Building models using GluonTS and encountered DeferredInitializationError ,NotImplementedError

[suyalong]

Hi, everyone, I am suya, a beginner using mxnet.
These days, when I built the Deepstate and Seq2Seq model, there were bugs that I really cannot solve. I tried many ways but couldn't solve it, please help!

This is my code:
weeks =[1424-7*i for i in range(15)]
iris = ['/content/IRI_DataSet/iri{}.csv'.format(i+1) for i in range(15)]
length = 55
data = []
iri_category = []
for file_path, w in zip(iris, weeks):
unit_columns = ['unit.' + str(i) for i in range(w, w+length)]
price_columns = ['price.' + str(i) for i in range(w, w+length)]
holiday_columns = ['holiday.' + str(i) for i in range(w, w+length)]
d_columns = ['D.' + str(i) for i in range(w, w+length)]
f_columns = ['F.' + str(i) for i in range(w, w+length)]
d = pd.read_csv(file_path)
U = d[unit_columns].values.astype(np.float32)
P = d[price_columns].values.astype(np.float32)
D = d[d_columns].values.astype(np.float32)
F = d[f_columns].values.astype(np.float32)
H = d[holiday_columns].values.astype(np.float32)
iri_category.append(d['unit.cate'].values)
combined = np.concatenate((U, P, D, F, H), axis=1).reshape(-1, 5, length)
data.append(combined)
iri_original = np.concatenate(data, axis=0)
iri_original[iri_original == np.inf] = 0

here, iri_original.shape = (120138, 5, 55)

then:
feat_static_cat = np.hstack(iri_category)
feat_static_cat = LabelEncoder().fit_transform(feat_static_cat)
feat_static_cat.reshape(len(feat_static_cat))

train_ds = ListDataset(
[{'target': iri_original[i, 0, :-prediction_length],
'feat_dynamic_real': iri_original[i, 1:2, :-prediction_length],
'feat_dynamic_cat': iri_original[i, 2:, :-prediction_length],
'feat_static_cat': feat_static_cat,
'start': start} for i in range(len(iri_original))],
freq=freq)

test_ds = ListDataset(
[{'target': iri_original[i, 0, :],
'feat_dynamic_real': iri_original[i, 1:2, :],
'feat_dynamic_cat': iri_original[i, 2:, :],
'feat_static_cat': feat_static_cat,
'start': start} for i in range(len(iri_original))],
freq=freq)

then build models:

estimator = DeepStateEstimator(
freq = freq,
prediction_length = prediction_length,
cardinality = [6],
embedding_dimension = [10],
num_layers = 1,
num_cells = 80,
#quantiles = [0.5],
trainer = Trainer(
epochs=100,
learning_rate=1e-4,
minimum_learning_rate=1e-10,
num_batches_per_epoch=32),
num_parallel_samples = 100,
#batch_size = 32,
use_feat_dynamic_real = True,
use_feat_static_cat = True,
)
and
estimator = Seq2SeqEstimator(
    freq = freq,
    prediction_length = prediction_length,
    cardinality = [2],
    embedding_dimension = 5,
    encoder = Seq2SeqEncoder("lstm"),
    decoder_mlp_layer = [1],
    decoder_mlp_static_dim = 16,
    scaler = NOPScaler(),
    context_length = 14,
    #quantiles = [0.5],
    trainer = Trainer(epochs = 10, num_batches_per_epoch= 32, learning_rate=0.001),
    num_parallel_samples = 100,
    batch_size = 32
)

However, there were problems:

1. DeferredInitializationError (Deepstate)
   0%| | 0/32 [00:00<?, ?it/s]infer_shape error. Arguments:
   data0: (32, 120138)
   data1: (32, 208, 1)
   data2: (32, 208, 2)
   data3: (32, 208, 4)
   data4: (32, 208, 1)

---

DeferredInitializationError Traceback (most recent call last)
/usr/local/lib/python3.7/dist-packages/mxnet/gluon/block.py in _build_cache(self, *args)
950 if name in params:
--> 951 params[name].data()
952 except DeferredInitializationError:
ValueError: Deferred initialization failed because shape cannot be inferred. MXNetError: Error in operator deepstatetrainingnetwork13_featureembedder0_squeeze0: [10:43:29] ../src/operator/tensor/./matrix_op-inl.h:2365: Check failed: dshape[axes[i]] == 1 (120138 vs. 1) : cannot select an axis to squeeze out which has size=120138 not equal to one

2. NotImplementedError (Seq2Seq)

0%| | 0/32 [00:00<?, ?it/s]

NotImplementedError Traceback (most recent call last)
in ()
----> 1 predictor = estimator.train(train_ds)

13 frames
/usr/local/lib/python3.7/dist-packages/gluonts/mx/block/encoder.py in hybrid_forward(self, F, target, static_features, dynamic_features)
63 shape (batch_size, sequence_length, num_feat_dynamic)
64 """
---> 65 raise NotImplementedError
66
67 def _assemble_inputs(

NotImplementedError: