with open("%s/args.pkl" % exp_dir, "rb") as f: print "load hparams from %s/args.pkl" % exp_dir hparams = cPickle.load(f) print hparams used_labs = hparams.facs.split(':') c_n = OrderedDict([(used_labs[0], 5), (used_labs[1], 3)]) # hack used_talabs = hparams.talab_facs.split(':') b_n = OrderedDict([(used_talabs[0], 48)]) # hack if dataset == None: dataset = hparams.dataset dt_iterator, dt_iterator_by_seqs, dt_seqs, dt_seq2lab_d = \ load_data(dataset, set_name, hparams.is_numpy, seqlist) FHVAE = load_model(hparams.model) if hasattr(hparams, "nmu2"): print "model trained with hierarchical sampling, nmu2=%s" % hparams.nmu2 nmu2 = hparams.nmu2 else: print "model trained with normal training, nmu2=%s" % hparams.tr_nseqs nmu2 = hparams.tr_nseqs tf.reset_default_graph() xin = tf.placeholder(tf.float32, shape=(None,)+hparams.tr_shape, name="xin") xout = tf.placeholder(tf.float32, shape=(None,)+hparams.tr_shape, name="xout") y = tf.placeholder(tf.int64, shape=(None,), name="y") n = tf.placeholder(tf.float32, shape=(None,), name="n") #cReg = tf.placeholder(tf.int64, shape=(None,len(used_labs)), name="cReg")
parser.add_argument("--seqlist", type=str, default=None, help="specify a list of sequences to evaluate; randomly sample 10 by default") parser.add_argument("--step", type=int, default=-1, help="step of the model to load. -1 for the best") args = parser.parse_args() exp_dir, set_name, seqlist, step = args.exp_dir, args.set_name, args.seqlist, args.step with open("%s/args.pkl" % exp_dir, "rb") as f: print "load arguments from %s/args.pkl" % exp_dir args = cPickle.load(f) print args dt_iterator, dt_iterator_by_seqs, dt_seqs, dt_seq2lab_d = \ load_data(args.dataset, set_name, args.is_numpy, seqlist) FHVAE = load_model(args.model) if hasattr(args, "nmu2"): print "model trained with hierarchical sampling, nmu2=%s" % args.nmu2 nmu2 = args.nmu2 else: print "model trained with normal training, nmu2=%s" % args.tr_nseqs nmu2 = args.tr_nseqs tf.reset_default_graph() xin = tf.placeholder(tf.float32, shape=(None,)+args.tr_shape, name="xin") xout = tf.placeholder(tf.float32, shape=(None,)+args.tr_shape, name="xout") y = tf.placeholder(tf.int64, shape=(None,), name="y") n = tf.placeholder(tf.float32, shape=(None,), name="n") model = FHVAE(xin, xout, y, n, nmu2)