Based on fast.ai dogs_cats_redux notebook in order to make my own entry into the Kaggle competition.
https://www.kaggle.com/c/the-nature-conservancy-fisheries-monitoring
My dir structure is similar, but not exactly the same:
utils
ncfish
data
train
test
In [1]:
#Verify we are in the lesson1 directory
%pwd
Out[1]:
In [2]:
%matplotlib inline
In [3]:
import os, sys
sys.path.insert(1, os.path.join(sys.path[0], '../utils'))
from utils import *
from vgg16 import Vgg16
from PIL import Image
from keras.preprocessing import image
from sklearn.metrics import confusion_matrix
Note: had to comment out vgg16bn in utils.py (whatever that is)
In [4]:
current_dir = os.getcwd()
LESSON_HOME_DIR = current_dir
DATA_HOME_DIR = current_dir+'/data'
categories = sorted([os.path.basename(x) for x in glob(DATA_HOME_DIR+'/train/*')])
In [ ]:
from shutil import copyfile
#Create directories
%cd $DATA_HOME_DIR
# did this once
%mkdir valid
%mkdir results
%mkdir -p sample/train
%mkdir -p sample/test
%mkdir -p sample/valid
%mkdir -p sample/results
%mkdir -p test/unknown
In [ ]:
# Create subdirectories
for c in categories:
%mkdir -p valid/{c}
%mkdir -p sample/train/{c}
%mkdir -p sample/valid/{c}
In [ ]:
%cd $DATA_HOME_DIR/train
In [ ]:
# how many images we talking about?
for c in categories:
g = glob(c+"/*.jpg")
print c, len(g)
This was original output:
ALB 1719
BET 200
DOL 117
LAG 67
NoF 465
OTHER 299
SHARK 176
YFT 734
In [ ]:
validation_ratio = 0.1
for c in categories:
g = glob(c+"/*.jpg")
shuf = np.random.permutation(g)
num_valid = int(validation_ratio*len(g))
for i in range(num_valid):
#print shuf[i], DATA_HOME_DIR+'/valid/' + shuf[i]
os.rename(shuf[i], DATA_HOME_DIR+'/valid/' + shuf[i])
In [ ]:
# Now, how many images we talking about?
for c in categories:
g = glob(c+"/*.jpg")
print c, len(g),
g = glob("../valid/"+c+"/*.jpg")
print len(g)
In [ ]:
# now create the sample train subset of 10 per category
for c in categories:
g = glob(c+"/*.jpg")
shuf = np.random.permutation(g)
for i in range(10):
#print shuf[i], DATA_HOME_DIR+'/sample/train/' + shuf[i]
copyfile(shuf[i], DATA_HOME_DIR+'/sample/train/' + shuf[i])
In [ ]:
%cd $DATA_HOME_DIR/valid
In [ ]:
# now create the sample valid subset of 2 per category
for c in categories:
g = glob(c+"/*.jpg")
shuf = np.random.permutation(g)
for i in range(2):
#print shuf[i], DATA_HOME_DIR+'/sample/valid/' + shuf[i]
copyfile(shuf[i], DATA_HOME_DIR+'/sample/valid/' + shuf[i])
In [ ]:
!ls {DATA_HOME_DIR}/train/*/* |wc -l
!ls {DATA_HOME_DIR}/valid/*/* |wc -l
!ls {DATA_HOME_DIR}/sample/train/*/* |wc -l
!ls {DATA_HOME_DIR}/sample/valid/*/* |wc -l
Training & 10% for Validation numbers
ALB 1548 171
BET 180 20
DOL 106 11
LAG 61 6
NoF 419 46
OTHER 270 29
SHARK 159 17
YFT 661 73
In [ ]:
# Create single 'unknown' class for test set
%cd $DATA_HOME_DIR/test_stg1
%mv *.jpg ../test/unknown/
In [ ]:
!ls {DATA_HOME_DIR}/test
In [5]:
%cd $DATA_HOME_DIR
#Set path to sample/ path if desired
path = DATA_HOME_DIR + '/'
#path = DATA_HOME_DIR + '/sample/'
test_path = DATA_HOME_DIR + '/test/' #We use all the test data
results_path=DATA_HOME_DIR + '/results/'
train_path=path + '/train/'
valid_path=path + '/valid/'
In [6]:
vgg = Vgg16()
In [7]:
#Set constants. You can experiment with no_of_epochs to improve the model
batch_size=64
no_of_epochs=1
In [8]:
#Finetune the model
batches = vgg.get_batches(train_path, batch_size=batch_size)
val_batches = vgg.get_batches(valid_path, batch_size=batch_size*2)
vgg.finetune(batches)
#Not sure if we set this for all fits
vgg.model.optimizer.lr = 0.01
In [9]:
#Notice we are passing in the validation dataset to the fit() method
#For each epoch we test our model against the validation set
latest_weights_filename = None
#latest_weights_filename='ft24.h5'
#vgg.model.load_weights(results_path+latest_weights_filename)
if you are training, stay here. if you are loading & creating submission skip down from here.
In [10]:
# if you have run some epochs already...
epoch_offset=1 # trying again from ft1
for epoch in range(no_of_epochs):
print "Running epoch: %d" % (epoch + epoch_offset)
vgg.fit(batches, val_batches, nb_epoch=1)
latest_weights_filename = 'ft%d.h5' % (epoch + epoch_offset)
vgg.model.save_weights(results_path+latest_weights_filename)
print "Completed %s fit operations" % no_of_epochs
In [ ]:
# only if you have to
latest_weights_filename='ft1.h5'
vgg.model.load_weights(results_path+latest_weights_filename)
In [ ]:
val_batches, probs = vgg.test(valid_path, batch_size = batch_size)
filenames = val_batches.filenames
expected_labels = val_batches.classes # 0 - 7
#Round our predictions to 0/1 to generate labels
#our_predictions = probs[:,0]
#our_labels = np.round(1-our_predictions)
our_labels = np.argmax(probs, axis=1)
In [ ]:
cm = confusion_matrix(expected_labels, our_labels)
plot_confusion_matrix(cm, val_batches.class_indices)
In [ ]:
#Helper function to plot images by index in the validation set
#Plots is a helper function in utils.py
def plots_idx(idx, titles=None):
plots([image.load_img(valid_path + filenames[i]) for i in idx], titles=titles)
#Number of images to view for each visualization task
n_view = 4
In [ ]:
#1. A few correct labels at random
correct = np.where(our_labels==expected_labels)[0]
print "Found %d correct labels" % len(correct)
idx = permutation(correct)[:n_view]
plots_idx(idx, our_predictions[idx])
In [ ]:
#2. A few incorrect labels at random
incorrect = np.where(our_labels!=expected_labels)[0]
print "Found %d incorrect labels" % len(incorrect)
idx = permutation(incorrect)[:n_view]
plots_idx(idx, our_predictions[idx])
In [ ]:
val_batches.class_indices
In [ ]:
#3a. The images we most confident were X, and are actually X
X='YFT'
Xi=val_batches.class_indices[X]
correct_cats = np.where((our_labels==Xi) & (our_labels==expected_labels))[0]
print "Found %d confident correct %s labels" % (len(correct_cats),X)
most_correct_cats = np.argsort(our_predictions[correct_cats])[::-1][:n_view]
plots_idx(correct_cats[most_correct_cats], our_predictions[correct_cats][most_correct_cats])
In [ ]:
#4a. The images we were most confident were cats, but are actually dogs
incorrect_cats = np.where((our_labels==0) & (our_labels!=expected_labels))[0]
print "Found %d incorrect cats" % len(incorrect_cats)
if len(incorrect_cats):
most_incorrect_cats = np.argsort(our_predictions[incorrect_cats])[::-1][:n_view]
plots_idx(incorrect_cats[most_incorrect_cats], our_predictions[incorrect_cats][most_incorrect_cats])
In [ ]:
#4b. The images we were most confident were dogs, but are actually cats
incorrect_dogs = np.where((our_labels==1) & (our_labels!=expected_labels))[0]
print "Found %d incorrect dogs" % len(incorrect_dogs)
if len(incorrect_dogs):
most_incorrect_dogs = np.argsort(our_predictions[incorrect_dogs])[:n_view]
plots_idx(incorrect_dogs[most_incorrect_dogs], our_predictions[incorrect_dogs][most_incorrect_dogs])
In [ ]:
#5. The most uncertain labels (ie those with probability closest to 0.5).
most_uncertain = np.argsort(np.abs(our_predictions-0.5))
plots_idx(most_uncertain[:n_view], our_predictions[most_uncertain])
In [ ]:
In [ ]:
batches, preds = vgg.test(test_path, batch_size = batch_size*2)
# Error allocating 3347316736 bytes of device memory (out of memory).
# got this error when batch-size = 128
# I see this pop up to 6GB memory with batch_size = 64 & this takes some time...
In [ ]:
#For every image, vgg.test() generates two probabilities
#based on how we've ordered the cats/dogs directories.
#It looks like column one is cats and column two is dogs
print preds[:5]
filenames = batches.filenames
print filenames[:5]
In [ ]:
#You can verify the column ordering by viewing some images
Image.open(test_path + filenames[1])
In [ ]:
#Save our test results arrays so we can use them again later
save_array(results_path + 'test_preds.dat', preds)
save_array(results_path + 'filenames.dat', filenames)
In [ ]:
#Load our test predictions from file
preds = load_array(results_path + 'test_preds.dat')
filenames = load_array(results_path + 'filenames.dat')
In [ ]:
#Grab the dog prediction column
isdog = preds[:,1]
print "Raw Predictions: " + str(isdog[:5])
print "Mid Predictions: " + str(isdog[(isdog < .6) & (isdog > .4)])
print "Edge Predictions: " + str(isdog[(isdog == 1) | (isdog == 0)])
In [ ]:
#play it safe, round down our edge predictions
#isdog = isdog.clip(min=0.05, max=0.95)
#isdog = isdog.clip(min=0.02, max=0.98)
isdog = isdog.clip(min=0.01, max=0.99)
In [ ]:
#Extract imageIds from the filenames in our test/unknown directory
filenames = batches.filenames
ids = np.array([int(f[8:f.find('.')]) for f in filenames])
In [ ]:
subm = np.stack([ids,isdog], axis=1)
subm[:5]
In [ ]:
%cd $DATA_HOME_DIR
submission_file_name = 'submission4.csv'
np.savetxt(submission_file_name, subm, fmt='%d,%.5f', header='id,label', comments='')
In [ ]:
from IPython.display import FileLink
%cd $LESSON_HOME_DIR
FileLink('data/'+submission_file_name)
In [ ]: