Since this PR has been merged into the master branch (and some code changes), TensorFlow/TFLite models can be exported using

python export.py --weights yolov5s.pt --include saved_model pb tflite [--int8] --img 640
python export.py --weights yolov5s.pt --include tfjs --img 640

and validated using

python detect.py --weights yolov5s_saved_model --img 640
                           yolov5s.pb
                           yolov5s-fp16.tflite
                           yolov5s-int8.tflite

After exporting TFLite models, https://github.com/zldrobit/yolov5/tree/tf-android can be used as an Android demo.

For Edge TPU model export, plz refer to https://github.com/ultralytics/yolov5/pull/3630.

python models/tf.py --weights weights/yolov5s.pt --cfg models/yolov5s.yaml --img 320

python3 models/tf.py --weights weights/yolov5s.pt --cfg models/yolov5s.yaml --tfl-int8 --source /data/dataset/coco/coco2017/train2017 --ncalib 100

python3 detect.py --weights weights/yolov5s.pb          --img 320
                            weights/yolov5s_saved_model
                            weights/yolov5s-fp16.tflite
                            weights/yolov5s-int8.tflite           --tfl-int8

python3 models/tf.py --weights weights/yolov5s.pt --cfg models/yolov5s.yaml --img 320 --tf-nms --agnostic-nms

# install tfjs converter: pip install tensorflowjs
tensorflowjs_converter \
    --input_format=tf_frozen_model \
    --output_node_names='Identity,Identity_1,Identity_2,Identity_3' \
    weights/yolov5s.pb \
    weights/web_model

sed -i 's/{"outputs": {"Identity.\?.\?": {"name": "Identity.\?.\?"}, "Identity.\?.\?": {"name": "Identity.\?.\?"}, "Identity.\?.\?": {"name": "Identity.\?.\?"}, "Identity.\?.\?": {"name": "Identity.\?.\?"}}}/'\
'{"outputs": {"Identity": {"name": "Identity"}, "Identity_1": {"name": "Identity_1"}, "Identity_2": {"name": "Identity_2"}, "Identity_3": {"name": "Identity_3"}}}/' \
./weights/web_model/model.json

EDIT:

• Change --img from 640 to 320.
• Remove --no-tfl-detect. It is now only used in tf-android-tfl-detect tf-edgetpu branch for Edge TPU model export.
• Deprecate TensorFlow 1.x support. If you still want to use TF 1.x, plz refer to the archived tf-export branch. Note this branch does not support TFLite model export.
• For EdgeTPU support, tf-android-tfl-detect branch is obsoleted by tf-edgetpu branch.
• Add --tf-raw-resize option to map resize ops to EdgeTPU, which accelerates inference (not necessary with Edge TPU compiler v16).
• Add instruction for model inference.
• Add TensorFlow.js model export and detection example tfjs-yolov5-example. Demo page on https://zldrobit.github.io/tfjs-yolov5-example/
• Remove --cfg since the config yaml is stored in the model weights now.

FAQ:

• Is TFLite NMS supported? NMS is not supported on TFLite GPU delegate by now (https://github.com/ultralytics/yolov5/discussions/2095). If you are interested in usage on CPU, another issue is working on this.
• Inconsistent model output shape with Java code (java.lang.IllegalArgumentException: Cannot copy to a TensorFlowLite tensor (serving_default_input_1:0) with 1769472 bytes from a Java Buffer with 1228800 bytes.): see https://github.com/ultralytics/yolov5/pull/1127#issuecomment-1062729008
• OpenCL not found with GPU delegate (Can not open OpenCL library on this device - dlopen failed: library "libOpenCL.so" not found): see https://github.com/ultralytics/yolov5/pull/1127#issuecomment-1056038455
• If you're interested in detecting objects in an image with Android, plz refer to https://github.com/ultralytics/yolov5/pull/1127#issuecomment-1057576360 and modify the project accordingly.
• TFLite backend is optimized for ARM processors by defualt. Running a YOLOv5 TFLite model on a X86 CPU is very slow as expected and inference with python detect.py is only for validation purposes. If you're interested in running YOLOv5 TFLite models on Android, plz refer to https://github.com/zldrobit/yolov5/tree/tf-android.

Usage:

# train
python segment/train.py --cfg yolov5(m/s/l)-seg 
# detect
python segment/predict.py --source

TODO:

• [x] Resolve conflict
• [x] Resolve pre-commit
• [x] Enable overlap-mask by default
• [x] Clean callback related residue code
• [ ] Write segmentation related readme
• [x] torch.deterministic solution for segment?
• [x] merge non_max_suppression_masks() with non_max_suppression()?
• [x] remove edict dependency?
• [x] DetectMultiBackend compatibility - PyTorch only
• [x] export.py compatability all non-TF
• [x] TensorBoard and W&B logging
• [x] CI - Test segmentation: Requires YOLOv5n-seg.pt and COCO128-seg dataset
• [x] export.py TF with TFDetectSegment()
• [x] Fix csv logging
• [x] cpu support
• [x] implement SegmentationModel class (requires retraining all models) to help identify seg models easily on load.
• [x] Fix Logging

Hello, @glenn-jocher

From your advice, Multi-process DistributedDataParallel should be better than the Single process we have now.

I have been trying to change it here at my fork on my ddp branch. (Apologize that it's messy). However, I've reached into many issues.

Since it's in testing, I didn't account for the device being cpu as of now.

What I did so far

• Added setup method to init_process_group and set torch.cuda device

• Called torch.multiprocessing.spawn on the modified train function

• ~~Created a new argument called world_size to be called when running script (we can change this to counting # of device later)~~

• Added condition checks for only 1 process to download weights file, remove the batch.jpg, saving checkpoints

• ~Added dist.barrier() while waiting for first process to do its job~

• ~Replaced all .to(device) to .to(rank) for each process.~

• ~Changed map_location for loading weights.~

• ~~Added more parameters to train function because the processes cannot see the global variables~~

• Added DistributedSampler for multiple gpu on dataset so they each get a different sample

• ~Turned off tensorboard as I needed to pass tb_writer to train as argument to be able to use it.~

Things to fix

• ~~Do not divide dataset for validation set in create_dataloader~~

• ~~Reduce the need to call world_size as argument to say that we want multiprocess~~

• ~Cleaning up~

• ~~Fixing the inconsistent output prints (All process printing at once makes it hard to track)~~

• ~Enable tensorboard again~

• ~~Splitting batch_size/learning rate/epoch for multiple GPU~~

• ~~Figure out why global variables are always recalled (I disabled print(hyp) because of this)~~

Problems

Since I am still learning, it is very likely I messed up the training. The learnt information from each epoch for each process may not be distributed among themselves because I tested training at it's 0 mAP. I read somewhere that a SyncBatch layer may be needed, but I am not sure how to add it.

Saving checkpoints is done by only the first process as multiprocess saving concurrently causes problem for strip_optimizer later on. I am not sure if this is the correct way.

I am testing it, and it is much slower than using just one GPU, but I figured that if the training is fixed, it can be a good boost for multi-gpu.

I also understand that this isn't as high in your priority list but maybe some guidance would be nice. Thank you.

tf-edgetpu branch is created from tf-android-tfl-detect branch by deleting the android directory. tf-android-tfl-detect is deprecated since tf-android branch could be used for deploying YOLOv5 models on Android devices.

Before export, install edgetpu-compiler (https://coral.ai/software/#debian-packages), and install pycoral (https://coral.ai/software/#pycoral-api).

Export EdgeTPU models using:

python export.py --weights yolov5s.pt --include edgetpu --img 320 --data data/coco128.yaml

--data is used for calibration in int8 quantization.

Detect objects using:

python detect.py --weights yolov5s-int8_edgetpu.tflite --img 320 --data data/coco128.yaml

--data is used to show real class names.

