WIP: Failing test case for initial state dependency

tests/test_TFNetworkRecLayer.py

@@ -325,6 +325,56 @@ def test_nativelstm2_initial_state_keep_epoch():
   _check_train_simple_network({
     "output": {"class": "rec", "unit": "nativelstm2", "loss": "mse", "initial_state": "keep_over_epoch"}})
 
+def test_lstmblock_initial_state_dependency():
+  beam_size = 2
+  network = {
+    "data_embed": {"class": "linear", "activation": None, "with_bias": False, "n_out": 6},
+    "lstm0_fw" : { "class": "rec", "unit": "nativelstm2", "n_out" : 5, "direction": 1, "from": ["data_embed"] },
+
+    "encoder_state": {"class": "get_last_hidden_state", "from": ["lstm0_fw"], 'key': 'c', "n_out": 5},
+
+    "output": {"class": "rec", "from": [], "unit": {
+      'output': {'class': 'choice', 'target': 'classes', 'beam_size': beam_size, 'from': ["output_prob"],
+                 "initial_output": 0},
+      "end": {"class": "compare", "from": ["output"], "value": 0},
+      'orth_embed': {'class': 'linear', 'activation': None, "with_bias": False, 'from': ['prev:output'], "n_out": 6},
+
+      "s": {"class": "rnn_cell", "unit": "LSTMBlock", "from": ["orth_embed"], "initial_state": {"c": "base:encoder_state", "h": 0}, "n_out": 5},  # h_t
+
+      "output_prob": {"class": "softmax", "from": ["s"], "target": "classes", "loss": "ce"}
+    }, "target": "classes", "max_seq_len": 7},
+  }
+
+  from GeneratingDataset import DummyDataset
+  seq_len = 5
+  n_data_dim = 2
+  n_classes_dim = 3
+  train_data = DummyDataset(input_dim=n_data_dim, output_dim=n_classes_dim, num_seqs=4, seq_len=seq_len)
+  train_data.init_seq_order(epoch=1)
+  cv_data = DummyDataset(input_dim=n_data_dim, output_dim=n_classes_dim, num_seqs=2, seq_len=seq_len)
+  cv_data.init_seq_order(epoch=1)
+
+  config = Config()
+  config.update({
+    "model": "/tmp/model",
+    "num_outputs": n_classes_dim,
+    "num_inputs": n_data_dim,
+    "network": network,
+    "start_epoch": 1,
+    "num_epochs": 2,
+    "batch_size": 10,
+    "nadam": True,
+    "learning_rate": 0.01,
+    "debug_add_check_numerics_ops": True
+  })
+
+  from tests.test_TFEngine import _cleanup_old_models
+  from TFEngine import Engine
+  _cleanup_old_models(config)
+  engine = Engine(config=config)
+  engine.init_train_from_config(config=config, train_data=train_data, dev_data=cv_data, eval_data=None)
+  engine.train()
+  engine.search(cv_data)
 
 def test_slow_TensorArray():
   """