#11197: MobilenetV2 bring-up in WH150 with new conv-api

models/experimental/functional_mobilenetv2/README.md

@@ -0,0 +1,26 @@
+# MobilenetV2
+The MobileNetV2 model is a convolutional neural network (CNN) architecture designed for efficient mobile and embedded vision applications. It was introduced in the paper ["MobileNetV2: Inverted Residuals and Linear Bottlenecks"](https://arxiv.org/abs/1801.04381). </br>
+The MobileNetV2 model has been pre-trained on the ImageNet dataset and can be used for various tasks such as image classification, object detection, and semantic segmentation. It has achieved state-of-the-art performance on several benchmarks 1 for mobile and embedded vision applications.
+
+## How to Run
+
+To run the demo, make sure to build the project, activate the environment, and set the appropriate environment variables.
+For more information, refer [installation and build guide](https://docs.tenstorrent.com/tt-metalium/latest/get_started/get_started.html#install-and-build).
+
+To run the functional Mobilenetv2 model on a single-chip:
+```sh
+pytest --disable-warnings models/experimental/functional_mobilenetv2/test/test_ttnn_mobilenetv2.py
+```
+
+To run the functional Mobilenetv2 model on a single-chip:
+```sh
+pytest --disable-warnings models/experimental/functional_mobilenetv2/demo/demo.py
+```
+
+## Supported Hardware
+- N150
+
+## Other Details
+
+- Inputs by default are random data in test_ttnn_mobilenetv2 and images can be fed as input in demo.py.
+- The model weights will be automatically downloaded from Google Drive using wget implemented in weights_download.sh.

models/experimental/functional_mobilenetv2/demo/demo.py

@@ -0,0 +1,101 @@
+# SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
+
+# SPDX-License-Identifier: Apache-2.0
+
+
+import pytest
+import ttnn
+import torch
+
+from tests.ttnn.utils_for_testing import assert_with_pcc
+
+from models.experimental.functional_mobilenetv2.reference.mobilenetv2 import Mobilenetv2
+from models.experimental.functional_mobilenetv2.tt.model_preprocessing import (
+    create_mobilenetv2_input_tensors,
+    create_mobilenetv2_model_parameters,
+)
+from models.experimental.functional_mobilenetv2.tt import ttnn_mobilenetv2
+import os
+from models.utility_functions import (
+    skip_for_grayskull,
+)
+from PIL import Image
+from torchvision import transforms
+
+
+@pytest.mark.parametrize("device_params", [{"l1_small_size": 32768}], indirect=True)
+@skip_for_grayskull()
+def test_mobilenetv2_demo(device, reset_seeds):
+    if not os.path.exists("models/experimental/functional_mobilenetv2/mobilenet_v2-b0353104.pth"):
+        os.system(
+            "bash models/experimental/functional_mobilenetv2/weights_download.sh"
+        )  # execute the weights_download.sh file
+
+    state_dict = torch.load("models/experimental/functional_mobilenetv2/mobilenet_v2-b0353104.pth")
+    ds_state_dict = {k: v for k, v in state_dict.items()}
+    torch_model = Mobilenetv2()
+
+    new_state_dict = {}
+
+    for (name1, parameter1), (name2, parameter2) in zip(torch_model.state_dict().items(), ds_state_dict.items()):
+        if isinstance(parameter2, torch.FloatTensor):
+            new_state_dict[name1] = parameter2
+
+    torch_model.load_state_dict(new_state_dict)
+    torch_model.eval()
+
+    transform = transforms.Compose(
+        [
+            transforms.Resize(128),
+            transforms.CenterCrop(128),
+            transforms.ToTensor(),
+            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
+        ]
+    )
+
+    img = Image.open("models/experimental/functional_mobilenetv2/demo/images/strawberry.jpg")
+
+    img_t = transform(img)
+    torch_input_tensor = torch.unsqueeze(img_t, 0)
+
+    # torch_input_tensor, ttnn_input_tensor = create_mobilenetv2_input_tensors()
+    torch_output_tensor = torch_model(torch_input_tensor)
+
+    parameters = create_mobilenetv2_model_parameters(torch_model, torch_input_tensor, device=device)
+
+    ttnn_input_tensor = torch.permute(torch_input_tensor, (0, 2, 3, 1))
+    ttnn_input_tensor = ttnn_input_tensor.reshape(
+        1,
+        1,
+        ttnn_input_tensor.shape[0] * ttnn_input_tensor.shape[1] * ttnn_input_tensor.shape[2],
+        ttnn_input_tensor.shape[3],
+    )
+    ttnn_input_tensor = ttnn.from_torch(ttnn_input_tensor, dtype=ttnn.bfloat16)
+
+    ttnn_model = ttnn_mobilenetv2.MobileNetV2(parameters, device, torch_model)
+    output_tensor = ttnn_model(device, ttnn_input_tensor)
+
+    #
+    # Tensor Postprocessing
+    #
+    output_tensor = ttnn.to_torch(output_tensor)
+    output_tensor = output_tensor.reshape(torch_output_tensor.shape)
+    output_tensor = output_tensor.to(torch_input_tensor.dtype)
+
+    with open("models/experimental/functional_mobilenetv2/demo/imagenet_classes.txt") as f:
+        classes = [line.strip() for line in f.readlines()]
+
+    # Get the predicted class index and probability
+    _, index = torch.max(output_tensor, 1)  # Get the index of the highest probability class
+    percentage = torch.nn.functional.softmax(output_tensor, dim=1)[0] * 100  # Calculate the class probabilities
+
+    # Print the predicted class and its probability
+    print("\033[1m" + f"Predicted class: {classes[index[0]]}")
+    print(
+        "\033[1m" + f"Probability: {percentage[index[0]].item():.2f}%"
+    )  # Format the probability with two decimal places
+
+    _, indices = torch.sort(output_tensor, descending=True)
+    [(classes[idx], percentage[idx].item()) for idx in indices[0][:5]]
+
+    assert_with_pcc(torch_output_tensor, output_tensor, pcc=0.95)