def save_img(object, url, domain=None, **kwargs):
if isinstance(object, np.ndarray):
normalized = _normalize_array(object, domain=domain)
object = PIL.Image.fromarray(normalized)
if isinstance(object, PIL.Image.Image):
object.save(url, **kwargs) # will infer format from url's url ext.
else:
raise ValueError("Can only save_img for numpy arrays or PIL.Images!")
def save_img(object, handle, **kwargs):
"""Save numpy array as image file on CNS."""
if isinstance(object, np.ndarray):
normalized = _normalize_array(object)
image = PIL.Image.fromarray(normalized)
elif isinstance(object, PIL.Image):
image = object
else:
raise ValueError("Can only save_img for numpy arrays or PIL.Images!")
image.save(handle, **kwargs) # will infer format from handle's url ext.
def save_img(object, handle, domain=None, **kwargs):
"""Save numpy array as image file on CNS."""
if isinstance(object, np.ndarray):
normalized = _normalize_array(object, domain=domain)
object = PIL.Image.fromarray(normalized)
if isinstance(object, PIL.Image.Image):
object.save(handle,
**kwargs) # will infer format from handle's url ext.
else:
raise ValueError("Can only save_img for numpy arrays or PIL.Images!")
return {
"type": "image",
"shape": object.size + (len(object.getbands()), ),
"url": handle.name,
}
def optimize_input(obj,
model,
param_f,
transforms,
lr=0.05,
step_n=512,
num_output_channels=4,
do_render=False,
out_name="out"):
sess = create_session()
# Set up optimization problem
size = 84
t_size = tf.placeholder_with_default(size, [])
T = render.make_vis_T(
model,
obj,
param_f=param_f,
transforms=transforms,
optimizer=tf.train.AdamOptimizer(lr),
)
tf.global_variables_initializer().run()
if do_render:
video_fn = out_name + '.mp4'
writer = FFMPEG_VideoWriter(video_fn, (size, size * 4), 60.0)
# Optimization loop
try:
for i in range(step_n):
_, loss, img = sess.run([T("vis_op"), T("loss"), T("input")])
if do_render:
#if outputting only one channel...
if num_output_channels == 1:
img = img[..., -1:] #print(img.shape)
img = np.tile(img, 3)
else:
#img=img[...,-3:]
img = img.transpose([0, 3, 1, 2])
img = img.reshape([84 * 4, 84, 1])
img = np.tile(img, 3)
writer.write_frame(_normalize_array(img))
if i > 0 and i % 50 == 0:
clear_output()
print("%d / %d score: %f" % (i, step_n, loss))
show(img)
except KeyboardInterrupt:
pass
finally:
if do_render:
print("closing...")
writer.close()
# Save trained variables
if do_render:
train_vars = sess.graph.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES)
params = np.array(sess.run(train_vars), object)
save(params, out_name + '.npy')
# Save final image
final_img = T("input").eval({t_size: 600})[..., -1:] #change size
save(final_img, out_name + '.jpg', quality=90)
out = T("input").eval({t_size: 84})
sess.close()
return out
