Merge pull request #4891 from slSeanWU/whisper-ci-fix

Add python 3.8 requirement for Whisper & update tests

espnet2/text/cleaner.py

@@ -13,7 +13,7 @@
 
 try:
     from whisper.normalizers import BasicTextNormalizer, EnglishTextNormalizer
-except ImportError:
+except (ImportError, SyntaxError):
     BasicTextNormalizer = None
 
 

test/espnet2/asr/decoder/test_whisper_decoder.py

@@ -1,3 +1,5 @@
+import sys
+
 import pytest
 import torch
 from packaging.version import parse as V
@@ -10,37 +12,41 @@
 
 # NOTE(Shih-Lun): needed for `persistent` param in
 #                 torch.nn.Module.register_buffer()
-is_torch_1_6_plus = V(torch.__version__) >= V("1.6.0")
+is_torch_1_7_plus = V(torch.__version__) >= V("1.7.0")
+is_python_3_8_plus = sys.version_info >= (3, 8)
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 @pytest.fixture()
 def whisper_decoder(request):
-    if not is_torch_1_6_plus:
-        return None
-
     return OpenAIWhisperDecoder(
         vocab_size=VOCAB_SIZE_WHISPER_MULTILINGUAL,
         encoder_output_size=384,
         whisper_model="tiny",
     )
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 @pytest.mark.timeout(50)
 def test_decoder_init(whisper_decoder):
-    if not is_torch_1_6_plus:
-        return
-
     assert (
         whisper_decoder.decoders.token_embedding.num_embeddings
         == VOCAB_SIZE_WHISPER_MULTILINGUAL
     )
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 @pytest.mark.timeout(50)
 def test_decoder_reinit_emb():
-    if not is_torch_1_6_plus:
-        return
-
     vocab_size = 1000
     decoder = OpenAIWhisperDecoder(
         vocab_size=vocab_size,
@@ -50,10 +56,11 @@ def test_decoder_reinit_emb():
     assert decoder.decoders.token_embedding.num_embeddings == vocab_size
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 def test_decoder_invalid_init():
-    if not is_torch_1_6_plus:
-        return
-
     with pytest.raises(AssertionError):
         decoder = OpenAIWhisperDecoder(
             vocab_size=VOCAB_SIZE_WHISPER_MULTILINGUAL,
@@ -63,11 +70,12 @@ def test_decoder_invalid_init():
         del decoder
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 @pytest.mark.timeout(50)
 def test_decoder_forward_backward(whisper_decoder):
-    if not is_torch_1_6_plus:
-        return
-
     hs_pad = torch.randn(4, 100, 384, device=next(whisper_decoder.parameters()).device)
     ys_in_pad = torch.randint(
         0, 3000, (4, 10), device=next(whisper_decoder.parameters()).device
@@ -78,11 +86,12 @@ def test_decoder_forward_backward(whisper_decoder):
     out.sum().backward()
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 @pytest.mark.timeout(50)
 def test_decoder_scoring(whisper_decoder):
-    if not is_torch_1_6_plus:
-        return
-
     hs_pad = torch.randn(4, 100, 384, device=next(whisper_decoder.parameters()).device)
     ys_in_pad = torch.randint(
         0, 3000, (4, 10), device=next(whisper_decoder.parameters()).device

test/espnet2/asr/encoder/test_whisper_encoder.py

@@ -1,3 +1,5 @@
+import sys
+
 import pytest
 import torch
 from packaging.version import parse as V
@@ -8,40 +10,45 @@
 
 # NOTE(Shih-Lun): needed for `return_complex` param in torch.stft()
 is_torch_1_7_plus = V(torch.__version__) >= V("1.7.0")
+is_python_3_8_plus = sys.version_info >= (3, 8)
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 @pytest.fixture()
 def whisper_encoder(request):
-    if not is_torch_1_7_plus:
-        return None
-
     encoder = OpenAIWhisperEncoder(whisper_model="tiny")
 
     return encoder
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 @pytest.mark.timeout(50)
 def test_encoder_init(whisper_encoder):
-    if not is_torch_1_7_plus:
-        return
-
     assert whisper_encoder.output_size() == 384
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 def test_encoder_invalid_init():
-    if not is_torch_1_7_plus:
-        return
-
     with pytest.raises(AssertionError):
         encoder = OpenAIWhisperEncoder(whisper_model="aaa")
         del encoder
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 @pytest.mark.timeout(50)
 def test_encoder_forward_no_ilens(whisper_encoder):
-    if not is_torch_1_7_plus:
-        return
-
     input_tensor = torch.randn(
         4, 3200, device=next(whisper_encoder.parameters()).device
     )
@@ -50,11 +57,12 @@ def test_encoder_forward_no_ilens(whisper_encoder):
     assert xs_pad.size() == torch.Size([4, 10, 384])
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 @pytest.mark.timeout(50)
 def test_encoder_forward_ilens(whisper_encoder):
-    if not is_torch_1_7_plus:
-        return
-
     input_tensor = torch.randn(
         4, 3200, device=next(whisper_encoder.parameters()).device
     )
@@ -67,11 +75,12 @@ def test_encoder_forward_ilens(whisper_encoder):
     assert torch.equal(olens.cpu(), torch.tensor([2, 3, 5, 10]))
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 @pytest.mark.timeout(50)
 def test_encoder_backward(whisper_encoder):
-    if not is_torch_1_7_plus:
-        return
-
     whisper_encoder.train()
     input_tensor = torch.randn(
         4, 3200, device=next(whisper_encoder.parameters()).device

test/espnet2/asr/frontend/test_whisper.py

@@ -1,3 +1,5 @@
+import sys
+
 import pytest
 import torch
 from packaging.version import parse as V
@@ -8,40 +10,45 @@
 
 # NOTE(Shih-Lun): required by `return_complex` in torch.stft()
 is_torch_1_7_plus = V(torch.__version__) >= V("1.7.0")
+is_python_3_8_plus = sys.version_info >= (3, 8)
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 @pytest.fixture()
 def whisper_frontend(request):
-    if not is_torch_1_7_plus:
-        return None
-
     with torch.no_grad():
         return WhisperFrontend("tiny")
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 @pytest.mark.timeout(50)
 def test_frontend_init():
-    if not is_torch_1_7_plus:
-        return
-
     frontend = WhisperFrontend("tiny")
     assert frontend.output_size() == 384
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 def test_frontend_invalid_init():
-    if not is_torch_1_7_plus:
-        return
-
     with pytest.raises(AssertionError):
         frontend = WhisperFrontend("aaa")
         del frontend
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 @pytest.mark.timeout(50)
 def test_frontend_forward_no_ilens(whisper_frontend):
-    if not is_torch_1_7_plus:
-        return
-
     input_tensor = torch.randn(
         4, 3200, device=next(whisper_frontend.parameters()).device
     )
@@ -50,11 +57,12 @@ def test_frontend_forward_no_ilens(whisper_frontend):
     assert feats.size() == torch.Size([4, 10, 384])
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus or not is_torch_1_7_plus,
+    reason="whisper not supported on python<3.8, torch<1.7",
+)
 @pytest.mark.timeout(50)
 def test_frontend_forward_ilens(whisper_frontend):
-    if not is_torch_1_7_plus:
-        return
-
     input_tensor = torch.randn(
         4, 3200, device=next(whisper_frontend.parameters()).device
     )

test/espnet2/text/test_whisper_token_id_converter.py

@@ -1,25 +1,41 @@
+import sys
+
 import pytest
 
 from espnet2.text.whisper_token_id_converter import OpenAIWhisperTokenIDConverter
 
 pytest.importorskip("whisper")
 
+is_python_3_8_plus = sys.version_info >= (3, 8)
+
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus, reason="whisper not supported on python<3.8"
+)
 @pytest.fixture(params=["whisper_multilingual"])
 def whisper_token_id_converter(request):
     return OpenAIWhisperTokenIDConverter(request.param)
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus, reason="whisper not supported on python<3.8"
+)
 def test_init_invalid():
     with pytest.raises(ValueError):
         OpenAIWhisperTokenIDConverter("whisper_aaa")
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus, reason="whisper not supported on python<3.8"
+)
 def test_init_en():
     id_converter = OpenAIWhisperTokenIDConverter("whisper_en")
     assert id_converter.get_num_vocabulary_size() == 50363
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus, reason="whisper not supported on python<3.8"
+)
 def test_ids2tokens(whisper_token_id_converter: OpenAIWhisperTokenIDConverter):
     tokens = whisper_token_id_converter.ids2tokens(
         [17155, 11, 220, 83, 378, 320, 311, 5503, 307, 1481, 13, 8239, 485]
@@ -42,6 +58,9 @@ def test_ids2tokens(whisper_token_id_converter: OpenAIWhisperTokenIDConverter):
     ]
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus, reason="whisper not supported on python<3.8"
+)
 def test_tokens2ids(whisper_token_id_converter: OpenAIWhisperTokenIDConverter):
     ids = whisper_token_id_converter.tokens2ids(
         [

test/espnet2/text/test_whisper_tokenizer.py

@@ -1,29 +1,48 @@
+import sys
+
 import pytest
 
 from espnet2.text.whisper_tokenizer import OpenAIWhisperTokenizer
 
 pytest.importorskip("whisper")
 
+is_python_3_8_plus = sys.version_info >= (3, 8)
+
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus, reason="whisper not supported on python<3.8"
+)
 @pytest.fixture(params=["whisper_multilingual"])
 def whisper_tokenizer(request):
     return OpenAIWhisperTokenizer(request.param)
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus, reason="whisper not supported on python<3.8"
+)
 def test_init_en():
     tokenizer = OpenAIWhisperTokenizer("whisper_en")
     assert tokenizer.tokenizer.tokenizer.vocab_size == 50257
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus, reason="whisper not supported on python<3.8"
+)
 def test_init_invalid():
     with pytest.raises(ValueError):
         OpenAIWhisperTokenizer("whisper_aaa")
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus, reason="whisper not supported on python<3.8"
+)
 def test_repr(whisper_tokenizer: OpenAIWhisperTokenizer):
     print(whisper_tokenizer)
 
 
+@pytest.mark.skipif(
+    not is_python_3_8_plus, reason="whisper not supported on python<3.8"
+)
 def test_tokenization_consistency(whisper_tokenizer: OpenAIWhisperTokenizer):
     s = "Hi, today's weather is nice. Hmm..."