[PYTORCH]Minor bug fixes (apache#5683)

* [PYTORCH]Minor bug fixes

* Review comment fix, testcase added

* Added testcase for bert model

python/tvm/relay/frontend/pytorch.py

@@ -34,6 +34,7 @@
 from .common import get_relay_op
 from .common import infer_shape as _infer_shape
 from .common import infer_value as _infer_value
+from .common import infer_value_simulated as _infer_value_simulated
 from .common import infer_type as _infer_type
 from ..prelude import Prelude, StaticTensorArrayOps
 
@@ -152,19 +153,33 @@ def _impl(inputs, input_types):
 
 def _arange():
     def _impl(inputs, input_types):
+        def _get_value(val, dtype):
+            if isinstance(val, _expr.Expr):
+                return _op.cast(val, _convert_data_type(dtype))
+            return _create_typed_const(val, dtype)
+
+        def _get_type(val, inp_type):
+            if isinstance(val, _expr.Expr):
+                dtype = str(_infer_type(val).checked_type)
+                return dtype if dtype != "float32" else "float"
+            return inp_type
+
         if len(inputs) == 5:
-            dtype = "float" if "float" in input_types[0:1] else _convert_dtype_value(inputs[1])
-            start = _create_typed_const(0, dtype)
-            stop = _create_typed_const(inputs[0], dtype)
-            step = _create_typed_const(1, dtype)
+            dtype0 = _get_type(inputs[0], input_types[0])
+            dtype = "float" if dtype0 == "float" else _convert_dtype_value(inputs[1])
+            start = _get_value(0, dtype)
+            stop = _get_value(inputs[0], dtype)
+            step = _get_value(1, dtype)
         elif len(inputs) == 7:
-            dtype = "float" if "float" in input_types[0:3] else _convert_dtype_value(inputs[3])
-            start = _create_typed_const(inputs[0], dtype)
-            stop = _create_typed_const(inputs[1], dtype)
-            step = _create_typed_const(inputs[2], dtype)
+            types = [_get_type(inputs[i], input_types[i]) for i in range(3)]
+            dtype = "float" if "float" in types else _convert_dtype_value(inputs[3])
+            start = _get_value(inputs[0], dtype)
+            stop = _get_value(inputs[1], dtype)
+            step = _get_value(inputs[2], dtype)
         else:
             msg = "Unknown number of arguments (%d) to parse." % (len(inputs))
             raise AssertionError(msg)
+
         return _op.transform.arange(start=start,
                                     stop=stop,
                                     step=step,
@@ -235,12 +250,18 @@ def _impl(inputs, input_types):
 
         begin = [0] * len(end)
         dim = int(inputs[1])
-        begin[dim] = int(inputs[2])
+        if isinstance(inputs[2], _expr.Call):
+            begin[dim] = np.asscalar(_infer_value(inputs[2], {}).asnumpy().astype(np.int))
+        else:
+            begin[dim] = int(inputs[2])
 
         if isinstance(inputs[3], str) and inputs[3].isdigit():
             end[dim] = min(end[dim], int(inputs[3]))
         else:
-            end[dim] = inputs[3]
+            if isinstance(inputs[3], _expr.Call):
+                end[dim] = np.asscalar(_infer_value(inputs[3], {}).asnumpy().astype(np.int))
+            else:
+                end[dim] = inputs[3]
 
         strides.append(int(inputs[4]))
         return _op.transform.strided_slice(data, begin, end, strides)
@@ -997,7 +1018,10 @@ def _impl(inputs, input_types):
 def _numtotensor():
     def _impl(inputs, input_types):
         val = inputs[0]
-        dtype = type(val)
+        dtype = input_types[0]
+
+        if isinstance(val, _expr.Expr):
+            return val
 
         if isinstance(val, tvm.tir.IntImm):
             val = val.__int__()
@@ -1019,16 +1043,22 @@ def _impl(inputs, input_types):
         data = inputs[0]
 
         if len(inputs) == 3:
-            new_shape = [inputs[1], _infer_shape(inputs[2])[0]]
+            shape_inp = [inputs[1], _infer_shape(inputs[2])[0]]
         else:
             if isinstance(inputs[1], list):
-                new_shape = inputs[1]
+                shape_inp = inputs[1]
             else:
-                new_shape = _infer_shape(inputs[1])
+                shape_inp = _infer_shape(inputs[1])
+        new_shape = shape_inp
+        for i, shape in enumerate(shape_inp):
+            if isinstance(shape, _expr.Expr):
+                val = _infer_value_simulated(shape, {})
+                new_shape[i] = np.asscalar(val.asnumpy())
 
         return _op.transform.reshape(data, new_shape)
     return _impl
 
+
 def _reshape():
     def _impl(inputs, input_types):
         data = inputs[0]

tests/python/frontend/pytorch/test_forward.py

@@ -381,28 +381,61 @@ def test_forward_arange():
     class Arange1(Module):
         def forward(self, *args):
             return torch.arange(5)
+
     class Arange2(Module):
         def forward(self, *args):
             return torch.arange(2.5)
+
     class Arange3(Module):
         def forward(self, *args):
             return torch.arange(1, 4)
+
     class Arange4(Module):
         def forward(self, *args):
             return torch.arange(1, 2.5, 0.5)
+
     class Arange5(Module):
         def forward(self, *args):
             return torch.arange(1, 2, 1, dtype=torch.int32)
+
     class Arange6(Module):
         def forward(self, *args):
             return torch.arange(start=1, end=6, step=2)
+
     class Arange7(Module):
         def forward(self, *args):
             return torch.arange(1, 4, dtype=torch.float32)
+
     class Arange8(Module):
         def forward(self, *args):
             return torch.arange(1, 2, 1, dtype=torch.int16)
 
+    class Arange9(Module):
+        def forward(self, *args):
+            end = torch.add(torch.tensor(4), 1)
+            return torch.arange(end) + torch.ones((5,), dtype=torch.int64)
+
+    class Arange10(Module):
+        def forward(self, *args):
+            end = torch.add(torch.tensor(4.0), torch.tensor(1.0))
+            return torch.arange(end) + torch.ones((5,), dtype=torch.float)
+
+    class Arange11(Module):
+        def forward(self, *args):
+            start = torch.add(torch.tensor(1), 1)
+            end = torch.add(torch.tensor(4), 1)
+            step = torch.add(torch.tensor(2), 1)
+            out = torch.arange(start, end, step)
+            return out + torch.ones((3,), dtype=torch.int64)
+
+    class Arange12(Module):
+        def forward(self, *args):
+            start = torch.add(torch.tensor(1), 1)
+            end = torch.add(torch.tensor(4), 1)
+            step = torch.add(torch.tensor(2.5), torch.tensor(4.1))
+            out = torch.arange(start, end, step)
+            return out + torch.ones((3,), dtype=torch.float)
+
     verify_model(Arange1().float().eval())
     verify_model(Arange2().float().eval())
     verify_model(Arange3().float().eval())
@@ -411,6 +444,11 @@ def forward(self, *args):
     verify_model(Arange6().float().eval())
     verify_model(Arange7().float().eval())
     verify_model(Arange8().float().eval())
+    verify_model(Arange9().float().eval())
+    verify_model(Arange10().float().eval())
+    verify_model(Arange11().float().eval())
+    verify_model(Arange12().float().eval())
+
 
 def test_forward_abs():
     torch.set_grad_enabled(False)
@@ -810,9 +848,15 @@ class View2(Module):
         def forward(self, *args):
             return args[0].view(args[0].shape[0], -1)
 
+    class View3(Module):
+        def forward(self, *args):
+            d1 = torch.tensor(3) * torch.tensor(10) * torch.tensor(10)
+            return args[0].view(args[0].shape[0], d1)
+
     input_data = torch.rand(input_shape).float()
     verify_model(View1().float().eval(), input_data=input_data)
     verify_model(View2().float().eval(), input_data=input_data)
+    verify_model(View3().float().eval(), input_data=input_data)
 
 def test_forward_select():
     torch.set_grad_enabled(False)
@@ -896,9 +940,17 @@ class Slice2(Module):
         def forward(self, *args):
             return args[0][0, :, :, :]
 
+    class Slice3(Module):
+        def forward(self, *args):
+            x0 = torch.tensor(2) - torch.tensor(1)
+            x1 = torch.tensor(3) + torch.tensor(1)
+            return args[0][:, x0:, :x1, :]
+
     input_data = torch.rand(input_shape).float()
     verify_model(Slice1().float().eval(), input_data=input_data)
     verify_model(Slice2().float().eval(), input_data=input_data)
+    verify_model(Slice3().float().eval(), input_data=input_data)
+
 
 def test_forward_mean():
     torch.set_grad_enabled(False)
@@ -2157,6 +2209,134 @@ def forward(self, *args):
     verify_model(MatMul1().float().eval(), input_data=[tensor1, tensor2])
 
 
+def test_forward_pretrained_bert_base_uncased():
+    ######################################################################
+    # This is an example how to run BERT models using TVM
+    # ---------------------------------------------------
+    """
+    Refer the bert example given in https://pypi.org/project/pytorch-pretrained-bert
+
+    # To get started, pretrained bert package needs to be installed as prerequisite.
+
+    .. code-block:: bash
+
+        # install bert package
+        pip install pytorch_pretrained_bert==0.6.2 --user
+    """
+
+    try:
+        from pytorch_pretrained_bert import BertTokenizer, BertForMaskedLM
+    except:
+        print("Torch pretrained bert package must be installed to run this script.")
+        return
+
+    ######################################################################
+    # Load the tokenizer and tokenize the input
+    # -----------------------------------------
+
+    # Load pre-trained model tokenizer (vocabulary)
+    tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+
+    # Tokenized input
+    text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
+    tokenized_text = tokenizer.tokenize(text)
+
+    # Mask a token that we will try to predict back with `BertForMaskedLM`
+    masked_index = 8
+    tokenized_text[masked_index] = '[MASK]'
+    assert tokenized_text == ['[CLS]', 'who', 'was', 'jim', 'henson', '?', '[SEP]', 'jim', '[MASK]', 'was', 'a', 'puppet',
+                              '##eer', '[SEP]']
+
+    # Convert token to vocabulary indices
+    indexed_tokens = tokenizer.convert_tokens_to_ids(tokenized_text)
+    # Define sentence A and B indices associated to 1st and 2nd sentences (see paper)
+    segments_ids = [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1]
+
+    # Convert inputs to PyTorch tensors
+    tokens_tensor = torch.tensor([indexed_tokens])
+    segments_tensors = torch.tensor([segments_ids])
+
+    ######################################################################
+    # Load a pretrained PyTorch model bert-base-uncased
+    # -------------------------------------------------
+
+    # Bert Model with a language modeling
+    model = BertForMaskedLM.from_pretrained('bert-base-uncased')
+    model.eval()
+
+    ######################################################################
+    # Predict all tokens with pytorch
+    # -------------------------------
+
+    with torch.no_grad():
+        torch_preds = model(tokens_tensor, segments_tensors)
+
+    ######################################################################
+    # Make TorchScripted model via jit trace
+    # --------------------------------------
+
+    scripted_model = torch.jit.trace(model, (tokens_tensor, segments_tensors)).eval()
+
+    ######################################################################
+    # Import the graph to Relay
+    # -------------------------
+    # Convert PyTorch graph to Relay graph. The input name can be arbitrary.
+
+    input_1 = 'input_ids'
+    input_2 = 'input.2'
+    shape_list = [(input_1, list(tokens_tensor.shape)),
+                  (input_2, list(segments_tensors.shape))]
+
+    mod, params = relay.frontend.from_pytorch(scripted_model, shape_list)
+
+    ######################################################################
+    # Compile the model with relay
+    # ----------------------------
+
+    target = 'llvm'
+    with relay.build_config(opt_level=3):
+        relay_graph, relay_lib, relay_params = relay.build(mod, target=target, params=params)
+
+    ######################################################################
+    # Execute on TVM
+    # --------------
+
+    ctx = tvm.context(target, 0)
+    relay_model = graph_runtime.create(relay_graph, relay_lib, ctx)
+    relay_model.set_input(**relay_params)
+    relay_model.set_input(input_1, tokens_tensor)
+    relay_model.set_input(input_2, segments_tensors)
+    relay_model.run()
+    compiled_output = relay_model.get_output(0).asnumpy()
+
+    ######################################################################
+    # Validate the outputs
+    # --------------------
+    # Compare the torch and tvm outputs
+
+    tvm.testing.assert_allclose(torch_preds, compiled_output, rtol=1e-3, atol=1e-3)
+
+    ######################################################################
+    # Process the output
+    # ------------------
+    # Process the model output to token.
+
+    # Torch output to token
+    torch_pred_idx = torch.argmax(torch_preds[0, masked_index]).item()
+    torch_pred_token = tokenizer.convert_ids_to_tokens([torch_pred_idx])[0]
+
+    # TVM output to token
+    tvm_pred_idx = compiled_output[0, masked_index].argmax()
+    tvm_pred_token = tokenizer.convert_ids_to_tokens([tvm_pred_idx])[0]
+
+    assert torch_pred_idx == tvm_pred_idx
+    assert torch_pred_token == tvm_pred_token
+
+    # Print the outputs
+    print('Torch top-1 id: {}, token: {}'.format(torch_pred_idx, torch_pred_token))
+    print('TVM   top-1 id: {}, token: {}'.format(tvm_pred_idx, tvm_pred_token))
+
+
 if __name__ == "__main__":
     # Single operator tests
     test_forward_add()
@@ -2284,3 +2464,6 @@ def forward(self, *args):
     from lstm_test import custom_lstm_test
 
     custom_lstm_test()
+
+    # Test bert model
+    test_forward_pretrained_bert_base_uncased()