def before_step(self, step, fetches, feeds, batch):
"""
Steps taken before the training step.
Parameters
----------
step
Training step.
fetches
Fetches for the next session.run call.
feeds
Feeds for the next session.run call.
batch
The batch to be iterated over.
"""
def to_image(obj):
if isinstance(obj, np.ndarray) and len(obj.shape) == 3:
batches, height, width = obj.shape
obj = obj.reshape(batches, height, width, 1)
if isinstance(obj, np.ndarray) and len(obj.shape) == 4:
return obj.transpose(0, 3, 1, 2)
else:
return obj
walk(feeds, to_image, inplace=True)
super().before_step(step, fetches, feeds, batch)
def plot_datum(
nested_thing,
savename="datum.png",
heuristics=default_heuristic,
plt_functions=PLOT_FUNCTIONS,
):
"""Plots all data in the nested_thing as best as can.
If heuristics is given, this determines how each leaf datum is converted
to something plottable.
Parameters
---------
nested_thing : dict or list
Some nested object.
savename : str
``Path/to/the/plot.png``.
heuristics : Callable
If given this should produce a string specifying
the kind of data of the leaf. If ``None`` determinde automatically.
See :func:`default_heuristic`.
plt_functions : dict of Callables
Maps a ``kind`` to a function which
can plot it. Each callable must be able to receive a the key, the
leaf object and the Axes to plot it in.
"""
class Plotter(object):
def __init__(self, kind_fn, savename):
self.kind_fn = kind_fn
self.savename = savename
self.buffer = []
def __call__(self, key, obj):
kind = self.kind_fn(key, obj)
self.buffer += [[kind, key, obj]]
def plot(self):
n_pl = len(self.buffer)
f = plt.figure(figsize=(5, 2 * n_pl))
gs = gridspec.GridSpec(n_pl, 1)
for i, [kind, key, obj] in enumerate(self.buffer):
ax = f.add_subplot(gs[i])
plt_functions[kind](key, obj, ax)
f.savefig(self.savename)
def __str__(self):
self.plot()
return "Saved Plot at {}".format(self.savename)
P = Plotter(heuristics, savename)
walk(nested_thing, P, pass_key=True)
print(P)
def __init__(self, root):
"""
Parameters
----------
root : str
Where to look for all the data.
"""
meta_path = os.path.join(root, "meta.yaml")
self.meta = meta = yaml.safe_load(open(meta_path, "r"))
labels = load_labels(os.path.join(root, "labels"))
self.loaders, self.loader_kwargs = setup_loaders(labels, meta)
self.labels = clean_keys(labels, self.loaders)
class Lenner:
def __init__(self):
self.l = None
self.visited = []
def __call__(self, key, label):
if self.l is None:
self.l = len(label)
else:
if len(label) != self.l:
raise ValueError(f"Label {key} has a different length "
"than the other labels.\n"
f"Already seen: {self.visited}")
self.visited += [key]
L = Lenner()
walk(self.labels, L, pass_key=True)
self.num_examples = L.l
self.append_labels = True
def convert_logs2numpy(logs):
def conditional_convert2np(log_item):
if isinstance(log_item, torch.Tensor):
log_item = log_item.detach().cpu().numpy()
return log_item
# convert to numpy
walk(logs, conditional_convert2np, inplace=True)
return logs
def show_example(dset, idx):
ex = dset[idx]
st.header("Keys")
walk(ex, display, pass_key=True)
st.header("Summary")
summary = pp2mkdtable(ex, jupyter_style=True)
# print markdown summary on console for easy copy and pasting in readme etc
print(summary)
st.markdown(summary)
def test_walk_pass_key_inplace():
dol = {"a": [1, 2], "b": {"c": {"d": 1}}, "e": 2}
ref = {"a": [-1, -2], "b": {"c": {"d": -1}}, "e": -2}
def fn(key, leaf):
return -leaf
walk(dol, fn, inplace=True, pass_key=True)
assert dol == ref
def get_example(self, idx):
"""Get the examples from the base dataset at defined at ``view[idx]``.
"""
def get_view(view):
return view[idx]
view = walk(self.views, get_view)
view_example = walk(view, self.base.__getitem__, walk_np_arrays=True)
return view_example
def after_step(self, step, results):
def convert(var_or_tens):
if hasattr(var_or_tens, 'cpu'):
var_or_tens = var_or_tens.cpu()
if isinstance(var_or_tens, torch.autograd.Variable):
return var_or_tens.data.numpy()
elif isinstance(var_or_tens, torch.Tensor):
return var_or_tens.numpy()
else:
return var_or_tens
walk(results, convert, inplace=True)
def example_callbacks(example, dset_handler):
'''Creates all interactive connections to update and visualize the content
of examples.'''
Cooon = Connector(dset_handler)
walk(example, Cooon, pass_key=True)
callbacks = Cooon.callbacks
callbacks['toggle_all'] = {
'args': Cooon.toggle_all_ins_and_outs,
'callback': ToggleAllCallback(len(callbacks), True)
}
return callbacks
def get_example(self, idx):
"""Get the examples from the base dataset at defined at ``view[idx]``. Load loaders if applicable.
"""
def get_view(view):
return view[idx]
view = walk(self.views, get_view)
view_example = walk(view, self.base.__getitem__, walk_np_arrays=True)
if len(self.loaders) > 0:
loaders_example = super().get_example(idx)
view_example.update(loaders_example)
return view_example
def update_config(config, additional_kwargs):
"""additional_kwargs are added in order of the keys' length, e.g. 'a'
is overriden by 'a/b'."""
keys = sorted(additional_kwargs.keys())
for k in keys:
set_value(config, k, additional_kwargs[k])
def replace(k):
if isinstance(k, str) and k[0] == "{" and k[-1] == "}":
k_ = k[1:-1].strip()
return retrieve(config, k_, default=k)
else:
return k
walk(config, replace, inplace=True)
def before_step(self, step, fetches, feeds, batch):
def convert(obj):
if isinstance(obj, np.ndarray):
try:
obj = torch.tensor(obj)
obj = obj.to(self.dtype)
if self.use_gpu:
obj = obj.cuda()
return obj
except Exception:
return obj
else:
return obj
walk(feeds, convert, inplace=True)
def test_meta_view_dset():
N = 100
V = 25
try:
super_root, base_root, view_root = _setup(".", N, V)
M = MetaViewDataset(view_root)
M.expand = True
M.append_labels = False
M.show()
assert len(M) == V
for kk in ["simple1", "simple", "complex"]:
assert kk in M.labels
if kk == "complex":
for i in range(2):
for k in ["attr1", "attr2", "image_", "keypoints"]:
assert k in M.labels[kk][i]
assert len(M.labels[kk][i][k]) == V
else:
for k in ["attr1", "attr2", "image_", "keypoints"]:
assert k in M.labels[kk]
assert len(M.labels[kk][k]) == V
d = M[0]
# For ex 0 this is the same for both complex and simple
single_ref = {"image": np.ones(shape=(64, 64, 3)), "index_": 0}
ref_simple = single_ref
ref_complex = [[single_ref] * 3] * 20
ref = {
"simple1": ref_simple,
"simple": ref_simple,
"complex": [ref_complex, ref_simple],
"index_": 0,
}
def tester(key, val):
assert np.all(val == retrieve(ref, key))
walk(d, tester, pass_key=True)
assert hasattr(M, "meta")
finally:
_teardown(super_root)
def _maybe_append_labels(self, datum, index):
if self.append_labels:
def label_getter(labels):
return labels[index]
labels = walk(self.labels, label_getter)
update(datum, {"labels_": labels})
def prepare_inputs_inplace(self, inputs):
'''Casts all input to torch Tensor and pushes them to the gpu.'''
before = time.time()
inputs = walk(inputs, np2pt, inplace=True)
if retrieve(self.config, "debug_timing", default=False):
self.logger.info("prepare of data needed {} s".format(time.time() - before))
def test_walk():
dol = {"a": [1, 2], "b": {"c": {"d": 1}}, "e": 2}
ref = {"a": [-1, -2], "b": {"c": {"d": -1}}, "e": -2}
def fn(leaf):
return -leaf
val = walk(dol, fn)
assert val == ref
def test_walk_pass_key():
dol = {"a": [1, 2], "b": {"c": {"d": 1}}, "e": 2}
ref = {"a": [-1, -2], "b": {"c": {"d": -1}}, "e": -2}
def fn(key, leaf):
return -leaf
val = walk(dol, fn, pass_key=True)
assert val == ref
def load_labels(root):
"""
Parameters
----------
root : str
Where to look for the labels.