# Install the Compiler
! curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | sudo apt-key add -
! echo "deb https://packages.cloud.google.com/apt coral-edgetpu-stable main" | sudo tee /etc/apt/sources.list.d/coral-edgetpu.list
! sudo apt-get update
! sudo apt-get install edgetpu-compiler	

# Install Branch https://github.com/zldrobit/yolov5/tree/tf-edgetpu 
!git clone https://github.com/zldrobit/yolov5.git
%cd yolov5
!git checkout tf-edgetpu
%pip install -qr requirements.txt
from yolov5 import utils
display = utils.notebook_init()  # checks

# Export file Compiled for Coral Edge TPU
%pip install --upgrade flatbuffers==1.12 # downgrade from v2 to v1.12
!python export.py --weights yolov5s.pt --include edgetpu --img 320 --data data/coco128.yaml

|input resolution| infercence time (batch 1)| |-|-| |320 | 31.7 ms| |448 | 74.3 ms| |576 | 155 ms| |640 | 271.6 ms|

EDIT:

• Merge master, so Edge TPU models can be exported with export.py and validated with detect.py
• Add --data option to show real class names in detection.
• 640x640 input resolution is supported.
• Edge TPU models contain bbox reconstruction, and this feature is not backward-compatible. Edge TPU models have to be re-exported!

TODO:

• [ ] Add TensorFlow Raw Resize back Edge TPU compiler v16 supports native Resize ops.

Hello,

I download the provided ckpt files and use python test.py --img-size 736 --conf_thres 0.001 for reproducing results, but I got different AP and AP50 when compare with the table in readme.

yolov3-spp: 44.1% AP, 64.4% AP50 (Table 45.5% AP, 65.2% AP50) Speed: 10.4/2.1/12.6 ms inference/NMS/total per 736x736 image at batch-size 16

yolov5s: 31.4% AP, 52.3% AP50 (Table 33.1% AP, 53.3% AP50) Speed: 2.2/2.1/4.4 ms inference/NMS/total per 736x736 image at batch-size 16

yolov5m: 39.9% AP, 60.7% AP50 (Table 41.5% AP, 61.5% AP50) Speed: 5.4/1.8/7.2 ms inference/NMS/total per 736x736 image at batch-size 16

yolov5l: 42.7% AP, 63.5% AP50 (Table 44.2% AP, 64.3% AP50) Speed: 11.3/2.2/13.5 ms inference/NMS/total per 736x736 image at batch-size 16

yolov5x: 45.7% AP, 65.9% AP50 (Table 47.1% AP, 66.7% AP50) Speed: 20.3/2.2/22.5 ms inference/NMS/total per 736x736 image at batch-size 16

Do the reported results are from the test.py or they are calculated by evaluation server?

Search before asking

• [X] I have searched the YOLOv5 issues and found no similar feature requests.

Description

I visualize where the model concerns using grad-cam. detection results: layer 17 layer 20 layer 23

detection results: layer 17 layer 20 layer 23

my code: in detect.py 1.add a function

def cam_show_img(img, feature_map, grads, out_name):
    H, W, _ = img.shape
    cam = np.zeros(feature_map.shape[1:], dtype=np.float32)		
    grads = grads.reshape([grads.shape[0],-1])					
    weights = np.mean(grads, axis=1)							
    for i, w in enumerate(weights):
        cam += w * feature_map[i, :, :]							
    cam = np.maximum(cam, 0)
    cam = cam / cam.max()
    cam = cv2.resize(cam, (W, H))

    heatmap = cv2.applyColorMap(np.uint8(255 * cam), cv2.COLORMAP_JET)
    cam_img = 0.3 * heatmap + 0.7 * img

    cv2.imwrite(out_name, cam_img)

change: https://github.com/ultralytics/yolov5/blob/f5b8f7d54c9fa69210da0177fec7ac2d9e4a627c/detect.py#L59-L61 to:

if visualize:
    # require grad
    for k, v in model.named_parameters():
        v.requires_grad = True  # train all layers
    compute_loss = ComputeLoss(model)

# Run inference
if pt and device.type != 'cpu':
    model(torch.zeros(1, 3, *imgsz).to(device).type_as(next(model.parameters())))  # run once

change: https://github.com/ultralytics/yolov5/blob/f5b8f7d54c9fa69210da0177fec7ac2d9e4a627c/detect.py#L71-L72 to

visualize = increment_path(save_dir / Path(path).stem, mkdir=True) if visualize else False
# pred = model(img, augment=augment, visualize=visualize)[0]

if visualize:
    # grad-cam
    pred = model(img, augment=augment, visualize=visualize)
    
    model.zero_grad()
    targets = torch.zeros(2, 6)
    loss, loss_items = compute_loss(pred[1], targets.to(device))
    loss.requires_grad_(True)
    loss.backward()

    _grads = model.grads_list
    _grads.reverse()
    _features = model.features_list

    # for g, f in zip(_grads, _features):
    #     print('grad', type(g), g.shape)
    #     print('feature', type(f), f.shape)
    
    for i in [17, 20, 23]:
        out_name = str(save_dir / f"{i}.jpg")
        cam_show_img(im0s, _features[i].cpu().detach().numpy()[0], _grads[i].cpu().detach().numpy()[0], out_name)

    pred = pred[0]
else:
    pred = model(img, augment=augment, visualize=visualize)[0]

in yolo.py 1.add grad_list and feature_list to store grad and feature https://github.com/ultralytics/yolov5/blob/f5b8f7d54c9fa69210da0177fec7ac2d9e4a627c/models/yolo.py#L66-L67

class Model(nn.Module):

    grads_list = []
    features_list = []

    def __init__(self, cfg='yolov5s.yaml', ch=3, nc=None, anchors=None):  # model, input channels, number of classes

2.add hook function in class Model

    def save_gradient(self, grad):
        # print('grad shape:', grad.shape)
        self.grads_list.append(grad)

3.store grad and feature in Model.forward_once()

            y.append(x if m.i in self.save else None)  # save output
            
            # save grad
            if visualize:
                if 'Detect' not in m.type:
                    x.register_hook(self.save_gradient)
                    self.features_list.append(x)
                else:
                    if isinstance(x, list):
                        for _x in x:
                            _x.register_hook(self.save_gradient)
                            self.features_list.append(_x)
                    elif isinstance(x, tuple):
                        for _x in x[1]:
                            _x.register_hook(self.save_gradient)
                            self.features_list.append(_x)

            if visualize:
                feature_visualization(x, m.type, m.i, save_dir=visualize)

Use case

a piece of code about Grad-Cam, which can visualize where the model concerns.

Additional

No response

Are you willing to submit a PR?

• [X] Yes I'd like to help by submitting a PR!

❔Question

Use export.py to generate yolov5s.onnx will get a negative number.

This is the code that executes the onnx part

session = onnxruntime.InferenceSession('./weights/yolov5s.onnx')

batch_size = session.get_inputs()[0].shape[0]
img_size_h = session.get_inputs()[0].shape[2]
img_size_w = session.get_inputs()[0].shape[3]

image_src = Image.open(image_path)
resized = letterbox_image(image_src, (img_size_w, img_size_h))
img_in = np.transpose(resized, (2, 0, 1)).astype(np.float32)  # HWC -> CHW
img_in = np.expand_dims(img_in, axis=0)
img_in /= 255.0

input_name = session.get_inputs()[0].name
# output, output_exist = session.run(['decoder.output_conv', 'lane_exist.linear2'], {"input.1": image_np})
outputs = session.run(None, {input_name: img_in})

There are already negative numbers in the outputs. Then after the result is processed, it will appear that part of it is correct, such as car in the figure. However, the top/bottom of the bicycle is right, the left/right is wrong, the left/right of the dog is right, and the top/bottom is wrong. What might cause this problem? Thanks.

Additional context

torch:1.5.1 torchvision:0.6.1 onnxruntime:1.3.0

This PR add models/tf.py to export TensorFlow saved_model, graph_def and TFLite models. This script support both yolov5 v2 (LeakyReLU activations) and v3 (Hardswish activations) models.

Export TensorFlow and TFLite models using:

PYTHONPATH=. python models/tf.py --weights weights/yolov5s.pt --cfg models/yolov5s.yaml --img 640

and use one of the following command to detect objects:

python detect.py --weights weights/yolov5s_saved_model --img 640
python detect.py --weights weights/yolov5s.pb --img 640
python detect.py --weights weights/yolov5s.tflite --img 640

This PR is tested sucessfully with PyTorchc 1.6, TensorFlow 1.15.3/2.3.0. However, TFLite export is only supported under TensorFlow 2.3.0.

🚀 Feature

Transformer is popular in NLP, and now is also applied on CV. I added C3TR just by replacing the sequential self.m in C3 with a Transformer block, which could reduce GFlOPs and make Yolo achieve a better result.

Motivation

• Dosovitskiy et al. proposed ViT
• Facebook applied Transformer on object detection as an encoder
• So I thought Transformer could make yolo better

Pitch

I add 3 classes in https://github.com/dingyiwei/yolov5/blob/Transformer/models/common.py :

class TransformerLayer(nn.Module):
    def __init__(self, c, num_heads):
        super().__init__()

        self.ln1 = nn.LayerNorm(c)
        self.q = nn.Linear(c, c, bias=False)
        self.k = nn.Linear(c, c, bias=False)
        self.v = nn.Linear(c, c, bias=False)
        self.ma = nn.MultiheadAttention(embed_dim=c, num_heads=num_heads)
        self.ln2 = nn.LayerNorm(c)
        self.fc1 = nn.Linear(c, c, bias=False)
        self.fc2 = nn.Linear(c, c, bias=False)

    def forward(self, x):
        x_ = self.ln1(x)
        x = self.ma(self.q(x_), self.k(x_), self.v(x_))[0] + x
        x = self.ln2(x)
        x = self.fc2(self.fc1(x)) + x
        return x


class TransformerBlock(nn.Module):
    def __init__(self, c1, c2, num_heads, num_layers):
        super().__init__()

        self.conv = None
        if c1 != c2:
            self.conv = Conv(c1, c2)
        self.linear = nn.Linear(c2, c2)
        self.tr = nn.Sequential(*[TransformerLayer(c2, num_heads) for _ in range(num_layers)])
        self.c2 = c2

    def forward(self, x):
        if self.conv is not None:
            x = self.conv(x)
        b, _, w, h = x.shape
        p = x.flatten(2)
        p = p.unsqueeze(0)
        p = p.transpose(0, 3)
        p = p.squeeze(3)
        e = self.linear(p)
        x = p + e

        x = self.tr(x)
        x = x.unsqueeze(3)
        x = x.transpose(0, 3)
        x = x.reshape(b, self.c2, w, h)
        return x


class C3TR(C3):
    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):
        super().__init__(c1, c2, n, shortcut, g, e)
        c_ = int(c2 * e)
        self.m = TransformerBlock(c_, c_, 4, n)

And I just put it as the last part of the backbone instead of a C3 block.

backbone:
  # [from, number, module, args]
  [[-1, 1, Focus, [64, 3]],  # 0-P1/2
   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
   [-1, 3, C3, [128]],
   [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
   [-1, 9, C3, [256]],
   [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
   [-1, 9, C3, [512]],
   [-1, 1, Conv, [1024, 3, 2]],  # 7-P5/32
   [-1, 1, SPP, [1024, [5, 9, 13]]],
   [-1, 3, C3TR, [1024, False]],  # 9    <---- here is my modifcation
  ]

