Returns
Output tensor for the stacked blocks.
"""
x = block2(x, filters, conv_shortcut=True, name=name + '_block1')
for i in range(2, blocks):
x = block2(x, filters, name=name + '_block' + str(i))
x = block2(x, filters, stride=stride1, name=name + '_block' + str(blocks))
return x
def ResNet50V2(include_top=True,
weights='imagenet',
input_tensor=None,
input_shape=None,
pooling=None,
classes=1000,
**kwargs):
def stack_fn(x):
x = stack2(x, 64, 3, name='conv2')
x = stack2(x, 128, 4, name='conv3')
x = stack2(x, 256, 6, name='conv4')
x = stack2(x, 512, 3, stride1=1, name='conv5')
return x
return ResNet(stack_fn, True, True, 'resnet50v2',
include_top, weights,
input_tensor, input_shape,
pooling, classes,
**kwargs)
# reference link: https://gist.github.com/drscotthawley/d1818aabce8d1bf082a6fb37137473ae
from keras.callbacks import Callback
def get_1cycle_schedule(lr_max=1e-3, n_data_points=8000, epochs=200, batch_size=40, verbose=0):
"""
Creates a look-up table of learning rates for 1cycle schedule with cosine annealing
See @sgugger's & @jeremyhoward's code in fastai library: https://github.com/fastai/fastai/blob/master/fastai/train.py
Wrote this to use with my Keras and (non-fastai-)PyTorch codes.
Note that in Keras, the LearningRateScheduler callback (https://keras.io/callbacks/#learningratescheduler) only operates once per epoch, not per batch
So see below for Keras callback
Keyword arguments:
lr_max chosen by user after lr_finder
n_data_points data points per epoch (e.g. size of training set)
epochs number of epochs
batch_size batch size
Output:
lrs look-up table of LR's, with length equal to total # of iterations
Then you can use this in your PyTorch code by counting iteration number and setting
optimizer.param_groups[0]['lr'] = lrs[iter_count]
"""
if verbose > 0:
print("Setting up 1Cycle LR schedule...")
pct_start, div_factor = 0.3, 25. # @sgugger's parameters in fastai code
lr_start = lr_max/div_factor
lr_end = lr_start/1e4
n_iter = (n_data_points * epochs // batch_size) + 1 # number of iterations
a1 = int(n_iter * pct_start)
a2 = n_iter - a1
# make look-up table
lrs_first = np.linspace(lr_start, lr_max, a1) # linear growth
lrs_second = (lr_max-lr_end)*(1+np.cos(np.linspace(0,np.pi,a2)))/2 + lr_end # cosine annealing
lrs = np.concatenate((lrs_first, lrs_second))
return lrs
class OneCycleScheduler(Callback):
"""My modification of Keras' Learning rate scheduler to do 1Cycle learning
which increments per BATCH, not per epoch
Keyword arguments
**kwargs: keyword arguments to pass to get_1cycle_schedule()
Also, verbose: int. 0: quiet, 1: update messages.
Sample usage (from my train.py):
lrsched = OneCycleScheduler(lr_max=1e-4, n_data_points=X_train.shape[0],
epochs=epochs, batch_size=batch_size, verbose=1)
"""
def __init__(self, **kwargs):
super(OneCycleScheduler, self).__init__()
self.verbose = kwargs.get('verbose', 0)
self.lrs = get_1cycle_schedule(**kwargs)
self.iteration = 0
def on_batch_begin(self, batch, logs=None):
lr = self.lrs[self.iteration]
K.set_value(self.model.optimizer.lr, lr) # here's where the assignment takes place
if self.verbose > 0:
print('\nIteration %06d: OneCycleScheduler setting learning '
'rate to %s.' % (self.iteration, lr))
self.iteration += 1
def on_epoch_end(self, epoch, logs=None): # this is unchanged from Keras LearningRateScheduler
logs = logs or {}
logs['lr'] = K.get_value(self.model.optimizer.lr)
self.iteration = 0
from keras.regularizers import l2
def create_model(input_shape, n_out):
input_tensor = Input(shape=input_shape)
base_model = ResNet50V2(include_top=False,
weights=None,
input_tensor=input_tensor)
base_model.load_weights('../input/keras-pretrain-model-weights/resnet50v2_weights_tf_dim_ordering_tf_kernels_notop.h5')
# x = Conv2D(32, kernel_size=(1,1), activation='relu')(base_model.output)
# x = Flatten()(x)
x = GlobalAveragePooling2D()(base_model.output)
x = Dropout(0.5)(x)
x = Dense(1024, activation='relu')(x)
x = Dropout(0.5)(x)
# x = Dense(1024, activation='relu')(x)
# x = Dropout(0.3)(x)
final_output = Dense(n_out, activation='sigmoid', name='final_output')(x)
model = Model(input_tensor, final_output)
return model
# create callbacks list
from keras.callbacks import (ModelCheckpoint, LearningRateScheduler,
EarlyStopping, ReduceLROnPlateau,CSVLogger)
from sklearn.model_selection import train_test_split
epochs = 18; batch_size = 64
checkpoint = ModelCheckpoint('../working/Resnet50_focal.h5', monitor='val_loss', verbose=1,
save_best_only=True, mode='min', save_weights_only = True)
reduceLROnPlat = ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=2,
verbose=1, mode='auto', epsilon=0.0001)
early = EarlyStopping(monitor="val_loss",
mode="min",
patience=7)
csv_logger = CSVLogger(filename='../working/training_log.csv',
separator=',',
append=True)
# split data into train, valid
indexes = np.arange(train_dataset_info.shape[0])
train_indexes, valid_indexes = train_test_split(indexes, test_size=0.15, random_state=8)
# create train and valid datagens
train_generator = data_generator.create_train(
train_dataset_info[train_indexes], batch_size, (SIZE,SIZE,3), augument=True, mix=False)
train_mixup = data_generator.create_train(
train_dataset_info[train_indexes], batch_size, (SIZE,SIZE,3), augument=True, mix=True)
train_generator_warmup = data_generator.create_train(
train_dataset_info[train_indexes], batch_size, (SIZE,SIZE,3), augument=False)
validation_generator = data_generator.create_valid(
train_dataset_info[valid_indexes], batch_size, (SIZE,SIZE,3), augument=False)
lrsched = OneCycleScheduler(lr_max=1e-3, n_data_points=len(train_indexes),
epochs=1, batch_size=32, verbose=0)
# callbacks_list = [checkpoint, csv_logger, lrsched]
callbacks_list = [checkpoint, csv_logger, reduceLROnPlat]
# callbacks_list = [checkpoint, csv_logger]
model = create_model(
input_shape=(SIZE,SIZE,3),
n_out=NUM_CLASSES)
# warm up model
for layer in model.layers:
layer.trainable = False
for i in range(-3,0):
model.layers[i].trainable = True
model.compile(
loss='binary_crossentropy',
optimizer=Adam(1e-3))
# optimizer=AdamAccumulate(lr=1e-3, accum_iters=2))
model.fit_generator(
train_generator_warmup,
steps_per_epoch=np.ceil(float(len(train_indexes)) / float(128)),
epochs=2,
max_queue_size=16, workers=WORKERS, use_multiprocessing=True,
verbose=1)
# train all layers
for layer in model.layers:
layer.trainable = True
callbacks_list = [checkpoint, csv_logger, reduceLROnPlat]
model.compile(loss='binary_crossentropy',
# loss=focal_loss,
# optimizer=Adam(lr=1e-4,decay=0.0))
optimizer=AdamAccumulate(lr=2e-4, accum_iters=2))
model.fit_generator(
# train_generator,
train_mixup,
steps_per_epoch=np.ceil(float(len(train_indexes)) / float(batch_size)),
validation_data=validation_generator,
validation_steps=np.ceil(float(len(valid_indexes)) / float(batch_size)),
epochs=epochs,
verbose=1,
max_queue_size=16, workers=WORKERS, use_multiprocessing=True,
callbacks=callbacks_list)
# model.compile(loss='binary_crossentropy',
# # loss=focal_loss,
# optimizer=Adam(lr=1e-4))
# # optimizer=AdamAccumulate(lr=1e-4, accum_iters=4))
# callbacks_list = [checkpoint, csv_logger, reduceLROnPlat]
# model.fit_generator(
# train_generator,
# steps_per_epoch=np.ceil(float(len(train_indexes)) / float(batch_size)),
# validation_data=validation_generator,
# validation_steps=np.ceil(float(len(valid_indexes)) / float(batch_size)),
# epochs=epochs,
# verbose=1,
# max_queue_size=16, workers=WORKERS, use_multiprocessing=True,
# callbacks=callbacks_list)
print(os.listdir('../working/'))
submit = pd.read_csv('../input/imet-2019-fgvc6/sample_submission.csv')
model.load_weights('../working/Resnet50_focal.h5')
predicted = []
'''Search for the best threshold regarding the validation set'''
BATCH = 512
fullValGen = data_generator.create_valid(
train_dataset_info[valid_indexes], BATCH, (SIZE,SIZE,3))
n_val = round(train_dataset_info.shape[0]*0.15)//BATCH
print(n_val)
lastFullValPred = np.empty((0, NUM_CLASSES))
lastFullValLabels = np.empty((0, NUM_CLASSES))
for i in tqdm(range(n_val+1)):
im, lbl = next(fullValGen)
scores = model.predict(im)
lastFullValPred = np.append(lastFullValPred, scores, axis=0)
lastFullValLabels = np.append(lastFullValLabels, lbl, axis=0)
print(lastFullValPred.shape, lastFullValLabels.shape)
def my_f2(y_true, y_pred):
assert y_true.shape[0] == y_pred.shape[0]
tp = np.sum((y_true == 1) & (y_pred == 1), axis=1)
tn = np.sum((y_true == 0) & (y_pred == 0), axis=1)
fp = np.sum((y_true == 0) & (y_pred == 1), axis=1)
fn = np.sum((y_true == 1) & (y_pred == 0), axis=1)
p = tp / (tp + fp + K.epsilon())
r = tp / (tp + fn + K.epsilon())
f2 = (1+beta_f2**2)*p*r / (p*beta_f2**2 + r + 1e-15)
return np.mean(f2)
def find_best_fixed_threshold(preds, targs, do_plot=True):
score = []
thrs = np.arange(0, 0.5, 0.01)
for thr in tqdm(thrs):
score.append(my_f2(targs, (preds > thr).astype(int) ))
score = np.array(score)
pm = score.argmax()
best_thr, best_score = thrs[pm], score[pm].item()
print(f'thr={best_thr:.3f}', f'F2={best_score:.3f}')
if do_plot:
plt.plot(thrs, score)
plt.vlines(x=best_thr, ymin=score.min(), ymax=score.max())
plt.text(best_thr+0.03, best_score-0.01, f'$F_{2}=${best_score:.3f}', fontsize=14);
plt.show()
return best_thr, best_score
best_thr, best_score = find_best_fixed_threshold(lastFullValPred, lastFullValLabels, do_plot=True)
# for i, name in tqdm(enumerate(submit['id'])):
# path = os.path.join('../input/imet-2019-fgvc6/test/', name)
# image = data_generator.load_image(path, (SIZE,SIZE,3))
# score_predict = model.predict(preprocess_input(image[np.newaxis]))
# # score_predict = model.predict((image[np.newaxis])/255)
# # print(score_predict)
# label_predict = np.arange(NUM_CLASSES)[score_predict[0]>=best_thr]
# # print(label_predict)
# str_predict_label = ' '.join(str(l) for l in label_predict)
# predicted.append(str_predict_label)
# submit['attribute_ids'] = predicted
# submit.to_csv('submission.csv', index=False)