Upload folder using huggingface_hub

ressources/e2e_model/squeezeformer

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2367109bab7fb7c395af10b5e2daf17941d8690c69050912ded400ba392ce9eb
-size 36523741
+oid sha256:7e044c35cc4ec67b5f31b6d2c5c0d88b35260ec4230b4681c06a125fb4c26946
+size 133

ressources/kenLM_model/kab_5k_6-gram.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:880abff36dd3c64ad9df109773ef3d49c7909e5de80cd63350b188e13e1df42c
-size 6568786
+oid sha256:5997204a01a5b77af19b3fb97a443abf99ce187e4cf6bb80ce20b6fb60e49a3a
+size 132

ressources/kenLM_model/kab_5k_6-gram_v2.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:88604fe3c4c653f822748386ad143ba4e3d9086577ed6d3ef205100a28d3d115
-size 87786060
+oid sha256:031b9844ae8c586a44b7b271629fe8ee242ff4e55519de98b83bee1e84d229db
+size 133

ressources/kenLM_model/kab_5k_trigram.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b89ecec10ddd295bca9abfc10a91a5488ad47217efd9ef8d672f3ba0f61765a4
-size 1102536
+oid sha256:a1d49064d3be8455bc04c87a95371fbc9823e01cb88a00ed32eb06a2d45982dd
+size 132

ressources/tokenizer/128_v7.model

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0cfd6e89442fcff8a9850d69a1b23fb6b71cc8dcc98952d5c3c144038a013bc2
-size 239191
+oid sha256:0b6ab83ef6996f24e285c1d1acb8fbec87b857412bfd65ebd7a656b9033fea15
+size 131

ressources/tokenizer/5K.model

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:eae00d2ddc1baf700a3247817d6371959c71f967d9fb91529a86efea1492dfed
-size 315116
+oid sha256:8f95318dba0f4ee4bbe5ab99d8148d577d603479ffd92f973b79bcb456732544
+size 131

ressources/train.csv

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:61709eed30b35a4a6403fb54968326d4d20731263ed1b76190146d11347eab27
-size 18971108
+oid sha256:063513ebb954b4fba411d96779a19598374cf1a237a0326e199cf40d7b3e9f2d
+size 133

src/inference_file.py

@@ -8,7 +8,7 @@ import kenlm
 from torchaudio.models.decoder import ctc_decoder
 from torchaudio.transforms import Resample
 from squeezeformer import MySqueezeformer
-
+import torch.ao.quantization
 
 # -------------------------
 # Paths
@@ -34,6 +34,11 @@ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 model = MySqueezeformer().to(device)
 
 acoustic_model_path = os.path.join(dirname, "../ressources/e2e_model/squeezeformer")
+if device == "cpu":
+    model = torch.ao.quantization.quantize_dynamic(
+        model, {torch.nn.Linear}, dtype=torch.qint8
+    )
+
 
 model.load_state_dict(torch.load(acoustic_model_path, map_location=device))