shashankprasanna

from tensor import Tensor
from math import trunc, mod
fn tensorprint[type: DType](t: Tensor[type])->None:
    let rank = t.rank()
    var dim0:Int=0
    var dim1:Int=0
    var dim2:Int=0
    if rank==0 or rank>3:
        print("Error: Tensor rank should be: 1,2, or 3. Tensor rank is ", rank)
        return

shashankprasanna

def f(x):
    return torch.sin(x)**2 + torch.cos(x)**2 

torch._dynamo.reset()
compiled_f = torch.compile(f, backend='inductor',
                              options={'trace.enabled':True,
                                       'trace.graph_diagram':True})

# device = 'cpu'
device = 'cuda'

shashankprasanna

import torch._dynamo
from torch.fx.passes.graph_drawer import FxGraphDrawer
from functorch.compile import make_boxed_func
from torch._functorch.aot_autograd import aot_module_simplified

def f(x):
    return torch.sin(x)**2 + torch.cos(x)**2

def inspect_backend(gm, sample_inputs): 
    # Forward compiler capture

shashankprasanna

def inspect_backend(gm, sample_inputs):
    code = gm.print_readable()
    with open("forward.svg", "wb") as file:
        file.write(FxGraphDrawer(gm,'f').get_dot_graph().create_svg())
    return gm.forward

torch._dynamo.reset()
compiled_f = torch.compile(f, backend=inspect_backend)

x = torch.rand(1000, requires_grad=True).to(device)

shashankprasanna

Architecture	NVIDIA GPU	Instance type	Instance name	Number of GPUs	GPU Memory (per GPU)	GPU Interconnect (NVLink / PCIe)	Thermal Design Power (TDP) from nvidia-smi	Tensor Cores (mixed-precision)	Precision Support	CPU Type	Nitro based
Ampere	A100	P4	p4d.24xlarge	8	40 GB	NVLink gen 3 (600 GB/s)	400W	Tensor Cores (Gen 3)	FP64, FP32, FP16, INT8, BF16, TF32	Intel Xeon Scalable (Cascade Lake)	Yes
Ampere	A10G	G5	g5.xlarge	1	24 GB	NA (

shashankprasanna

def image_preprocess(img, reps=1):
    img = np.asarray(img.resize((224, 224)))
    img = np.stack([img]*reps)
    img = tf.keras.applications.resnet50.preprocess_input(img)
    return img

from PIL import Image 
import numpy as np
import json

shashankprasanna

from sagemaker.tensorflow.model import TensorFlowModel, TensorFlowPredictor

prefix = 'keras_models'
s3_model_path = sess.upload_data(path='model.tar.gz', key_prefix=prefix)

model = TensorFlowModel(model_data=s3_model_path, 
                        framework_version='1.15',
                        role=role,
                        predictor_cls = TensorFlowPredictor,
                        sagemaker_session=sess)

shashankprasanna

from smdebug.trials import create_trial

def tensor_df(tname):
    tval = trial.tensor(tname).values()
    df   = pd.DataFrame.from_dict(tval,orient='index',columns=[tname])
    df_tval = df.reset_index().rename(columns={'index':'steps'})
    return df_tval

def trial_perf_curves(job_name, tname, experiment_name):
    debug_data = f's3://{bucket_name}/{experiment_name}/{job_name}/debug-output'

shashankprasanna

from sagemaker.analytics import ExperimentAnalytics

experiment_name = training_experiment.experiment_name

trial_component_analytics = ExperimentAnalytics(
    sagemaker_session=sagemaker_session, 
    experiment_name=experiment_name,
)
trial_comp_ds = trial_component_analytics.dataframe()

shashankprasanna

for trial_hyp in trial_hyperparameter_set:
    # Combine static hyperparameters and trial specific hyperparameters
    hyperparams = {**static_hyperparams, **trial_hyp}
    
    # Create unique job name with hyperparameter and time
    time_append = int(time.time())
    hyp_append = "-".join([str(elm) for elm in trial_hyp.values()])
    job_name = f'cifar10-training-{hyp_append}-{time_append}'
    
    # Create a Tracker to track Trial specific hyperparameters