Returns
-------
labels : dict
All labels as ``np.memmap`` s.
"""
regex = re.compile(r".*-\*-.*-\*-.*\.npy")
label_files = _get_label_files(root)
class Loader:
def __init__(self):
self.labels = {}
def __call__(self, key_path, path):
if isinstance(path, str) and regex.match(path):
f = os.path.basename(path)
f_ = f[: -len(".npy")]
key_, shape, dtype = f_.split("-*-")
shape = tuple([int(s) for s in shape.split("x")])
key_path = key_path.split("/")
if len(key_path) == 1:
key = key_
else:
key = "/".join(key_path[:-1] + [key_])
mmap = np.memmap(path, mode="c", shape=shape, dtype=dtype)
set_value(self.labels, key, mmap)
L = Loader()
walk(label_files, L, pass_key=True)
return L.labels
def before_step(self, *args, **kwargs):
"""Checks if something changed and if yes runs the callback."""
try:
updates = yaml.full_load(open(self.ufile, "r"))
if self.last_updates is not None:
changes = {}
def is_changed(key, val, changes=changes):
if contains_key(key, updates):
other_val = retrieve(key, updates)
change = np.any(val != other_val)
else:
# This key is new -> Changes did happen!
change = True
changes[key] = change
self.logger.debug("Pre CHANGES: {}".format(changes))
walk(self.last_updates, is_changed, pass_key=True)
self.logger.debug("Post CHANGES: {}".format(changes))
if np.any(list(changes.values())):
self.callback(updates)
self.logger.debug("Runtime inputs received.")
self.logger.debug("{}".format(updates))
self.last_updates = updates
else:
if updates is not None:
self.callback(updates)
self.logger.info("Runtime inputs received.")
self.logger.debug("{}".format(updates))
self.last_updates = updates
except Exception as e:
self.logger.error("Something bad happend :(")
self.logger.error("{}".format(e))
self.logger.error(traceback.format_exc())
def display_controls(datasource_1_value):
'''Makes sure, that the slider changes the displayed example.
'''
# Get example
ex = Ht[int(datasource_1_value)]
# display leaf variables
de = DisplayElements()
walk(ex, de, pass_key=True)
content = de.elements
connector = Connector()
walk(ex, connector, pass_key=True)
for key, connection in connector.callbacks.items():
print(key, connection)
app.callback(**connection['args'])(connection['callback'])
return html.Div(content, )
def after_step(self, step, results):
"""
Steps taken after the training step.
:param step: Training step.
:param results: Result of the session.
:return:
"""
super().after_step(step, results)
def to_image(k, obj):
if (
"weights" not in k
and isinstance(obj, np.ndarray)
and len(obj.shape) == 4
):
return obj.transpose(0, 2, 3, 1)
else:
return obj
walk(results, to_image, inplace=True, pass_key=True)
def clean_keys(labels, loaders):
"""Removes all loader information from the keys.
Parameters
----------
labels : dict(str, numpy.memmap)
Labels contain all load-easy dataset relevant data.
Returns
-------
labels : dict(str, numpy.memmap)
The original labels, with keys without the ``:loader`` part.
