Added reproducer

Author: nikita-savelyevv

optimum/intel/openvino/modeling_base.py

@@ -471,6 +471,7 @@ def _reshape(
         height: int = None,
         width: int = None,
     ):
+        return model
         shapes = {}
         for inputs in model.inputs:
             shapes[inputs] = inputs.get_partial_shape()

reshape_reproducer.py

@@ -0,0 +1,67 @@
+import nncf
+import numpy as np
+import logging
+import tempfile
+
+from transformers import AutoModelForSequenceClassification, AutoTokenizer, TrainingArguments, default_data_collator
+from datasets import load_dataset
+import evaluate
+from optimum.intel import OVModelForSequenceClassification, OVConfig, OVTrainer
+
+
+nncf.nncf_logger.setLevel(logging.ERROR)
+
+
+def reshape_model(model):
+    shapes = {}
+    for inputs in model.inputs:
+        shapes[inputs] = inputs.get_partial_shape()
+        shapes[inputs][0] = -1
+        shapes[inputs][1] = -1
+    model.reshape(shapes)
+
+
+def get_num_fqs(model):
+    num_fake_quantize = 0
+    for node in model.get_ops():
+        if "FakeQuantize" in node.get_type_name():
+            num_fake_quantize += 1
+    return num_fake_quantize
+
+
+model_id = "distilbert-base-uncased"
+for _ in range(5):
+    # n_samples = 16
+    n_samples = (np.random.randint(1000) % 16) + 1
+
+    model = AutoModelForSequenceClassification.from_pretrained(model_id)
+    tokenizer = AutoTokenizer.from_pretrained(model_id)
+    ov_config = OVConfig()
+    dataset = load_dataset("glue", "sst2")
+    dataset = dataset.map(
+        lambda examples: tokenizer(examples["sentence"], padding="max_length", max_length=128), batched=True
+    )
+    train_dataset = dataset["train"].select(range(n_samples))
+    eval_dataset = dataset["validation"].select(range(n_samples))
+    metric = evaluate.load("glue", "sst2")
+    with tempfile.TemporaryDirectory() as tmp_dir:
+        trainer = OVTrainer(
+            model=model,
+            ov_config=ov_config,
+            task="sequence-classification",
+            args=TrainingArguments(tmp_dir, num_train_epochs=1.0, do_train=True, do_eval=True),
+            train_dataset=train_dataset,
+            eval_dataset=eval_dataset,
+            compute_metrics=lambda p: metric.compute(predictions=np.argmax(p.predictions, 1), references=p.label_ids),
+            tokenizer=tokenizer,
+            data_collator=default_data_collator,
+        )
+        trainer.train()
+        trainer.evaluate()
+        trainer.save_model()
+
+        ov_model = OVModelForSequenceClassification.from_pretrained(tmp_dir)
+        fqs_before_reshape = get_num_fqs(ov_model.model)
+        reshape_model(ov_model.model)
+        fqs_after_reshape = get_num_fqs(ov_model.model)
+        print(f"Number of FQ nodes before reshape: {fqs_before_reshape}, after reshape: {fqs_after_reshape}")