def load_fc_weights_from_vgg16bn(model):
"Load weights for model from the dense layers of the Vgg16BN model."
# See imagenet_batchnorm.ipynb for info on how the weights for
# Vgg16BN can be generated from the standard Vgg16 weights.
from vgg16bn import Vgg16BN
vgg16_bn = Vgg16BN()
_, fc_layers = utils.split_at(vgg16_bn.model, Convolution2D)
utils.copy_weights(fc_layers, model.layers)
def hybrid_ensemble(number=1):
batch_size=32
target_size=(224, 224)
reset_valid()
set_valid(number=3)
gen = image.ImageDataGenerator(rotation_range=10, width_shift_range=0.05,
height_shift_range=0.05, width_zoom_range=0.1, zoom_range=0.1,
shear_range=0.1, channel_shift_range=10)
trn_batches = gen.flow_from_directory(train_path, target_size=target_size,
batch_size=batch_size, shuffle=True, class_mode='categorical')
val_batches = gen.flow_from_directory(valid_path, target_size=target_size,
batch_size=batch_size, shuffle=False, class_mode='categorical')
VGGbn = Vgg16BN()
VGGbn.compile(Adam(), loss='categorical_crossentropy', metrics=['accuracy'])
VGGbn.fit_generator(trn_batches, trn_batches.n, nb_epoch=1,
validation_data=val_batches, nb_val_samples=val_batches.n)
last_conv_idx = [index, for index, layer in enumerate(VGGbn.model.layers) \
if type(layer) is Convolution2D][-1]
Conv_layers = VGGbn.model.layers[:last_conv_idx + 1]
Conv_model = Sequential(Conv_layers)
trn_batches = gen.flow_from_directory(train_path, target_size=target_size,
batch_size=batch_size, shuffle=False, class_mode='categorical')
conv_features = Conv_model.predict_generator(trn_batches,
trn_batches.nb_sample)
conv_val_feat = Conv_model.predict_generator(val_batches,
val_batches.nb_sample)
predarray = []
for n in xrange(number):
reset_valid()
set_valid(number=3)
FC_model = Sequential(create_FCbn_layers(p=0.3))
FC_model.compile(Adam(), loss='categorical_crossentropy',
metrics=['accuracy'])
FC_model.fit(conv_features, trn_batches.labels, batch_size=batch_size,
nb_epoch=1, validation_data=(conv_val_feat, val_batches.labels))
gen_t = image.ImageDataGenerator()
tst_batches = gen_t.flow_from_directory(test_path,
batch_size=batch_size, shuffle=False, class_mode=None)
conv_tst_feat = Conv_model.predict_generator(tst_batches,
tst_batches.nb_sample)
preds = FC_model.predict(conv_tst_feat, batch_size=batch_size*2)
predarray.append(preds)
def main():
path, datadir = path_setup()
# create model with correct number of outputs
# this pops off the existing last layer, sets remaining layers to not trainable
# and puts on a layer with the correct number of outputs
from vgg16bn import Vgg16BN
vgg = Vgg16BN()
vgg.ft(settings.NUM_CLASSES)
model = vgg.model
#load training data
trn = ut.load_array(path, settings.TRAIN_FOLDER_NAME)
val = ut.load_array(path, settings.VALIDATE_FOLDER_NAME)
tst = ut.load_array(path, settings.TEST_FOLDER_NAME)
# gen = image.ImageDataGenerator()
from keras import optimizers
model.compile(optimizer=optimizers.Adam(1e-3),
loss='categorical_crossentropy',
metrics=['accuracy'])
(val_classes, trn_classes, val_labels, trn_labels, val_filenames,
filenames, test_filenames) = ut.get_classes(path)
# see https://stackoverflow.com/questions/41771965/error-when-checking-model-input-expected-convolution2d-input-1-to-have-shape-n
trn = trn.transpose(0, 3, 1, 2)
val = val.transpose(0, 3, 1, 2)
# tst = tst.transpose(0, 3, 1, 2)
model.fit(trn,
trn_labels,
batch_size=settings.BATCH_SIZE,
nb_epoch=settings.NUMBER_EPOCHS,
validation_data=(val, val_labels))
file_best_weights = os.path.join(path, settings.CHECKPOINTFILE_2)
print("saving bestweights to " + file_best_weights)
model.save_weights(file_best_weights)
def load_dense_weights_from_vgg16bn(model):
from vgg16bn import Vgg16BN
vgg16_bn = Vgg16BN()
_, dense_layers = split_model(vgg16_bn.model)
copy_weights(dense_layers, model.layers)
save_array(path+'results/test_640.dat', test) # In[6]: trn = load_array(path+'results/trn_640.dat') val = load_array(path+'results/val_640.dat') # We can now create our VGG model - we'll need to tell it we're not using the normal 224x224 images, which also means it won't include the fully connected layers (since they don't make sense for non-default sizes). We will also remove the last max pooling layer, since we don't want to throw away information yet. # In[74]: vgg640 = Vgg16BN((360, 640)).model vgg640.pop() vgg640.input_shape, vgg640.output_shape vgg640.compile(Adam(), 'categorical_crossentropy', metrics=['accuracy']) # We can now pre-compute the output of the convolutional part of VGG. # In[75]: conv_val_feat = vgg640.predict(val, batch_size=32, verbose=1) conv_trn_feat = vgg640.predict(trn, batch_size=32, verbose=1) # In[76]:
from vgg16bn import Vgg16BN vgg = Vgg16BN()
from utils import *
#Define path
path = "data/"
#path = "data/sample/"
results_path = path + 'results/'
input_shape = (224, 224)
#Define conv model
from vgg16bn import Vgg16BN
vgg16 = Vgg16BN()
vgg16.ft(8)
conv_layers, fc_layers = split_at(vgg16.model, Convolution2D)
conv_model = Sequential(conv_layers)
batch_size = 64
gen = image.ImageDataGenerator()
train_datagen = gen.flow_from_directory(path + 'train/',
target_size=input_shape,
batch_size=batch_size)
test_datagen = gen.flow_from_directory(path + 'test/',
target_size=input_shape,
batch_size=batch_size,
shuffle=False)
s = FeatureSaver(train_datagen=train_datagen, test_datagen=test_datagen)
f = h5py.File(results_path + 'vgg_conv.h5', 'w')
s.save_train(conv_model, f, num_epochs=1)
s.save_test(conv_model, f)
img = crop(img, img_row)
img = rotate(img, img_row)
img = img.resize((IMG_WIDTH, IMG_HEIGHT))
yield np.expand_dims(np.moveaxis(np.asarray(img), 2, 0), 0)
def crop(img, img_row):
return img.crop((img_row['xl'], img_row['yu'], img_row['xr'],img_row['yd']))
def rotate(img, img_row):
result = img if(img_row['xr'] - img_row['xl'] > img_row['yd'] - img_row['yu']) else img.transpose(Image.ROTATE_90)
return result
# ## Neural net
# In[4]:
vgg640 = Vgg16BN((IMG_HEIGHT, IMG_WIDTH)).model
vgg640.pop()
vgg640.compile(Adam(), 'categorical_crossentropy', metrics=['accuracy'])
#compute features
vgg_features = vgg640.predict_generator(crop_rotate_scale_gen(), val_samples = len(os.listdir(DATA_PATH))
vgg_features = np.moveaxis(vgg_features, 1, 3)
#serialize vgg features
np.save('{}/vgg16_cropped_block5.npy'.format(FEATURE_PATH), vgg_features)
target_size = (224, 224)
# train/valid batch generators
gen = image.ImageDataGenerator(rotation_range=10, width_shift_range=0.05,
height_shift_range=0.05, width_zoom_range=0.1, zoom_range=0.1,
shear_range=0.1, channel_shift_range=10)
# does it matter that I don't set dim_ordering='tf' ?
trn_batches = gen.flow_from_directory(train_path, target_size=target_size,
batch_size=batch_size, shuffle=True, class_mode='categorical')
val_batches = gen.flow_from_directory(valid_path, target_size=target_size,
batch_size=batch_size, shuffle=False, class_mode='categorical')
# load VGG16BN model & its weights
VGGbn = Vgg16BN()
VGGbn.compile(Adam(), loss='categorical_crossentropy', metrics=['accuracy'])
# (maybe) train the model at low η to train the Conv layers a bit
VGGbn.fit_generator(trn_batches, trn_batches.n, nb_epoch=1,
validation_data=val_batches, nb_val_samples=val_batches.n)
# find out how many epochs at what η to do this until it's ~optimal
# separate Conv layers & create new ConvNet (w/ vgg weights)
last_conv_idx = [index, for index, layer in enumerate(VGGbn.model.layers) \
if type(layer) is Convolution2D][-1]
Conv_layers = VGGbn.model.layers[:last_conv_idx + 1]
# create new ConvNet from VGG16BN conv layers
Conv_model = Sequential(Conv_layers)
def main():
path, datadir = path_setup()
# create model with correct number of outputs
# this pops off the existing last layer, sets remaining layers to not trainable
# and puts on a layer with the correct number of outputs
from vgg16bn import Vgg16BN
vgg = Vgg16BN()
vgg.ft(settings.NUM_CLASSES)
model = vgg.model
# now load wth the best weights (created by train_2_VGG_BN.py
file_best_weights = os.path.join(path, settings.CHECKPOINTFILE_2)
print("loading bestweights from " + file_best_weights)
try:
model.load_weights(file_best_weights)
except OSError:
print('WHOAH!!!! Run train_2_VGG_BN.py in order to create ' +
settings.CHECKPOINTFILE_2)
return
#load training data
trn = ut.load_array(path, settings.TRAIN_FOLDER_NAME)
val = ut.load_array(path, settings.VALIDATE_FOLDER_NAME)
tst = ut.load_array(path, settings.TEST_FOLDER_NAME)
# see https://stackoverflow.com/questions/41771965/error-when-checking-model-input-expected-convolution2d-input-1-to-have-shape-n
trn = trn.transpose(0, 3, 1, 2)
val = val.transpose(0, 3, 1, 2)
tst = tst.transpose(0, 3, 1, 2)
(val_classes, trn_classes, val_labels, trn_labels, val_filenames,
trn_filenames, tst_filenames) = ut.get_classes(path)
# lets strip off the FC layers and just calculate the stacked conv layer outputs for the network
# then create a fc network of appropriate input and output size) and use above conv outputs
# to train it, much faster since most of the back prop compute is in the conv layers.
# this means the only training is happenning in the FC layers
conv_layers, fc_layers = ut.split_at(model, Conv2D)
conv_model = Sequential(conv_layers)
#whats it look like
plot_model(conv_model,
to_file=settings.CONV_LAYERS_MODEL_PNG,
show_shapes=True)
# display_summary("CONV model size is", conv_model)
# get the predictions
conv_trn_feat = load_CNN_Codes(conv_model, trn, path,
settings.CONV_TRN_FEAT)
conv_val_feat = load_CNN_Codes(conv_model, val, path,
settings.CONV_VAL_FEAT)
conv_tst_feat = load_CNN_Codes(conv_model, tst, path,
settings.CONV_TST_FEAT)
print("Feature shape:" + str(conv_val_feat.shape))
#********************************************************************************
# the follwoing bits create and train a fully connected top layer using CNN_Codes
fc_top_layer_model = create_and_train_FC_top_layer_using_CNN_Codes(
conv_layers, conv_trn_feat, conv_val_feat, path, trn_labels,
val_labels)
#********************************************************************************
#tape the fully connected top layer on to the conv base layer and train the whole thingt once more (conv layers are not trainable)
tape_FC_Top_layer_onto_conv_base_and_train(conv_model, fc_top_layer_model,
trn, trn_labels, val,
val_labels)
###########################################
#there are thousands of images but only a few image sizes(<10), assumme each image size corresponds to a particular
# boat and that each boat is going for a particular type of fish, then can we assumme a higher percentage of pix for
# that kind of fish on that particular boat? (Probably, but the neural net has already taken this into account, so
# no accuracy boost)
create_new_model_that_includes_size_of_image_as_input(
conv_trn_feat, conv_val_feat, model, path, trn_filenames, trn_labels,
val_filenames, val_labels)
import os
os.chdir('Desktop/input')
from utils import *
from vgg16bn import Vgg16BN
path = ""
batch_size = 64
(val_classes, trn_classes, val_labels, trn_labels, val_filenames, filenames,
test_filenames) = get_classes(path)
trn = get_data(path + 'train', (540, 960))
val = get_data(path + 'valid', (540, 960))
test = get_data(path + 'test', (540, 960))
vgg640 = Vgg16BN((540, 960)).model
vgg640.pop()
vgg640.input_shape, vgg640.output_shape
vgg640.compile(Adam(), 'categorical_crossentropy', metrics=['accuracy'])
##
conv_val_feat = vgg640.predict(val, batch_size=32, verbose=1)
conv_trn_feat = vgg640.predict(trn, batch_size=32, verbose=1)
conv_test_feat = vgg640.predict(test, batch_size=32, verbose=1)
##
save_array(path + 'results/conv_val_960.dat', conv_val_feat)
save_array(path + 'results/conv_trn_960.dat', conv_trn_feat)
save_array(path + 'results/conv_test_960.dat', conv_test_feat)
##
conv_val_feat = load_array(path + 'results/conv_val_960.dat')
conv_trn_feat = load_array(path + 'results/conv_trn_960.dat')
conv_test_feat = load_array(path + 'results/conv_test_960.dat')
conv_layers, _ = split_at(vgg640, Convolution2D)
def vgg(output_size, dropout=0.5):
model = Vgg16BN((224, 224), True,
dropout=dropout).model #NB: batch normalization added
model.pop()
model.add(Dense(output_size, activation='softmax'))
return model
# val = get_batches(path+'valid', shuffle=False, batch_size=batch_size, class_mode=None, target_size=(360,640))
if not os.path.exists(path.split('stage1')[0] + 'results'):
os.makedirs(path.split('stage1')[0] + 'results')
import glob
import pandas as pd
# save_array(path+'results/trn_640.dat', trn)
# save_array(path+'results/val_640.dat', val)
# save_array(path+'results/test_640.dat', test)
nb_train= len(glob.glob(path+'*/*.dcm'))
# nb_train_samples = 500
# nb_val= len(glob.glob(path+'valid/*/*.jpg'))
# nb_test= len(glob.glob('input/test/*/*.jpg'))
import numpy as np
vgg512 = Vgg16BN((512, 512), use_preprocess=True).model
vgg512.pop()
print(vgg512.input_shape, vgg512.output_shape)
vgg512.compile(Adam(), 'categorical_crossentropy', metrics=['accuracy'])
import time
train_csv_table = pd.read_csv('../input/stage1_labels.csv')
t0=time.time()
print(len([x[0] for x in os.walk(path)][1:]))
print(len(glob.glob(path+'*/*.dcm')))
patient_ids = [x[0].split('/')[-1] for x in os.walk(path)][1:]
for an_index,an_id in enumerate(patient_ids):
if an_index %100==0:
print(time.time()-t0)
t0 = time.time()