"""
class Cleaner:
def __init__(self):
self.to_delete = []
self.to_set = []
def __call__(self, key, val):
k, l = loader_from_key(key)
if l is not None:
self.to_set += [[k + "_", retrieve(labels, key)]]
self.to_delete += [key]
C = Cleaner()
walk(labels, C, pass_key=True)
for key, val in C.to_set:
set_value(labels, key, val)
for key in C.to_delete:
pop_keypath(labels, key)
for k_ in list(loaders.keys()):
if k_ in labels:
k = k_ + "_"
labels[k] = labels[k_]
del labels[k_]
return labels
def __init__(self, root):
super().__init__(root)
base_import = retrieve(self.meta, "base_dset")
base_kwargs = retrieve(self.meta, "base_kwargs")
self.base = get_obj_from_str(base_import)(**base_kwargs)
self.base.append_labels = False
views = retrieve(self.meta, "views", default="view")
def get_label(key):
return retrieve(self.labels, key)
self.views = walk(views, get_label)
if not os.path.exists(os.path.join(root, ".constructed.txt")):
def constructor(name, view):
folder_name = name
savefolder = os.path.join(root, "labels", folder_name)
os.makedirs(savefolder, exist_ok=True)
for key, label in tqdm(self.base.labels.items(),
desc=f"Exporting View {name}"):
savepath = os.path.join(root, "labels", name)
label_view = np.take(label, view, axis=0)
store_label_mmap(label_view, savepath, key)
walk(self.views, constructor, pass_key=True)
with open(os.path.join(root, ".constructed.txt"), "w+") as cf:
cf.write("Do not delete, this reduces loading times.\n"
"If you need to re-render the view, you can safely "
"delete this file.")
# Re-initialize as we need to load the labels again.
super().__init__(root)
def prepare_logs(self, inputs, predictions, losses, model, granularity):
'''Logs need to be differentiated into ``images`` and ``scalars``. This
function casts everything we want to log to numpy and stores it
correctly in the output log dictionary.
'''
# sample variational part
output_sample = model(inputs["pt"], mode="sample_appearance")
output_sample = {'image': output_sample}
output_sample.update(model.saved_tensors)
losses_sample = self.criterion(inputs["pt"], output_sample)
sample_images = {
"images_prediction_sample": pt2np(output_sample["image"]),
}
# concatenate logs
logs = {
"images": {
"appearance": inputs["np"]["appearance"],
"target": inputs["np"]["target"],
"pose": inputs["np"]["pose"],
"images_prediction": pt2np(predictions['image']),
},
"scalars": {
**losses[granularity]
},
}
# convert to numpy
def conditional_convert2np(log_item):
if isinstance(log_item, torch.Tensor):
log_item = log_item.detach().cpu().numpy()
return log_item
walk(logs, conditional_convert2np, inplace=True)
return logs
def test_meta_dset():
N = 100
try:
root = _setup(".", N)
M = MetaDataset(root)
M.expand = True
M.show()
assert len(M) == N
for k in ["attr1", "attr2", "image_", "keypoints"]:
assert k in M.labels
assert len(M.labels[k]) == N
d = M[0]
ref = {
"image": np.ones(shape=(64, 64, 3)),
"index_": 0,
"labels_": {
"image_": os.path.join(root, "images", "000.png"),
"attr1": 0,
"attr2": np.zeros((2)),
"keypoints": np.ones((17, 2)),
},
}
def tester(key, val):
assert np.all(val == retrieve(ref, key))
walk(d, tester, pass_key=True)
assert hasattr(M, "meta")
finally:
_teardown(root)
def test_walk_npon():
import numpy as np
dol = {"a": np.array([1, 2]), "b": {"c": {"d": 1}}, "e": 2}
ref = {"a": np.array([-1, -2]), "b": {"c": {"d": -1}}, "e": -2}
def fn(leaf):
return -leaf
val = walk(dol, fn, walk_np_arrays=True)
assert np.all(val["a"] == ref["a"])
del val["a"]
del ref["a"]
assert val == ref
def run(self, fetches, feed_dict):
"""Runs all fetch ops and stores the results.
Args:
fetches (dict): name: Callable pairs.
feed_dict (dict): Passed as kwargs to all fetch ops
Returns:
dict: name: results pairs.
"""
def fn(fetch_fn):
return fetch_fn(self.model, **feed_dict)
results = walk(fetches, fn)
return results
def run(self, fetches, feed_dict):
"""Runs all fetch ops and stores the results.
Parameters
----------
fetches : dict
name: Callable pairs.
feed_dict : dict
Passed as kwargs to all fetch ops
Returns
-------
dict
name: results pairs.
"""
def fn(fetch_fn):
return fetch_fn(self.model, **feed_dict)
results = walk(fetches, fn)
return results
def make_feeds(self, batch):
# copy of batches
feeds = walk(batch, lambda val: val)
return feeds
def sizes(t):
return walk(t, sizes_)