I conducted experiments on 2 Nvidia GTX 1080Ti cards, where depth_multiple and width_multiple are the same as Yolov5s. Here are my experimental results with img-size 640. For convenience I named the method in this issue as Yolov5TRs.

|Model|Params|GFLOPs| |---|---|---| |Yolov5s|7266973|17.0| |Yolov5TRs|7268765|16.8|

|Model|Dataset|TTA|[email protected]|[email protected]:.95|Speed (ms)| |---|---|---|---|---|---| |Yolov5s|coco (val)|N|0.558|0.365|4.4| |Yolov5TRs|coco (val)|N|0.568|0.363|4.4| |Yolov5s|coco (test-dev)|N|0.559|0.365|4.6| |Yolov5TRs|coco (test-dev)|N|0.567|0.365|4.5| |Yolov5s|coco (test-dev)|Y|0.568|0.378|12.0| |Yolov5TRs|coco (test-dev)|Y|0.571|0.375|11.0|

We can see that Yolov5TRs get higher scores in [email protected] with a faster speed. (I'm not sure why my results of Yolov5s are different from which shown in README. The model was downloaded from release v4.0) When depth_multiple and width_multiple are set to larger numbers, C3TR should be more lightweight than C3. Since I do not have so much time on it and my machine is not very strong, I did not run experiments on M, L and X. Maybe someone could conduct the future experiments:smile:

❔Question

I am trying to convert a trained yolov5s model to an SNPE format in order to be able to run it on a Snapdragon chip. Unfortunately, Qualcomm's ONNX to SNPE converter fails on the Detect level with the following error message

ValueError: Unable to permute shape [1, 3, 64, 64, 2] to NSC ordering
2021-09-14 15:15:37,327 - 183 - ERROR - Node Mul_268: Unable to permute shape [1, 3, 64, 64, 2] to NSC ordering

I can imagine, it may have something to do with the fact that SNPE currently supports 4D input data, where the first dimension is batch SNPE doc and yolov5 Detect layer has 5D reshape.

Would it be possible to modify Detect layer so that no 5D reshape is performed?

Fixing DDP mode. #177 Work in Progress, But most of the hard things have already been done! There are lots of commits. If every thing is settled down, I will merge them into two commits!

Search before asking

• [X] I have searched the YOLOv5 issues and found no similar bug report.

YOLOv5 Component

No response

Bug

I trained a model, but UL crashed when uploading it:

      Epoch    GPU_mem   box_loss   obj_loss   cls_loss  Instances       Size
      99/99      73.7G   0.009337   0.004801          0        621        640: 100%|██████████| 37/37 00:31
                 Class     Images  Instances          P          R      mAP50   mAP50-95: 100%|██████████| 3/3 00:06
                   all       1556       1860      0.867       0.76       0.82      0.387
Ultralytics: Uploading checkpoint Xr7RjMjjuMXh2Vto2x9q

24 epochs completed in 0.251 hours.
Optimizer stripped from yolov5/runs/train/exp/weights/last.pt, 14.4MB
Ultralytics: Training completed successfully ✅
Ultralytics: Uploading final Xr7RjMjjuMXh2Vto2x9q
Ultralytics: View model at https://hub.ultralytics.com/models/Xr7RjMjjuMXh2Vto2x9q 🚀
Ultralytics: Unhandled server error. Retrying 10x for 3600s. If this issue persists please visit https://github.com/ultralytics/hub/issues for assistance. (500 #4)
Results saved to yolov5/runs/train/exp

The contents of yolov5/runs/train/exp:

Please advise how can I upload the model?

Environment

No response

Minimal Reproducible Example

No response

Additional

No response

Are you willing to submit a PR?

• [ ] Yes I'd like to help by submitting a PR!

Search before asking

• [X] I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

I am training YOLOv5l on COCO taking the weights which were shared. I noticed that the mAP_0.5 increases to 65% on COCO after training for 300 epochs. Can someone explain why this is happening? If that is the case, what should be ideal number of Epoch to get best result in COCO

I don't suspect Overfitting as both train and val loss are decreasing. image

Additional

Search before asking

• [X] I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

why can't i gain the segmentation by use python val.py --save-json(the results val_batch{batch_i}_pred.jpg has segmentation label)

Additional

No response

Search before asking

• [X] I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

I am running Yolov5 in Azure Machine Learning instead of Google Colab. Microsoft has a full featured integrated capability of creating docker containers, and run it on GPU's in the cloud straight from Visual Studio Code. 0.23 USD/ hr for a T80 Tesla on low priority infrastructure. It even has full github ML Ops integrated. See image below where I show it is working...but with some hacks in your code. It is much easier to use than Google CoLab, but getting it to work with this repository is a bit hard. I particular as I do not know how to avoid mixing up our code with this repository, but that is maybe because Python is not that OO as C# to have it perfectly separated.

I modified this in the requirements.txt: opencv-python-headless>=4.1.1 #opencv-python>=4.1.1

My script is a py file in the Yolo folder. Basically this is all needed to fully create a container and sent it to the Azure ML cloud: `from pathlib import Path

from azureml.core import ( Environment, Experiment, Run, ScriptRunConfig, Workspace, Dataset, Datastore, ) from azureml.data import OutputFileDatasetConfig

from azureml.core.compute import AmlCompute, ComputeTarget from azureml.core.authentication import AzureCliAuthentication

def find_or_create_compute_target( workspace, name, vm_size="Standard_NC6", #"Standard_D8_v3", min_nodes=0, max_nodes=1, idle_seconds_before_scaledown=900, vm_priority="lowpriority", ):

if name in workspace.compute_targets:
    return ComputeTarget(workspace=workspace, name=name)
else:
    config = AmlCompute.provisioning_configuration(
        vm_size=vm_size,
        min_nodes=min_nodes,
        max_nodes=max_nodes,
        vm_priority=vm_priority,
        idle_seconds_before_scaledown=idle_seconds_before_scaledown,
    )
    target = ComputeTarget.create(workspace, name, config)
    target.wait_for_completion(show_output=True)

return target

workspace_config = 'config.json' requirements_file = 'requirements.txt' compute_target_name = 'LowPrioStdNC6' experiment_name = 'train-yolo-7' script_path = 'segment/train.py' blobstorage = 'workspaceblobstore' blobcontainerpath = 'UI/2023-01-05_###_UTC'

cli_auth = AzureCliAuthentication()

Authenticate with your AzureML Resource via its config.json file

ws = Workspace.from_config(path=workspace_config, auth=cli_auth)

ws = Workspace.from_config(path="v4a/config.json", auth=cli_auth)

The experiment in this workspace under which our runs will be grouped

If an experiment with the given name doesn't exist, it will be created

exp = Experiment(ws, experiment_name)

The compute cluster you want to run on and its settings.

If it doesn't exist, it'll be created.

compute_target = find_or_create_compute_target(ws, compute_target_name)

The Environment lets us define any dependencies needed to make our script run

env = Environment.from_pip_requirements("my-pip-env", requirements_file)

datastore = Datastore.get(ws, blobstorage) dataset = Dataset.File.from_files(path=(datastore, blobcontainerpath))

output = OutputFileDatasetConfig(name='outputs', destination=(datastore, 'outputdataset'))

script_args = ['--device', '0', '--img', '640', '--cfg', 'v4a/data/yolov5s-seg.yaml', '--hyp', 'data/hyps/hyp.scratch-high.yaml', '--batch', '16', '--epochs', '100', '--data', 'v4a/data/###.yaml', '--weights', 'v4a/data/yolov5s-seg.pt', '--workers', '24', '--name', '#####', '--project', output.as_mount(), '--inputs', dataset.as_named_input('inputs').as_mount()] #'--outputs', output.as_mount()]

