def get_attributes(self, predictions):
predictions = {}
for label, predict in self.predict.items():
prediction = predict(
np.array(preprocess(image,
self.config['image_shape']))[None])[0]
predictions[label] = self.labels[label][np.argmax(prediction)]
return predictions
def normalize_sample(quantity):
# Use the mean and deviance to produce normalized data for an arbitrary image
pixel_mean = np.load('pixel_mean.npy')
pixel_deviance = np.load('pixel_deviation.npy')
for record in list(db.ebay.aggregate([{"$sample": {"size": quantity}}])):
# print(record['title'])
im = preprocess(Image.open(BytesIO(record['image'])),
config['image_shape'],
mask=False,
letterboxing=False)
standardized = (im - pixel_mean) / pixel_deviance
print(standardized)
def find_normalizers():
samples = 5000
image_list = []
for listing in db.ebay.aggregate([{"$sample": {"size": samples}}]):
image_list.append(
np.array(
preprocess(Image.open(BytesIO(listing['image'])),
config['image_shape'],
mask=False,
letterboxing=False)))
images = np.array(image_list)
print(images.shape)
print(np.max(images.std(axis=0)))
np.save('pixel_mean.npy', images.mean(axis=0))
np.save('pixel_deviation.npy', images.std(axis=0))
def prepare(data, label):
# Convert list of dictionaries into two lists of values
binaries, tags = [
i for i in zip(*[(stim['image'], stim[label]) for stim in data])
]
# Convert a list of binaries to a 4D stimulus array
stim_list = []
for binary in binaries:
img = preprocess(Image.open(BytesIO(binary)),
config['image_shape'])
normalized = (np.array(img) - pixel_mean) / pixel_deviance
stim_list.append(np.transpose(normalized, axes=(1, 0, 2)))
stim = np.stack(stim_list)
# Convert a text label to a onehot encoding
exp = np.stack([
np.eye(network.classifications[label])[labels[label].index(tag)]
for tag in tags
])
return stim, exp
import numpy as np
import sys
from Tagger.image_formatter import preprocess
# Run once after scraping to remove records with invalid binary data
mongo_client = MongoClient(host='localhost', port=27017) # Default port
db = mongo_client.deep_fashion
count = 0
for document in db.ebay.find({}):
try:
shape = np.shape(
np.array(preprocess(Image.open(BytesIO(document['image'])))))
if shape != (224, 224, 3):
count += 1
# Show status
sys.stdout.write("\r\x1b[KRemoved: " + str(count))
sys.stdout.flush()
db.ebay.remove({'image_url': document['image_url']})
except Exception as err:
print("\nException: " + document['image_url'])
print(err)
count += 1
db.ebay.remove({'image_url': document['image_url']})
def checkpoint(check_iteration):
print()
total_loss = 0
# Take first n records with all attributes
query = [{
"$match": {label: {
"$exists": True
}
for label in labels.keys()}
}, {
"$project": {
'image_url': 1,
'image': 1,
**{label: 1
for label in labels.keys()}
}
}, {
"$limit": config['batch_size']
}]
# Exhaust the generator into a list
samples = list(db.ebay.aggregate(query))
# Print the predicted values for a single sample
example = random.choice(samples)
# Evaluate losses for all five classifiers
for label in labels.keys():
# Reformat sampled data into input/output values
stimulus, expected = prepare(samples, label)
loss = sess.run(network.loss[label],
feed_dict={
network.stimulus: stimulus,
network.expected[label]: expected
})
img = preprocess(Image.open(BytesIO(example['image'])),
config['image_shape'],
mask=config['mask'],
letterboxing=config['letterbox'])
standard = (img - pixel_mean) / pixel_deviance
prediction = sess.run(network.predict[label],
feed_dict={
network.stimulus:
np.transpose(standard,
axes=(1, 0, 2))[None]
})[0]
expectation = np.eye(
network.classifications[label])[labels[label].index(
example[label])]
if config['plot']:
history[label].append((int(check_iteration), float(loss)))
print("\t" + label + ": " + str(loss) + '\t' +
np.array2string(prediction, precision=2) + ' ' +
str(expectation))
total_loss += loss
print(example['image_url'])
# Store the graph
if save:
save_all.save(sess, save_path + 'model.ckpt')
# Plotting
if config['plot']:
with open(save_path + 'history.json', 'w') as histfile:
json.dump(history, histfile)
plt.gca().clear()
plt.title('Loss')
for index, label in enumerate(labels.keys()):
plt.plot(*zip(*history[label]),
marker='.',
color=colors[index],
label=label)
plt.legend(loc='upper left', prop={'size': 6})
# Draw
backend = plt.rcParams['backend']
if backend in matplotlib.rcsetup.interactive_bk:
fig_manager = getattr(matplotlib,
'_pylab_helpers').Gcf.get_active()
if fig_manager is not None:
canvas = fig_manager.canvas
if canvas.figure.stale:
canvas.draw()
canvas.start_event_loop(0.00001)
# If converged, return false to end the program. Wouldn't this be nice?
return total_loss > config['epsilon']