def make_series(model_root_folder,
plot_interval=100,
limit=None,
no_new=False,
dont_average_embeddings=False,
**run_model_args):
average_embeddings = not dont_average_embeddings
if average_embeddings:
suffix = 'eval-averaged.pkl'
else:
suffix = 'eval-None-False.pkl' # holdout from when we had include_synsets and normalize_components
store_fname = join(model_root_folder, suffix)
print store_fname
try:
stats = pandas.read_pickle(store_fname)
print "read pickle from %s" % store_fname
except:
stats = pandas.DataFrame()
print 'created new frame'
if no_new:
return stats
models = models_in_folder(model_root_folder)
model_nums = sorted(models.keys())
latest_num = model_nums[-1] if model_nums else -1
latest_num = -1
print 'plotting every %i' % plot_interval
to_plot = [n for n in model_nums if n % plot_interval == 0 and n != latest_num]
if 1 in model_nums:
to_plot = [1] + to_plot
if limit is not None:
to_plot = [n for n in to_plot if n <= limit]
vocab_container = None
print model_root_folder
for n in to_plot:
if n in stats.index:
print 'already has %i' % n
continue
try:
print 'loading %i' % n
with gzip.open(models[n]) as f:
model = cPickle.load(f)
except Exception as e:
print e
continue
# load the vocabulary if not already cached
if not vocab_container:
vocab_container = get_vocab_container(model)
embeddings = model.embeddings if not average_embeddings else model.averaged_embeddings()
this_stats = run_model(embeddings, vocab_container, **run_model_args)
stats = pandas.concat([stats, pandas.DataFrame([this_stats], index=[n])]).sort()
stats.to_pickle(store_fname)
return stats
def dump_params_from_directory(model_directory, dump_filename='params.json'):
try:
models = models_in_folder(model_directory)
first_model = min(models)
model_path = models[first_model]
with gzip.open(model_path, 'rb') as f:
model = cPickle.load(f)
except StopIteration:
print 'no models found in %s' % model_directory
return
except IOError as e:
print 'IO error for %s' % model_directory
print e
return
dump_params(model.other_params, join(model_directory, dump_filename))
def dump_params_from_directory(model_directory, dump_filename='params.json'):
try:
models = models_in_folder(model_directory)
first_model = min(models)
model_path = models[first_model]
with gzip.open(model_path, 'rb') as f:
model = cPickle.load(f)
except StopIteration:
print 'no models found in %s' % model_directory
return
except IOError as e:
print 'IO error for %s' % model_directory
print e
return
dump_params(model.other_params, join(model_directory, dump_filename))
def models_to_delete(directory, retain_every=10):
models = models_in_folder(directory)
if not models:
print 'no models found'
return {}, set()
to_keep = set([n for n in models if n % retain_every == 0])
# don't delete the first one, so we can use it as a reference point
to_keep.add(min(models))
# the last one may be currently being written
to_keep.add(max(models))
return dict((model, path) for model, path in models.items()
if model not in to_keep), to_keep
def models_to_delete(directory, retain_every=10):
models = models_in_folder(directory)
if not models:
print 'no models found'
return {}, set()
to_keep = set([n for n in models
if n % retain_every == 0])
# don't delete the first one, so we can use it as a reference point
to_keep.add(min(models))
# the last one may be currently being written
to_keep.add(max(models))
return dict((model, path)
for model, path in models.items()
if model not in to_keep), to_keep
parser.add_argument('--semantic_blocks_to_run', type=int, default=1)
args = vars(parser.parse_args())
# if we're only running semantic or syntactic, rho and y init must be 0 to
# isolate the loss function to the syntactic or semantic loss
if args['dont_run_semantic'] or args['dont_run_syntactic']:
print 'not running joint model, setting y and rho to 0'
args['rho'] = 0
args['y_init'] = 0
base_dir = args['base_dir']
# see if this model's already been run. If it has, load it and get the
# params
models = models_in_folder(base_dir)
if models:
model_num = max(models.keys())
print 'loading existing model %s' % models[model_num]
with gzip.open(models[model_num]) as f:
model = cPickle.load(f)
model_loaded = True
args = model.other_params
if 'vsgd' not in args: # backward compatibility
args['vsgd'] = False
if 'simple_joint' not in args: # backward compatibility
args['simple_joint'] = False
# rewrite in case we've copied the model file into this folder
args['base_dir'] = base_dir
else:
parser.add_argument('--existing_semantic_model', help='use this existing trained model as the semantic model')
parser.add_argument('--semantic_learning_rate', type=float, default=0.01)
parser.add_argument('--semantic_tensor_n_hidden', type=int, default=50)
# parser.add_argument('--semantic_block_size', type=int, default=100000)
# parser.add_argument('--sem_validation_num_nearest', type=int, default=50, help='when running semantic validation after each round, look at the intersection of top N words in wordnet and top N by embedding for a given test word')
# parser.add_argument('--sem_validation_num_to_test', type=int, default=500, help='in semantic validation after each round, the number of test words to sample')
parser.add_argument('--semantic_blocks_to_run', type=int, default=1)
args = vars(parser.parse_args())
print args
base_dir = args['base_dir']
# see if this model's already been run. If it has, load it and get the
# params
models = models_in_folder(base_dir)
if models:
model_num = max(models.keys())
print 'loading existing model %s' % models[model_num]
with gzip.open(models[model_num]) as f:
model = cPickle.load(f)
model_loaded = True
args = model.other_params
# rewrite in case we've copied the model file into this folder
args['base_dir'] = base_dir
else:
model_loaded = False
# dump the params
with open(os.path.join(args['base_dir'], 'params.json'), 'w') as f:
json.dump(args, f)
def load_models(base_path, indices=range(0, 1050, 50)):
model_paths = utils.models_in_folder(base_path)
return { index: load_model(path) for index, path in model_paths.iteritems()
if index in indices }
