ported pytorch weights

config.json

@@ -1,13 +1,13 @@
 {
   "architectures": [
-    "MCTCModel"
+    "MCTCForCTC"
   ],
   "attention_head_dim": 384,
   "attention_probs_dropout_prob": 0.3,
   "bos_token_id": 0,
   "conv_channels": null,
   "conv_dropout": 0.3,
-  "conv_glu_dim": 2,
+  "conv_glu_dim": 1,
   "conv_kernel": [
     7
   ],
@@ -24,13 +24,15 @@
   "input_channels": 1,
   "input_feat_per_channel": 80,
   "intermediate_size": 6144,
-  "layer_norm_eps": 1e-12,
+  "layer_norm_eps": 1e-05,
   "layerdrop": 0.3,
+  "max_position_embeddings": 920,
   "model_type": "mctc",
   "num_attention_heads": 4,
   "num_conv_layers": 1,
   "num_hidden_layers": 36,
   "pad_token_id": 1,
+  "position_embedding_type": "relative_key",
   "torch_dtype": "float32",
   "transformers_version": "4.18.0.dev0",
   "use_cache": true,

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:38147e5a3277e391f5a2398cb24d10d2e05cd0549a174766410a2b68813fdf6b
-size 4083428601
+oid sha256:d420bb6793a540c8beb636ac255e232aec9d5e2bf11c1b7db9762c6fb131818c
+size 4236102557