Adding min_size to classification and video models

[f-parfenov]

Fixes #5199, I have added the min_size to classification and video model.

The unit test has been updated and passes for all targeted models.

To obtain the min_size values (including the newer models and the video models) I have adapted the following script from #3944 :

import torch

import test_models as TM
from torchvision.prototype import models

def binarysearch(min_, max_, model):
    while min_ < max_:
        mid = min_ + (max_ - min_) // 2
        success = True
        try:
            _ = model(torch.zeros((1, 3, mid, mid)))
        except Exception as e:
            success = False
        if success:
            max_ = mid      # try smaller
        else:
            min_ = mid + 1  # try bigger
    return min_ # or max_ since max_ == min_ now.

def binarsearch_video(min_, max_, model):
    while min_ < max_:
        mid = min_ + (max_ - min_) // 2
        success = True
        try:
            _ = model(torch.zeros((1, 3, 1, mid, mid)))
        except Exception as e:
            success = False
        if success:
            max_ = mid      # try smaller
        else:
            min_ = mid + 1  # try bigger
    return min_ # or max_ since max_ == min_ now.


if __name__ == "__main__":
    classification_prototype_models = (TM.get_models_from_module(models)
    + TM.get_models_from_module(models.quantization))

    min_dim = 1
    max_dim = 224


    for model in classification_prototype_models:
        eval_model = model(pretrained=False).eval()
        _ = eval_model(torch.zeros((1, 3, max_dim, max_dim)))

        min_viable = binarysearch(min_dim, max_dim, eval_model)
        print(f'Model: {model.__name__} \tMin viable shape = (1, 3, {min_viable}, {min_viable})')


    video_prototype_models = TM.get_models_from_module(models.video)


    for video_model in video_prototype_models:
        eval_model = video_model().eval()
        _ = eval_model(torch.zeros((1, 3, 1, max_dim, max_dim)))

        min_viable = binarsearch_video(min_dim, max_dim, eval_model)
        print(f'Model: {video_model.__name__} \tMin viable shape = (1, 3, 1, {min_viable}, {min_viable})')

Which gave me the following values:

Model: alexnet 	Min viable shape = (1, 3, 63, 63)
Model: densenet121 	Min viable shape = (1, 3, 29, 29)
Model: densenet161 	Min viable shape = (1, 3, 29, 29)
Model: densenet169 	Min viable shape = (1, 3, 29, 29)
Model: densenet201 	Min viable shape = (1, 3, 29, 29)
Model: efficientnet_b0 	Min viable shape = (1, 3, 1, 1)
Model: efficientnet_b1 	Min viable shape = (1, 3, 1, 1)
Model: efficientnet_b2 	Min viable shape = (1, 3, 1, 1)
Model: efficientnet_b3 	Min viable shape = (1, 3, 1, 1)
Model: efficientnet_b4 	Min viable shape = (1, 3, 1, 1)
Model: efficientnet_b5 	Min viable shape = (1, 3, 1, 1)
Model: efficientnet_b6 	Min viable shape = (1, 3, 1, 1)
Model: efficientnet_b7 	Min viable shape = (1, 3, 1, 1)
Model: googlenet 	Min viable shape = (1, 3, 15, 15)
Model: inception_v3 	Min viable shape = (1, 3, 75, 75)
Model: mnasnet0_5 	Min viable shape = (1, 3, 1, 1)
Model: mnasnet0_75 	Min viable shape = (1, 3, 1, 1)
Model: mnasnet1_0 	Min viable shape = (1, 3, 1, 1)
Model: mnasnet1_3 	Min viable shape = (1, 3, 1, 1)
Model: mobilenet_v2 	Min viable shape = (1, 3, 1, 1)
Model: mobilenet_v3_large 	Min viable shape = (1, 3, 1, 1)
Model: mobilenet_v3_small 	Min viable shape = (1, 3, 1, 1)
Model: regnet_y_400mf 	Min viable shape = (1, 3, 1, 1)
Model: regnet_y_800mf 	Min viable shape = (1, 3, 1, 1)
Model: regnet_y_1_6gf 	Min viable shape = (1, 3, 1, 1)
Model: regnet_y_3_2gf 	Min viable shape = (1, 3, 1, 1)
Model: regnet_y_8gf 	Min viable shape = (1, 3, 1, 1)
Model: regnet_y_16gf 	Min viable shape = (1, 3, 1, 1)
Model: regnet_y_32gf 	Min viable shape = (1, 3, 1, 1)
Model: regnet_y_128gf 	Min viable shape = (1, 3, 1, 1)
Model: regnet_x_400mf 	Min viable shape = (1, 3, 1, 1)
Model: regnet_x_800mf 	Min viable shape = (1, 3, 1, 1)
Model: regnet_x_1_6gf 	Min viable shape = (1, 3, 1, 1)
Model: regnet_x_3_2gf 	Min viable shape = (1, 3, 1, 1)
Model: regnet_x_8gf 	Min viable shape = (1, 3, 1, 1)
Model: regnet_x_16gf 	Min viable shape = (1, 3, 1, 1)
Model: regnet_x_32gf 	Min viable shape = (1, 3, 1, 1)
Model: resnet18 	Min viable shape = (1, 3, 1, 1)
Model: resnet34 	Min viable shape = (1, 3, 1, 1)
Model: resnet50 	Min viable shape = (1, 3, 1, 1)
Model: resnet101 	Min viable shape = (1, 3, 1, 1)
Model: resnet152 	Min viable shape = (1, 3, 1, 1)
Model: resnext50_32x4d 	Min viable shape = (1, 3, 1, 1)
Model: resnext101_32x8d 	Min viable shape = (1, 3, 1, 1)
Model: wide_resnet50_2 	Min viable shape = (1, 3, 1, 1)
Model: wide_resnet101_2 	Min viable shape = (1, 3, 1, 1)
Model: shufflenet_v2_x0_5 	Min viable shape = (1, 3, 1, 1)
Model: shufflenet_v2_x1_0 	Min viable shape = (1, 3, 1, 1)
Model: shufflenet_v2_x1_5 	Min viable shape = (1, 3, 1, 1)
Model: shufflenet_v2_x2_0 	Min viable shape = (1, 3, 1, 1)
Model: squeezenet1_0 	Min viable shape = (1, 3, 21, 21)
Model: squeezenet1_1 	Min viable shape = (1, 3, 17, 17)
Model: vgg11 	Min viable shape = (1, 3, 32, 32)
Model: vgg11_bn 	Min viable shape = (1, 3, 32, 32)
Model: vgg13 	Min viable shape = (1, 3, 32, 32)
Model: vgg13_bn 	Min viable shape = (1, 3, 32, 32)
Model: vgg16 	Min viable shape = (1, 3, 32, 32)
Model: vgg16_bn 	Min viable shape = (1, 3, 32, 32)
Model: vgg19 	Min viable shape = (1, 3, 32, 32)
Model: vgg19_bn 	Min viable shape = (1, 3, 32, 32)
Model: vit_b_16 	Min viable shape = (1, 3, 224, 224)
Model: vit_b_32 	Min viable shape = (1, 3, 224, 224)
Model: vit_l_16 	Min viable shape = (1, 3, 224, 224)
Model: vit_l_32 	Min viable shape = (1, 3, 224, 224)
Model: googlenet 	Min viable shape = (1, 3, 15, 15)
Model: inception_v3 	Min viable shape = (1, 3, 75, 75)
Model: mobilenet_v2 	Min viable shape = (1, 3, 1, 1)
Model: mobilenet_v3_large 	Min viable shape = (1, 3, 1, 1)
Model: resnet18 	Min viable shape = (1, 3, 1, 1)
Model: resnet50 	Min viable shape = (1, 3, 1, 1)
Model: resnext101_32x8d 	Min viable shape = (1, 3, 1, 1)
Model: shufflenet_v2_x0_5 	Min viable shape = (1, 3, 1, 1)
Model: shufflenet_v2_x1_0 	Min viable shape = (1, 3, 1, 1)
Model: r3d_18 	Min viable shape = (1, 3, 1, 1, 1)
Model: mc3_18 	Min viable shape = (1, 3, 1, 1, 1)
Model: r2plus1d_18 	Min viable shape = (1, 3, 1, 1, 1)

@datumbox

cc @datumbox @bjuncek

[facebook-github-bot]

💊 CI failures summary and remediations

As of commit a38ef88 (more details on the Dr. CI page):

---

✅ None of the CI failures appear to be your fault 💚

---

• 1/1 broken upstream at merge base e047623 since Jan 14

---

🚧 1 ongoing upstream failure:

These were probably caused by upstream breakages that are not fixed yet.

• unittest_prototype since Jan 14 (adf8466)
  • 🔁 rerun

---

This comment was automatically generated by Dr. CI (expand for details).

Please report bugs/suggestions to the (internal) Dr. CI Users group.

Click here to manually regenerate this comment.

[datumbox]

@f-parfenov The changes look good. I only left 2 minor comments, let me know what you think.

[datumbox]

LGTM, thanks @f-parfenov! The failing test on prototype is unrelated. I'll merge once all other tests pass.