Merge branch 'dev' into feature/exec_hlslib_nodes

src/finn/backend/fpgadataflow/utils.py

@@ -6,11 +6,15 @@ from finn.core.datatype import DataType
 from finn.core.utils import pack_innermost_dim_as_hex_string
 
 
-def numpy_to_hls_code(ndarray, dtype, hls_var_name, pack_innermost_dim=True):
+def numpy_to_hls_code(
+    ndarray, dtype, hls_var_name, pack_innermost_dim=True, no_decl=False
+):
     """Return C++ code representation of a numpy ndarray with FINN DataType
     dtype, using hls_var_name as the resulting C++ variable name. If
     pack_innermost_dim is specified, the innermost dimension of the ndarray
-    will be packed into a hex string using array2hexstring.
+    will be packed into a hex string using array2hexstring. If no_decl is
+    set to True, no variable name and type will be generated as part of the
+    emitted string.
     """
     hls_dtype = dtype.get_hls_datatype_str()
     if type(ndarray) != np.ndarray or ndarray.dtype != np.float32:
@@ -47,5 +51,8 @@ def numpy_to_hls_code(ndarray, dtype, hls_var_name, pack_innermost_dim=True):
     strarr = np.array2string(ndarray, separator=", ", formatter={"all": elem2str})
     np.set_printoptions(**orig_printops)
     strarr = strarr.replace("[", "{").replace("]", "}")
-    ret = ret + " = \n" + strarr + ";"
+    if no_decl:
+        ret = strarr + ";"
+    else:
+        ret = ret + " = \n" + strarr + ";"
     return ret

tests/test_brevitas_cnv.py

+import os
+import pkg_resources as pk
+
+import brevitas.onnx as bo
+import numpy as np
+import torch
+from models.CNV import CNV
+
+import finn.core.onnx_exec as oxe
+from finn.core.modelwrapper import ModelWrapper
+from finn.transformation.fold_constants import FoldConstants
+from finn.transformation.infer_shapes import InferShapes
+
+export_onnx_path = "test_output_cnv.onnx"
+# TODO get from config instead, hardcoded to Docker path for now
+trained_cnv_checkpoint = (
+    "/workspace/brevitas_cnv_lfc/pretrained_models/CNV_1W1A/checkpoints/best.tar"
+)
+
+
+def test_brevitas_cnv_w1a1_export():
+    cnv = CNV(weight_bit_width=1, act_bit_width=1, in_bit_width=1, in_ch=3).eval()
+    bo.export_finn_onnx(cnv, (1, 3, 32, 32), export_onnx_path)
+    model = ModelWrapper(export_onnx_path)
+    assert model.graph.node[2].op_type == "Sign"
+    assert model.graph.node[3].op_type == "Conv"
+    conv0_wname = model.graph.node[3].input[1]
+    assert list(model.get_initializer(conv0_wname).shape) == [64, 3, 3, 3]
+    assert model.graph.node[4].op_type == "Mul"
+    os.remove(export_onnx_path)
+
+
+def test_brevitas_cnv_w1a1_export_exec():
+    cnv = CNV(weight_bit_width=1, act_bit_width=1, in_bit_width=1, in_ch=3).eval()
+    checkpoint = torch.load(trained_cnv_checkpoint, map_location="cpu")
+    cnv.load_state_dict(checkpoint["state_dict"])
+    bo.export_finn_onnx(cnv, (1, 3, 32, 32), export_onnx_path)
+    model = ModelWrapper(export_onnx_path)
+    model = model.transform(InferShapes())
+    model = model.transform(FoldConstants())
+    model.save(export_onnx_path)
+    fn = pk.resource_filename("finn", "data/cifar10/cifar10-test-data-class3.npz")
+    input_tensor = np.load(fn)["arr_0"].astype(np.float32)
+    assert input_tensor.shape == (1, 3, 32, 32)
+    # run using FINN-based execution
+    input_dict = {"0": input_tensor}
+    output_dict = oxe.execute_onnx(model, input_dict)
+    produced = output_dict[list(output_dict.keys())[0]]
+    # do forward pass in PyTorch/Brevitas
+    input_tensor = torch.from_numpy(input_tensor).float()
+    expected = cnv.forward(input_tensor).detach().numpy()
+    assert np.isclose(produced, expected, atol=1e-3).all()
+    os.remove(export_onnx_path)
+
+
+def test_brevitas_trained_cnv_w1a1_pytorch():
+    # load pretrained weights into CNV-w1a1
+    cnv = CNV(weight_bit_width=1, act_bit_width=1, in_bit_width=1, in_ch=3).eval()
+    checkpoint = torch.load(trained_cnv_checkpoint, map_location="cpu")
+    cnv.load_state_dict(checkpoint["state_dict"])
+    fn = pk.resource_filename("finn", "data/cifar10/cifar10-test-data-class3.npz")
+    input_tensor = np.load(fn)["arr_0"]
+    input_tensor = torch.from_numpy(input_tensor).float()
+    assert input_tensor.shape == (1, 3, 32, 32)
+    # do forward pass in PyTorch/Brevitas
+    produced = cnv.forward(input_tensor).detach().numpy()
+    expected = np.asarray(
+        [
+            [
+                3.7939777,
+                -2.3108773,
+                0.06898145,
+                0.55185133,
+                0.37939775,
+                -1.9659703,
+                -0.3104164,
+                -2.828238,
+                2.6902752,
+                0.48286998,
+            ]
+        ],
+        dtype=np.float32,
+    )
+    assert np.isclose(produced, expected, atol=1e-3).all()

tests/test_npy2hls.py

@@ -41,3 +41,7 @@ def test_numpy_to_hls_code():
     {{ap_uint<4>("f", 16), ap_uint<4>("f", 16)},
      {ap_uint<4>("7", 16), ap_uint<4>("d", 16)}};"""
     assert remove_all_whitespace(ret) == remove_all_whitespace(eB)
+    ret = numpy_to_hls_code(B, DataType.UINT2, "test", True, True)
+    eB = """{{ap_uint<4>("f", 16), ap_uint<4>("f", 16)},
+     {ap_uint<4>("7", 16), ap_uint<4>("d", 16)}};"""
+    assert remove_all_whitespace(ret) == remove_all_whitespace(eB)