def load_lfw_dataset_EW(use_raw=False, dimx=45, dimy=45):
df_attrs = pd.read_csv(ATTRS_NAME, sep='\t', skiprows=1)
df_attrs.columns = list(df_attrs.columns)[1:] + ["NaN"]
df_attrs = df_attrs.drop("NaN", axis=1)
imgs_with_attrs = set(map(tuple, df_attrs[["person", "imagenum"]].values))
# read photos
photos = []
mapping = []
index = 0
with ZipFile(IMAGES_NAME_POC) as f:
for m in tqdm_utils.tqdm_notebook_failsafe(f.namelist()):
# prepare image
img = decode_image_from_raw_bytes(f.open(m).read())
img = cv2.resize(img, (dimx, dimy))
# parse person
fname = os.path.split(m)[-1]
fname_splitted = fname[:-4].replace('_', ' ').split()
if fname_splitted[-1][-1].isdigit():
photo_number = int(fname_splitted[-1])
photos.append(img)
mapping.append([index, '_'.join(fname_splitted)])
index += 1
photos = np.stack(photos).astype('uint8')
return photos, mapping
def on_epoch_begin(self, epoch, logs=None):
print('\nEpoch %d/%d' % (epoch + 1, self.epochs))
if "steps" in self.params:
self.use_steps = True
self.target = self.params['steps']
else:
self.use_steps = False
self.target = self.params['samples']
self.prog_bar = tqdm_utils.tqdm_notebook_failsafe(total=self.target)
self.log_values_by_metric = defaultdict(list)
def iterate_minibatches(inputs, targets, batchsize, shuffle=False):
assert len(inputs) == len(targets)
if shuffle:
indices = np.random.permutation(len(inputs))
for start_idx in tqdm_utils.tqdm_notebook_failsafe(range(0, len(inputs) - batchsize + 1, batchsize)):
if shuffle:
excerpt = indices[start_idx:start_idx + batchsize]
else:
excerpt = slice(start_idx, start_idx + batchsize)
yield inputs[excerpt], targets[excerpt]
def test_validation_loss(decoder, s, generate_batch, val_img_embeds,
val_captions_indexed):
np.random.seed(300)
random.seed(300)
val_loss = 0
batches_for_eval = 1000
for _ in tqdm_utils.tqdm_notebook_failsafe(range(batches_for_eval)):
val_loss += s.run(
decoder.loss,
generate_batch(val_img_embeds, val_captions_indexed, 32, 20))
val_loss /= 1000.
return val_loss
def download_file(url, file_path):
r = requests.get(url, stream=True)
total_size = int(r.headers.get('content-length'))
try:
with open(file_path, 'wb', buffering=16 * 1024 * 1024) as f:
bar = tqdm_utils.tqdm_notebook_failsafe(total=total_size,
unit='B',
unit_scale=True)
bar.set_description(os.path.split(file_path)[-1])
for chunk in r.iter_content(32 * 1024):
f.write(chunk)
bar.update(len(chunk))
bar.close()
except Exception:
print("Download failed")
finally:
if os.path.getsize(file_path) != total_size:
os.remove(file_path)
print("Removed incomplete download")
def download_file(url, file_path):
r = requests.get(url, stream=True)
total_size = int(r.headers.get('content-length'))
bar = tqdm_utils.tqdm_notebook_failsafe(total=total_size, unit='B', unit_scale=True)
bar.set_description(os.path.split(file_path)[-1])
incomplete_download = False
try:
with open(file_path, 'wb', buffering=16 * 1024 * 1024) as f:
for chunk in r.iter_content(4 * 1024 * 1024):
f.write(chunk)
bar.update(len(chunk))
except Exception as e:
raise e
finally:
bar.close()
if os.path.exists(file_path) and os.path.getsize(file_path) != total_size:
incomplete_download = True
os.remove(file_path)
if incomplete_download:
raise Exception("Incomplete download")
def load_lfw_dataset(use_raw=False, dx=80, dy=80, dimx=45, dimy=45):
# read attrs
df_attrs = pd.read_csv(ATTRS_NAME, sep='\t', skiprows=1)
df_attrs = pd.DataFrame(df_attrs.iloc[:, :-1].values,
columns=df_attrs.columns[1:])
imgs_with_attrs = set(map(tuple, df_attrs[["person", "imagenum"]].values))
# read photos
all_photos = []
photo_ids = []
with tarfile.open(RAW_IMAGES_NAME if use_raw else IMAGES_NAME) as f:
for m in tqdm_utils.tqdm_notebook_failsafe(f.getmembers()):
if m.isfile() and m.name.endswith(".jpg"):
# prepare image
img = decode_image_from_raw_bytes(f.extractfile(m).read())
img = img[dy:-dy, dx:-dx]
img = cv2.resize(img, (dimx, dimy))
# parse person
fname = os.path.split(m.name)[-1]
fname_splitted = fname[:-4].replace('_', ' ').split()
person_id = ' '.join(fname_splitted[:-1])
photo_number = int(fname_splitted[-1])
if (person_id, photo_number) in imgs_with_attrs:
all_photos.append(img)
photo_ids.append({
'person': person_id,
'imagenum': photo_number
})
photo_ids = pd.DataFrame(photo_ids)
all_photos = np.stack(all_photos).astype('uint8')
