Merge branch 'fix_pytorch_pruner' into 1.0b

docs/tutorials/pytorch/question-answering/bert-large-uncased-whole-word-masking-finetuned-squad.ipynb

@@ -116,7 +116,7 @@
     "import transformers\n",
     "from dataclasses import dataclass, field\n",
     "from datasets import load_dataset, load_metric\n",
-    "from nlp_toolkit import metrics , QuantizationConfig\n",
+    "from intel_extension_for_transformers import metrics , QuantizationConfig\n",
     "from transformers import (\n",
     "    AutoConfig,\n",
     "    AutoModelForQuestionAnswering,\n",
@@ -161,7 +161,7 @@
    "outputs": [],
    "source": [
     "from transformers import  is_torch_tpu_available\n",
-    "from nlp_toolkit.optimization.trainer import NLPTrainer\n",
+    "from intel_extension_for_transformers.optimization.trainer import NLPTrainer\n",
     "from transformers.trainer_utils import PredictionOutput\n",
     "\n",
     "if is_torch_tpu_available():\n",

docs/tutorials/pytorch/question-answering/orchestrate_optimizations.ipynb

@@ -72,7 +72,7 @@
     "import timeit\n",
     "import torch\n",
     "import transformers\n",
-    "from nlp_toolkit import (\n",
+    "from intel_extension_for_transformers import (\n",
     "    metrics,\n",
     "    PrunerConfig,\n",
     "    PruningConfig,\n",

docs/tutorials/pytorch/question-answering/pruning.ipynb

@@ -66,7 +66,7 @@
     "import transformers\n",
     "from dataclasses import dataclass, field\n",
     "from datasets import load_dataset, load_metric\n",
-    "from nlp_toolkit import metrics, OptimizedModel, PrunerConfig, PruningConfig, PruningMode\n",
+    "from intel_extension_for_transformers import metrics, OptimizedModel, PrunerConfig, PruningConfig, PruningMode\n",
     "from trainer_qa import QuestionAnsweringTrainer\n",
     "from transformers import (\n",
     "    AutoConfig,\n",

docs/tutorials/pytorch/text-classification/orchestrate_optimizations.ipynb

@@ -69,7 +69,7 @@
     "import transformers\n",
     "from dataclasses import dataclass, field\n",
     "from datasets import load_dataset, load_metric\n",
-    "from nlp_toolkit import (\n",
+    "from intel_extension_for_transformers import (\n",
     "    metrics,\n",
     "    PrunerConfig,\n",
     "    PruningConfig,\n",
@@ -78,7 +78,7 @@
     "    OptimizedModel,\n",
     "    objectives\n",
     ")\n",
-    "from nlp_toolkit.optimization.trainer import NLPTrainer\n",
+    "from intel_extension_for_transformers.optimization.trainer import NLPTrainer\n",
     "from torch.utils.data import DataLoader\n",
     "from tqdm.auto import tqdm\n",
     "from transformers import (\n",

docs/tutorials/pytorch/text-classification/pruning.ipynb

@@ -67,13 +67,13 @@
     "import transformers\n",
     "from dataclasses import dataclass, field\n",
     "from datasets import load_dataset, load_metric\n",
-    "from nlp_toolkit import (\n",
+    "from intel_extension_for_transformers import (\n",
     "    metrics,\n",
     "    OptimizedModel,\n",
     "    PrunerConfig,\n",
     "    PruningConfig,\n",
     ")\n",
-    "from nlp_toolkit.optimization.trainer import NLPTrainer\n",
+    "from intel_extension_for_transformers.optimization.trainer import NLPTrainer\n",
     "from transformers import (\n",
     "    AutoConfig,\n",
     "    AutoModelForSequenceClassification,\n",

examples/deployment/neural_engine/emotion/distilbert_base_uncased/requirements.txt

@@ -1,4 +1,4 @@
-neural-compressor
+neural-compressor==1.14.2
 transformers==4.23.1
 accelerate
 datasets >= 1.8.0

examples/deployment/neural_engine/imagenet/vit/requirements.txt

@@ -1,4 +1,4 @@
-neural-compressor
+neural-compressor==1.14.2
 transformers==4.23.1
 accelerate
 datasets >= 1.8.0

examples/deployment/neural_engine/mrpc/bert_base/requirements.txt

@@ -1,4 +1,4 @@
-neural-compressor
+neural-compressor==1.14.2
 transformers==4.23.1
 accelerate
 datasets >= 1.8.0

examples/deployment/neural_engine/mrpc/bert_base_cased/requirements.txt

@@ -1,4 +1,4 @@
-neural-compressor
+neural-compressor==1.14.2
 transformers==4.23.1
 accelerate
 datasets >= 1.8.0

examples/deployment/neural_engine/mrpc/bert_mini/requirements.txt

@@ -1,4 +1,4 @@
-neural-compressor
+neural-compressor==1.14.2
 transformers==4.23.1
 accelerate
 datasets >= 1.8.0

examples/deployment/neural_engine/mrpc/distilbert_base_uncased/requirements.txt

@@ -1,4 +1,4 @@
-neural-compressor
+neural-compressor==1.14.2
 transformers==4.23.1
 accelerate
 datasets >= 1.8.0

examples/deployment/neural_engine/mrpc/roberta_base/requirements.txt

@@ -1,4 +1,4 @@
-neural-compressor
+neural-compressor==1.14.2
 transformers==4.23.1
 accelerate
 datasets >= 1.8.0

examples/deployment/neural_engine/sparse/bert_mini/requirements.txt

@@ -1,4 +1,4 @@
-neural-compressor
+neural-compressor==1.14.2
 transformers==4.23.1
 accelerate
 datasets >= 1.8.0

examples/deployment/neural_engine/sparse/distilbert_base_uncased/requirements.txt

@@ -1,4 +1,4 @@
-neural-compressor
+neural-compressor==1.14.2
 transformers==4.23.1
 accelerate
 datasets >= 1.8.0

examples/deployment/neural_engine/squad/bert_large/requirements.txt

@@ -1,4 +1,4 @@
-neural-compressor
+neural-compressor==1.14.2
 transformers==4.23.1
 accelerate
 datasets >= 1.8.0

examples/deployment/neural_engine/sst2/bert_mini/requirements.txt

@@ -1,4 +1,4 @@
-neural-compressor
+neural-compressor==1.14.2
 transformers==4.23.1
 accelerate
 datasets >= 1.8.0

examples/deployment/neural_engine/sst2/distilbert_base_uncased/requirements.txt

@@ -1,4 +1,4 @@
-neural-compressor
+neural-compressor==1.14.2
 transformers==4.23.1
 accelerate
 datasets >= 1.8.0

examples/deployment/neural_engine/sst2/minilm_l6_h384_uncased/requirements.txt

@@ -1,4 +1,4 @@
-neural-compressor
+neural-compressor==1.14.2
 transformers==4.23.1
 accelerate
 datasets >= 1.8.0

examples/optimization/pytorch/huggingface/pytorch_pruner/README.md

@@ -85,27 +85,22 @@ we have provided several pruning examples, which are trained on different datase
 ### [SQuAD](https://github.com/intel/neural-compressor/tree/master/examples/pytorch/nlp/huggingface_models/question-answering/pruning)
 We can train a sparse model with NxM (2:4) pattern:
 ```
-python3 ./run_qa_no_trainer.py \
+python3 ./run_glue_no_trainer.py \
             --model_name_or_path "/path/to/dense_finetuned_model/" \
-            --pruning_config "./bert_mini_2in4.yaml" \
-            --dataset_name "squad" \
-            --max_seq_length "384" \
-            --doc_stride "128" \
+            --pruning_config "./bert_mini_mrpc_4x1.yaml" \
+            --task_name "mrpc" \
             --per_device_train_batch_size "8" \
             --weight_decay "1e-7" \
             --learning_rate "1e-4" \
             --num_train_epochs 10 \
-            --teacher_model_name_or_path "/path/to/dense_finetuned_model/" \
             --distill_loss_weight "8.0"
 ```
 We can also choose 4x1 as our pruning pattern:
 ```
-python ./run_qa_no_trainer.py \
+python ./run_glue_no_trainer.py \
         --model_name_or_path "/path/to/dense_finetuned_model/" \
-        --pruning_config "./bert_mini_4x1.yaml" \
-        --dataset_name "squad" \
-        --max_seq_length "384" \
-        --doc_stride "128" \
+        --pruning_config "./bert_mini_mrpc_4x1.yaml" \
+        --task_name "mrpc" \
         --per_device_train_batch_size "16" \
         --per_device_eval_batch_size "16" \
         --num_warmup_steps "1000" \
@@ -115,18 +110,15 @@ python ./run_qa_no_trainer.py \
         --num_train_epochs 10 \
         --weight_decay  "1e-7" \
         --output_dir "pruned_squad_bert-mini" \
-        --teacher_model_name_or_path "/path/to/dense_finetuned_model/" \
         --distill_loss_weight "4.5"
 ```
 Dense model training is also supported as following (by setting --do_prune to False):
 ```
 python \
-    ./run_qa_no_trainer.py \
+    ./run_glue_no_trainer.py \
     --model_name_or_path "prajjwal1/bert-mini" \
-    --pruning_config "./bert_mini_4x1.yaml" \
-    --dataset_name "squad" \
-    --max_seq_length "384" \
-    --doc_stride "128" \
+    --task_name "mrpc" \
+    --pruning_config "./bert_mini_mrpc_4x1.yaml" \
     --per_device_train_batch_size "8" \
     --per_device_eval_batch_size "16" \
     --num_warmup_steps "1000" \

examples/optimization/pytorch/huggingface/pytorch_pruner/requirements.txt

@@ -0,0 +1,7 @@
+datasets
+tqdm
+torch >= 1.10.0
+transformers>=4.12.0
+accelerate
+huggingface_hub
+intel_extension_for_transformers

examples/optimization/pytorch/huggingface/pytorch_pruner/run_glue_no_trainer.py

@@ -20,7 +20,7 @@
 from pathlib import Path
 import sys
 
-from nlp_toolkit.optimization.pytorch_pruner.pruning import Pruning
+from intel_extension_for_transformers.optimization.pytorch_pruner.pruning import Pruning
 
 sys.path.insert(0, './')
 import datasets

examples/optimization/tensorflow/huggingface/language-modeling/quantization/inc/ptq/requirements.txt

@@ -1,6 +1,6 @@
 datasets >= 1.17
 sentencepiece != 0.1.92
 protobuf
-intel-tensorflow >= 2.5
+intel-tensorflow == 2.10.0
 transformers >= 4.12.0
 sklearn

examples/optimization/tensorflow/huggingface/multiple-choice/quantization/inc/requirements.txt

@@ -1,5 +1,5 @@
 datasets >= 1.17
 sentencepiece != 0.1.92
 protobuf
-intel-tensorflow >= 2.5
+intel-tensorflow == 2.10.0
 transformers >= 4.12.0

examples/optimization/tensorflow/huggingface/text-classification/auto_distillation/requirements.txt

@@ -1,5 +1,6 @@
+git+https://github.com/intel/neural-compressor.git@old_api_examples
 datasets >= 1.17
 sentencepiece != 0.1.92
 protobuf
-intel-tensorflow >= 2.5
+intel-tensorflow == 2.10.0
 transformers>=4.12.0

examples/optimization/tensorflow/huggingface/text-classification/distillation/requirements.txt

@@ -1,5 +1,6 @@
+git+https://github.com/intel/neural-compressor.git@old_api_examples
 datasets >= 1.17
 sentencepiece != 0.1.92
 protobuf
-intel-tensorflow >= 2.5
+intel-tensorflow == 2.10.0
 transformers>=4.12.0

examples/optimization/tensorflow/huggingface/text-classification/pruning/requirements.txt

@@ -1,5 +1,6 @@
+git+https://github.com/intel/neural-compressor.git@old_api_examples
 datasets >= 1.17
 sentencepiece != 0.1.92
 protobuf
-intel-tensorflow >= 2.5
+intel-tensorflow == 2.10.0
 transformers>=4.12.0

examples/optimization/tensorflow/huggingface/text-classification/quantization/inc/ptq/requirements.txt

@@ -1,5 +1,5 @@
 datasets >= 1.17
 sentencepiece != 0.1.92
 protobuf
-intel-tensorflow >= 2.5
+intel-tensorflow == 2.10.0
 transformers >= 4.12.0

examples/optimization/tensorflow/huggingface/token-classification/quantization/inc/requirements.txt

@@ -2,5 +2,5 @@ datasets >= 1.17
 sentencepiece != 0.1.92
 seqeval
 protobuf
-intel-tensorflow >= 2.5
+intel-tensorflow == 2.10.0
 transformers >= 4.12.0

intel_extension_for_transformers/backends/neural_engine/test/pytest/requirements.txt

@@ -1,2 +1,3 @@
 onnx
 onnxruntime
+tensorflow==2.10.0

intel_extension_for_transformers/backends/neural_engine/test/pytest/test_graph_dispatch.py

@@ -43,7 +43,7 @@ def test_graph_dispatch(self):
         # validate int8 sparse graph tuning
         int8_model_path = "/home/tensorflow/localfile/nlptoolkit_ut_model/bert_mini_int8_original_IR"
         if is_win():
-            int8_model_path = "C:/Users/sdp/Documents/models/localfile/nlptoolkit_ut_model/bert_mini_int8_original_IR"
+            int8_model_path = "D:\\dataset\\nlptoolkit_ut_model\\bert_mini_int8_original_IR"
         self.assertTrue(os.path.exists(int8_model_path),
             'INT8 IR model is not found, please set your own model path!')
         int8_model = compile(int8_model_path)
@@ -59,7 +59,7 @@ def test_graph_dispatch(self):
         # validate onednn graph tuning
         fp32_model_path = "/home/tensorflow/localfile/nlptoolkit_ut_model/bert_mini_sst2_1x4_fp32.onnx"
         if is_win():
-            fp32_model_path = "C:/Users/sdp/Documents/models/localfile/nlptoolkit_ut_model/bert_mini_sst2_1x4_fp32.onnx"
+            fp32_model_path = "D:\\dataset\\nlptoolkit_ut_model\\bert_mini_sst2_1x4_fp32.onnx"
         self.assertTrue(os.path.exists(fp32_model_path),
             'FP32 ONNX model is not found, please set your own model path!')
         fp32_model = compile(fp32_model_path)

intel_extension_for_transformers/backends/neural_engine/test/pytest/test_pattern_dispatch.py

@@ -46,7 +46,7 @@ def test_pattern_dispatch(self):
         # validate pattern tuning
         fp32_model_path = "/home/tensorflow/localfile/nlptoolkit_ut_model/bert_mini_sst2_1x4_fp32.onnx"
         if is_win():
-            fp32_model_path = "C:/Users/sdp/Documents/models/localfile/nlptoolkit_ut_model/bert_mini_sst2_1x4_fp32.onnx"
+            fp32_model_path = "D:\\dataset\\nlptoolkit_ut_model\\bert_mini_sst2_1x4_fp32.onnx"
         self.assertTrue(os.path.exists(fp32_model_path),
             'FP32 ONNX model is not found, please set your own model path!')
         fp32_model = compile(fp32_model_path)

intel_extension_for_transformers/backends/neural_engine/test/pytest/test_qkv_merge.py

@@ -36,8 +36,8 @@ def test_qkv_merge_1(self):
         model_path = "/home/tensorflow/localfile/nlptoolkit_ut_model/onnx_best_acc_distilbert.onnx"
         pattern_config = "/home/tensorflow/localfile/nlptoolkit_ut_model/qkv_merge_pattern_config"
         if is_win():
-            model_path = "C:/Users/sdp/Documents/models/localfile/nlptoolkit_ut_model/onnx_best_acc_distilbert.onnx"
-            pattern_config = "C:/Users/sdp/Documents/models/localfile/nlptoolkit_ut_model/qkv_merge_pattern_config"
+            model_path = "D:\\dataset\\nlptoolkit_ut_model\\onnx_best_acc_distilbert.onnx"
+            pattern_config = "D:\\dataset\\nlptoolkit_ut_model\\qkv_merge_pattern_config"
         graph = compile(model_path, config=pattern_config)
         self.assertEqual(100, len(graph.nodes))
 

intel_extension_for_transformers/backends/neural_engine/test/pytest/test_quant_onnx_execute.py

@@ -16,10 +16,14 @@
 # limitations under the License.
 
 import os
+import sys
 import unittest
 import numpy as np
 from intel_extension_for_transformers.backends.neural_engine.compile import compile
 
+
+def is_win():
+    return sys.platform.startswith('win')
 class TestQuantOnnxExecute(unittest.TestCase):
     @classmethod
     def setUpClass(self):
@@ -35,15 +39,19 @@ def test_onnx_qlinear_compile(self):
         input_1 = np.random.uniform(low=0, high=1, size=[1, 128]).astype('int32')
         input_2 = np.random.uniform(low=0, high=1, size=[1, 128]).astype('int32')
         # compile and execute qlinear model
-        qlinear_model_path = "/tf_dataset2/models/nlp_toolkit/qlinear/bert_mini_sst2_qlinear.onnx"
+        qlinear_model_path = "/home/tensorflow/localfile/nlptoolkit_ut_model/qlinear/bert_mini_sst2_qlinear.onnx"
+        if is_win():
+            qlinear_model_path = "D:\\dataset\\nlptoolkit_ut_model\\qlinear\\bert_mini_sst2_qlinear.onnx"
         os.environ['GLOG_minloglevel'] = '2'
         self.assertTrue(os.path.exists(qlinear_model_path),
             'ONNX QLinear model is not found, please set your own model path!')
         qlinear_model = compile(qlinear_model_path)
         qlinear_output_dict = qlinear_model.inference([input_0, input_1, input_2])
         qlinear_output = list(qlinear_output_dict.values())[0]        
         # compile and execute qdq model
-        qdq_model_path = "/tf_dataset2/models/nlp_toolkit/qlinear/bert_mini_sst2_qdq.onnx"
+        qdq_model_path = "/home/tensorflow/localfile/nlptoolkit_ut_model/qlinear/bert_mini_sst2_qdq.onnx"
+        if is_win():
+            qdq_model_path = "D:\\dataset\\nlptoolkit_ut_model\\qlinear\\bert_mini_sst2_qdq.onnx"
         self.assertTrue(os.path.exists(qdq_model_path),
             'ONNX QDQ model is not found, please set your own model path!')
         qdq_model = compile(qdq_model_path)

intel_extension_for_transformers/backends/neural_engine/test/pytest/test_transpose.py

@@ -38,7 +38,7 @@ def test_transpose(self):
         os.environ['GLOG_minloglevel'] = '2'
         root_dir = '/home/tensorflow/localfile/nlptoolkit_ut_model/'
         if is_win():
-            root_dir = 'C:/Users/sdp/Documents/models/localfile/nlptoolkit_ut_model/'
+            root_dir = 'D:\\dataset\\nlptoolkit_ut_model\\'
         model_dir = root_dir + 'onnx_best_acc_distilbert.merged.untransposed'
         pattern_config = root_dir + 'pattern_config'
         self.assertTrue(os.path.exists(model_dir),

intel_extension_for_transformers/optimization/mixture/auto_distillation.py

@@ -34,7 +34,7 @@
 torch = LazyImport("torch")
 
 def distributed_log_wrapper(func, msg):
-    if self.framework != "pytorch" or not torch.distributed.is_initialized() or torch.distributed.get_rank() == 0:
+    if not torch.distributed.is_initialized() or torch.distributed.get_rank() == 0:
         func(msg)
 
 
@@ -135,7 +135,8 @@ def reload_tf_model(model):
                 "Metrics of model architecture {} is {}.".format(model_arch_paras, metrics)
             )
             self.search_results[tuple(model_arch_paras.values())] = metrics
-            if self.framework != "pytorch" or not torch.distributed.is_initialized() or torch.distributed.get_rank() == 0:
+            if (self.framework != "pytorch" or not torch.distributed.is_initialized() 
+              or torch.distributed.get_rank() == 0):
                 self.advisor.feedback(sum(self.metrics_conversion(metrics)))
                 print(f'res_save_path: {res_save_path}, save_path = {save_path}')
                 os.makedirs(save_path, exist_ok=True)

requirements.txt

@@ -1,4 +1,4 @@
-neural_compressor
+neural-compressor==1.14.2
 numpy
 packaging
 protobuf --no-binary=protobuf