Remove onnx and onnxruntime from requirements.txt (#640)

* Remove onnx and onnxruntime from requirements.txt

egs/librispeech/ASR/lstm_transducer_stateless2/streaming-onnx-decode.py

@@ -42,6 +42,11 @@
 import logging
 from typing import List, Optional, Tuple
 
+from icefall import is_module_available
+
+if not is_module_available("onnxruntime"):
+    raise ValueError("Please 'pip install onnxruntime' first.")
+
 import onnxruntime as ort
 import sentencepiece as spm
 import torch

egs/librispeech/ASR/pruned_transducer_stateless3/onnx_check.py

@@ -24,6 +24,11 @@
 import argparse
 import logging
 
+from icefall import is_module_available
+
+if not is_module_available("onnxruntime"):
+    raise ValueError("Please 'pip install onnxruntime' first.")
+
 import onnxruntime as ort
 import torch
 

egs/librispeech/ASR/pruned_transducer_stateless3/test_onnx.py

@@ -21,6 +21,11 @@
 """
 import os
 
+from icefall import is_module_available
+
+if not is_module_available("onnxruntime"):
+    raise ValueError("Please 'pip install onnxruntime' first.")
+
 import onnxruntime as ort
 import torch
 from conformer import (

egs/librispeech/ASR/pruned_transducer_stateless6/model.py

@@ -21,7 +21,6 @@
 import torch
 import torch.nn as nn
 from encoder_interface import EncoderInterface
-from multi_quantization.prediction import JointCodebookLoss
 from scaling import ScaledLinear
 
 from icefall.utils import add_sos
@@ -74,6 +73,14 @@ def __init__(
             encoder_dim, vocab_size, initial_speed=0.5
         )
         self.simple_lm_proj = ScaledLinear(decoder_dim, vocab_size)
+
+        from icefall import is_module_available
+
+        if not is_module_available("multi_quantization"):
+            raise ValueError("Please 'pip install multi_quantization' first.")
+
+        from multi_quantization.prediction import JointCodebookLoss
+
         if num_codebooks > 0:
             self.codebook_loss_net = JointCodebookLoss(
                 predictor_channels=encoder_dim,

egs/librispeech/ASR/pruned_transducer_stateless6/vq_utils.py

@@ -28,18 +28,21 @@
 import numpy as np
 import torch
 import torch.multiprocessing as mp
-import multi_quantization as quantization
 
+from icefall import is_module_available
+
+if not is_module_available("multi_quantization"):
+    raise ValueError("Please 'pip install multi_quantization' first.")
+
+import multi_quantization as quantization
 from asr_datamodule import LibriSpeechAsrDataModule
 from hubert_xlarge import HubertXlargeFineTuned
-from icefall.utils import (
-    AttributeDict,
-    setup_logger,
-)
 from lhotse import CutSet, load_manifest
 from lhotse.cut import MonoCut
 from lhotse.features.io import NumpyHdf5Writer
 
+from icefall.utils import AttributeDict, setup_logger
+
 
 class CodebookIndexExtractor:
     """

egs/wenetspeech/ASR/pruned_transducer_stateless2/onnx_check.py

@@ -40,6 +40,11 @@
 import argparse
 import logging
 
+from icefall import is_module_available
+
+if not is_module_available("onnxruntime"):
+    raise ValueError("Please 'pip install onnxruntime' first.")
+
 import onnxruntime as ort
 import torch
 

egs/wenetspeech/ASR/pruned_transducer_stateless2/onnx_pretrained.py

@@ -49,6 +49,12 @@
 import k2
 import kaldifeat
 import numpy as np
+
+from icefall import is_module_available
+
+if not is_module_available("onnxruntime"):
+    raise ValueError("Please 'pip install onnxruntime' first.")
+
 import onnxruntime as ort
 import torch
 import torchaudio

icefall/__init__.py

@@ -50,6 +50,7 @@
     get_executor,
     get_texts,
     is_jit_tracing,
+    is_module_available,
     l1_norm,
     l2_norm,
     linf_norm,

icefall/ngram_lm.py

@@ -17,7 +17,7 @@
 from collections import defaultdict
 from typing import List, Optional, Tuple
 
-import kaldifst
+from icefall.utils import is_module_available
 
 
 class NgramLm:
@@ -36,6 +36,11 @@ def __init__(
           is_binary:
             True if the given file is a binary FST.
         """
+        if not is_module_available("kaldifst"):
+            raise ValueError("Please 'pip install kaldifst' first.")
+
+        import kaldifst
+
         if is_binary:
             lm = kaldifst.StdVectorFst.read(fst_filename)
         else:
@@ -85,6 +90,8 @@ def _get_next_state_and_cost_without_backoff(
         self, state: int, label: int
     ) -> Tuple[int, float]:
         """TODO: Add doc."""
+        import kaldifst
+
         arc_iter = kaldifst.ArcIterator(self.lm, state)
         num_arcs = self.lm.num_arcs(state)
 

icefall/utils.py

@@ -976,3 +976,17 @@ def display_and_save_batch(
     y = sp.encode(supervisions["text"], out_type=int)
     num_tokens = sum(len(i) for i in y)
     logging.info(f"num tokens: {num_tokens}")
+
+
+# `is_module_available` is copied from
+# https://github.com/pytorch/audio/blob/6bad3a66a7a1c7cc05755e9ee5931b7391d2b94c/torchaudio/_internal/module_utils.py#L9
+def is_module_available(*modules: str) -> bool:
+    r"""Returns if a top-level module with :attr:`name` exists *without**
+    importing it. This is generally safer than try-catch block around a
+    `import X`.
+
+    Note: "borrowed" from torchaudio:
+    """
+    import importlib
+
+    return all(importlib.util.find_spec(m) is not None for m in modules)

requirements.txt

@@ -3,9 +3,4 @@ kaldialign
 sentencepiece>=0.1.96
 tensorboard
 typeguard
-multi_quantization
-onnx
-onnxruntime
---extra-index-url https://pypi.ngc.nvidia.com
 dill
-kaldifst

test/test_ngram_lm.py

@@ -16,6 +16,12 @@
 # limitations under the License.
 
 import graphviz
+
+from icefall import is_module_available
+
+if not is_module_available("kaldifst"):
+    raise ValueError("Please 'pip install kaldifst' first.")
+
 import kaldifst
 
 from icefall import NgramLm, NgramLmStateCost