# preserve photo_ids order!
all_attrs = photo_ids.merge(df_attrs,
on=('person',
'imagenum')).drop(["person", "imagenum"],
axis=1)
return all_photos, all_attrs
def reading_thread(zip_fn):
zf = zipfile.ZipFile(zip_fn)
for fn in tqdm_utils.tqdm_notebook_failsafe(zf.namelist()):
if kill_read_thread.is_set():
break
if os.path.splitext(fn)[-1] in extensions:
buf = zf.read(fn) # read raw bytes from zip for fn
img = decode_image_from_buf(buf) # decode raw bytes
img = crop_and_preprocess(img, input_shape,
preprocess_for_model)
while True:
try:
q.put((os.path.split(fn)[-1], img),
timeout=1) # put in queue
except queue.Full:
if kill_read_thread.is_set():
break
continue
break
read_thread_completed.set() # read all images
# # Submit to Coursera
# In[25]:
# token expires every 30 min
COURSERA_TOKEN = "*****************"
COURSERA_EMAIL = "d****************m"
# In[26]:
from submit import submit_char_rnn
samples = [generate_sample(' Al') for i in tqdm_utils.tqdm_notebook_failsafe(range(25))]
submission = (history, samples)
submit_char_rnn(submission, COURSERA_EMAIL, COURSERA_TOKEN)
# # Try it out!
#
# __Disclaimer:__ This part of assignment is entirely optional. You won't receive bonus points for it. However, it's a fun thing to do. Please share your results on course forums.
#
# You've just implemented a recurrent language model that can be tasked with generating any kind of sequence, so there's plenty of data you can try it on:
#
# * Novels/poems/songs of your favorite author
# * News titles/clickbait titles
# * Source code of Linux or Tensorflow
# * Molecules in [smiles](https://en.wikipedia.org/wiki/Simplified_molecular-input_line-entry_system) format
# * Melody in notes/chords format
# you can load "weights_{epoch}" and continue training
# uncomment the next line if you need to load weights
# saver.restore(s, os.path.abspath("weights"))
"""Look at the training and validation loss, they should be decreasing!"""
# actual training loop
MAX_LEN = 20 # truncate long captions to speed up training
# to make training reproducible
np.random.seed(42)
random.seed(42)
for epoch in range(n_epochs):
train_loss = 0
pbar = tqdm_utils.tqdm_notebook_failsafe(range(n_batches_per_epoch))
counter = 0
for _ in pbar:
train_loss += s.run([decoder.loss, train_step],
generate_batch(train_img_embeds,
train_captions_indexed, batch_size,
MAX_LEN))[0]
counter += 1
pbar.set_description("Training loss: %f" % (train_loss / counter))
train_loss /= n_batches_per_epoch
val_loss = 0
for _ in range(n_validation_batches):
val_loss += s.run(
decoder.loss,
def sample_probas(bsize):
plt.title('Generated vs real data')
plt.hist(np.exp(discriminator.predict(sample_data_batch(bsize)))[:,1],
label='D(x)', alpha=0.5,range=[0,1])
plt.hist(np.exp(discriminator.predict(generator.predict(sample_noise_batch(bsize))))[:,1],
label='D(G(z))',alpha=0.5,range=[0,1])
plt.legend(loc='best')
plt.show()
"""### Training
Main loop.
We just train generator and discriminator in a loop and plot results once every N iterations.
"""
from IPython import display
for epoch in tqdm_utils.tqdm_notebook_failsafe(range(3000)):
feed_dict = {
real_data:sample_data_batch(100),
noise:sample_noise_batch(100)
}
for i in range(5):
s.run(disc_optimizer,feed_dict)
s.run(gen_optimizer,feed_dict)
if epoch %100==0:
display.clear_output(wait=True)
sample_images(2,3,True)
sample_probas(1000)
for _ in range(10):
print(generate_sample())
# with prefix conditioning
for _ in range(10):
print(generate_sample(' Trump'))
"""# Submit to Coursera"""
# token expires every 30 min
COURSERA_TOKEN = "XXXXXXXXXXXX"
COURSERA_EMAIL = "XXXXXXXXXXXXXXXXXXXXXXXXXX"
from submit import submit_char_rnn
samples = [
generate_sample(' Al')
for i in tqdm_utils.tqdm_notebook_failsafe(range(25))
]
submission = (history, samples)
submit_char_rnn(submission, COURSERA_EMAIL, COURSERA_TOKEN)
"""# Try it out!
__Disclaimer:__ This part of assignment is entirely optional. You won't receive bonus points for it. However, it's a fun thing to do. Please share your results on course forums.
You've just implemented a recurrent language model that can be tasked with generating any kind of sequence, so there's plenty of data you can try it on:
* Novels/poems/songs of your favorite author
* News titles/clickbait titles
* Source code of Linux or Tensorflow
* Molecules in [smiles](https://en.wikipedia.org/wiki/Simplified_molecular-input_line-entry_system) format
* Melody in notes/chords format
* IKEA catalog titles
* Pokemon names
* Cards from Magic, the Gathering / Hearthstone
If you're willing to give it a try, here's what you wanna look at: