def _write_traces(self, split_name, train_steps, edit_traces, loss_traces):
trace_dir = join(self.workspace.traces, split_name)
trace_path = join(trace_dir, '{}.txt'.format(train_steps))
makedirs(trace_dir)
with codecs.open(trace_path, 'w', encoding='utf-8') as f:
for edit_trace, loss_trace in zip(edit_traces, loss_traces):
f.write(unicode(edit_trace))
f.write('\n')
f.write(unicode(loss_trace))
f.write('\n\n')
def show(title, directory=''):
"""If in IPython, show, otherwise, save to file."""
import matplotlib.pyplot as plt
if in_ipython():
plt.show()
else:
# ensure directory exists
makedirs(directory)
plt.savefig(os.path.join(directory, title) + '.png')
# close all figures to conserve memory
plt.close('all')
def save(self, path):
makedirs(path)
# Store the latest random state
self.random_state = RandomState()
# save model
torch.save(self.model.state_dict(), join(path, 'model'))
torch.save(self.optimizer.state_dict(), join(path, 'optimizer'))
# pickle remaining attributes
d = {
attr: getattr(self, attr)
for attr in ['train_steps', 'random_state', 'max_grad_norm']
}
with open(join(path, 'metadata.pkl'), 'wb') as f:
pickle.dump(d, f)
def _log_traces(self, episodes, label, train_step):
trace_dir = join(self.trace_dir, label)
makedirs(trace_dir)
trace_path = join(trace_dir, str(train_step))
episode_traces = [MiniWoBEpisodeTrace(ep) for ep in episodes]
# save machine-readable version
with open(trace_path + '.json', 'w', 'utf8') as f:
trace_dicts = [trace.to_json_dict() for trace in episode_traces]
# print(trace_dicts[0])
json.dump(str(trace_dicts), f, indent=2)
# save pretty-printed version
with open(trace_path + '.txt', 'w', 'utf8') as f:
for i, trace in enumerate(episode_traces):
f.write("=" * 25 + " EPISODE {} ".format(i) + "=" * 25)
f.write('\n\n')
f.write(trace.dumps())
f.write('\n\n')
# save screenshots
for i, ep in enumerate(episodes):
if not self._has_screenshot(ep):
continue
img_path = trace_path + '-img'
makedirs(img_path)
actions = []
for j, experience in enumerate(ep):
state, action = experience.state, experience.action
path = join(img_path, '{}-{}-{}.png'.format(train_step, i, j))
state.screenshot.save(path)
actions.append(action.to_dict())
# write action summary
path = join(img_path, '{}-{}.json'.format(train_step, i))
with open(path, 'w') as fout:
json.dump(actions, fout)
def __init__(self, absolute_path, sync=True):
self._name = os.path.basename(os.path.normpath(absolute_path))
self._absolute_path = absolute_path
self._sync = sync
makedirs(absolute_path)
self._subdirs = []
def clear_cache(self):
shutil.rmtree(self.cache_dir)
from gtd.io import makedirs
makedirs(self.cache_dir)
def tensorboard_dir(self):
p = self.get_tensorboard_dir(self.save_dir)
makedirs(p)
return p
def checkpoint_dir(self):
p = self.get_checkpoint_dir(self.save_dir)
makedirs(p)
return p
def tensorboard_dir(self):
p = self.get_tensorboard_dir(self.save_dir)
makedirs(p)
return p
def checkpoint_dir(self):
p = self.get_checkpoint_dir(self.save_dir)
makedirs(p)
return p
def clear_cache(self):
shutil.rmtree(self.cache_dir)
makedirs(self.cache_dir)
def __init__(self, fxn, cache_dir, serialize, deserialize):
super(FileMemoized, self).__init__(fxn)
self.cache_dir = cache_dir
self.serialize = serialize
self.deserialize = deserialize
makedirs(cache_dir)
def __init__(self, absolute_path, sync=True):
self._name = os.path.basename(os.path.normpath(absolute_path))
self._absolute_path = absolute_path
self._sync = sync
makedirs(absolute_path)
self._subdirs = []
