Qualcomm AI Engine Direct - GA Model Enablement (deit)

Summary
- support e2e script / test case for GA deit model
- perf: 8a8w 2.7ms/inf
- acc: top1/5 ~= 85%/95%

Author: chenweng-quic

backends/qualcomm/tests/test_qnn_delegate.py

@@ -3880,6 +3880,41 @@ def test_conv_former(self):
                 self.assertGreaterEqual(msg["top_1"], 60)
                 self.assertGreaterEqual(msg["top_5"], 80)
 
+    def test_deit(self):
+        if not self.required_envs([self.image_dataset]):
+            self.skipTest("missing required envs")
+        cmds = [
+            "python",
+            f"{self.executorch_root}/examples/qualcomm/oss_scripts/deit.py",
+            "--dataset",
+            self.image_dataset,
+            "--artifact",
+            self.artifact_dir,
+            "--build_folder",
+            self.build_folder,
+            "--device",
+            self.device,
+            "--model",
+            self.model,
+            "--ip",
+            self.ip,
+            "--port",
+            str(self.port),
+        ]
+        if self.host:
+            cmds.extend(["--host", self.host])
+
+        p = subprocess.Popen(cmds, stdout=subprocess.DEVNULL)
+        with Listener((self.ip, self.port)) as listener:
+            conn = listener.accept()
+            p.communicate()
+            msg = json.loads(conn.recv())
+            if "Error" in msg:
+                self.fail(msg["Error"])
+            else:
+                self.assertGreaterEqual(msg["top_1"], 75)
+                self.assertGreaterEqual(msg["top_5"], 90)
+
     def test_dino_v2(self):
         if not self.required_envs([self.image_dataset]):
             self.skipTest("missing required envs")

examples/qualcomm/oss_scripts/deit.py

@@ -0,0 +1,147 @@
+# Copyright (c) Qualcomm Innovation Center, Inc.
+# All rights reserved
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import getpass
+import json
+import os
+from multiprocessing.connection import Client
+
+import numpy as np
+import torch
+from executorch.backends.qualcomm._passes.qnn_pass_manager import (
+    get_capture_program_passes,
+)
+from executorch.backends.qualcomm.quantizer.quantizer import QuantDtype
+from executorch.examples.qualcomm.utils import (
+    build_executorch_binary,
+    get_imagenet_dataset,
+    make_output_dir,
+    parse_skip_delegation_node,
+    setup_common_args_and_variables,
+    SimpleADB,
+    topk_accuracy,
+)
+from transformers import AutoConfig, AutoModelForImageClassification
+
+
+def get_instance():
+    module = (
+        AutoModelForImageClassification.from_pretrained("facebook/deit-base-distilled-patch16-224")
+        .eval()
+        .to("cpu")
+    )
+
+    return module
+
+
+def main(args):
+    skip_node_id_set, skip_node_op_set = parse_skip_delegation_node(args)
+
+    os.makedirs(args.artifact, exist_ok=True)
+    config = AutoConfig.from_pretrained("facebook/deit-base-distilled-patch16-224")
+    data_num = 100
+    height = config.image_size
+    width = config.image_size
+    inputs, targets, input_list = get_imagenet_dataset(
+        dataset_path=f"{args.dataset}",
+        data_size=data_num,
+        image_shape=(height, width),
+        crop_size=(height, width),
+    )
+
+    # Get the Deit model.
+    model = get_instance()
+    pte_filename = "deit_qnn"
+
+    # lower to QNN
+    passes_job = get_capture_program_passes()
+    build_executorch_binary(
+        model,
+        inputs[0],
+        args.model,
+        f"{args.artifact}/{pte_filename}",
+        dataset=inputs,
+        skip_node_id_set=skip_node_id_set,
+        skip_node_op_set=skip_node_op_set,
+        quant_dtype=QuantDtype.use_8a8w,
+        passes_job=passes_job,
+        shared_buffer=args.shared_buffer,
+    )
+
+    if args.compile_only:
+        return
+
+    workspace = f"/data/local/tmp/{getpass.getuser()}/executorch/{pte_filename}"
+    pte_path = f"{args.artifact}/{pte_filename}.pte"
+
+    adb = SimpleADB(
+        qnn_sdk=os.getenv("QNN_SDK_ROOT"),
+        build_path=f"{args.build_folder}",
+        pte_path=pte_path,
+        workspace=workspace,
+        device_id=args.device,
+        host_id=args.host,
+        soc_model=args.model,
+    )
+    adb.push(inputs=inputs, input_list=input_list)
+    adb.execute()
+
+    # collect output data
+    output_data_folder = f"{args.artifact}/outputs"
+    make_output_dir(output_data_folder)
+
+    adb.pull(output_path=args.artifact)
+
+    # top-k analysis
+    predictions = []
+    for i in range(data_num):
+        predictions.append(
+            np.fromfile(
+                os.path.join(output_data_folder, f"output_{i}_0.raw"), dtype=np.float32
+            )
+        )
+
+    k_val = [1, 5]
+    topk = [topk_accuracy(predictions, targets, k).item() for k in k_val]
+    if args.ip and args.port != -1:
+        with Client((args.ip, args.port)) as conn:
+            conn.send(json.dumps({f"top_{k}": topk[i] for i, k in enumerate(k_val)}))
+    else:
+        for i, k in enumerate(k_val):
+            print(f"top_{k}->{topk[i]}%")
+
+
+if __name__ == "__main__":
+    parser = setup_common_args_and_variables()
+    parser.add_argument(
+        "-a",
+        "--artifact",
+        help="path for storing generated artifacts and output by this example. Default ./deit_qnn",
+        default="./deit_qnn",
+        type=str,
+    )
+
+    parser.add_argument(
+        "-d",
+        "--dataset",
+        help=(
+            "path to the validation folder of ImageNet dataset. "
+            "e.g. --dataset imagenet-mini/val "
+            "for https://www.kaggle.com/datasets/ifigotin/imagenetmini-1000)"
+        ),
+        type=str,
+        required=True,
+    )
+
+    args = parser.parse_args()
+    try:
+        main(args)
+    except Exception as e:
+        if args.ip and args.port != -1:
+            with Client((args.ip, args.port)) as conn:
+                conn.send(json.dumps({"Error": str(e)}))
+        else:
+            raise Exception(e)