A run configuration is how you define what youd like to run

We give it the directory where our code is, the script we want to run, the environment, and the compute info

run_config = ScriptRunConfig( source_directory=".",#Path(script_path).parent, script="./segment/train.py", #Path(script_path).name, arguments=script_args, compute_target=compute_target, environment=env, )

Submit our configured run under our experiment

run = exp.submit(run_config)

Wait for the run to finish

run.wait_for_completion(show_output=True)

Delete our compute target so it doesn't cost us money

#compute_target.delete()`

The input and outputs are explained in this Jupyter notebook: https://github.com/Azure/MachineLearningNotebooks/blob/master/how-to-use-azureml/work-with-data/datasets-tutorial/scriptrun-with-data-input-output/how-to-use-scriptrun.ipynb

As our dataset is too large we need to mount the dataset from an azure blob storage. I can however not see how to change the path. I see in the yaml file a path can be set, but that is not working.

Is it possible to support this natively ? I am a C# programmer and do not get the idea of OO in Python. So I did some modifications which are not so nice:

Another not so nice:

It is currently processing, question is if I can see the outputs of the training somewhere:

The output is also saved to a azure blob:

Additional

Would be nice to support Microsoft Azure Machine Learning natively including azure blob storage for large files. It will save the environment by not having to pump data around.

Search before asking

• [X] I have searched the YOLOv5 issues and found no similar bug report.

YOLOv5 Component

Export

Bug

Hello,

it's compatibility with Android environment problem. The capture below shows the warning I received from trying to run the exported tflite model on my Android mobile using the GPU delegate. Not only this, but the model is quite slow ~700ms. I tested on different android phones.

if I use the mobilenetv2 tflite for example I dont get this warning and the model is much faster.

My thoughts is maybe we should rethink tf.py implementation to get more ops supported by tflite on GPU following mobileNet implementation.

Even inference on my computer (M1 Mac Pro) using tflite exported model is slow ~140ms compared to pytorch model and onnx.

If you run into the same problem please share your thoughts/solutions

Thank you

Environment

No response

Minimal Reproducible Example

No response

Additional

No response

Are you willing to submit a PR?

• [ ] Yes I'd like to help by submitting a PR!

Search before asking

• [X] I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

In file \yolov5-master\yolov5-master\classify\train.py

no parameter --save-period

parser.add_argument('--model', type=str, default='yolov5s-cls.pt', help='initial weights path')
parser.add_argument('--data', type=str, default='imagenette160', help='cifar10, cifar100, mnist, imagenet, ...')
parser.add_argument('--epochs', type=int, default=10, help='total training epochs')
parser.add_argument('--batch-size', type=int, default=64, help='total batch size for all GPUs')
parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=224, help='train, val image size (pixels)')
parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
parser.add_argument('--cache', type=str, nargs='?', const='ram', help='--cache images in "ram" (default) or "disk"')
parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
parser.add_argument('--project', default=ROOT / 'runs/train-cls', help='save to project/name')
parser.add_argument('--name', default='exp', help='save to project/name')
parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
parser.add_argument('--pretrained', nargs='?', const=True, default=True, help='start from i.e. --pretrained False')
parser.add_argument('--optimizer', choices=['SGD', 'Adam', 'AdamW', 'RMSProp'], default='Adam', help='optimizer')
parser.add_argument('--lr0', type=float, default=0.001, help='initial learning rate')
parser.add_argument('--decay', type=float, default=5e-5, help='weight decay')
parser.add_argument('--label-smoothing', type=float, default=0.1, help='Label smoothing epsilon')
parser.add_argument('--cutoff', type=int, default=None, help='Model layer cutoff index for Classify() head')
parser.add_argument('--dropout', type=float, default=None, help='Dropout (fraction)')
parser.add_argument('--verbose', action='store_true', help='Verbose mode')
parser.add_argument('--seed', type=int, default=0, help='Global training seed')
parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
return parser.parse_known_args()[0] if known else parser.parse_args()

Additional

No response